JP7464346B2 - Object attribute expression generation model capable of generating object attribute expressions, object attribute estimation device and method - Google Patents

Description

The present invention relates to a technology that handles distributed representations of attributes of objects, such as user attributes.

Personality, which is a characteristic that holistically captures mental indicators such as a person's character, personality, temperament, individuality, and even values, is known to have a significant impact on behavior in various behavioral domains, such as the purchase and selection of various products and services. For this reason, in recent years, there have been many attempts to understand users' personalities in order to predict their behavior and personalize the products and services they provide.

Today, human personality is generally understood according to a concept called "trait theory." This trait theory states that human personality is composed of multiple traits and can be expressed by the level (score) of each trait. For example, the "Big Five," a representative example of a personality model based on trait theory, states that human personality is composed of five traits: intellectual curiosity (O: Openness to experience), conscientiousness (C: Conscientiousness), extraversion (E: Extraversion), agreeableness (A: Agreeableness), and emotional stability (N: Neuroticism), and expresses human character as a set of scores for these five traits (OCEAN).

Questionnaire surveys are often used as a method of measuring such personality traits. For example, the "Big Five Scales," a representative questionnaire for measuring the Big Five, has subjects answer a set of questions consisting of a total of 60 items. When measuring intellectual curiosity (O), for example, subjects are asked to answer questions such as "creative" and "versatile" on a seven-point scale from "not at all applicable (1 point)" to "very applicable (7 points)," and the score for intellectual curiosity (O) is determined by the total score for all questions.

On the other hand, given that conducting questionnaire surveys of a huge number of users takes a lot of time and money, there is also widespread research into technology that first administers questionnaire surveys to a small number of users and then uses the responses to these surveys and the user's behavioral data to estimate the user's personality.

Non-Patent Document 1 is a survey paper on such technologies, and introduces a technology that estimates a user's personality using various user behavioral data, such as the content posted to SNS (Social Networking Service), the number of friends connected on SNS, the call history of a mobile device, movement patterns derived based on device location information, and even the speed and intonation of speech during phone calls, as well as the volume of the voice.

A. Vinciarelli and G. Mohammadi, "A Survey of Personality Computing," IEEE Transactions on Affective Computing, vol. 5, no. 3, pp. 273-291, 2014, <https://doi.org/10.1109/TAFFC.2014.2330816> Donkers, T., Loepp, B., & Ziegler, J., "Sequential user-based recurrent neural network recommendations", RecSys'17 - Proceedings of the 11th ACM Conference on Recommender Systems, pp. 152-160, 2017, ＜https://doi.org/10.1145/3109859.3109877＞

As explained above, various techniques have been developed to measure personality, but these techniques share common issues that need to be resolved: (a) coarse measurement granularity (information granularity) and (b) low reliability.

First, there is the coarseness of the measurement granularity mentioned above in (a). This poses a significant challenge to technologies for grasping personality based on "trait theory", including technologies for estimating a user's personality from the user's behavioral data. For example, the Big Five traits were originally determined by aggregating a huge number of personality-descriptive words extracted from dictionaries into five groups of words with similar concepts using statistical methods, and then based on the results of this aggregation, in a way that would be easy for humans to interpret.

However, in the process of aggregation, small or subtle differences in the original meanings of the personality expressions are lost. For example, conscientiousness (C), one of the Big Five, is aggregating personality expressions such as "serious," "hardworking," and "organized" into a single trait, and the measurement result of this conscientiousness (C) is merely a single score, and the differences in meaning of each personality expression before aggregation are lost from the measurement results.

Next, there is the low reliability of (b) above, which is a particularly notable issue for technologies that involve measuring personality through questionnaire surveys. For example, there is usually considerable variation in the way subjects respond to questions in a questionnaire. Specifically, it is not rare for subjects to answer based on how they "should be" rather than on their own original thoughts, or to consider answering itself to be a tedious task and give false answers. As a result, the reliability of personality measurement results is reduced.

Even if the expected (honest and serious) response attitude is guaranteed, the questions and response methods set in a questionnaire survey are unlikely to measure all of the areas of the characteristics to be measured, and the reliability of the measurement results is inevitably low. For example, there are more than 20 types of Big Five questionnaires, but the number of questions, question wording, and response methods (such as answering on a 5-point scale or yes/no) are significantly different for each. As a result, it is widely known that the measurement results of the same subject differ depending on the type of questionnaire answered, and there have been cases where the correlation r of the measurement results of agreeableness (A) between two questionnaires is below 0.40.

Furthermore, even in technology that estimates a user's personality from the user's behavioral data, the estimation model used is usually trained and constructed using measurement results from questionnaire surveys, so the low reliability of (b) described above is a serious issue that needs to be resolved.

On the other hand, although it is not a method for measuring personality, particularly in the field of product and service recommendation technology, User Representation Learning (URL) technology, which makes it possible to express user attributes without conducting a questionnaire survey, is currently being actively researched.

Specifically, this URL technology is a technology that acquires distributed representations (usually representations using vectors (lists of many numerical values) with several hundred dimensions) of user attributes such as music and movie preferences and purchasing trends from user behavioral data such as music playback, video viewing, ad clicks, and even item purchases. Therefore, although it is for user attributes, the measurement granularity (information granularity) is sufficiently high (fine), and as a result, it can be said that URL technology can solve the problem of the coarse measurement granularity mentioned above in (a).

Here, URL technologies are broadly divided into static URL technologies, such as matrix factorization, which use non-time-series behavioral data that does not reflect the sequential relationships in the behavioral history, and sequential URL technologies, which use time-series behavioral data that reflects the sequential relationships in the behavioral history.

The latter Sequential URL technology differs from the former Static URL technology in that it is capable of extracting dynamic user attributes that can change separately from static user attributes that do not change. For example, Non-Patent Document 2 discloses an example of this Sequential URL technology in which time-series behavioral data on a user's movie viewing is input into a GRU (Gated Recurrent Unit), which is a type of Recurrent Neural Network (RNN), to acquire a distributed representation of the user attributes.

Specifically, the technology described in Non-Patent Document 2 discloses a GRU cell that receives as input the distributed representation of a user output from a distributed representation extraction unit that receives the user's identification information as input, and the distributed representation of a movie output from a distributed representation extractor that receives the identification information of a movie watched by the user as input. Non-Patent Document 2 claims that by training the GRU cell and the distributed representation extractor with a sufficient amount of data, the distributed representation of the user output from the distributed representation extraction unit reflects user attributes that are static and do not change easily, such as the user's movie preferences.

However, the distributed representation of user attributes obtained by such Sequential URL technology is naturally highly dependent on the behavioral domain (area) to which the behavioral data used belongs, and is not at all interpretable as the user's personality (a characteristic that captures mental indicators holistically). For example, when behavioral data related to the behavioral domain of movie viewing is used, only user attributes related to movie viewing, such as the user's movie preferences, can be extracted. For this reason, no attempt has been made to estimate a user's personality using Sequential URL technology to date.

Furthermore, the static URL technology mentioned above is not designed to separate or extract the degree of dependency of user attributes on behavioral domains or whether they are static or dynamic. Therefore, it is impossible to interpret the distributed representation of a user generated by static URL technology as the personality of that user.

The present invention aims to provide a target attribute representation generation model, a target attribute estimation device, and a target attribute estimation method that are capable of generating attribute representations in which a given attribute, such as a target's personality, is expressed with higher reliability.

According to the present invention, there is provided an object attribute representation generation model for causing a computer to function to map a representation vector of an object to another representation vector that is an element of another representation vector space, comprising:
A mapping unit that applies a trained mapping operator to the input expression vector to generate another expression vector and outputs the another expression vector as information related to the predetermined attribute of the target , or that is a trained neural network algorithm that receives the expression vector as an input and outputs the another expression vector as information related to the predetermined attribute of the target.
The computer functions as
The mapping operator or the neural network algorithm is trained such that, with respect to a strength index of a predetermined relationship between the objects, which is related or may be related to the degree of similarity between the objects with respect to a predetermined attribute, the greater the degree of similarity between the objects related to the two different representation vectors obtained by mapping, or the greater the likelihood that the strength index of the predetermined relationship will be, the smaller the degree of separation between these two different representation vectors.
The present invention provides an object attribute representation generation model , characterized in that :

As one embodiment of the object attribute representation generation model according to the present invention, it is also preferable that the mapping operator or the neural network algorithm is trained such that the stronger the objects related to the two separate representation vectors obtained by mapping are, the smaller the degree of separation between these two separate representation vectors will be, with respect to a strength index of the specified relationship that is positively correlated or likely to be positively correlated with the degree of similarity between the objects for a specified attribute.

