Model generation program, model generation method, and information processing device

Abstract

To improve the accuracy of models predicting human selection behavior.SOLUTION: A computer acquires first feature amount data indicative of a plurality of first feature amounts corresponding to a plurality of options, and historical data associating human attribute values with human selection outcomes. The computer also generates second feature amount data indicative of a plurality of second feature amounts converted using the first feature amount corresponding to a reference option for each of a plurality of classes where each of a plurality of classes has a different reference option specified from among a plurality of options. In addition, the computer generates a prediction model including a plurality of first models which calculate prediction results predicting selection behavior and correspond to a plurality of classes, and a second model which calculates classification results to the plurality of classes on the basis of an attribute value, using historical data and second feature amount data of each of the plurality of classes.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present invention relates to a model generation program, a model generation method, and an information processing device. [Background technology]

[0002] The computer may generate a model that predicts a person's choice behavior using data collected from the real world. The generated model may be a discrete choice model that predicts the choice behavior of selecting one option from multiple options. The prediction result may indicate the selection probability of each of the multiple options.

[0003] Selection behavior is influenced by features assigned to each of a plurality of options. For example, the selection behavior of selecting one of a plurality of transportation modes is influenced by the fare of each of the plurality of transportation modes. The computer may use the generated model to perform a human behavior simulation that simulates people's selection behavior under a particular situation. For example, the computer may predict the number of users of each of a plurality of transportation modes under a particular fare setting.

[0004] A behavioral characteristic prediction system has been proposed that calculates features that affect the cancellation of a communication service based on a communication log of a base station, and trains a model that predicts a user's cancellation behavior by using the calculated features as explanatory variables.A model estimation method has also been proposed that generates a model that shows the behavior of a person who comes into contact with advertising information to select and purchase a product, and calculates the selection probability of each of multiple products.

[0005] Also, a behavioral selection learning device has been proposed that updates the weights for each factor such as transportation cost and travel time for a person each time the person selects a transportation mode, and uses the updated weights to present options that will increase satisfaction the next time the person travels.Also, a recommendation system has been proposed that trains a prediction model based on the behavior logs of various people to predict the known probability that a store's existence is known and the selection probability that a person who knows the store will select the store, and recommends stores with low known probability and high selection probability.

[0006] In addition, an information processing system has been proposed that converts a person's behavior log into a first feature, converts a person's vital data into a second feature, and uses the two types of feature to generate a predictive model that predicts the effectiveness of each lifestyle improvement measure. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] International Publication No. 2016 / 017086 [Patent Document 2] JP 2017-16273 A [Patent Document 3] International Publication No. 2018 / 198323 [Patent Document 4] Patent Publication No. 2021-22243 [Patent Document 5] Patent Publication No. 2022-122584 Summary of the Invention [Problem to be solved by the invention]

[0008] A person's choice behavior may be influenced by the subjective reference feature of the person in addition to the objective feature of the option. The reference feature is sometimes called a reference point. For example, prospect theory in behavioral economics proposes loss aversion, which states that a loss that is less favorable than a reference point has a greater impact on subjective value than a gain that is more favorable than the reference point. Therefore, it is conceivable that a computer could generate a model that takes into account the reference feature.

[0009] However, the reference feature varies from person to person and is often a latent variable that is not directly observed from the outside. Therefore, it is not easy to directly generate a model that has the reference feature as an explanatory variable. Therefore, in one aspect, the present invention aims to improve the accuracy of a model that predicts a person's selection behavior. [Means for solving the problem]

[0010] In one embodiment, a model generation program is provided that causes a computer to execute the following processes: Acquire first feature amount data indicating a plurality of first feature amounts corresponding to a plurality of options, and history data in which a person's attribute value is associated with a selection result of the person for the plurality of options; Generate second feature amount data indicating a plurality of second feature amounts converted from the plurality of first feature amounts using the first feature amount corresponding to the reference option for each of a plurality of classes in which a reference option is different, the reference option being specified from among the plurality of options; Generate a prediction model using the history data and the second feature amount data for each of the plurality of classes, the prediction model being a first model for calculating a prediction result predicting selection behavior for the plurality of options, the prediction model including a plurality of first models corresponding to the plurality of classes and a second model for calculating a classification result into the plurality of classes based on an attribute value.

[0011] In one aspect, a computer-implemented model generating method is provided.In another aspect, an information processing device is provided having a storage unit and a processing unit. Effect of the Invention

[0012] On the one hand, it improves the accuracy of models that predict people's choice behavior. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram illustrating an information processing device according to a first embodiment. [Diagram 2] FIG. 11 illustrates an example of hardware of an information processing apparatus according to a second embodiment. [Diagram 3] FIG. 13 is a diagram illustrating an example of a first discrete choice model. [Figure 4] FIG. 13 is a diagram illustrating an example of a second discrete choice model. [Diagram 5] FIG. 13 is a diagram illustrating an example of calculation of selection probability using a plurality of classes with different reference points. [Figure 6] FIG. 1 illustrates an example of a predictive model including a multiple class discrete choice model. [Figure 7] FIG. 13 is a diagram illustrating an example of input and output data of a discrete choice model. [Figure 8] FIG. 13 is a diagram illustrating an example of a class estimation model. [Figure 9] FIG. 2 is a block diagram showing an example of functions of the information processing device; [Figure 10] FIG. 13 is a diagram illustrating an example of a behavior history table. [Figure 11] 13 is a flowchart illustrating an example of a procedure for generating a model. [Figure 12] 13 is a flowchart illustrating an example of a procedure for behavior prediction. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, the present embodiment will be described with reference to the drawings. [First embodiment] A first embodiment will be described.

