JP7611867B2 - USER BEHAVIOR PREDICTION DEVICE, USER BEHAVIOR PREDICTION METHOD, AND USER BEHAVIOR PREDICTION PROGRAM - Google Patents

Description

The present invention relates to a user behavior prediction device, a user behavior prediction method, and a user behavior prediction program, and in particular to a user behavior prediction device, a user behavior prediction method, and a user behavior prediction program that predicts user behavior in a target domain (also called a goal domain) using behavior history data from a source domain.

There is a technology that predicts user behavior in a target domain using behavior history data from the original domain (Non-Patent Document 1). Here, a domain refers to, for example, a service related to user behavior, such as e-commerce.

Li, B., Yang, Q., & Xue, X. (2009, June). Can movies and books collaborate? cross-domain collaborative filtering for sparsity reduction. In Twenty-First international joint conference on artificial intelligence.

However, in Non-Patent Document 1, there was a problem in that it was not possible to select a behavior predictor that was useful for the target domain from those trained in the original domain.

The present invention aims to realize user behavior prediction in a target domain by selecting a part of a behavior predictor trained in the original domain that is useful for the target domain.

The user behavior device according to the present invention includes a first behavior predictor learning unit that learns a first behavior predictor having a first expression learning unit and a first predictor for predicting an unobserved behavior of a first user from a behavior of a first user with respect to a first item that constitutes first behavior history data;
an expression extraction unit that extracts a first vector expression from the first expression learning device of the first behavior predictor and outputs a statistical value of the first vector expression as a first expression profile;
a predictor extraction unit that extracts the first predictor of the first behavior predictor as a predictor;
a second behavior predictor learning unit that learns a second behavior predictor having a second expression learner and a second predictor for predicting an unobserved behavior from a behavior of a second user with respect to a second item constituting second behavior history data, and in this learning, freezes parameters of the predictor extracted by the predictor extraction unit and uses them as the second predictor of the second behavior predictor, and additionally learns the second expression learner;
an expression extraction and comparison unit that extracts a second vector expression from the second expression learning device of the second behavior predictor, sets a statistical value of the second vector expression as a second expression profile, compares the second expression profile with the first expression profile, and evaluates the usefulness of the second behavior predictor based on a similarity between the first and second expression profiles;
a behavior prediction unit that outputs a prediction result based on the second behavior predictor when the second behavior predictor is evaluated to be highly useful by the expression extraction comparison unit;
The user behavior prediction device includes:

the first behavior predictor learning unit, the expression extraction unit, and the predictor extraction unit each perform operations in a plurality of different domains;
The behavior prediction unit may ensemble prediction results based on the second behavior predictors evaluated as having high usefulness and output the ensembled prediction results as prediction results.

The behavior prediction unit may use majority voting as the ensemble method.

The user behavior prediction device may further include an expression profile storage unit that stores the first expression profile, and a predictor storage unit that stores the extracted predictor.

The first expression profile and the second expression profile may be generated by extracting statistical values from the first vector representation and the second vector representation corresponding to a portion of the samples.

The first expression profile and the second expression profile may be generated by clustering the first vector representation and the second vector representation as statistical values and determining the representative points and the number of samples belonging to the cluster.

The number of samples may be a predetermined number or more.

The expression extraction and comparison unit may evaluate the inter-distribution dissimilarity between the first expression profile and the second expression profile.

The expression extraction and comparison unit may use the transport distance between the first expression profile and the second expression profile as the inter-distribution dissimilarity.

A user behavior prediction method according to the present invention is a user behavior prediction method for a user behavior prediction device that uses first behavior history data of a first user to predict a behavior of a second user based on second behavior history data of the second user, the method comprising:
training a first behavior predictor having a first representation learner and a first predictor for predicting an unobserved behavior of the first user from the behavior of the first user with respect to a first item constituting the first behavior history data;
extracting a first vector representation from the first representation learner of the first behavior predictor and outputting statistics of the first vector representation as a first representation profile;
Extracting the first predictor of the first behavior predictors as a predictor;
A second behavior predictor having a second representation learner and a second predictor is trained to predict an unobserved behavior from the behavior of the second user with respect to a second item constituting the second behavior history data, and in this training, parameters of the predictor extracted are frozen and used as the second predictor of the second behavior predictor, and the second representation learner is additionally trained;
extracting a second vector representation from the second representation learner of the second behavior predictor, defining a statistical value of the second vector representation as a second representation profile, comparing the second representation profile with the first representation profile, and evaluating the usefulness of the second behavior predictor based on a similarity between the first and second representation profiles;
The user behavior prediction method outputs a prediction result based on the second behavior predictor when the usefulness of the second behavior predictor is evaluated to be high.