In the above embodiment, the mapping operator or the neural network algorithm may be
Two separate expression vectors relating to objects that are stronger in terms of the strength index of the predetermined relationship, but that are different or distant in terms of an attribute other than the predetermined attribute, are trained to have a smaller degree of separation between these two separate expression vectors; and/or
It is also preferable that two separate expression vectors relating to objects that are weaker in terms of the strength indicator of the specified relationship, but have the same or a close relationship in terms of an attribute other than the specified attribute, are trained to have a greater degree of separation between these two separate expression vectors.

Furthermore, regarding the strength index of a predetermined relationship according to the present invention, when the subject is a human being and the predetermined attribute is personality,
The strength indicator of the predetermined relationship is
(a) An indicator that is stronger for subjects who are related by blood or for subjects who are closely related by blood;
(a) An indicator that indicates that the stronger the relationship between objects that are friends or acquaintances, or the deeper the relationship between objects that are friends or acquaintances, the stronger the relationship between objects that are friends or acquaintances,
(c) An indicator that indicates that the closer the relationship between objects is, the stronger the relationship is.
(D) An index that makes objects that have good compatibility, or objects that have good compatibility, stronger;
(E) An indicator that is stronger between subjects with identical specific gene sequences, or between subjects with similar specific gene sequences, and/or
(f) It is also preferable that the index be stronger for subjects whose features extracted from electroencephalograms are more similar.

In another embodiment of the object attribute representation generation model according to the present invention, it is also preferable that the mapping operator or the neural network algorithm is trained such that the weaker the objects related to the two different representation vectors obtained by mapping are, the smaller the degree of separation between these two different representation vectors will be, with respect to a strength indicator of the specified relationship that is negatively correlated or may be negatively correlated with the degree of similarity between the objects for a specified attribute.

Furthermore, in the above embodiment using a predetermined relationship strength indicator related to negative correlation, the mapping operator or the neural network algorithm may further include:
Two separate expression vectors relating to objects that are weaker in terms of the strength index of the predetermined relationship, but have different or distant relationships in terms of an attribute other than the predetermined attribute, are trained to have a smaller degree of separation between the two separate expression vectors; and/or
It is also preferable that two separate expression vectors relating to objects that are stronger in terms of the strength indicator of the specified relationship and that have the same or a close relationship in terms of an attribute other than the specified attribute are trained to have a greater degree of separation between the two separate expression vectors.

In addition, regarding the representation vector before mapping according to the present invention, when the target is a user and the predetermined attribute is a personality,
The expression vector is a user attribute expression vector that is generated by a user attribute information generation model and is information that expresses a user attribute including the personality of the user,
The above user attribute information generation model is as follows:
A plurality of domain-specific regression neural network (RNN) cells are set for each of a plurality of behavioral domains, and the domain-specific RNN cells receive domain behavior information, which is information related to the behavior of the user in the behavioral domain, and generate new domain-specific hidden state information by reflecting the domain behavior information on domain-specific hidden state information, which is hidden state information generated by the RNN cells themselves at a previous point in time;
a user representation generator trained in conjunction with the plurality of domain-specific RNN cells to generate a user attribute representation vector of the user from information identifying the user;
It is also preferable that the model functions as a domain-independent RNN cell that receives the generated domain-specific hidden state information and the generated user attribute representation vector, and generates new domain-independent hidden state information by reflecting the domain-specific hidden state information and the user attribute representation vector on domain-independent hidden state information, which is hidden state information that the computer itself generated at a previous point in time.

Furthermore, regarding the representation vector before mapping according to the present invention, when the target is a user and the predetermined attribute is a personality,
It is also preferable that the expression vector is a user attribute expression vector that expresses the user attributes of the user and is calculated by a method related to a user representation learning (URL) technology using information related to a predetermined behavior of the user. In addition, it is also preferable that the method related to the user representation learning technology is specifically a matrix factorization method.

The present invention also provides a target attribute estimation device that generates a different expression vector using the target attribute expression generation model described above and outputs the generated different expression vector as information related to a specific attribute of the target.

According to the present invention, further comprising:
By performing training to obtain the above-mentioned trained mapping operator or trained neural network algorithm , constructing the object attribute representation generation model according to any one of claims 1 to 9;
A computer -implemented target attribute estimation method is provided, which is characterized by generating the other expression vector using the constructed target attribute expression generation model, and outputting the generated other expression vector as information related to a specified attribute of the target.

The object attribute representation generation model, object attribute estimation device, and object attribute estimation method of the present invention can generate attribute representations that express specific attributes, such as the personality of an object, with higher reliability.

1 is a schematic diagram showing an embodiment of an object attribute expression generation model and a user attribute information generation model according to the present invention, and an embodiment of an object attribute estimation device equipped with these models. 1 is a schematic diagram for explaining an embodiment of a mapping process performed in a mapping unit of an object attribute representation generation model according to the present invention; FIG. FIG. 1 is a schematic diagram showing an embodiment of a user attribute information generation model according to the present invention. FIG. 13 is a schematic diagram showing another embodiment of a user attribute information generation model according to the present invention.

The following describes in detail an embodiment of the present invention with reference to the drawings.

[Object attribute expression generation model, user attribute information generation model]
FIG. 1 is a schematic diagram showing an embodiment of an object attribute expression generation model and a user attribute information generation model according to the present invention, and an embodiment of an object attribute estimation device equipped with these models.

First, the user attribute information generation model 2 shown in Figure 1 is a machine learning model that generates and outputs a "user attribute representation vector r_u," which is information related to the attributes (user attributes) of a target, from "domain behavior information," which is information related to the behavior of a target (a user in this embodiment) in a specific behavioral domain (area), as will be described in detail later.

The output "user attribute representation vector r_u", which will also be described in detail later, is information that includes "information related to a specific attribute (personality in this embodiment)" of the target (user. Note that hereafter, "representation vector" refers to a multi-dimensional (e.g., tens to hundreds of dimensions) vector handled by known embedded representation (embedding) technology, such as word embedded representation, and is a vector quantity that is a sequence of many numerical values.

Also in FIG. 1, the object attribute expression generation model 1 of this embodiment is as follows:
(a) receiving the generated “user attribute expression vector r_u” from the user attribute information generation model 2;
(b) This is a machine learning model that generates and outputs a “target attribute representation vector” (in this embodiment, a personality representation vector M(r_u)) that expresses a specific attribute (personality) of the target (user) from the received “user attribute representation vector r_u.”

In this way, the target attribute expression generation model 1 can be considered as a model that separates and extracts the "information related to a specific attribute (personality)" contained in the received "user attribute expression vector r_u" as a "target attribute expression vector (personality expression vector) M(r_u)."

In this regard, the object attribute representation generation model 1 has the following features:
(A) A mapping unit 10 that outputs an "object attribute representation vector (personality representation vector) M(r_u)" which is another representation vector (in another representation vector space) obtained by mapping a "user attribute representation vector r_u" of an object (user) which is an element of a certain representation vector space, as "information related to a predetermined attribute (personality) of the object (user)"
It is a model that causes a computer to function as a system that can extract and separate target attribute expressions (user personality expressions) as described above.

Furthermore, the mapping unit 10 has the following features:
(B) It is a subject that has been given specific training (learning) using a "strength index of a predetermined relationship between objects" that is related or may be related to the degree of similarity between objects (users) regarding a predetermined attribute (personality). Incidentally, as the "strength index of a predetermined relationship between objects", which will be described in detail later, for example, a blood relationship index, a friend/acquaintance relationship index, a closeness index, a compatibility index, a gene sequence index, an electroencephalogram feature amount index, and even a bad relationship index, etc. can be adopted.