[0015] FIG. 1 is a diagram illustrating an information processing apparatus according to a first embodiment. The information processing device 10 of the first embodiment generates a prediction model 17 that predicts a person's selection behavior for a plurality of options. The generation of the prediction model 17 may be called machine learning. The information processing device 10 or another information processing device may predict selection behavior using the generated prediction model 17. The information processing device 10 may be a client device or a server device. The information processing device 10 may be called a computer, a model generation device, a machine learning device, a behavior prediction device, or a simulation device.

[0016] The information processing device 10 includes a storage unit 11 and a processing unit 12. The storage unit 11 may be a volatile semiconductor memory such as a random access memory (RAM). The storage unit 11 may also be a non-volatile storage such as a hard disk drive (HDD) or a flash memory.

[0017] The processing unit 12 is, for example, a processor such as a central processing unit (CPU), a graphics processing unit (GPU), or a digital signal processor (DSP). However, the processing unit 12 may include an electronic circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The processor executes a program stored in a memory such as a RAM (which may be the storage unit 11). A collection of processors may be called a multiprocessor or simply a "processor."

[0018] The storage unit 11 stores feature amount data 13 and history data 14. The information processing device 100 may accept input of the feature amount data 13 and history data 14 from a user, or may receive the feature amount data 13 and history data 14 from another information processing device.

[0019] The feature data 13 indicates a first feature corresponding to each of a plurality of options. Usually, each person selects one option from the plurality of options. When the selection behavior is performed repeatedly, each person may select different options at different times (e.g., different days). An example of the options is transportation such as a bus or a train. Another example of the options is how to use the transportation, such as a boarding location and an alighting location.

[0020] The first feature is usually a numerical value, typically a non-negative integer. The first feature may be determined from the nature of the option, or may be artificially set for the option. The larger the value of the first feature, the more advantageous the gain for the person may be, or the larger the value of the first feature, the more disadvantageous the cost for the person may be. Examples of the first feature include transportation fares and travel times.

[0021] The history data 14 associates the values ​​of a person's attributes with the person's selection results for multiple options. The history data 14 may include multiple records showing selection results of different people, or may include multiple records showing selection results of the same person at different times. The history data 14 is, for example, data collected from the real world, and may include data collected from sensor devices such as automatic ticket gates and smartphones.

[0022] An attribute is an information item that represents a characteristic of a person. It is preferable that an attribute is related to a selection behavior. The value of an attribute is, for example, a numerical value, or a code or character string that can be converted into a numerical value. Examples of an attribute include age and gender. The value of an attribute is identified, for example, by matching registered information registered in advance with an identifier read by a sensor device. The selection result indicates, for example, one selected option.

[0023] As an example, the feature amount data 13 indicates that the first feature amount corresponding to option A is 300, and the first feature amount corresponding to option B is 500. Furthermore, the history data 14 indicates that a 30-year-old person selected option B, and a 20-year-old person selected option A.

[0024] The processing unit 12 generates a prediction model 17 using the feature amount data 13 and the history data 14. The generation of the prediction model 17 corresponds to, for example, machine learning that adjusts the values ​​of parameters included in the prediction model 17. The prediction model 17 predicts the selection behavior of each person for multiple options. The prediction result may indicate any one option that is most likely to be selected, or may include the selection probability of each of the multiple options.

[0025] Here, in order to improve the prediction accuracy of the prediction model 17, the processing unit 12 may reflect human cognitive biases such as loss aversion in the prediction model 17. For example, the prediction model 17 may evaluate an option in which the first feature amount is greater than the reference feature amount and an option in which the first feature amount is smaller than the reference feature amount using different evaluation rules based on the prospect theory. The reference feature amount may be called a reference point. However, the reference feature amount differs from person to person and is not usually observed directly from the outside. Therefore, the processing unit 12 generates the prediction model 17 in the following manner.

[0026] First, the processing unit 12 defines a plurality of classes having different reference options. The reference option is specified from among the plurality of options indicated by the feature amount data 13. The class may be interpreted as corresponding to a group of people who consider the reference option to be the most reasonable among the plurality of options, or may be interpreted as corresponding to a group of people who have the highest frequency of selecting the reference option. The processing unit 12 may define a class corresponding to all options, or may define only a class corresponding to some options. As an example, the processing unit 12 defines a class 15a having option A as a reference option and a class 15b having option B as a reference option.

[0027] The processing unit 12 generates feature data converted from the feature data 13 for each of the multiple classes. The converted feature data includes multiple second features corresponding to the multiple options. At this time, the processing unit 12 uses the first feature corresponding to the reference option as a reference feature to convert the first feature into the second feature.

