JP7589832B2 - Information transition prediction device, method, and program - Google Patents

Description

One aspect of the present invention relates to an information transition prediction device, method, and program that, for example, models the evolution of opinions of a group and predicts the evolution of opinions of individuals.

In groups, people exchange opinions on a variety of topics, such as policies, new products, and sports games, and as a result of this exchange (interaction), each person's opinion may change.

Predicting the evolution of people's opinions is an important task in a wide range of fields, such as marketing and planning election strategies. Predictive models are generally used to predict the evolution of opinions. A widely used predictive model is, for example, a theoretical model called the opinion dynamics model. This type of theoretical model is analyzed using, for example, an agent-based model or a differential equation solver. For example, the interaction mechanism is described mathematically, and the evolution of an individual's opinion is simulated based on the predefined interaction mechanism.

Well-known theoretical models of opinion dynamics include the DeGroot model and the Bounded confidence model (BCM). In the framework of these theoretical models, it is assumed that each individual (hereafter referred to as user u) has an opinion x _u (t) ∈ [−1, 1], which is expressed as a continuous value from -1 to 1, about a specific topic (politics, products, events, etc.) at each time t. Here, -1 represents the strongest negative opinion, 1 represents the strongest positive opinion, and 0 represents a neutral opinion. Below, an overview of each model is explained using mathematical formulas.

The DeGroot model assumes that opinions change in discrete time according to the following rules:

Here, x _v (t-1) is the opinion of user v one time point earlier, and a _uv is a parameter that represents the strength of influence of user v on user u. The DeGroot model is based on the assumption that the opinion x _u (t) of each user u at a certain time t is determined by the weighted sum of the opinions of the user u and other users at the immediately preceding time t-1.

In contrast, the bounded confidence model (BCM) is a model that takes into account conformity bias. Conformity bias is a bias that ignores opinions that are far from one's own. If two users have sufficiently similar opinions, these users will influence each other and become closer in opinion. If not, the users do not interact with each other and each user's original opinion is maintained. In BCM, the opinion update rule is defined as follows:

Here, |Δ| is the absolute value of Δ, and ε is a parameter that represents the threshold of difference in opinion. Also, 1{C} is an indicator function that takes the value 1 if condition C is true and 0 otherwise.

Theoretical models such as the DeGroot model and BCM are highly interpretable and can reflect knowledge from the social sciences when designing interaction mechanisms. However, they have limitations in their predictive accuracy due to the difficulty in handling large amounts of data.

In recent years, social media such as Twitter (registered trademark), social networking services (SNS), electronic bulletin boards, and blogs have rapidly spread as a means of expressing personal opinions. These social media have become valuable sources of information for analyzing the formation of opinions in groups. When using this data to calibrate theoretical models, it is necessary to manually select models and set parameters, which is time-consuming and laborious.

Recently, however, attempts to utilize social media posting histories to model and predict opinion evolution have been attracting attention, and several approaches based on machine learning models have been proposed in particular. For example, Non-Patent Document 1 describes a method that uses a probabilistic generative model called a point process. This method learns the parameters of a point process model based on the history of opinion evolution estimated from social media posts. However, it lacks flexibility because it uses specific assumptions about the interaction mechanism.

Non-Patent Document 2 discloses a method that uses a deep learning model. This method extracts opinions from text posted on social media and estimates the opinions of each individual in an end-to-end manner. This method makes it possible to flexibly model interactions by utilizing a deep learning model. However, on the other hand, there is an issue that it cannot take into account prior knowledge about the mechanism of interactions.

Abir De et al., “Learning and Forecasting Opinion Dynamics in Social Networks.” In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016. Lixing Zhu et al., “Neural opinion dynamics model for the prediction of user-level stance dynamics.” Information Processing & Management 57, 2 (2020), 102031.

However, when trying to introduce prior knowledge from the social sciences into the framework of deep learning, the following challenges arise. Generally, in a physics-informed neural network that takes into account the laws of physics, parameter estimation is performed using supervised learning. This requires a sufficient amount of observational data on the variables to be predicted. "Opinions" are inherently unobservable, and it is usually difficult to secure a sufficient amount of observational data.