More specifically, the mapping unit 10:
(C) This “indicator of the strength of a specified relationship between objects” is trained so that the greater the degree of similarity of a specified attribute (personality) between objects (users) related to the two “object attribute expression vectors (personality expression vectors)” obtained by mapping, or the greater the possibility of such a degree of similarity, the smaller the degree of separation between these two “object attribute expression vectors (personality expression vectors).”

Furthermore, as a result of such training, the object attribute representation generation model 1 (mapping unit 10) can use the mapping M based on the "strength index of a specified relationship between objects" to tailor the received "user attribute representation vector r_u" into an "object attribute representation vector (personality representation vector) M(r_u)" in which the degree of separation between the objects becomes smaller as the specified attributes (personalities) become more similar.

This "target attribute expression vector (personality expression vector) M(r_u)" is precisely information that expresses a specified attribute (personality) with greater reliability. In other words, the target attribute expression generation model 1 can generate an expression vector as an attribute expression that expresses a specified attribute (personality) with greater reliability.

In addition, since this "target attribute expression vector (personality expression vector) M(r_u)" is a multidimensional vector (e.g., tens to hundreds of dimensions) as described above, the measurement granularity (information granularity) in expressing a given attribute (personality) is sufficiently high (fine). For example, in the case of personality, the information granularity is much higher than that of conventional personality indices such as the Big Five (which express personality in five dimensions).

Below, the above will be explained in more detail, assuming that the "target" is a user (human) and the "predetermined attribute" is personality. Of course, the present invention does not limit the "target" and "predetermined attribute" to those mentioned above; for example, the "target" could be a product and the "predetermined attribute" could be the overall performance or quality of that product (in this case, for example, the product's manufacturing version, time, usage history, environment, etc. could be used as related indicators). However, for ease of explanation, we will limit it to the above.

For a more detailed explanation, we will use the user attribute classification graph shown in the upper part of Figure 1, which classifies user attributes. In general, user attributes (in a broad sense, including personality) are:
Attributes can be classified into four groups based on whether they are (a) dynamic (time-dependent or time-varying) or static (time-independent or time-unchangeable) and (b) domain-dependent (affecting only a specific behavioral domain or relating to a specific behavioral domain) or domain-independent (can affect behavior in various behavioral domains or not relating to only a specific behavioral domain).

Specifically, the four groups are groups related to the first to fourth quadrants in the user attribute classification graph (top of Figure 1) consisting of the horizontal axis related to (a) above and the vertical axis related to (b) above. Incidentally, personality, which is one of the user attributes, is static and domain-independent, so it belongs to the group related to the first quadrant.

Also in FIG. 1, the user attribute representation vector r_u generated and output by the user attribute information generation model 2 is described in detail later, but it represents the user attributes belonging to the first quadrant of the above user attribute classification graph. The user attributes belonging to this first quadrant include personality as their main component, but also include other user attributes such as (user) gender, age, occupation, income, and place of residence (to the extent that they affect behavior in each behavioral domain).

Thus, the user attribute expression vector r_u output by the user attribute information generation model 2 can be treated as an expression vector that represents the user's personality, but it is highly desirable to further improve the reliability of this expression.

On the other hand, the personality expression vector M(r_u) generated and output by the target attribute expression generation model 1 is obtained as a result of the mapping unit 10, trained as described above in (C), applying mapping M to such a user attribute expression vector r_u, and is therefore a more reliable expression vector that expresses the user's personality. In other words, the personality expression vector M(r_u) is an expression vector that separates and extracts (or emphasizes) personality components from the user attributes belonging to the first quadrant in the above user attribute classification graph.

Incidentally, the personality embedding vector that represents the user's personality was created independently by the inventor of this application.

Here, in word embedding representation, which is an essential technique in natural language processing using artificial intelligence, each word is represented by a word embedding vector of, for example, several hundred dimensions (a string of numerous numerical values), and the closer the meaning of words are, the smaller the degree of separation between these vectors. In other words, a word embedding vector can be thought of as having the meaning of the corresponding word embedded in it. It is also possible to perform calculations between word embedding vectors, and one famous example is addition and subtraction, such as "king" - "man" + "woman" = "queen."

In a similar manner, the personality expression vector (personality embedding vector) created by the inventors of the present application can be considered to have the personality characteristics of the corresponding user embedded in it, so it is expected that even more useful information regarding personality can be derived and generated by using the inverse of the distance between vectors or the dot product as a measure of personality similarity (between users) or by performing calculations between vectors.

Furthermore, although the personality embedding vector as a distributed representation cannot be directly interpreted by humans, it is expected that it can be used, for example, as input data in machine learning models for various applications, enabling artificial intelligence to gain a detailed understanding of a user's personality and more accurately predict the user's behavior and personalize services.

Next, we will use Figure 2 to provide a more detailed explanation of the training in the mapping unit 10 described above in (C).

Figure 2 is a schematic diagram for explaining one embodiment of the mapping process performed by the mapping unit 10 of the object attribute representation generation model 1.

According to the embodiment shown in FIG. 2, the mapping M performed by the mapping unit 10 trained as described above in (C) in the target attribute representation generation model 1 is a mapping (M: V_u → V_p) from the user attribute representation vector space V_u spanned by the user attribute representation vectors r_u for various users (output from the user attribute information generation model 2) to the personality representation vector space V_p spanned by the personality representation vectors M(r_u) for the various users (generated by the target attribute representation generation model 1).

In this embodiment, the training of the mapping unit 10 in (C) above uses the following as the "strength index of a predetermined relationship":
(a) Blood relationship index: An index is adopted that is stronger for subjects who are related by blood, or for subjects who are closely related by blood. This blood relationship index was set based on the findings in personality psychology that "personality is half determined by genetics," that "people who are related by blood (have a genetic connection) tend to have similar personalities," and that "the more genes they have in common, the more similar their personalities will be (for example, identical twins are more similar in personality than fraternal twins)."

In fact, based on the above findings, it can be understood that the blood relationship index is an index that is positively correlated or has the potential to be positively correlated with the degree of similarity between users in terms of personality (in the sense that the stronger the index, the greater the degree of similarity tends to be).

In this embodiment, the mapping unit 10 is trained so that, with respect to such blood relationship indices, the stronger the users associated with the two personality expression vectors M(r_u) obtained by mapping, the smaller the degree of separation between these two personality expression vectors M(r_u) (e.g., the distance defined in the space V_p).

As a result, as shown in Figure 2, for example, the three user attribute representation vectors r_u of users "Father 1", "Elder Brother 1", and "Younger Brother 1", who are related by blood, are separated from each other in the user attribute representation vector space V_u (because they do not only represent personality), but the three corresponding personality representation vectors M(r_u) (of users "Father 1", "Elder Brother 1", and "Younger Brother 1") after mapping M are applied are close to each other in the personality representation vector space V_p (reflecting their blood relationship). Also, users "Father 2", "Older Sister 2", and "Younger Brother 2", who are related by a different blood relationship, have a similar behavior.

In this way, by acquiring the blood relationship information of the user related to the expression vector to which mapping M is applied, and training the mapping unit 10 that performs mapping M based on the above blood relationship index, the personality expression vectors M(r_u) of users with similar personalities can be made to be close to each other in the personality expression vector space V_p. In other words, it becomes possible to generate a personality expression vector M(r_u) that expresses the user's personality with higher reliability.

The mapping unit 10 that performs such a mapping M may, for example, apply (multiply) a matrix as a mapping operator to the input user attribute expression vector r_u, or may add a bias vector after multiplying the matrix. The value thus obtained may also be input to a sigmoid function, tanh function, ReLu function, or the like, and a function value may be output. Furthermore, it may be a forward propagation type neural network, particularly a multi-layer perceptron.

In addition, such training (learning) of the mapping unit 10 can be performed, for example, by setting a loss function so that the degree of separation between the representation vectors M(r_u) after mapping M is smaller for blood-related users than for unrelated users, and the difference in the degree of separation between the two is as large as possible.