[0028] The second feature may be a relative value from the reference feature. The second feature corresponding to the reference option may be 0. For example, the processing unit 12 calculates the second feature by subtracting the reference feature from the first feature. The larger the value of the second feature, the more advantageous the gain for the person may be, or the larger the value of the second feature, the more disadvantageous the cost for the person may be. When the first feature represents a cost and the second feature represents a gain, the processing unit 12 may invert the positive and negative of the numerical values.

[0029] As an example, the processing unit 12 generates feature amount data 16a corresponding to class 15a and feature amount data 16b corresponding to class 15b. The second feature amount included in the feature amount data 16a is calculated by subtracting the first feature amount "300" of option A used as a reference feature amount from the first feature amount of each of the multiple options. The second feature amount included in the feature amount data 16b is calculated by subtracting the first feature amount "500" of option B used as a reference feature amount from the first feature amount of each of the multiple options.

[0030] The processing unit 12 generates a prediction model 17 using the history data 14 and the transformed feature data of each of the multiple classes. The prediction model 17 includes multiple first models corresponding to the multiple classes and a model 19 that is a second model. As an example, the prediction model 17 includes a model 18a corresponding to a class 15a in which option A is the reference option, and a model 18b corresponding to a class 15b in which option B is the reference option.

[0031] The first model calculates a prediction result that predicts a selection behavior for a plurality of options. The prediction result of the first model may indicate any one option that is most likely to be selected, or may include a selection probability for each of the plurality of options. The first model of a certain class evaluates the plurality of options relatively with respect to a reference option of the class. The first model may be called a discrete choice model, and may be based on prospect theory.

[0032] The first model may include a function having a variable indicating the second feature amount. This function may be called a utility function or a value function. The utility function calculates an evaluation value indicating the utility that a person obtains from an option from the second feature amount of the option. The value function calculates an evaluation value indicating the value of an option to a person from the second feature amount of the option. The function may further include a variable indicating an attribute of the person and a variable indicating the environment, such as the weather, at the time of selection.

[0033] The slope of the function may be asymmetric between an interval where the value of the variable is greater than the second feature of the reference option (e.g., 0) and an interval where the value of the variable is less than the second feature of the reference option. Typically, the slope of the function is greater in an interval where the value of the variable is less favorable than the reference option than in an interval where the value of the variable is more favorable than the reference option. The function may be a nonlinear function including one or more variables and one or more parameters.

[0034] The first model may convert the multiple evaluation values ​​corresponding to the multiple options into prediction results. For example, the first model uses a softmax function to convert the multiple evaluation values ​​calculated by the above function into multiple selection probabilities.

[0035] In generating the prediction model 17, the processing unit 12 trains a first model of a certain class using transformed feature data corresponding to the class. As an example, the processing unit 12 inputs each second feature included in the feature data 16a to the model 18a to calculate a prediction result of the class 15a. The processing unit 12 also inputs each second feature included in the feature data 16b to the model 18b to calculate a prediction result of the class 15b. The processing unit 12 may adjust the values ​​of parameters included in each of the multiple first models using the prediction results of the multiple classes and the selection results included in the history data 14.

[0036] A second model, model 19, calculates a classification result into multiple classes based on the values ​​of the person's attributes. The classification result may indicate which one the person is most likely to belong to, or may include multiple class probabilities corresponding to the multiple classes. Model 19 may be a neural network that calculates the classification result from the values ​​of the attributes.

[0037] The prediction model 17 determines how to use the prediction results of the multiple classes based on the classification results calculated by the model 19. For example, the prediction model 17 outputs the prediction result of any one of the multiple classes having the highest class probability among the multiple classes. Also, for example, the prediction model 17 synthesizes the prediction results of the multiple classes using the class probability and outputs the synthesized prediction result. The synthesized prediction result is, for example, a weighted average of the multiple prediction results.

[0038] In generating the prediction model 17, the processing unit 12 trains the model 19 using the history data 14. The processing unit 12 may train the model 19 together with the first model. For example, the processing unit 12 inputs attribute values ​​included in the history data 14 to the model 19 to calculate a classification result. The processing unit 12 adjusts parameter values ​​included in the model 19 using prediction results of multiple classes, the classification result, and the selection result included in the history data 14. The processing unit 12 may compare a prediction result selected based on the classification result or a synthesized prediction result with a correct selection result to adjust the parameter values.

[0039] As described above, the information processing device 10 of the first embodiment acquires feature data 13 and history data 14. The feature data 13 indicates a plurality of first feature amounts corresponding to a plurality of options. The history data 14 associates the value of a person's attribute with the person's selection result. The information processing device 10 generates feature data 16a, 16b for each of classes 15a, 15b having different reference options designated from the plurality of options. The feature data 16a, 16b each indicate a plurality of second feature amounts converted from the plurality of first feature amounts using the first feature amount corresponding to the reference option.

[0040] The information processing device 10 uses the history data 14 and the feature amount data 16a and 16b to generate a prediction model 17 including models 18a and 18b corresponding to the classes 15a and 15b and a model 19. The models 18a and 18b each calculate a prediction result of predicting a selection behavior. The model 19 calculates a classification result into the classes 15a and 15b based on the attribute value.