The user behavior prediction program according to the present invention is
On the computer,
A process of learning a first behavior predictor having a first representation learner and a first predictor for predicting an unobserved behavior of a first user from a behavior of the first user with respect to a first item constituting the first behavior history data;
extracting a first vector representation from the first representation learner of the first behavior predictor and outputting statistics of the first vector representation as a first representation profile;
extracting the first predictor from the first behavior predictor as a predictor;
a process of training a second behavior predictor having a second expression learner and a second predictor for predicting an unobserved behavior from a behavior of a second user with respect to a second item constituting second behavior history data, freezing parameters of the predictor extracted in the training and using the parameters as the second predictor of the second behavior predictor, and additionally training the second expression learner;
a process of extracting a second vector representation from the second representation learner of the second behavior predictor, setting a statistical value of the second vector representation as a second representation profile, comparing the second representation profile with the first representation profile, and evaluating the usefulness of the second behavior predictor based on a similarity between the first and second representation profiles;
outputting a prediction result based on the second behavior predictor when the usefulness of the second behavior predictor is evaluated to be high;
This is a user behavior prediction program for executing the above.

According to the present invention, it is possible to select a part of the behavior predictor trained in the original domain that is useful for the target domain, and to realize prediction of the user's behavior in the target domain.

1 is a diagram illustrating a functional configuration of an embodiment of a user behavior prediction device according to the present invention. 1 is a diagram showing a model of a first behavior predictor learned by a first behavior predictor learning unit of a user behavior prediction device; FIG. 11 is a diagram showing a model of a second behavior predictor learned by a second behavior predictor learning unit of the user behavior prediction device; FIG. 3 is a flowchart showing a user behavior prediction method in the present embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
Fig. 1 is a diagram showing a functional configuration of an embodiment of a user behavior prediction device according to the present invention. Fig. 2 is a diagram showing a model of a first behavior predictor trained by a first behavior predictor learning unit. Fig. 3 is a diagram showing a model of a second behavior predictor trained by a second behavior predictor learning unit.
As shown in FIG. 1, the user behavior prediction device 10 includes a first behavior predictor learning unit 101, an expression extraction unit 102, a predictor extraction unit 103, an expression profile storage unit 104, a predictor storage unit 105, a second behavior predictor learning unit 106, an expression extraction comparison unit 107, and a behavior prediction unit 108.

ユーザＩＤは表２に示すワンホットベクトルに変換されて図２に示す第１の表現学習器に入力され、アイテムＩＤは表３に示すワンホットベクトルに変換されて図２に示す第１の表現学習器に入力される。

The first behavior predictor learning unit 101 learns a first behavior predictor for predicting unobserved behavior from the behavior of the first user with respect to the first item, which constitutes the first behavior history data of the original domain, using the model shown in FIG. 2. As shown in FIG. 2, the first behavior predictor includes a first expression learning unit and a first predictor. The form of the first expression learning unit and the first predictor is not particularly limited, but may be, for example, any linear transformation or neural network. The first behavior history data input to the first behavior predictor learning unit 101 is composed of a user ID, an item ID, and a behavior (type). The behavior (type) may be omitted if it is uniquely determined corresponding to the original domain. The behavior (type) may also be an explicit evaluation of the user ID with respect to the item ID, such as a five-level evaluation. Table 1 shows an example of the first behavior history data.

The user ID is converted into a one-hot vector shown in Table 2 and input into the first representation learning device shown in Figure 2, and the item ID is converted into a one-hot vector shown in Table 3 and input into the first representation learning device shown in Figure 2.

The first representation learner outputs a vector representation of each user and a vector representation of each item based on the input user ID and item ID. Table 4 shows the vector representation of each user, and Table 5 shows the vector representation of each item. Elements (also called dimensions) 1 and element 2 are numerical values indicating the characteristics of the user, and elements 3 and 4 are numerical values indicating the characteristics of the item. Note that, although two elements are shown here as the vector representation of each user and each item, three or more elements may be used.