<Loss function reflecting the presence or absence of blood ties>
For example, the loss function for a specific user u1 is expressed as follows: loss(u1, u2, u3)
=Σ _u2 Σ _u3 max[0, margin－M(r_u1)・M(r_u2)＋M(r_u1)・M(r_u3)]
may be adopted. Here, u2 is a blood relative of user u1, and u3 is a non-blood relative of user u1. M(r_ui) is the personality expression vector of user u1. Furthermore, Σ _u2 is the sum of all blood relatives of user u1, and Σ _u3 is the sum of all non-blood relatives of user u1. Furthermore, max[0, α] is a function that takes α if α is positive and 0 if α is non-positive, and margin is a positive constant for adjusting the value after the brackets [].

The loss function loss in the above formula (1) takes a smaller value as the inner product M(r_u1)·M(r_u2) between the user u1's relatives becomes larger than the inner product M(r_u1)·M(r_u3) between the user u1's non-relatives. Here, this inner product is an amount that can be considered the reciprocal of the degree of separation between the representation vectors M(r_u) after mapping M, and therefore the loss function loss is a function suitable for training the mapping unit 10.

The loss function that can be used to train the mapping unit 10 is not limited to the above formula (1). For example, in the above formula (1), the inner product is used as a quantity related to the degree of separation between the representation vectors M(r_u) after mapping M, but it is also possible to set the loss function using cosine similarity, Euclidean distance, or the like instead. In addition, the blood-related user u2 and the non-blood-related user u3 do not need to be all blood-related users or non-blood-related users of the user u1, but may be a part of them.

In addition, in the training (learning) of the mapping unit 10, in order to calculate the loss function as described above, it is necessary to obtain information on the user's blood ties. However, once the training is completed, the blood ties information is no longer necessary when deriving the personality expression vector M(r_u).

<Loss function reflecting the strength (depth) of blood ties>
The loss function loss in the above formula (1) is, so to speak, a loss function that reflects the presence or absence of a blood relationship between users, but as a modification, it is also possible to set a loss function that reflects the strength of the blood relationship between users. For example, in terms of blood relationship, since identical twins' genetic information matches each other, it can be assumed that the similarity of their personalities is the highest or most likely to be the highest. On the other hand, cousins, uncles and aunts, etc. are of course blood relatives, but the proportion of shared genetic information is lower than that of siblings and parents and children, and it is considered that the similarity of their personalities is relatively low. Therefore, it is possible to set a loss function based on the fact that the closer the blood relationship, the greater (tends to be) the similarity of their personalities.

Specifically, the loss function loss' is expressed as follows: loss(u1,u2,u2')
=Σ _u2 Σ _u2' max[0, margin－M(r_u1)・M(r_u2)＋M(r_u1)・M(r_u2')]
Here, both u2 and u2' are blood relatives of user u1, but u2 is a user who is more closely related to u1 than u2'. Also, Σ _u2 is the sum of all users u2 (blood relatives who are more closely related to u1), and Σ _u3 is the sum of all users u2' (blood relatives who are not as closely related to u1 as u2).

In this way, by reflecting the strength of the blood relationship in the loss function used to train (learn) the mapping unit 10, it is possible to further increase the reliability of the personality expression vector M(r_u) derived from the mapping unit 10 after training. In other words, it is possible to generate an expression vector that is expected to be more accurate as a representation of personality.

<Use of relationships other than blood ties>
The above describes the training (setting of the loss function) using the blood relationship index (A) as the "predetermined relationship strength index", but the "predetermined relationship strength index" is not limited to this blood relationship index. For example,
(a) Friend/acquaintance relationship index: an index that considers users who are friends or acquaintances to be stronger, or users who have a deeper friend/acquaintance relationship to be stronger,
(c) Closeness index: An index that gives a stronger relationship between users who are close to each other, or between users who are closer to each other.
(D) Compatibility index: An index that makes users who have good compatibility stronger, or users who have good compatibility stronger;
(E) Gene sequence index: an index that indicates that users who have the same specific gene sequence, or users who have similar specific gene sequences, are stronger; or
(F) EEG feature index: An index that gives a stronger relationship to users with similar features extracted from EEG can be adopted as a "predetermined relationship strength index." In other words, in any of the above (A) to (F), it is possible to set a loss function such as the above formula (1) relating to the presence or absence of the relationship, or the above formula (2) relating to the strength of the relationship.

By providing a range of options for the "specified relationship strength index" in this way, for example, if information on the user's blood ties cannot be obtained, it is possible to obtain information on the user's friend and acquaintance relationships, which is available, and apply this information to training the mapping unit 10.

Here, the "depth of a friend/acquaintance relationship" in the friend/acquaintance relationship index in (i) above can be a quantification of information such as the frequency of communication on a specific SNS (Social Networking Service) (e.g., frequency of message exchange or chat), the form of communication (e.g., non-real-time message exchange only, real-time chat, video calls, etc.), and the duration of communication (the period during which the "friend" relationship was registered, the total duration of calls, etc.). In other words, it can also be generated automatically from the SNS information.

Moreover, more than one of the above (a) to (f), for example all of them, may be adopted as the "strength index of a specified relationship." For example, it is also possible to set a loss function taking into account both the strength of the blood relationship index and the strength of the gene sequence index.

The strength indices (i) to (d) above are set based on the knowledge that "people with similar personalities are more likely to become friends, acquaintances, or close friends, and tend to get along well with each other." Furthermore, the strength indices (e) and (f) above can be set based on the results of recent research into the relationship between personality and biometric information, as described below.

For example, the human dopamine D4 receptor gene contains a repeat sequence consisting of a specific 48-base unit, and the number of repeats varies from person to person (polymorphism) between 2 and 11 times. Furthermore, it is known that there are many SNPs (single nucleotide polymorphisms) within the repeat sequence. Two research groups in the field have reported that people with a higher number of repeats tend to be more curious and fearless, and that mice with knockout dopamine D4 receptors have a lower curiosity and are less likely to exhibit exploratory behavior. Furthermore, it has been revealed that there are eight types of polymorphisms in the dog dopamine D4 receptor gene, caused by deletions or insertions of certain sequences. For example, it is known that there are corresponding differences in genotype between aggressive Shiba Inu and docile golden retrievers.

Furthermore, numerous studies have shown that features extracted from brain waves show significant correlations with specific personality traits.

Based on the research results and findings described above, it is also preferable to employ the gene sequence index (e) and the electroencephalogram feature index (f) above, measure the user's gene sequence and electroencephalogram feature in advance, and then train the mapping unit 10 so that the degree of separation between corresponding personality expression vectors is smaller for users who, for example, have similar dopamine D4 receptor repetition numbers or repeat sequence patterns, or who have similar electroencephalogram feature values that show a correlation with personality.

All of the "specified relationship strength indexes" (a) to (f) described above are indices that are positively correlated or have the potential to be positively correlated with the degree of personality similarity between users (in the sense that the stronger the index, the greater the degree of similarity tends to be). In this case, the mapping unit 10 can be trained so that the stronger the relationship between the users associated with the two personality expression vectors obtained by mapping, the smaller the degree of separation between these two personality expression vectors becomes with respect to such a "specified relationship strength index."

On the other hand, unlike the above (A) to (F), it is also possible to adopt a "predetermined relationship strength index" that is negatively correlated or has the potential to be negatively correlated with the degree of similarity between users in terms of personality (in the sense that the stronger the index, the smaller the degree of similarity tends to be). For example,
(G) Bad relationship index: An index may be adopted that equates users who do not get along well with each other, or users who do not get along well with each other, with a higher level of strength.

When adopting a "specified relationship strength index" related to such a negative correlation, the mapping unit 10 can be trained so that the degree of separation between the two personality expression vectors obtained by mapping becomes smaller the weaker the users related to the two personality expression vectors obtained by mapping are in terms of this "specified relationship strength index."

As a further modification of the "specified relationship strength index," it is also possible to employ a "specified relationship strength index" that does not show a positive or negative correlation with the degree of personality similarity, but that is related or potentially related to the degree of personality similarity. For example, a "specified relationship strength index" that maximizes or minimizes the degree of personality similarity when it takes a specific index value can also be used to train the mapping unit 10.