[0041] As a result, the information processing device 10 or another information processing device can execute a human behavior simulation that predicts people's selection behavior using the generated prediction model 17. Furthermore, the information processing device 10 or another information processing device can predict a change in people's selection behavior when the feature amount of each option changes. Therefore, the prediction model 17 can provide useful information when considering a measure to change the feature amount of each option.

[0042] Furthermore, the prediction model 17 can calculate prediction results that take into account the reference feature amount of each person, which is difficult to observe directly. Therefore, the prediction accuracy of the prediction model 17 is improved compared to a model that does not take into account the reference feature amount or a model that regards the reference feature amount of all people as the same.

[0043] The prediction model 17 may combine the prediction results calculated by the models 18a and 18b using class probabilities included in the classification result of the model 19. This improves the prediction accuracy for people who have reference features that do not completely match the features of any of the options. The information processing device 10 may determine the parameter values ​​included in the models 18a and 18b together with the parameter values ​​included in the model 19 using the history data 14. This improves the accuracy of the final prediction result output by the prediction model 17.

[0044] Furthermore, the information processing device 10 may update the parameter values ​​included in the models 18a and 18b based on the prediction results calculated from the feature amount data 16a and 16b, respectively, and the selection results included in the history data 14. This allows the models 18a and 18b to be trained to evaluate each option according to a relative value from the reference feature amount. Furthermore, the model 19 may be a neural network. This improves the classification accuracy of the model 19 and allows the information processing device 10 to efficiently train the model 19.

[0045] [Second embodiment] Next, a second embodiment will be described. The information processing device 100 of the second embodiment generates a prediction model that predicts a person's selection behavior for a plurality of options. Moreover, the information processing device 100 predicts a selection behavior using the generated prediction model. However, the generated prediction model may be used by another information processing device. The information processing device 100 may be a client device or a server device. The information processing device 100 may be called a computer, a model generation device, a machine learning device, a behavior prediction device, or a simulation device. The information processing device 100 corresponds to the information processing device 10 of the first embodiment.

[0046] Here, the predictive model of the second embodiment provides a virtual environment that reproduces the real world with high accuracy based on IoT (Internet of Things) data collected from sensor devices. Such a virtual environment is sometimes called a social digital twin. This virtual environment makes it possible to simulate changes in the real world when the design of a social infrastructure is changed. Such a simulation is sometimes called a digital rehearsal.

[0047] In the second embodiment, as an example, a human behavior simulation is assumed that predicts changes in people's travel methods when the fare for a transportation facility is changed. For example, a shared mobility service allows users to rent a vehicle at a base near the departure point and return the vehicle at a base near the destination, instead of traveling in a vehicle they own.

[0048] In this case, since the population density is not uniform, there is a possibility that the returned vehicles will be unevenly distributed among the bases. In order to reduce the uneven distribution of the number of vehicles, a measure to change the fee depending on the return base can be considered. The prediction model generated by the second embodiment can be applied to a simulation of the change in the return base when the fee is changed.

[0049] FIG. 2 illustrates an example of hardware of an information processing apparatus according to the second embodiment. The information processing device 100 has a CPU 101, a RAM 102, a HDD 103, a GPU 104, an input interface 105, a medium reader 106, and a communication interface 107, all connected via a bus. The CPU 101 corresponds to the processing unit 12 in the first embodiment. The RAM 102 or the HDD 103 corresponds to the storage unit 11 in the first embodiment.

[0050] The CPU 101 is a processor that executes instructions of a program. The CPU 101 loads the program and data stored in the HDD 103 into the RAM 102 and executes the program. The information processing device 100 may have multiple processors.

[0051] The RAM 102 is a volatile semiconductor memory that temporarily stores programs executed by the CPU 101 and data used in calculations by the CPU 101. The information processing device 100 may have a type of volatile memory other than a RAM.

[0052] The HDD 103 is a non-volatile storage that stores software programs such as an operating system (OS), middleware, and application software, and data. The information processing device 100 may have other types of non-volatile storage, such as a flash memory or a solid state drive (SSD).

[0053] The GPU 104 performs image processing in cooperation with the CPU 101, and outputs an image to a display device 111 connected to the information processing device 100. The display device 111 is, for example, a CRT (Cathode Ray Tube) display, a liquid crystal display, an organic EL (Electro Luminescence) display, or a projector. Other types of output devices, such as a printer, may be connected to the information processing device 100.

[0054] The GPU 104 may be used as a general purpose computing on graphics processing unit (GPGPU). The GPU 104 may execute a program in response to an instruction from the CPU 101. The information processing device 100 may include a volatile semiconductor memory other than the RAM 102 as a GPU memory.

[0055] The input interface 105 receives an input signal from an input device 112 connected to the information processing device 100. The input device 112 is, for example, a mouse, a touch panel, or a keyboard. A plurality of input devices may be connected to the information processing device 100.