On the other hand, it is also possible to assign correct opinion labels to each and every post on social media. However, assigning correct labels to each and every post would be a time-consuming and costly task.

This invention has been made in light of the above-mentioned circumstances, and aims to provide technology that makes it possible to incorporate prior knowledge in the social sciences when training a learning model, and that enables the utilization of large-scale social media post data without requiring a lot of time or cost, thereby enabling improved accuracy in predicting changes in information.

In order to solve the above problem, one aspect of the information transition prediction device or prediction method according to the present invention prepares a model in which a predetermined theoretical model is rewritten by a differential equation that represents interactions regarding information transmission by multiple users, and generates a first learning model with constraints that approximates the transition of the user's transmitted information using this model, and generates a second learning model that predicts and outputs the user's transmitted information based on multiple past posts on social media. Then, during learning, for the first learning model, a parameter is estimated that minimizes the overall likelihood defined by the likelihood of the predicted value of transmitted information for the posted information reflecting the output of the second learning model and the constraint term represented by the differential equation. Then, during prediction, attribute information and date and time of the user to be predicted are input to the trained first learning model in which the parameter is set, and a prediction result of the transmitted information of the user to be predicted at the date and time is obtained from the first learning model.

According to one aspect of the present invention, parameters of a first learning model are learned under the constraints of a differential equation representing interactions between users and with information predicted from posted information by a second learning model reflected.

Therefore, it is possible to predict future posts by users by incorporating prior knowledge from the social sciences that underpins the theoretical model of opinion dynamics and by utilizing a large amount of past posted information. This makes it possible to predict future posts by users with high accuracy while reducing the significant amount of time and cost required for the task of labeling correct answers.

In other words, according to one aspect of the present invention, it is possible to provide technology that makes it possible to incorporate prior knowledge in the social sciences when training a learning model, and that makes it possible to utilize large-scale social media post data without requiring a lot of time or cost, thereby improving the accuracy of predicting changes in information.

FIG. 1 is a diagram showing an example of the overall configuration of a social media network system including an information transition prediction device according to an embodiment of the present invention. FIG. 2 is a block diagram showing an example of a hardware configuration of an information transition prediction device according to an embodiment of the present invention. FIG. 3 is a block diagram showing an example of a software configuration of an information transition prediction device according to an embodiment of the present invention. FIG. 4 is a flowchart showing the procedure and contents of the model learning process executed by the control unit of the information transition prediction device shown in FIG. FIG. 5 is a flowchart showing the procedure and contents of the opinion transition prediction process executed by the control unit of the information transition prediction device shown in FIG. FIG. 6 is a diagram showing an outline of the configuration of a learning model used in the information transition prediction device shown in FIG. FIG. 7 is a diagram showing an example of social media posting history information acquired by the information transition prediction device shown in FIG. 3 .

Below, an embodiment of the present invention is described with reference to the drawings.

[One embodiment]
(overview)
In one embodiment of the present invention, a prediction method using semi-supervised learning of a constrained neural network is described. The prediction method using semi-supervised learning predicts the opinions of individuals at any time in the future, given a social media posting history in which some opinions are labeled with correct answer labels.

It is also possible to use unsupervised learning methods. Prediction methods using unsupervised learning predict the opinions of individuals at any time in the future based on their social media posting history, to which no correct opinion labels have been assigned.

Figure 6 shows an example of the configuration of a learning model used in one embodiment of the present invention. As shown in the figure, the learning model used in one embodiment is composed of a constrained neural network MD1 as a first learning model and a language model MD2 as a second learning model. The constrained neural network MD1 has a function of approximating changes in opinions over time. The language model MD2 has a function of predicting individual user opinions based on social media posts.

In one embodiment, when estimating parameters of the above-mentioned constrained neural network MD1, the parameters are learned under the constraints of a differential equation that represents the interaction mechanism of the exchange of opinions among multiple users.

(Configuration example)
(1) System FIG. 1 is a diagram showing an example of the overall configuration of a social media network system according to an embodiment of the present invention.

In one embodiment, the social media network system enables multiple user terminals UT1 to UTn to access a server device SV via a network NW, thereby enabling multiple users to register and view posted information via the server device SV.