<Using user attributes other than personality>
As described above, it has been possible to train (learn) the mapping unit 10 using various types of "predetermined relationship strength indexes." Here, as another preferred form of training, training using user attributes other than personality, such as place of residence, occupation, age, etc., will be described.

First, a case will be described in which the "indices of strength of a predetermined relationship" relating to a positive correlation such as those in (a) to (f) above are used. In this case, the mapping unit 10:
(a) It is also preferable that two personality expression vectors relating to users who have a stronger relationship with respect to such a "specified relationship strength index" and who have a different or distant relationship with respect to a user attribute other than personality (e.g., place of residence) are trained to reduce the degree of separation between these two personality expression vectors.

For example, when selecting training data (including user attribute expression vectors) related to related users, it is also preferable to select training data related to users who live in different or distant places (as user attributes other than personality) as much as possible.

Moreover, the mapping unit 10
(b) It is also preferable that two personality expression vectors relating to users who are weaker in terms of such a "specified relationship strength index" and who have the same or a close relationship in terms of a user attribute other than personality (e.g., place of residence) are trained to have a greater degree of separation between these two personality expression vectors.

For example, when selecting training data (including user attribute expression vectors) related to unrelated users during training, it is also preferable to select training data related to users who have the same or nearby place of residence (as a user attribute other than personality) whenever possible.

Here, the mapping unit 10 may undergo either or both of the above-mentioned training (a) and training (b). In any case, by selecting a user attribute that is not generally related to personality (considered to be not related to personality) as a user attribute other than personality and using it for training, it becomes possible to reduce or remove the expression of the user attribute other than personality in the personality expression vector. In other words, it is possible to derive a more reliable personality expression vector.

For example, even if the mapping unit 10 is trained to separate two personality expression vectors associated with users with different personalities, if the users live in different places (e.g., Hokkaido and Okinawa, respectively), the degree of separation between these two personality expression vectors may reflect the difference in place of residence. In response to this, by selecting users who live in the same place of residence (e.g., both in Hokkaido) as users associated with such two personality expression vectors, it is possible to reduce or eliminate the expression of place of residence in these two personality expression vectors.

Next, a case where the "indicator of strength of a predetermined relationship" related to the negative correlation as in (G) above is used will be described. In this case, the mapping unit 10
(c) It is also preferable that two personality expression vectors relating to users who are weaker in terms of such a "specified relationship strength index" and who have a different or distant relationship in terms of a user attribute other than personality (e.g., place of residence) are trained to reduce the degree of separation between these two personality expression vectors.

Moreover, the mapping unit 10
(d) It is also preferable that two personality expression vectors relating to users who have a stronger relationship with respect to such a "specified relationship strength index" and who have the same or a similar relationship with respect to a user attribute other than personality (e.g., place of residence) are trained to have a greater degree of separation between these two personality expression vectors.

Furthermore, the mapping unit 10 may undergo either or both of the training (c) and the training (d) above. In any case, by carrying out such training, the representation of user attributes other than the personality in the personality expression vector is reduced or removed, making it possible to derive a more reliable personality expression vector.

[One embodiment of a user attribute information generation model]
FIG. 3 is a schematic diagram showing an embodiment of a user attribute information generation model according to the present invention.

The user attribute information generation model 2 of this embodiment shown in FIG.
(a) Using “behavior domain information” which is information on multiple behavior domains (two in FIG. 3: “item purchase” and “video viewing”) of a user whose user attributes are to be estimated,
(b) A machine learning model that generates a “user attribute expression vector”, which is information expressing the user attributes of the user to be estimated.

Specifically, the user attribute information generation model 2 is
(A) A plurality of domain-specific regression neural network (RNN) cells (two in FIG. 3 , a DSL (Domain Specific Layer) 1 cell 21 and a DSL 2 cell 22) are set for each of a plurality of behavioral domains, and receive “domain behavior information” of a user to be estimated in the behavioral domain, and generate new “domain-specific hidden state information” by reflecting the “domain behavior information” on “domain-specific hidden state information” that is hidden state information generated by the RNN cells themselves at a previous point in time;
(B) a user expression generation unit (in FIG. 3, a user distributed expression extraction unit 20u) that generates a “user attribute expression vector” that is information expressing a user to be estimated from “user identification information” that is information identifying the user to be estimated;
(C) This is a machine learning model that causes a computer to function as a domain-independent RNN cell (in FIG. 3 , DIL (Domain Independent layer) 20) that receives the “domain-specific hidden state information” generated in (A) above and the “user attribute representation vector” generated in (B) above, and generates new “domain-independent hidden state information” by reflecting the “domain-specific hidden state information” and the “user attribute representation vector” on the “domain-independent hidden state information” that is the hidden state information that the computer itself generated at a previous point in time.

In this embodiment, the "behavior domain information" in (A) above is behavior expression information generated by a behavior expression generation unit (in FIG. 3, the item dispersed representation extraction unit 21i and the video dispersed representation extraction unit 22m) that receives information indicating the content of the behavior (such as purchase or viewing) of the estimated target user in the behavior domain (for example, identification information of a purchased item or a viewed video). Therefore, "behavior domain information" (behavior expression information) is generated each time a behavior occurs in each behavior domain, and as a result, the information as a whole forms a time series data group.

The user expression generation unit (user distributed expression extraction unit 20u) in (B) above is trained together with multiple domain-specific RNN cells (DSL1 cell 21 and DSL2 cell 22), which will be described in detail later, and as a result, after the training, it executes a domain-independent expression generation calculation (distributed expression extraction calculation) that is not biased toward a specific behavioral domain. As a result, the "user attribute expression vector" generated here is understood as the user attribute information of the user to be estimated.

In other words, the "user attribute representation vector" and the above-mentioned "domain-independent hidden state information" are made domain-independent by receiving "domain-specific hidden state information" from cells related to multiple or many different behavioral domains, and among these, the "user attribute representation vector" in particular can be considered as information related to static and domain-independent user attributes of the estimation target user, or, in terms of the user attribute classification graph at the top of Figure 1, user attribute information belonging to the first quadrant. Therefore, the "user attribute representation vector" is truly a representation of static and domain-independent user attributes including "personality".

Furthermore, this "user attribute expression vector" is a multidimensional (e.g., tens to hundreds of dimensions) expression vector, e.g., a list of many numerical values, output by the user expression generation unit (user distributed expression extraction unit 20u) that receives the "user identification information" of the user to be estimated (e.g., a one-hot vector indicating the user), and therefore the "user attribute expression vector" has a very high (fine) measurement granularity (information granularity) regarding the user attributes.

The domain-specific RNN cells (DSL1 cell 21 and DSL2 cell 22) and domain-independent RNN cell (DIL cell 20) described above can be configured with well-known RNNs such as a GRU (Gated Recurrent Unit) or a LSTM (Long-Short Term Memory).

In addition, in the user attribute information generation model 2, the number of domain-specific RNN cells, i.e., the number of behavioral domains handled, is naturally not limited to two as shown in FIG. 3, and can be three or more. In addition, the behavioral domains handled are not limited to the item purchases and video viewing shown in FIG. 3, and various domains related to behavior that depend on or can be influenced by the user attributes of the behavior subject can be adopted as behavioral domains. For example, when time-series data of the behavioral domain "ad click" is acquired, a domain-specific DNN cell that incorporates domain behavior information related to "ad click" may be added to configure the user attribute information generation model 2.

The configuration of the user attribute information generation model 2 of this embodiment will be described in more detail below. As shown in FIG. 3, the user attribute information generation model 2 includes:
(A) A DSL1 cell 21 as a domain-specific RNN cell related to the behavior domain “item purchase”, a user distributed representation extraction unit 21u, an item distributed representation extraction unit 21i, and an output layer 21o;
(a) A DSL2 cell 22 as a domain-specific RNN cell related to the behavior domain “video viewing”, a user distributed expression extraction unit 22u, a video distributed expression extraction unit 22m, and an output layer 22o;
(c) The DIL cell 20 as a domain-independent RNN cell, a user distributed expression extraction unit 20u as a user expression generation unit, and an output layer 20o are included as functional components embodied by a computer (a program installed in the computer).