表現抽出部１０２は、抽出した第１のベクトル表現の統計値を第１の表現プロファイルとして、表現プロファイル蓄積部１０４に蓄積する。 The expression extraction unit 102 extracts the first vector representation shown in Table 6, which is output from the first expression learner of the model of the first behavior predictor shown in Fig. 2 and in which the vector representation of the user and the vector representation of the item are linked together. The first vector representation of the first expression learner in which the vector representation of the user and the vector representation of the item are linked together is output to the first predictor.
The expression extraction unit 102 may extract the first vector representation by simply combining elements 1 and 2 and elements 3 and 4 as described above, or may extract the first vector representation by averaging the elements, for example, averaging elements 1 and 3, and elements 2 and 4. In this way, when the first vector representation is extracted by averaging elements 1 and 3, and elements 2 and 4, respectively, the second vector representation described later is also extracted by averaging elements 1 and 3, and elements 2 and 4, respectively.

The expression extraction unit 102 stores the extracted statistical value of the first vector expression as a first expression profile in the expression profile storage unit 104 .

The statistical value can be calculated, for example, by any of the following methods.
(1) Extract a first vector representation, which corresponds to a portion of the samples.
(2) Clustering the first vector representation, and outputting the representative points and the number of samples belonging to each cluster. Clustering refers to grouping vector representations based on the similarity between the vector representations, and a group of similar objects created by clustering is called a cluster.
Table 7 shows an example of the first expression profile. In Table 7, the first expression profile is calculated by the method (2) above, and the representative point is the average value for each element (the element's numerical value when there is only one), and the number of items (number of samples) belonging to the domain (original domain) "1" that constitutes the cluster is shown. From the viewpoint of anonymization, the number of samples may be limited to a predetermined number or more, for example, 2 or more, preferably 5 or more.

As shown in FIG. 2 , the predictor extraction unit 103 extracts the first predictor of the first behavior predictor as a predictor, and stores it in the predictor storage unit 105 .
The expression profile storage unit 104 stores the first expression profile output from the expression extraction unit 102 .
The predictor storage unit 105 stores the predictor output from the predictor extraction unit 103 .

The second behavior predictor learning unit 106 learns a second behavior predictor for predicting an unobserved behavior from the behavior of a second user with respect to a second item, which constitutes the second behavior history data of the target domain, using the model shown in Fig. 3. As shown in Fig. 3, the second behavior predictor includes a second expression learning device and a second predictor.
In learning the second behavior predictor, the second behavior predictor learning unit 106 extracts a predictor (corresponding to the first predictor) from the predictor storage unit 105, freezes the parameters of the predictor, uses it as a second predictor for the second behavior predictor, and performs additional learning (fine-tuning) on the part of the second expression learning unit enclosed by the dashed dotted line.

第２の行動予測器学習部１０６においても、既に説明した第１の行動予測器学習部１０１と同様に、ユーザＩＤとアイテムＩＤは、それぞれワンホットベクトルに変換されて図３に示す第２の表現学習器に入力される。
第２の表現学習器は、第１の表現学習器と同様に、入力されたユーザＩＤ、アイテムＩＤに基づいて各ユーザのベクトル表現と各アイテムのベクトル表現とを出力する。
そして、表現抽出比較部１０７は、図３に示す第２の行動予測器のモデルの第２の表現学習器から出力される、ユーザのベクトル表現とアイテムのベクトル表現とが結びつけられた、表９に示す第２のベクトル表現を抽出する。第２の表現学習器のユーザのベクトル表現とアイテムのベクトル表現とが結びつけられた第２のベクトル表現は、第２の予測器に出力される。

The second behavior history data input to the second behavior predictor learning unit 106 is composed of a user ID, an item ID, and a behavior (type). The behavior (type) may be omitted if it is uniquely determined corresponding to the domain. The behavior (type) may also be an explicit evaluation of the user ID for the item ID, such as a five-point scale. Table 8 shows an example of the second behavior history data.

In the second behavior predictor learning unit 106, similarly to the first behavior predictor learning unit 101 already described, the user ID and the item ID are each converted into a one-hot vector and input to the second representation learning unit shown in FIG. 3.
The second expression learner, like the first expression learner, outputs a vector representation of each user and a vector representation of each item based on the input user ID and item ID.
Then, the expression extraction and comparison unit 107 extracts the second vector representation shown in Table 9, which is output from the second expression learner of the model of the second behavior predictor shown in Fig. 3 and in which the vector representation of the user and the vector representation of the item are linked together. The second vector representation of the second expression learner in which the vector representation of the user and the vector representation of the item are linked together is output to the second predictor.