The user terminals UT1 to UTn consist of, for example, a smartphone, a tablet terminal, or a personal computer, and are equipped with a browser and specific applications to use the above-mentioned social media network.

The server device SV is composed of a server computer located, for example, on the Web or on the cloud, and transmits and receives the posted information sent from the user terminals UT1 to UTn, and manages the system. The server device SV also has a large-capacity data storage device, and manages the posted information sent from each of the user terminals UT1 to UTn as posting history information using the data storage device.

The network NW comprises, for example, a wide area network with the Internet at its core, and an access network for accessing this wide area network. For example, a public communication network using wired or wireless connections, or a LAN (Local Area Network) using wired or wireless connections may be used as the access network.

(2) Information Transition Prediction Device PD
In one embodiment of the system, an information transition prediction device PD is provided to model the evolution of the opinions of a set of users and predict the change in the opinion of any user. The information transition prediction device PD is, for example, a personal computer or a server computer.

Figures 2 and 3 are block diagrams showing examples of the hardware and software configurations, respectively, of the information transition prediction device PD.

The information transition prediction device PD includes a control unit 1 that uses a hardware processor such as a central processing unit (CPU). A storage unit having a program storage unit 2 and a data storage unit 3, a communication interface (hereinafter, the interface is referred to as I/F) unit 4, and an input/output I/F unit 5 are connected to the control unit 1 via a bus 8.

Under the control of the control unit 1, the communication I/F unit 4 performs data communication with the server device SV using a communication protocol defined by the network NW.

An input device 6 and an output device 7 are connected to the input/output I/F unit 5. The input device 6 includes, for example, a keyboard and a mouse, and is used to input various operation information related to the information transition prediction process. The output device 7 includes, for example, a display device, and displays the prediction information generated by the control unit 1. Note that a printer or an external storage device may also be used as the output device 7.

The program storage unit 2 is configured by combining a non-volatile memory such as a solid state drive (SSD) as a storage medium that can be written to and read from at any time, and a non-volatile memory such as a read only memory (ROM), and stores application programs necessary for executing various control processes according to one embodiment, in addition to middleware such as an operating system (OS). Hereinafter, the OS and each application program will be collectively referred to as a program.

The data storage unit 3 is, for example, a combination of a non-volatile memory such as an SSD that can be written to and read from at any time as a storage medium, and a volatile memory such as a RAM (Random Access Memory), and includes a post history storage unit 31, a language model storage unit 32, a constrained learning model storage unit 33, a parameter storage unit 34, and a prediction information storage unit 35 as the main storage units required to implement one embodiment.

The post history memory unit 31 is used to store social media post history information obtained from the server device SV.

The language model memory unit 32 stores the language model MD2 generated by the control unit 1. The language model MD2 is used to predict changes in user opinions based on the posting history information.

The constrained learning model memory unit 33 stores information regarding the configuration of the constrained neural network MD1. The constrained neural network MD1 approximates the time changes of opinions based on the posting history information.

The parameter memory unit 34 is used to store the optimization parameters of the above-mentioned constrained learning model obtained by the control unit 1.

The prediction information memory unit 35 is used to store information representing the predicted results of changes in individual opinions obtained by the control unit 1.

The control unit 1 includes, as processing functions necessary for implementing one embodiment, a posting history management processing unit 11, a posting history acquisition processing unit 12, a parameter estimation processing unit 13, a prediction processing unit 14, and a prediction information output processing unit 15. All of these processing units 11 to 15 are realized by causing the hardware processor of the control unit 1 to execute application programs stored in the program storage unit 2.

In addition, the above application programs may be stored in advance in the program storage unit 2, or may be downloaded from an external application server, etc., when necessary and stored in the program storage unit 2.

When, for example, a system administrator inputs various operation information for the posting history information of the server device SV on the input device 6 of the information transition prediction device PD, the posting history management processing unit 11 performs information operations on the posting history information of the server device SV in accordance with this operation information. Types of information operations include, for example, registering, modifying, and deleting information. Note that the information operation function for the above-mentioned posting history information does not necessarily need to be possessed by the information transition prediction device PD, and may be possessed by a management terminal attached to the server device SV.