Figure 3 shows the processing executed by the functional components (a) to (c) (the functional blocks on the left side of Figure 3) as described above, deployed on a time axis with time progression pointing to the right. Note that the three time axes set for each of (a) to (c) above are time axes for points in time when each has its own unique value.

As for the notation of time points on each time axis, for example, time point (t2+1) for the DSL2 cell 22 means the time point at which the DSL2 cell 22 next receives domain behavior information (video embedding vector) after receiving domain behavior information (video embedding vector) at time point t2. Here, the next reception time point is naturally (t2+2). Furthermore, from the perspective of time point (t2+1), time point t2 is the "previous time point." Furthermore, for the convenience of the following explanation, the DSL2 cell 22 that performs processing at this time point (t2+1) will be referred to as the DSL2 cell 22(t2+1). Furthermore, similar notations of processing time points and notations including the processing time will be used hereafter for the DSL1 cell 21 and the DIL cell 20.

Hereinafter, each of the above-mentioned functional components will be specifically described. In FIG. 3, the item distributed expression extraction unit 21i(t1), which is the item distributed expression extraction unit 21i at time t1, receives an event that "item item1 has been purchased by user UserA (the user to be estimated)" at time t1,
(a) Receive a one-hot vector i1 ^(t1) which is the item identification information of item item1,
(b) The received one-hot vector i1 ^(t1) is applied (multiplied) by the matrix W _i ^DSL1 as an item distributed representation extraction operator to generate an item representation vector (domain behavior information) r_i1 ^(t1) for item item1.

In addition, the user distributed expression extraction unit 21u(t1), which is the user distributed expression extraction unit 21u at the time t1, receives the one-hot vector i1 ^(t1) in (a) above,
(c) receiving a one-hot vector u a ^(t1) which is the user identification information of user UserA;
(d) The received one-hot vector ua ^(t1) is applied (multiplied) by the matrix W _u ^DSL1 as a user distributed expression extraction operator to generate a user attribute expression vector r_ua ^(t1) of user UserA.

Similarly, in FIG. 3, the DSL1 cell 21 (t1) at time t1 is
(e) receiving the item representation vector (domain behavior information) r_i1 ^(t1) generated in (b) above and the user attribute representation vector r_ua ^(t1) generated in (d) above,
(f) (which will be described in detail later) also receives domain-independent hidden state information h DIL,( ^{tI) that was recently generated (at time tI in FIG. 3) from time t1 when the domain action information (r_i1 ( t1)) of (e)} is received,
(g) Generate new domain-specific hidden state information h ^DSL1,(t1) by reflecting the received item representation vector (domain behavior information) r_i1 ^(t1) , user attribute representation vector r_ua ^(t1) , and domain-independent hidden state information h ^DIL,(tI) to the domain-specific hidden state information h ^{DSL1,(t1-1) generated by itself at the previous point in time (t1-1)} .

The above describes the processing at time t1 in the item shared representation extraction unit 21i, the user shared representation extraction unit 21u, and the DSL1 cell 21. However, of course, the same processing as (a) to (g) above is also performed for processing at other times, for example, the processing at the item shared representation extraction unit 21i(t1+1), the user shared representation extraction unit 21u(t1+1), and the DSL1 cell 21(t1+1) relating to the next time (t1+1).

Furthermore, the same processes as those described above (a) to (g) are also executed for the video dispersed expression extraction unit 22m, the user dispersed expression extraction unit 22u, and the DSL2 cell 22 related to other behavioral domains (video viewing), for example, the video dispersed expression extraction unit 22m(t2), the user dispersed expression extraction unit 22u(t2), and the DSL2 cell 22(t2) related to time t2. However, in this case, for example, the video dispersed expression extraction unit 22m(t2) at time t2,
(b') The matrix W _m ^DSL2 as a video distributed expression extraction operator is applied (accumulated) to the received one-hot vector m1 ^(t2) (indicating the video Mov1 viewed by user UserA) to generate a video expression vector (domain behavior information) r_m1 ^(t2) of the video Mov1. The DSL2 cell 22 (t2) incorporates this domain behavior information r_m1 ^(t2) and performs processing.

Similarly, in FIG. 3, the user distributed expression extraction unit 20u(tI), which is the user distributed expression extraction unit (user expression generation unit) 20u at time tI,
(h) Receive a one-hot vector u a ^{(t I )} , which is the user identification information of user UserA,
(i) The received one-hot vector ua ^(tI) is applied (multiplied) by a matrix W _u ^DIL as a user distributed expression extraction operator to generate a user attribute expression vector r_ua ^(tI) of user UserA.

Furthermore, DIL cell 20(tI), which is the DIL cell 20 at time tI, is
(j) receiving the user attribute expression vector r_ua ^(tI) generated in (i) above,
(k) receiving domain-specific hidden state information h ^{DSL2,(t 2 )} generated most recently (at time t 2 in FIG. 3 );
(l) Generate new domain-independent hidden state information h ^DIL,(tI) by reflecting the received user attribute representation vector r_ua ^(tI) and the domain-specific hidden state information h ^DSL2,(t2) on the domain-independent hidden state information h ^DIL,(tI-1) generated by itself at the previous point in time (tI-1).

Note that the time tI related to the above-described processes (h) to (l) can be, for example, a time (one of) that is set periodically (every time a predetermined period of time has elapsed), but in this embodiment, a more preferable setting is the time immediately after domain-specific hidden state information is generated in one of the domain-specific RNN cells (in FIG. 3, either DSL1 cell 21 or DSL2 cell 22).

In other words, in the case described above, for example, when the DSL2 cell 22(t2) generates the domain-specific hidden state information h ^DSL2,(t2) at time t2, the above-mentioned processes (h) to (l) are initiated immediately thereafter (i.e., at time tI).Furthermore, as also shown in FIG. 3, when the DSL1 cell 21(t1) generates the domain-specific hidden state information h ^DSL1,(t1) at time t1, the DSL1 cell 21(t1) receives the domain-specific hidden state information h ^DSL1,(t1) immediately after it is generated (at time tI+1), and processes similar to the above-mentioned processes (h) to (l) are executed.

In this manner, in this embodiment, each time domain-specific hidden state information is generated in any domain-specific RNN cell, the DIL cell 20 receives the domain-specific hidden state information and generates "domain-independent hidden state information."

In FIG. 3, the user UserA, who is the subject of estimation, performs the following actions in this order: "viewing video Mov1 (at time t2)" → "purchasing item item1 (at time t1)" → "viewing video Mov2 (at time t2+1)". The DIL cell 20 sequentially receives domain-specific hidden state information generated by the domain-specific RNN cell which sequentially receives the expression vectors (domain behavior information) related to these actions. This takes into account the order in which these domain-specific hidden state information was generated (received) and the contextual information, making it possible to generate "domain-independent hidden states" that are not dependent solely on the content of actions in a specific behavioral domain (domain-independent).

Furthermore, in this embodiment, as described in (f) above, such characteristic "domain-independent hidden states" are appropriately used as inputs to domain-specific RNN cells (DSL1 cell 21 and DSL2 cell 22). As a result, the domain-specific hidden state information later received by DIL cell 20 also reflects the above-mentioned characteristics, which promotes the domain independence of the "domain-independent hidden states." In other words, DIL cell 20 can further perform information processing that reflects the overall attributes of the estimated target user, independent of a specific behavioral domain.

Next, we will explain the training (learning) of user attribute information generation model 2.

First, in training the DSL1 cell 21, the item distributed representation extraction unit 21i, and the user distributed representation extraction unit 21u, time series data of one-hot vectors, which are identification information of items purchased by each user for a large number of users, are prepared and input sequentially to the item distributed representation extraction unit 21i. In addition, the one-hot vector, which is the user identification information of the user, is input to the user distributed representation extraction unit 21u. As a result, the DSL1 cell 21 updates the domain-specific hidden state information at the previous time point at each time point to generate domain-specific hidden state information at the current time point, and in response, the output layer 21o outputs a prediction result of the purchased item (identification information of the item predicted to be purchased by the user at the next time point, or the probability of each candidate item being purchased) (in this embodiment, the output layer 21o is configured to perform such output).