The expression extraction comparison unit 107 obtains a statistical value of the extracted second vector expression as a second expression profile, compares this second expression profile with the first expression profile retrieved from the expression profile storage unit 104, and evaluates the usefulness of the second behavior predictor based on the similarity between the first and second expression profiles.
The statistical values obtained by the expression extraction and comparison unit 107 are calculated in the same manner as in the expression extraction unit 102. Specifically, the expression extraction and comparison unit 107 clusters the second vector expressions, and obtains a second expression profile using the representative points and the number of samples belonging to the clusters as statistical values.
Table 10 shows an example of the second expression profile. In Table 10, the representative point of the second expression profile is shown as the average value for each element (the numerical value of the element when the number of cases is 1), and the number of cases (number of samples) belonging to the domain (target domain) "2" constituting the cluster is shown for the cluster. The number of samples is the same as the number of samples in the first expression profile, and from the viewpoint of anonymization, it may be limited to a predetermined number or more, for example, 2 or more, preferably 5 or more.

When evaluating the similarity between the first and second expression profiles, the expression extraction comparison unit 107 can use, for example, the inter-distribution dissimilarity between the distribution of the first expression profile and the distribution of the second expression profile (the smaller the inter-distribution dissimilarity, the higher the similarity). The evaluation of the similarity based on the inter-distribution dissimilarity can be performed, for example, by evaluating the similarity based on the shortness of the transport distance (the shorter the transport distance, the higher the similarity). When the distribution of the first expression profile and the distribution of the second expression profile can be expressed as point clouds as shown in Tables 9 and 10, the transport distance between the point clouds can be evaluated by the Synchorn algorithm.
A method for calculating the transport distance is already known and is described, for example, in https://www.slideshare.net/joisino/ss-249394573.

When the expression extraction and comparison unit 107 evaluates the usefulness of the second behavior predictor to be high, the behavior prediction unit 108 outputs, as a prediction result, an item with a high prediction probability for a specified user based on the second behavior predictor.
If the expression extraction and comparison unit 107 evaluates the usefulness as not being high, the second behavior predictor learning unit 106 additionally learns (fine-tunes) the second predictor using the second behavior history data.

As described above, in this embodiment, the first behavior predictor trained in the original domain is decomposed into the first expression learner and the first predictor, and only the first predictor is transferred to the second behavior predictor for the target domain, and the second expression learner is additionally trained (fine-tuned). The first expression profile for the original domain is compared with the second expression profile for the target domain, and it is determined that the original domain and the target domain in which the two expression profiles are similar have similar user and item characteristics, and the first predictor trained in the original domain is used to predict behavior in the target domain.

In this embodiment, the first behavior predictor learning unit 101, the expression extraction unit 102, and the predictor extraction unit 103 may be executed in a plurality of different original domains. In this case, the expression profile storage unit 104 and the predictor storage unit 105 store the first expression profile and the predictor (first predictor) extracted in each original domain together with information on the correspondence with each original domain.
The second behavior predictor learning unit 106 uses a predictor (first predictor) to learn for a target domain and generate a second expression vector, and the expression extraction and comparison unit 107 evaluates the usefulness of the second behavior predictor based on the similarity between the first expression profile and the second expression profile of each original domain. The behavior prediction unit 108 can ensemble prediction results based on the second behavior predictor evaluated as having high usefulness by the expression extraction and comparison unit 107 and output the prediction results as prediction results.
Ensemble refers to combining prediction results from multiple models to obtain more accurate prediction results. A typical example of an ensemble method is to calculate the average prediction probability from multiple second behavior predictor models and output a prediction result based on the average prediction probability.
As an ensemble method, for example, a majority vote is used, and a group of items having a higher prediction probability than a predetermined value for a specific user is output as a prediction result for each model of the second behavior predictor, and the result of the majority vote (the prediction result of the largest number of items among the group of items having a prediction probability higher than a predetermined value) is output as a final prediction result. As an ensemble method, an average of the prediction probabilities by the models of multiple second behavior predictors for a specific user is calculated for each candidate item, and the prediction result of the item with the averaged high prediction probability is output. Furthermore, instead of the average, a method called stacking may be used to create a new model for calculating the overall prediction probability by inputting the prediction probabilities by the models of multiple second behavior predictors, and the prediction result of the item with the calculated high overall prediction probability may be output.
When ensembling and outputting as a prediction result, multiple prediction results may be output. For example, when one or more item groups are ranked and output for a user, multiple prediction results are output. Also, there may be multiple items tied for first place for the most numerous item among the item groups having a prediction probability higher than a predetermined value. In this case, one may be randomly selected and output as a prediction result, or all of the multiple items may be output.