The post history acquisition processing unit 12 acquires the necessary post history information from the server device SV via the communication I/F unit 4 when constructing a learning model consisting of a constrained neural network MD1 and a language model MD2, and performs processing to store the acquired post history information in the post history memory unit 31.

The parameter estimation processing unit 13 performs a process of rewriting the theoretical model into the form of a differential equation, and generates a constrained neural network MD1 using the rewritten model. The parameter estimation processing unit 13 also reads the posting history information from the posting history storage unit 31, and constructs a language model MD2 based on the reading posting history information.

The parameter estimation processing unit 13 estimates parameters of the constrained neural network MD1 using the constraint conditions defined by the differential equation and the prediction information of the language model MD2, and stores the estimated parameters in the parameter storage unit 34. An example of the parameter estimation process will be described in detail in the operation example.

The prediction processing unit 14 sets the optimization parameters stored in the parameter memory unit 34 to the constrained neural network MD1 stored in the constrained learning model memory unit 33, thereby constructing a learned constrained neural network MD1.

The prediction processing unit 14 receives the user ID of the user u to be predicted and the date and time t to be predicted, which are input to the input device 6, via the input/output I/F unit 5, and inputs the received user u to be predicted and the date and time t to be predicted to the trained constrained neural network MD1. The value output from the trained constrained neural network MD1 is then stored in the prediction information storage unit 35 as a prediction result.

The prediction information output processing unit 15 reads prediction information from the prediction information storage unit 35 in response to a request to output a prediction result input by the input device 6, and outputs the read prediction information to the output device 7 via the input/output I/F unit 5.

(Example of operation)
Next, an example of the operation of the information transition prediction device PD configured as above will be described.

(I) Model Learning Process Fig. 4 is a flowchart showing an example of the processing procedure and processing contents of the model learning process executed by the control unit 1 of the information transition prediction device PD. Note that in this example, a case where a semi-supervised learning method is used as a constrained neural network will be described as an example, but it can be easily extended to a case where an unsupervised learning method is used.

The control unit 1 of the information transition prediction device PD determines whether or not the learning mode has been instructed in step S10. In this state, if, for example, a system administrator instructs the execution of the learning mode, the control unit 1 of the information transition prediction device PD executes the model learning process as follows.

(1) Acquisition of Post History Information The control unit 1 of the information transition prediction device PD first accesses the server device SV under the control of the post history acquisition processing unit 12, and receives social media post history information from the server device SV via the communication I/F unit 4. Then, the received post history information is stored in the post history storage unit 31. At this time, for example, a topic to be acquired (e.g. politics, products, events) may be specified, and post information related to the specified topic may be selectively acquired. An example of the acquired post history information is shown in FIG. 7.

The social media posting history information to be analyzed consists of a series of posted sentences related to a specific topic, and is represented as {(u _i , t _i , d _i )} ^I _{i = 1} , where u _i is a user, d _i is a social media post related to a specific topic, t _i is the posting time, and I is the number of data. Let U be the number of users.

Here, consider the case where some sentences are given a correct opinion label y _i . The correct label can be, for example, an opinion category y _i ∈ {-1, 0, 1}, or a score y _i ∈ {1, 2, 3, 4, 5} indicating the degree of positivity. For example, politically related tweets are collected during the election period, and labeled data is generated by labeling positive posts about a political party with y _i = 1, negative posts with y _i = -1, and neutral posts with y _i = 0. Let Ds = {(u _i , t _i , y _i )} ^Is _i=1 be the history of posted sentences given a correct opinion label, and Du = {(u _i , t _i , d _i )} ^Iu _i=1 be the history of posted sentences without a correct opinion label. Here, Is indicates the number of labeled data, and Iu indicates the number of unlabeled data.

(2) Parameter Estimation of Learning Model (2-1) Rewriting of Theoretical Model In step S11, the control unit 1 of the information transition prediction device PD rewrites the theoretical model in the form of a differential equation in order to introduce knowledge in the social science field that is the basis of the theoretical model of opinion dynamics into the machine learning model under the control of the parameter estimation processing unit 13. For example, theoretical models such as the DeGroot model and the Bounded confidence model (BCM) are generally rewritten using the following differential equation.