Here, the difference between the output item prediction result and the correct item identification information (the one-hot vector of the item actually purchased) is considered as the loss, and training is performed using the well-known RNN error backpropagation method to adjust each parameter in DSL1 cell 21 (such as parameters that determine the weight matrix and bias vector) and the parameters that determine the matrix W _i ^DSL1 and matrix W _u ^DSL1 .

Training of the DSL2 cell 22, the video distributed expression extraction unit 22m, and the user distributed expression extraction unit 22u can be performed in the same manner as the training of the DSL1 cell 21, etc. Specifically, in this case, the difference between the prediction result of the viewed video output from the output layer 22o (identification information of the video predicted to be viewed by the user at the next time point, or the probability of viewing each candidate video) and the identification information of the correct video (one-hot vector of the video actually viewed) is treated as a loss, and training is performed to adjust each parameter in the DSL2 cell 22 (parameters determining the weight matrix and bias vector, etc.) and parameters determining the matrix W _m ^DSL2 and the matrix W _u ^DSL2 using a known RNN backpropagation method. Furthermore, training can be performed in the same manner when a third or subsequent domain-specific RNN cell system is set.

Next, in training the DIL cell 20 and the user distributed expression extraction unit 20u, the above-mentioned loss in each domain-specific RNN cell (each of the DSL1 cell 21 and the DSL2 cell 22) is back-propagated to the DIL cell 20 and the user distributed expression extraction unit 20u through the link between each domain-specific RNN cell and the DIL cell 20 to perform training. For example, in Fig. 3, the loss, which is the difference between the predicted result of the purchase item based on the input up to time t1 in the DSL1 cell 21 and the identification information of the correct item, is back-propagated through the link that transmits the domain-independent hidden state information hDIL ^,(tI) , and is used to train each parameter in the DIL cell 20 (parameters that determine the weight matrix and bias vector, etc.) and parameters that determine ^the matrix _WuDIL of the user distributed expression extraction unit 20u.

In other words, in this embodiment, training of the DIL cell 20 and the user distributed expression extraction unit 20u is performed only by backpropagation of errors from the domain-specific RNN cells (DSL1 cell 21 and DSL2 cell 22) without using teacher data unique to the domain-independent RNN cells.

The DIL cell 20 and the user distributed expression extraction unit 20u trained (learned) as described above receive domain-specific hidden state information from a plurality of domain-specific RNN cells (DSL1 cell 21 and DSL2 cell 22) as appropriate, and further become domain-independent in the process of providing each domain-specific RNN cell with appropriate domain-independent hidden state information. Here, in particular, the user distributed expression extraction unit 20u (its matrix W _u ^DIL ) receives user identification information and outputs information related to the user's static and domain-independent user attributes (i.e., including the user's personality).

[Another embodiment of the user attribute information generation model]
FIG. 4 is a schematic diagram showing another embodiment of a user attribute information generation model according to the present invention.

As shown in FIG. 4, the user attribute information generation model 3 of this embodiment is a model that uses matrix factorization to generate a user attribute expression vector for the user to be estimated, and outputs it to the target attribute expression generation model 1'. Similarly to the target attribute expression generation model 1, the target attribute expression generation model 1' is equipped with a mapping unit 10' trained using, for example, a blood relationship index, and applies a mapping M' to the received user attribute expression vector to generate a personality expression vector for the user.

The matrix factorization method used to generate the user attribute expression vector is a well-known technique, and is a representative example of the static URL technique, which uses non-time-series behavioral data, among the user expression learning (URL) techniques that have been actively researched, particularly in the field of product and service recommendation technology.

Specifically, in this embodiment, the user attribute information generation model 3 follows the matrix factorization method as follows:
(a) From the item purchase history, which is a record of the fact that each of a large number (m people) of users purchased each of various (n types) items, an m × n "item purchase matrix" is generated, in which the matrix elements are values indicating whether or not a user in each row purchased an item in each column;
(b) The generated “item purchase matrix” is decomposed into an m × k “user matrix” and a k × n “item matrix” (where 0 < k < m) by dimensionality reduction processing.
(c) In the "user matrix" obtained by the decomposition, a vector formed of the matrix elements for each row is output as a user attribute expression vector for the user corresponding to that row.

In Figure 4 (which shows a 4x5 "item purchase matrix" with an extremely small number of users and items for easy understanding), the vector formed by the matrix elements of the first row in the 4x3 "user matrix" is output as the user attribute representation vector r_p1 for user p1, which corresponds to this first row.

The user attribute representation vector r_p1 of user p1 output from user attribute information generation model 3 has not undergone dynamic/static, or even domain-dependent/domain-independent, separation or extraction of user attributes due to the nature of the matrix factorization method (static URL technology). In other words, it is an expression vector in a state where grouping into the first to fourth quadrants of the user attribute classification graph in the upper part of Figure 1 has not been completed.

The target attribute representation generation model 1' receives this user attribute representation vector r_p1 and applies a mapping M' to it, so that it can generate a personality representation vector M'(r_p1) that is the result of extracting (or revealing) the personality components of user p1 from this user attribute representation vector r_p1.

In this embodiment, the user attribute information generation model 3 generates a user attribute representation using an "item purchase matrix" related to the behavioral domain "item purchase." However, it is of course possible to use matrices related to other behavioral domains, such as "video viewing" or "advertisement clicking."

Furthermore, the user attribute expression vector input to the target attribute expression generation model according to the present invention is naturally not limited to a model that implements the above-mentioned matrix factorization method or one output from the above-mentioned user attribute information generation model 2 (Figure 3). For example, if the user attribute expression vector is from a "user attribute information generation model using URL technology" that can output information that includes the first quadrant in the user attribute classification graph in the upper part of Figure 1, the target attribute expression generation model that receives it can generate a personality expression vector.

[Personality estimation device, program, and method]
Returning to FIG. 1 , we will now explain the personality estimation device 9 that is equipped with a personality estimation model 8, which is a model that combines the user attribute information generation model 2 and the target attribute expression generation model 1 described in detail above, and that estimates the personality of the user to be estimated.

Specifically, in the functional block diagram at the bottom of FIG. 1, the input unit 91 of the personality estimation device 9 has a communication function, and acquires time-series information relating to the behavior of a large number of users in each behavioral domain, for example, from an externally installed behavioral domain-related management server (for example, a web shopping management server or a video distribution management server), and outputs it to the training unit 92. In addition, blood relationship information for these users is also acquired, for example, from an externally installed questionnaire survey result management server, and output to the training unit 92. Furthermore, the input unit 91 receives information relating to the estimation target user, and outputs it to the personality estimation unit 93.

The training unit 92 generates training data from the received time-series information related to the behavior of the multiple users in each behavioral domain, and uses this to train (learn) the user attribute information generation model 2 portion of the personality estimation model 8. It also generates training data from the received blood relationship information of these users and the user attribute expression vectors output from the user attribute information generation model 2 portion, and uses this to train (learn) the target attribute expression generation model 1 portion of the personality estimation model 8.

The personality estimation unit 93 generates user identification information (one-hot vector) of the estimated user from the received information related to the estimated user, inputs this to the trained (learned) personality estimation model 8 (the user attribute information generation model 2 part), and obtains the personality expression vector of the estimated user from the personality estimation model 8 (the target attribute expression generation model 1 part). As already explained, this personality expression vector can be regarded as a distributed expression of the estimated user's personality, and is personality information with high measurement granularity (fineness) and high reliability.

The output unit 94 transmits the received personality embedding vector to an external information processing device (if the device has a communication function) as personality information of the estimated target user, or displays the vector (if the device has a display function).

The training unit 92 and the personality estimation unit 93 are the main functional components that implement one embodiment of the personality estimation method according to the present invention, and can also be considered as functions of a processor/memory that stores one embodiment of the personality estimation program according to the present invention. As a result, the personality estimation device 9 may be a dedicated device for personality estimation, but it can also be, for example, a cloud server, a non-cloud server device, a personal computer (PC), a notebook or tablet computer, or a smartphone, etc., that is equipped with the personality estimation program according to the present invention.