[0056] The medium reader 106 is a reading device that reads the program and data recorded on the recording medium 113. The recording medium 113 is, for example, a magnetic disk, an optical disk, or a semiconductor memory. The magnetic disk includes a flexible disk (FD) and a HDD. The optical disk includes a compact disc (CD) and a digital versatile disc (DVD). The medium reader 106 copies the program and data read from the recording medium 113 to another recording medium such as the RAM 102 or the HDD 103. The read program may be executed by the CPU 101.

[0057] The recording medium 113 may be a portable recording medium. The recording medium 113 may be used for distributing programs and data. The recording medium 113 and the HDD 103 may be called computer-readable recording media.

[0058] The communication interface 107 communicates with other information processing devices via the network 114. The communication interface 107 may be a wired communication interface connected to a wired communication device such as a switch or a router, or a wireless communication interface connected to a wireless communication device such as a base station or an access point.

[0059] Next, the basic structure of the discrete choice model will be explained. FIG. 3 is a diagram illustrating an example of the first discrete choice model. A discrete choice model is a model that predicts the selection behavior of an individual person in which he or she selects one of multiple options. FIG. 3 shows an example of a simple discrete choice model. Choice data 131 is provided to the discrete choice model. Choice data 131 includes the cost of each of multiple options representing different modes of transportation. The cost is a feature value that is more disadvantageous to the user as the value increases, such as the fare or travel time.

[0060] This discrete choice model includes a utility function 132 and a softmax function 133. The utility function 132 includes variables of option attributes. The option attributes are attributes possessed by the options, and correspond to the cost of each option included in the option data 131 here. The utility function 132 may further include variables of personal attributes and variables of environmental attributes. The personal attributes are attributes possessed by the person to be predicted, such as age and gender. The environmental attributes are attributes possessed by the environment at the time of decision-making, such as temperature. However, for the sake of simplicity in the second embodiment, the utility function 132 and utility functions described later include variables of option attributes, and do not include variables of personal attributes or environmental attributes.

[0061] The utility function 132 calculates utility from the value of the variable. Utility is an evaluation value that indicates the satisfaction of the option for the person to be predicted, and the larger the value, the higher the satisfaction. The utility function 132 is, for example, a polynomial including a parameter w that acts on the variable of the option attribute. In the second embodiment, there are three options: bus, train, and taxi. The utility function 132 calculates the utility of the bus from the cost of the bus. Similarly, the utility function 132 calculates the utility of the train from the cost of the train, and calculates the utility of the taxi from the cost of the taxi.

[0062] The softmax function 133 calculates a probability distribution 134 from the utilities of the multiple options. The probability distribution 134 includes the selection probability of each of the multiple options. The selection probability of each option is a numerical value between 0 and 1 (0% and 100%). The sum of the selection probabilities of the multiple options is 1 (100%). It can be said that the softmax function 133 performs a relative evaluation of the utilities of the multiple options. As an example, the probability distribution 134 indicates that the probability of selecting a bus is 42%, the probability of selecting a train is 33%, and the probability of selecting a taxi is 25%.

[0063] The discrete choice model may use a relative evaluation function other than the softmax function 133 to calculate the probability distribution 134 from the utilities of multiple options. The discrete choice model calculates the probability distribution 134 for each individual. The selection behavior of a group is represented, for example, by an average probability distribution obtained by averaging the probability distributions 134 of multiple people included in the group.

[0064] The value of the parameter w included in the utility function 132 is determined by machine learning using behavioral history data collected from the real world. The behavioral history data includes selection results indicating options actually selected by each of a plurality of people. When the utility function 132 further includes variables of personal attributes and variables of environmental attributes, the behavioral history data further includes values ​​of the personal attributes and the environmental attributes. The behavioral history data related to transportation is extracted, for example, from the use history of electronic money.

[0065] The machine learning determines the value of the parameter w from the behavior history data so as to maximize the log-likelihood. The log-likelihood function calculates the logarithm of the selection probability corresponding to the option actually selected for each of a plurality of selection actions, and calculates the sum of these logarithms as the log-likelihood. The machine learning may determine the value of the parameter w by an iterative method such as the steepest descent method.

[0066] For example, the machine learning first sets an initial value for the parameter w. The machine learning inputs option data 131 to a utility function 132, and calculates a probability distribution 134 based on the current value of the parameter w. The machine learning compares the probability distribution 134 with the behavioral history data to calculate a log-likelihood, and updates the value of the parameter w so that the log-likelihood increases. The machine learning repeats updating the value of the parameter w until a stopping condition is satisfied. The stopping condition may be that the number of updates to the value of the parameter w exceeds a threshold. Alternatively, the stopping condition may be that the value of the parameter w has converged, such as the amount of update to the value of the parameter w being less than a threshold.

[0067] The simple discrete choice model above assumes that utility varies linearly with the cost of the option. However, due to the existence of cognitive biases, this assumption may not be strictly valid. When evaluating the value of each option, people have a subjective baseline cost that serves as a standard for them. This baseline cost is sometimes called a reference point.

[0068] According to prospect theory in behavioral economics, the change in utility relative to cost is asymmetric between an area more favorable than the reference point and an area less favorable than the reference point. Normally, the change in utility is greater in an area less favorable than the reference point compared to an area more favorable than the reference point. This reflects people's tendency to overestimate the loss of being less favorable than the reference point. Therefore, it is conceivable that by reflecting this kind of cognitive bias in a discrete choice model, prediction accuracy can be improved.