Here, t is time, _xu (t) is the opinion of user u at time t, and Λ is a parameter of the differential equation. Φ(·) is a function that represents the interaction mechanism and is designed based on the assumptions of the theoretical model. For example, when using the assumptions of the bounded confidence model (equation (2)), Φ(·) is defined by the following equation.

When using the DeGroot model (equation (1)), Φ(·) is defined as follows:

Here, a _uv is a parameter representing the strength of interaction between user u and user v. Note that, although two representative theoretical models are used in this example, the present invention is also applicable to other theoretical models.

ここで、θ_f は制約付きニューラルネットのパラメータである。また、ニューラルネットワークの出力を
ｘ_u^ (t) ＝ｆ(・；θ_f )
と定義しておく。 (2-2) Approximation of Neural Network Next, in step S12, the parameter estimation processing unit 13 approximates the variable x _u (t) representing the opinion by a neural network f(·;θ _f ) that inputs the time t and the ID or profile of the user u as follows:

Here, _θf is the parameter of the constrained neural network. Also, the output of the neural network is _xu ^(t) = f(·; _θf )
Let us define it as follows.

(2-3) Building a Language Model The parameter estimation processing unit 13 then builds a language model in step S13. The language model is used to extract opinions from social media posts, and is a model that takes a posted sentence d on a specific topic as input and an opinion o ^∼ as output, and is expressed as follows:

Here, g(·) is a language model and _θg is a parameter of the language model. The language model g(·) can be freely designed to suit the problem. For example, one possible approach to design the language model g(·) is to fine-tune a pre-trained deep learning model.

(2-4) Parameter Estimation In the learning mode, the parameter Λ of the differential equation, the parameter θ _f of the constrained neural network MD1, and the parameter θ _g of the language model MD2 are estimated so as to minimize the likelihood L defined by the following equation (8).

Here, _Ls is the likelihood for a social media posting history with a correct answer label, _Lu is the likelihood for a posting history without a correct answer label, and _Lode is a term representing a constraint by a differential equation. α and λ are hyperparameters representing the importance of each term.

If we remove the first term _Ls ( _θf ; _Ds ) from equation (8), equation (8) above corresponds to the likelihood in unsupervised learning. The first term on the right-hand side of equation (8) is the error between the predicted value _xu ^(t) of the opinion by the neural network and the correct label _yi of the opinion, and is defined as follows:

Here, L(.) is the cross entropy, σ(.) is a function that embeds the continuous value _xu ^(.), which represents the strength and polarity (positive or negative) of an opinion, into the label space, and a sigmoid function is used when there are two types of correct opinion labels, and a softmax function is used when there are multiple classes. Also, Ds = {(u _i , t _i , y _i )} ^Is _i=1 is the posting history with the correct label, and Is is the number of labeled data.

なお、この例では、誤差の指標として平均二乗誤差を用いたが、他の誤差関数を用いることもできる。 For the unlabeled posting history Du = {(u _i , t _i , d _i )} ^Iu _{i = 1} , the prediction result by the language model MD2 is used instead of the correct label. Specifically, the squared error between the output x _u ^ (t) of the constrained neural network MD1 and the output o ^̂ of the language model MD2 is minimized.

In this example, the mean square error is used as an index of error, but other error functions can also be used.

The third term on the right side of the above equation (8) is defined as follows:

Here, {T ₁ , ...,T _J } are representative points on a predefined time axis. L _ode represents the error of the differential equation at J representative points. In this example, the mean square error is used as an error index, but other error functions can also be used. The above equation (11) is obtained by taking the difference between the left and right sides of the differential equation (equation (3)) and substituting the time T _J of the representative point. By adding the constraint term in the above equation (11) to the likelihood L, it is possible to learn the parameter θ _f of the constrained neural network MD1 such that the output x _u ^ (t) follows the differential equation defined in equation (3).

In addition, the differential dx _u ^ (t) /dt in the first term on the right-hand side of equation (11) can be easily calculated by using automatic differentiation implemented in deep learning packages such as PyTorch.

Although any method may be used to optimize the parameter _θf , it is preferable to use the backpropagation method in order to efficiently calculate the derivative included in the equation (11).