The user behavior prediction method will be specifically described below with reference to the flowchart in Fig. 4. In the following description, a case where the user behavior prediction method is executed by the user behavior prediction device 10 will be described, but the present invention is not limited to the configuration of the user behavior prediction device 10.
FIG. 4 is a flowchart showing a user behavior prediction method according to this embodiment.

In step S11, the first behavior predictor learning unit 101 learns a first behavior predictor for predicting unobserved behavior using the model shown in FIG. 2 from the behavior of the first user with respect to the first item, which constitutes the first behavior history data of the original domain.

In step S12, the predictor extraction unit 103 extracts the first predictor of the first behavior predictor as a predictor. Then, the predictor extraction unit 103 stores the extracted predictor (first predictor) in the predictor storage unit 105.

In step S13, the second behavior predictor learning unit 106 retrieves a predictor from the predictor storage unit 105, freezes the parameters of the predictor (first predictor), uses it as the second predictor of the second behavior predictor, and performs additional learning (fine-tuning) of a part of the second expression learning unit (the part surrounded by the dashed line in Figure 3), thereby performing learning.

In step S14, the expression extraction unit 102 extracts a first vector expression from the first behavior predictor, generates a statistical value of the extracted first vector expression as a first expression profile, and accumulates it in the expression profile accumulation unit 104. Step S14 may be performed before or simultaneously with step S12, between step S12 and step S13, or simultaneously with step S13.

In step S15, the expression extraction and comparison unit 107 extracts a second vector expression from the second behavior predictor, obtains a statistical value of the extracted second vector expression as a second expression profile, and compares this second expression profile with the first expression profile.

In step S16, the expression extraction comparison unit 107 evaluates the usefulness of the second behavior predictor based on the similarity between the first and second expression profiles.
If the expression extraction comparison unit 107 evaluates that the usefulness of the second behavior predictor is high, the process proceeds to step S17, and if the expression extraction comparison unit 107 evaluates that the usefulness of the second behavior predictor is not high, the process proceeds to step S18.

In step S17, the behavior prediction unit 108 outputs, as a prediction result, an item with a high prediction probability for a given user based on the second behavior predictor.
In step S18, the second behavior predictor learning unit 106 additionally learns (fine-tunes) the second predictor with the second behavior history data.

According to this embodiment, it is possible to select a part of the behavior predictor trained in the original domain that is useful for the target domain, making it easier to match the target domain and the original domain for cross-domain recommendations.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments. Furthermore, the effects described in the above-described embodiments are merely a list of the most favorable effects resulting from the present invention, and the effects of the present invention are not limited to those described in the embodiments.

The functions of the user behavior prediction device 10 can be realized by hardware, software, or a combination of these. The above-mentioned user behavior prediction method can also be realized by hardware, software, or a combination of these. Here, being realized by software means being realized by a computer reading and executing a program.
When the functions of the user behavior prediction device 10 or the user behavior prediction method are realized by software, the programs constituting this software are installed in an information processing device (computer). These programs may be recorded on a removable medium such as a CD-ROM and distributed to users, or may be distributed by being downloaded to the user's computer via a network. Furthermore, these programs may be provided to the user's computer as a Web service via a network without being downloaded.

REFERENCE SIGNS LIST 10 User behavior prediction device 101 First behavior predictor learning unit 102 Expression extraction unit 103 Predictor extraction unit 104 Expression profile storage unit 105 Predictor storage unit 106 Second behavior predictor learning unit 107 Expression extraction and comparison unit 108 Behavior prediction unit