[0069] FIG. 4 is a diagram illustrating an example of the second discrete choice model. This discrete choice model uses a person's reference point to convert option data 131 into relative value data 135 to calculate a probability distribution 137 for that person. Here, we assume that the cost corresponding to the reference point is 300. This reference point is the cost that the person considers to be the most "normal" for using transportation.

[0070] The relative value data 135 includes the relative cost of each of the multiple options. The relative cost is a cost relative to a reference point, and is calculated, for example, by subtracting the reference point from the original cost. For example, the relative cost of a bus is 100-300=-200, the relative cost of a train is 300-300=0, and the relative cost of a taxi is 500-300=+200.

[0071] The discrete choice model includes a utility function 136 and a softmax function 133. The utility function 136 is sometimes called a value function. Like the utility function 132 described above, the utility function 136 includes variables for option attributes, which correspond to the relative costs of the individual options contained in the relative value data 135.

[0072] The utility function 136 calculates the utility of each of a plurality of options. The utility function 136 is, for example, a nonlinear function in order to express the psychology that the change in utility relative to the cost is large near the reference point and the change in utility relative to the cost is small away from the reference point.

[0073] Here, the utility function 136 includes parameters w and w' that act on the variables of the option attributes. The parameter w is applied to options whose relative cost is equal to or less than 0. The parameter w' is applied to options whose relative cost is greater than 0. Usually, the value of the parameter w is different from the value of the parameter w'. For example, the value of the parameter w' is greater than the value of the parameter w. This expresses the asymmetry of the utility function 136 before and after the reference point.

[0074] The softmax function 133 calculates a probability distribution 137 from the utilities of multiple options. As an example, the probability distribution 137 indicates that the probability of selecting a bus is 60%, the probability of selecting a train is 30%, and the probability of selecting a taxi is 10%. In the probability distribution 137, the probability of selecting a bus is higher and the probability of selecting a taxi is lower than in the probability distribution 134. The probability distribution 137 reflects loss aversion, that is, people dislike options that are less favorable than a reference point. Usually, the probability distribution 137 predicts a person's selection behavior with higher accuracy than the probability distribution 134.

[0075] However, the exact reference point differs from person to person and is a latent variable that each person possesses internally, making it difficult to observe directly from the outside. Therefore, it is not easy to generate a discrete choice model that has the exact reference point as an explanatory variable. On the other hand, approximating the reference points of all people with the same fixed value may not provide sufficient prediction accuracy. Therefore, the information processing device 100 generates a prediction model that takes into account the existence of the reference point as follows.

[0076] FIG. 5 is a diagram showing an example of calculation of selection probability using a plurality of classes with different reference points. The information processing device 100 adopts the following assumption regarding the reference point. Each of a plurality of people considers one of a plurality of options to be a standard option, and has a reference point close to the cost of the standard option. Therefore, in probability calculation, the information processing device 100 approximates the reference point with the cost of one of the options. Also, people with similar values ​​of personal attributes are likely to take similar selection behavior.

[0077] Therefore, the information processing device 100 defines a plurality of classes corresponding to a plurality of options. A certain class corresponds to a set of people who consider a certain option to be a standard option, and can also be interpreted as a set of people who most frequently select that option. The information processing device 100 approximates the reference point of people belonging to that class by the cost of that option.

[0078] In the second embodiment, the information processing device 100 defines a bus class, a train class, and a taxi class. In probability calculation, a person belonging to the bus class has the cost of a bus as a reference point. A person belonging to the train class has the cost of a train as a reference point. A person belonging to the taxi class has the cost of a taxi as a reference point.

[0079] The information processing device 100 classifies people into classes based on the values ​​of personal attributes such as age and sex. However, people with the same values ​​of personal attributes do not necessarily consider the same option to be standard. Therefore, instead of exclusively assigning each of a plurality of people to one class, the information processing device 100 calculates the class probability for each of a plurality of classes. The class probability for a certain class represents the probability that a person belongs to that class.

[0080] The information processing device 100 generates a discrete choice model for each of a plurality of classes. The discrete choice model of a certain class calculates a probability distribution from relative value data converted using a reference point of the class. This discrete choice model is based on prospect theory and has two types of parameters that switch depending on whether the relative cost is positive or negative. The parameter values ​​are usually different between discrete choice models of different classes.

[0081] Discrete choice model 141 is a discrete choice model for a bus class, and calculates probability distribution 144 when the reference point is the cost of a bus. Discrete choice model 142 is a discrete choice model for a train class, and calculates probability distribution 145 when the reference point is the cost of a train. Discrete choice model 143 is a discrete choice model for a taxi class, and calculates probability distribution 146 when the reference point is the cost of a taxi.

[0082] The information processing device 100 calculates the probability distribution 147 of the person by combining the probability distributions of the multiple classes according to the class probabilities. The class probabilities function as weights for the multiple classes. The probability distribution 147 is a weighted average of the probability distributions of the multiple classes. Here, the information processing device 100 calculates the probability distribution 147 by combining the probability distribution 144 of the bus class, the probability distribution 145 of the train class, and the probability distribution 146 of the taxi class.