(2-5) Saving Parameters In step S15, the parameter estimation processing unit 13 stores a set of the optimization parameters θ _f ^ of the constrained neural network MD1 estimated as described above and the optimization parameters θ _g ^ of the language model MD2 in the parameter storage unit .

(II) Prediction Processing The control unit 1 of the information transition prediction device PD uses the learning model constructed as described above to predict future transmitted information of a specific individual user as follows.
FIG. 5 is a flowchart showing an example of the procedure and contents of the prediction process executed by the control unit 1 of the information transition prediction device PD.

In order to predict future information to be transmitted by a particular individual user, for example, when a system administrator specifies a prediction mode on the input device 6, the control unit 1 of the information change prediction device PD receives the specified input of the prediction mode via the input/output I/F unit 5 and sets the prediction mode.

(1) Prediction When the prediction mode is set, the control unit 1 of the information transition prediction device PD, under the control of the prediction processing unit 14, first receives, in step S21, attribute information of the user u to be predicted, such as a user ID or profile, and the date and time t to be predicted, which are inputted into the input device 6, via the input/output I/F unit 5. Next, in step S22, the prediction processing unit 14 reads the optimization parameter θ _f ^ from the parameter storage unit 34, and sets the read optimization parameter θ _f ^ to the constrained neural network MD1. That is, a trained constrained neural network MD1 is constructed.

In step S23, the prediction processing unit 14 then inputs the identification ID of the individual user u to be predicted and the time t to be predicted, which are input from the input device 6, to the trained constrained neural network MD1. The prediction processing unit 14 then receives the prediction result of the transmitted information of the individual user u to be predicted at the prediction time t, which is output from the trained constrained neural network MD1, and stores information representing the received prediction result in the prediction information storage unit 35.

(2) Output of prediction result In step S24, the control unit 1 of the information transition prediction device PD, under the control of the prediction information output processing unit 15, autonomously reads out information representing the prediction result at a predetermined timing from the prediction information storage unit 35 or in response to an input of an output request from the input device 6. Then, the prediction information output processing unit 15 generates, for example, display data of the information representing the read out prediction result, outputs the generated display data from the input/output I/F unit 5 to the output device 7, and displays it on the output device 7. The display data includes, for example, the user ID of the prediction target user u, the prediction target date and time t, and information representing the prediction result.

Thus, the prediction result of the outgoing information of the prediction target user u at the prediction target date and time t is displayed on the output device 7. The information representing the prediction result may not only be displayed on the output device 7, but may also be printed out on the output device 7 or stored in an external storage medium such as a USB memory, or may even be transmitted from the communication I/F unit 4 to another user terminal or the like via the network NW.

(Action and Effects)
As described above, in one embodiment, the system is based on a learning model described using differential equations representing interactions between a plurality of users who transmit information, and includes a constrained neural network that approximates the transition of the transmitted information of the user, and a language model that predicts and outputs the transmitted information of the user based on posted information on social media. During learning, parameters that minimize the likelihood of the predicted value of the transmitted information for the posted information reflecting the output of the language model and the constraint term by the differential equation are estimated, and during prediction, attribute information of the user to be predicted and a date and time are input as explanatory variables to the trained constrained neural network in which the estimated parameters are set, and information output from the constrained neural network as a target variable is output as a predicted result of the transmitted information of the user to be predicted at the date and time.

Therefore, according to one embodiment, a learning model is constructed by combining a constrained neural network and a language model that predicts outgoing information based on a large amount of posted information, and the parameters of the learning model are trained under the constraints of a differential equation that represents interactions between users and in a state that reflects the outgoing information predicted from a large amount of posted information on social media.

This makes it possible to predict what users will post in the future by incorporating prior knowledge from the social sciences that underpins the theoretical model of opinion dynamics, and by utilizing a large amount of past posted information. In other words, it is possible to predict what users will post in the future with a high degree of accuracy, without requiring a significant amount of time and cost for annotation.

[Other embodiments]
In the embodiment described above, an example was given in which all the functions of the information transition prediction device PD are provided in a terminal such as a personal computer used by a user. However, all or some of the functions of the information transition prediction device PD may be provided in a server device provided on the Web or the cloud. For example, a parameter estimation processing function during learning may be provided in the server device, and the terminal may download and use a trained constrained neural network from the server device during prediction. In this way, the processing load on the terminal can be reduced.