Claims (11)

a first behavior predictor learning unit configured to learn a first behavior predictor having a first expression learning device and a first predictor for predicting an unobserved behavior of a first user from a behavior of the first user with respect to a first item constituting the first behavior history data;
an expression extraction unit that extracts a first vector expression from the first expression learning device of the first behavior predictor and outputs a statistical value of the first vector expression as a first expression profile;
a predictor extraction unit that extracts the first predictor of the first behavior predictor as a predictor;
a second behavior predictor learning unit that learns a second behavior predictor having a second expression learner and a second predictor for predicting an unobserved behavior from a behavior of a second user with respect to a second item constituting second behavior history data, and in this learning, freezes parameters of the predictor extracted by the predictor extraction unit and uses them as the second predictor of the second behavior predictor, and additionally learns the second expression learner;
an expression extraction and comparison unit that extracts a second vector expression from the second expression learning device of the second behavior predictor, sets a statistical value of the second vector expression as a second expression profile, compares the second expression profile with the first expression profile, and evaluates the usefulness of the second behavior predictor based on a similarity between the first and second expression profiles;
a behavior prediction unit that outputs a prediction result based on the second behavior predictor when the second behavior predictor is evaluated to be highly useful by the expression extraction comparison unit;
A user behavior prediction device comprising: the first behavior predictor learning unit, the expression extraction unit, and the predictor extraction unit each perform operations in a plurality of different domains;
The user behavior prediction device according to claim 1 , wherein the behavior prediction unit ensembles prediction results based on the second behavior predictors evaluated as having high usefulness, and outputs the ensembled prediction results as prediction results. The user behavior prediction device according to claim 2, wherein the behavior prediction unit uses majority voting as the ensemble method. The user behavior prediction device according to any one of claims 1 to 3, further comprising an expression profile storage unit that stores the first expression profile, and a predictor storage unit that stores the extracted predictor. The user behavior prediction device according to any one of claims 1 to 4, wherein the first expression profile and the second expression profile are generated by extracting the first vector representation and the second vector representation corresponding to a portion of samples as statistical values. The user behavior prediction device according to any one of claims 1 to 4, wherein the first expression profile and the second expression profile are generated by clustering the first vector representation and the second vector representation as statistical values and determining the representative points and the number of samples belonging to the clusters. The user behavior prediction device according to claim 6, wherein the number of samples is equal to or greater than a predetermined number. The user behavior prediction device according to any one of claims 1 to 7, wherein the expression extraction and comparison unit evaluates the inter-distribution dissimilarity between the first expression profile and the second expression profile. The user behavior prediction device according to claim 8, wherein the expression extraction and comparison unit uses a transport distance between the first expression profile and the second expression profile as the inter-distribution dissimilarity. A user behavior prediction method for a user behavior prediction device that uses first behavior history data of a first user to predict a behavior of a second user based on second behavior history data of the second user, comprising:
training a first behavior predictor having a first representation learner and a first predictor for predicting an unobserved behavior of the first user from the behavior of the first user with respect to a first item constituting the first behavior history data;
extracting a first vector representation from the first representation learner of the first behavior predictor and outputting statistics of the first vector representation as a first representation profile;
Extracting the first predictor of the first behavior predictors as a predictor;
A second behavior predictor having a second representation learner and a second predictor is trained to predict an unobserved behavior from the behavior of the second user with respect to a second item constituting the second behavior history data, and in this training, parameters of the predictor extracted are frozen and used as the second predictor of the second behavior predictor, and the second representation learner is additionally trained;
extracting a second vector representation from the second representation learner of the second behavior predictor, defining a statistical value of the second vector representation as a second representation profile, comparing the second representation profile with the first representation profile, and evaluating the usefulness of the second behavior predictor based on a similarity between the first and second representation profiles;
A user behavior prediction method comprising: when the usefulness of the second behavior predictor is evaluated to be high, outputting a prediction result based on the second behavior predictor. On the computer,
A process of learning a first behavior predictor having a first representation learner and a first predictor for predicting an unobserved behavior of a first user from a behavior of the first user with respect to a first item constituting the first behavior history data;
extracting a first vector representation from the first representation learner of the first behavior predictor and outputting statistics of the first vector representation as a first representation profile;
extracting the first predictor from the first behavior predictor as a predictor;
a process of training a second behavior predictor having a second expression learner and a second predictor for predicting an unobserved behavior from a behavior of a second user with respect to a second item constituting second behavior history data, freezing parameters of the predictor extracted in the training and using the parameters as the second predictor of the second behavior predictor, and additionally training the second expression learner;
a process of extracting a second vector representation from the second representation learner of the second behavior predictor, setting a statistical value of the second vector representation as a second representation profile, comparing the second representation profile with the first representation profile, and evaluating the usefulness of the second behavior predictor based on a similarity between the first and second representation profiles;
outputting a prediction result based on the second behavior predictor when the usefulness of the second behavior predictor is evaluated to be high;
A user behavior prediction program for executing the above.

Images