[0083] The bus selection probability included in probability distribution 147 is the weighted average of the bus selection probabilities included in probability distributions 144, 145, and 146 weighted by the class probability. Similarly, the train selection probability included in probability distribution 147 is the weighted average of the train selection probabilities included in probability distributions 144, 145, and 146 weighted by the class probability. The taxi selection probability included in probability distribution 147 is the weighted average of the taxi selection probabilities included in probability distributions 144, 145, and 146 weighted by the class probability.

[0084] FIG. 6 is a diagram illustrating an example of a prediction model including a multiple class discrete choice model. The information processing device 100 converts the option data 131 into relative value data for each class. For the bus class, the information processing device 100 relativizes the cost of each option using the cost of a bus as a reference point. Similarly, for the train class, the information processing device 100 relativizes the cost of each option using the cost of a train as a reference point. For the taxi class, the information processing device 100 relativizes the cost of each option using the cost of a taxi as a reference point.

[0085] The information processing device 100 calculates a probability distribution 144 from the relative value data of the bus class using a discrete choice model 141. The discrete choice model 141 includes a utility function and a softmax function. The utility function of the discrete choice model 141 includes parameters w1 and w1'. The parameter w1 is used to calculate the utility of the bus. The parameter w1' is used to calculate the utility of the train and the utility of the taxi.

[0086] Similarly, the information processing device 100 calculates a probability distribution 145 from the relative value data of the train class using a discrete choice model 142. The discrete choice model 142 includes a utility function and a softmax function. The utility function of the discrete choice model 142 includes parameters w2 and w2'. The parameter w2 is used to calculate the utility of the bus and the utility of the train. The parameter w2' is used to calculate the utility of the taxi.

[0087] The information processing device 100 calculates a probability distribution 146 from the relative value data of the taxi class using a discrete choice model 143. The discrete choice model 143 includes a utility function and a softmax function. The utility function of the discrete choice model 143 includes a parameter w3. The parameter w3 is used to calculate the utility of the bus, the utility of the train, and the utility of the taxi.

[0088] Moreover, the information processing device 100 inputs personal data 151 to a class estimation model 152 to calculate class probabilities 153, 154, and 155. The personal data 151 includes age and gender. As will be described later, in the second embodiment, the information processing device 100 implements the class estimation model 152 as a neural network. However, the class estimation model 152 may be a type of model other than a neural network.

[0089] Class probabilities 153, 154, and 155 are each a numerical value between 0 and 1. The sum of class probabilities 153, 154, and 155 is 1. Class probability 153 represents the probability of belonging to a bus class. Class probability 154 represents the probability of belonging to a train class. Class probability 155 represents the probability of belonging to a taxi class.

[0090] The information processing device 100 calculates a probability distribution 147 by weighting and combining the probability distributions 144, 145, and 146 with the class probabilities 153, 154, and 155. The probability distribution 147 is the sum of the product of the probability distribution 144 and the class probability 153, the product of the probability distribution 145 and the class probability 154, and the product of the probability distribution 146 and the class probability 155.

[0091] Probability distribution 144 corresponds to the conditional probability assuming that the reference point of a person with personal data 151 is the cost of a bus. Probability distribution 145 corresponds to the conditional probability assuming that the reference point of that person is the cost of a train. Probability distribution 146 corresponds to the conditional probability assuming that the reference point of that person is the cost of a taxi. The product of probability distribution 144 and class probability 153 corresponds to the posterior probability of the bus class. The product of probability distribution 145 and class probability 154 corresponds to the posterior probability of the train class. The product of probability distribution 146 and class probability 155 corresponds to the posterior probability of the taxi class.

[0092] The prediction models include discrete choice models 141, 142, and 143 and a class estimation model 152. In machine learning, the information processing device 100 uses collected behavioral history data to collectively train parameters w1, w1', w2, w2', and w3 included in the discrete choice models 141, 142, and 143 and parameters included in the class estimation model 152.

[0093] For example, the information processing device 100 initializes parameters w1, w1', w2, w2', and w3 and parameters of a class estimation model 152. The information processing device 100 calculates probability distributions 144, 145, and 146 using option data 131 indicating the cost of each past option. The information processing device 100 inputs personal data 151 to the class estimation model 152 for each of multiple selection actions recorded in the behavior history data, and calculates a probability distribution 147. The information processing device 100 calculates a log likelihood by comparing the probability distribution 147 with an actual selection result for each of the multiple selection actions.

[0094] To increase the log likelihood, the information processing device 100 updates the parameters w1, w1', w2, w2', and w3 and the parameter values ​​of the class estimation model 152. The information processing device 100 repeats updating the parameter values ​​until a stopping condition is satisfied.

[0095] The information processing device 100 may use the trained prediction model to predict the selection behavior of a new person not listed in the behavior history data under the existing cost. In this case, the information processing device 100 does not change the option data 131, inputs the personal data 151 of the new person to the class estimation model 152, and calculates the probability distribution 147.