In addition, the types and configurations of the first learning model and the second learning model, the processing procedures and processing contents of the parameter estimation processing during learning, the processing procedures and processing contents of the prediction processing during prediction, the type of information to be predicted, etc. can be modified in various ways without departing from the spirit of this invention.

Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in every respect. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, in implementing the present invention, specific configurations according to the embodiments may be appropriately adopted.

In short, this invention is not limited to the above-described embodiment as it is, and in the implementation stage, the components can be modified and embodied without departing from the gist of the invention. Furthermore, various inventions can be formed by appropriate combinations of multiple components disclosed in the above-described embodiment. For example, some components may be deleted from all of the components shown in the embodiment. Furthermore, components from different embodiments may be appropriately combined.

PD...information transition prediction device PT1 to PTn...user terminal SV...server device NW...network MD1...constrained neural network MD2...language model 1...control unit 2...program storage unit 3...data storage unit 4...communication I/F unit 5...input/output I/F unit 6...input device 7...output device 8...bus 11...post history management processing unit 12...post history acquisition processing unit 13...parameter estimation processing unit 14...prediction processing unit 15...prediction information output processing unit 31...post history storage unit 32...language model storage unit 33...constrained learning model storage unit 34...parameter storage unit 35...prediction information storage unit

Claims (7)

An information transition prediction device for predicting a transition of information transmitted by a user in the future, comprising:
a first learning model with constraints that approximates the transition of the transmitted information of the user using a theoretical model described by differential equations that represent interactions regarding information transmitted by the multiple users;
a second learning model that predicts and outputs the user's transmitted information based on a plurality of pieces of past posted information on social media;
a parameter estimation processing unit that estimates, during learning, parameters for the first learning model that minimize an overall likelihood defined by a likelihood of a predicted value of the transmitted information for the posted information reflecting an output of the second learning model and a constraint term represented by the differential equation;
a prediction processing unit that inputs attribute information of a specific user to be predicted and a date and time into the first learning model in which the parameters have been set and obtains a prediction result of the transmitted information of the specific user at the date and time from the first learning model at the time of prediction;
and an output processing unit that outputs the prediction result. The method further includes a post information acquisition processing unit that selectively acquires the post information related to a specific topic from a server device that accumulates and manages the post information,
The information transition prediction device according to claim 1 , wherein the parameter estimation processing unit constructs the second learning model based on the acquired posted information. the first learning model is configured by a constrained neural network including the constraint term represented by the differential equation;
The second learning model is configured by a language model that receives the posted information related to the specific topic as an input, extracts the transmitted information related to the specific topic from the posted information, and outputs the transmitted information.
The information transition prediction device according to claim 2. The information transition prediction device of claim 1, wherein the parameter estimation processing unit estimates the parameters that minimize a first likelihood of a predicted value for the posted information with a correct label, a second likelihood of a predicted value for the posted information using the output of the second learning model as a correct label, and a third likelihood defined by a constraint term represented by the differential equation. the second likelihood and the constraint term are weighted by a hyperparameter that represents an importance to the third likelihood;
The information transition prediction device according to claim 4. An information transition prediction method executed by an information processing device that predicts a transition of information transmitted by a user in the future, comprising:
A process of rewriting a predetermined theoretical model using differential equations that represent interactions related to information transmission by a plurality of users;
generating a first learning model with constraints that approximates the transition of the transmitted information of the user;
generating a second learning model that predicts and outputs the user's transmitted information based on a plurality of pieces of past posted information on social media;
a step of estimating, during learning, parameters for the first learning model that minimize an overall likelihood defined by a likelihood of a predicted value of the transmitted information for the posted information reflecting an output of the second learning model and a constraint term represented by the differential equation;
At the time of prediction, inputting attribute information and a date and time of a specific user to be predicted into the first learning model having been trained and having the parameters set, and acquiring a prediction result of the transmitted information of the specific user to be predicted at the date and time from the first learning model;
and outputting the prediction result. A program for causing the information processing device to execute each process in each step of the information transition prediction method described in claim 6.

Images