Of course, the user attribute information generation model and the target attribute expression generation model constituting the personality estimation model 8 are not limited to those described above, and may be, for example, the user attribute information generation model 3 (Figure 4) and the target attribute expression generation model 1' (Figure 4).
In addition, the installed personality estimation model 8 may be the target attribute expression generation model itself, and the personality estimation device 9 may receive a user attribute expression vector from an external device equipped with a user attribute information generation model, and generate a personality expression vector.

As explained in detail above, according to the present invention, it is possible to map an expression vector of a target (e.g., a user) to another expression vector in another expression vector space, and in this case, a training process is performed for a strength index (e.g., a blood relationship index) that is (possibly) related to the degree of similarity in a specific attribute of the target (e.g., the user's personality) such that the greater the (possibly) the degree of similarity is, the smaller the degree of separation between the corresponding other expression vectors becomes. This makes it possible for the other expression vector after mapping to be an attribute representation that more reliably represents a specific attribute (personality) of the target (user).

For example, as one application example of the present invention, it is possible to generate personality information with higher information granularity and reliability from behavioral domain information of the estimated target user without using or relying on the results of personality measurements by questionnaire surveys that have been used in the past. Furthermore, by using the user's personality information generated in this way, for example in the field of marketing, it is possible to personalize the products and services provided and to implement, for example, suitable or effective recommendations.

In addition, for example, detailed and reliable personality information of children (e.g., estimated from their behavior) generated by the present invention can be used to provide children with high-quality education that is tailored to their individual personalities. In other words, the present invention can contribute to Goal 4 of the United Nations-led Sustainable Development Goals (SDGs), which is to "Ensure inclusive and equitable quality education and promote lifelong learning opportunities for all."

Furthermore, detailed and reliable personality information of adults (e.g., estimated from their behavior) generated by the present invention can be used to provide adults with decent work that does not harm the environment and high quality work that suits their individual personalities. In other words, the present invention can contribute to Goal 8 of the UN-led SDGs, "Promote inclusive and sustainable economic growth, employment and decent work for all."

Furthermore, for example, the detailed and highly reliable personality information (e.g., estimated from the content of the consumer's behavior) generated by the present invention, as well as the consumer's lifestyle history and consumption activity history, can be utilized to provide the consumer with education on sustainable consumption and lifestyles that are in line with the consumer's personality and current lifestyle. In other words, the present invention can also contribute to the achievement of Goal 12 of the United Nations-led SDGs, "Ensure sustainable consumption and production patterns."

With respect to the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical ideas and viewpoints of the present invention can be easily made by a person skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents.

1, 1' Object attribute expression generation model 10, 10' Mapping unit 2, 3 User attribute information generation model 20 DIL cell (Domain-independent RNN model)
20o, 21o, 22o Output layer 20u User distributed expression extraction unit (user expression generation unit)
21 DSL1 cell (domain specific RNN model)
21i Item distributed representation extraction unit 21u, 22u User distributed representation extraction unit 22 DSL2 cell (domain-specific RNN model)
22m Video distributed expression extraction unit 8 Personality estimation model 9 Personality estimation device 91 Input unit 92 Training unit 93 Personality estimation unit 94 Output unit

Claims (11)

An object attribute representation generation model that causes a computer to function to map a representation vector of an object to another representation vector that is an element of another representation vector space,
A mapping unit that applies a trained mapping operator to the input expression vector to generate another expression vector and outputs the another expression vector as information related to the predetermined attribute of the target , or that is a trained neural network algorithm that receives the expression vector as an input and outputs the another expression vector as information related to the predetermined attribute of the target.
The computer functions as
The mapping operator or the neural network algorithm is trained such that , with respect to a strength index of a predetermined relationship between the objects, which is related or may be related to the degree of similarity between the objects with respect to the predetermined attribute, the greater the strength index of the predetermined relationship is, or the greater the likelihood of it being able to increase the degree of similarity between the objects associated with the two different representation vectors obtained by mapping, the smaller the degree of separation between the two different representation vectors is .
A target attribute representation generation model characterized by : The object attribute representation generation model of claim 1, characterized in that the mapping operator or the neural network algorithm is trained such that the stronger the objects related to the two different representation vectors obtained by mapping are, the smaller the degree of separation between the two different representation vectors becomes, with respect to a strength index of the specified relationship that is positively correlated or likely to be positively correlated with the degree of similarity between the objects with respect to the specified attribute. The mapping operator or the neural network algorithm is
Two separate expression vectors relating to objects that are stronger in terms of the strength index of the predetermined relationship, and that are different or distant in terms of an attribute other than the predetermined attribute, are trained to have a smaller degree of separation between the two separate expression vectors; and/or
The object attribute representation generation model according to claim 2, characterized in that it is trained so that the degree of separation between two different representation vectors relating to objects that are weaker in terms of the strength index of the specified relationship, and that have the same or a close relationship with respect to an attribute other than the specified attribute, is greater. The object attribute representation generation model according to any one of claims 1 to 3, characterized in that the object is a human being, the predetermined attribute is personality, and the predetermined relationship strength index is an index that is stronger between objects related by blood or objects with a closer blood relationship, an index that is stronger between objects related by friends or acquaintances or objects with a closer friend or acquaintance relationship, an index that is stronger between objects related by closeness or objects with a closer relationship, an index that is stronger between objects with good compatibility or objects with good compatibility, an index that is stronger between objects with the same specific gene sequence or objects with similar specific gene sequences, and/or an index that is stronger between objects with similar features extracted from electroencephalograms. The object attribute representation generation model of claim 1, characterized in that the mapping operator or the neural network algorithm is trained such that the weaker the objects related to the two different representation vectors obtained by mapping are, the smaller the degree of separation between the two different representation vectors becomes, with respect to a strength index of the specified relationship that is negatively correlated or may be negatively correlated with the degree of similarity between the objects for the specified attribute. The mapping operator or the neural network algorithm is
Two separate expression vectors relating to objects that are weaker in terms of the strength index of the predetermined relationship, and that are different or distant in terms of an attribute other than the predetermined attribute, are trained to have a smaller degree of separation between the two separate expression vectors; and/or
The object attribute representation generation model of claim 5, characterized in that the model is trained so that the degree of separation between two different representation vectors relating to objects that are stronger in terms of the strength index of the specified relationship and that have the same or a close relationship in terms of an attribute other than the specified attribute becomes greater. the subject is a user, and the predetermined attribute is a personality;
The expression vector is a user attribute expression vector that is generated by a user attribute information generation model and is information that expresses a user attribute including the personality of the user,
The user attribute information generation model is
A plurality of domain-specific regression neural network (RNN) cells are set for each of a plurality of behavioral domains, and the domain-specific RNN cells receive domain behavior information, which is information related to the behavior of the user in the behavioral domain, and generate new domain-specific hidden state information by reflecting the domain behavior information on domain-specific hidden state information, which is hidden state information generated by the RNN cells themselves at a previous point in time;
a user representation generator trained in conjunction with the plurality of domain-specific RNN cells to generate a user attribute representation vector of the user from information identifying the user;
The object attribute representation generation model according to any one of claims 1 to 6, characterized in that the model causes a computer to function as a domain-independent RNN cell that receives the generated domain-specific hidden state information and the generated user attribute representation vector, and generates new domain-independent hidden state information by reflecting the domain-specific hidden state information and the user attribute representation vector on domain-independent hidden state information, which is hidden state information generated by the computer itself at a previous point in time. the subject is a user, and the predetermined attribute is a personality;
The target attribute representation generation model according to any one of claims 1 to 6, characterized in that the representation vector is a user attribute representation vector that expresses the user attributes of the user and is calculated by a method related to User Representation Learning (URL) technology using information related to a predetermined behavior of the user. The target attribute expression generation model described in claim 8, characterized in that the method related to the user expression learning technology is a matrix factorization method. A target attribute estimation device that generates the different expression vector using a target attribute expression generation model according to any one of claims 1 to 9, and outputs the generated different expression vector as information related to the predetermined attribute of the target. Constructing the object attribute representation generation model according to any one of claims 1 to 9 by performing training to obtain a trained mapping operator or a trained neural network algorithm ;
A target attribute estimation method implemented by a computer, characterized in that a different expression vector is generated using the constructed target attribute expression generation model, and the generated different expression vector is output as information related to the specified attribute of the target.

Images