[0096] Furthermore, the information processing device 100 may use the trained prediction model to predict selection behavior when the cost of an option is changed. In this case, the information processing device 100 modifies the cost included in the option data 131 and generates relative value data for each of a plurality of classes from the modified option data 131.

[0097] At this time, the reference point of a certain class is the revised cost corresponding to the standard option of that class. For example, if the cost of a bus is revised from 100 to 200, the information processing device 100 may regard the reference point of the bus class as 200 and relativize the cost of each option. However, the information processing device 100 may use the cost before revision as the reference point. For example, even if the cost of a bus is revised from 100 to 200, the information processing device 100 may regard the reference point of the bus class as 100 and relativize the cost of each option.

[0098] The information processing device 100 applies trained discrete choice models 141, 142, and 143 to the changed relative value data to calculate probability distributions 144, 145, and 146. The information processing device 100 inputs personal data 151 of known or new individuals described in the behavior history data to a class estimation model 152 to calculate a probability distribution 147 after cost correction.

[0099] FIG. 7 is a diagram showing an example of input and output data of a discrete choice model. As an example, the information processing device 100 uses a utility function shown in Equation (1). In Equation (1), V represents utility, and Δx represents relative gain from a reference point. The relative gain is an index value that is more advantageous as its value increases, and is, for example, a numerical value obtained by inverting the positive and negative of the relative cost. w, w', α, β, and λ are parameters whose values ​​are determined through machine learning.

[0100] This utility function calculates the utility V using different formulas depending on whether the relative payoff Δx is non-negative or negative. The parameters w and α are used when the relative payoff Δx is non-negative. The parameters w', β, and λ are used when the relative payoff Δx is negative.

[0101]

number

[0102] Moreover, the information processing device 100 uses a softmax function shown in Equation (2). In Equation (2), p k represents the selection probability of option k, and V k represents the utility of option k. exp represents the exponential function of Napier's constant e, and Σ i represents the sum of the values ​​for all options.

[0103]

number

[0104] Here, an example will be described in which a probability distribution 145 of the train class is calculated using the discrete choice model 142. The information processing device 100 converts the option data 131 into relative value data 148. The relative value data 148 includes the relative cost of each of a plurality of options. This relative cost is obtained by subtracting the cost of the train, 300, from the original cost. The relative cost of the bus is -200, the relative cost of the train is 0, and the relative cost of the taxi is +200.

[0105] The information processing device 100 converts the relative value data 148 into payoff data 149. The payoff data 149 includes the relative payoff of each of a plurality of options. The relative payoff is obtained by inverting the positive and negative of the relative cost. The relative payoff of the bus is +200, the relative payoff of the train is 0, and the relative payoff of the taxi is -200.

[0106] The discrete choice model 142 substitutes the relative payoff of the bus for Δx in formula (1) to calculate the utility of the bus. Similarly, the discrete choice model 142 substitutes the relative payoff of the train for Δx to calculate the utility of the train, and substitutes the relative payoff of the taxi for Δx to calculate the utility of the taxi. The discrete choice model 142 calculates the utility of the bus by substituting V i The utility of the bus, the utility of the train, and the utility of the taxi are substituted into (i=1, 2, 3) to calculate the probability of choosing the bus, the probability of choosing the train, and the probability of choosing a taxi. Probability distribution 145 includes these three selection probabilities.

[0107] FIG. 8 is a diagram illustrating an example of a class estimation model. As an example, the class estimation model 152 is a neural network including an input layer, one intermediate layer, and an output layer. The input layer is two-dimensional, the intermediate layer is five-dimensional, and the output layer is three-dimensional. However, the number of dimensions of the input layer depends on the number of personal attributes, and the number of dimensions of the output layer depends on the number of classes. The number of dimensions of the intermediate layer may be adjusted according to the number of dimensions of the input layer and the number of dimensions of the output layer.

[0108] The input layer includes nodes 161 and 162. Node 161 represents age, and node 162 represents gender. A non-negative integer representing age is input to node 161. A numeric value representing gender is input to node 162. For example, males are converted to 1, and females are converted to 0.

[0109] The output layer includes nodes 163, 164, and 165. Node 163 represents class probability 153 of bus, node 164 represents class probability 154 of train, and node 165 represents class probability 155 of taxi. The output layer uses a softmax function as an activation function. Thereby, the numerical values ​​calculated in nodes 163, 164, and 165 represent probabilities. The neural network includes parameters indicating the weights of edges between nodes. The information processing device 100 determines the values ​​of these parameters by machine learning.

[0110] Next, the functions and processing procedures of the information processing device 100 will be described. FIG. 9 is a block diagram illustrating an example of functions of the information processing device. The information processing device 100 has an option storage unit 121, a behavior history storage unit 122, a model storage unit 123, a model generation unit 124, and a prediction unit 125. The option storage unit 121, the behavior history storage unit 122, and the model storage unit 123 are implemented, for example, using the RAM 102 or the HDD 103. The model generation unit 124 and the prediction unit 125 are implemented, for example, using the CPU 101 or the GPU 104 and a program.

[0111] The option storage unit 121 stores option data 131. The option data 131 is created by, for example, a user. The option data 131 may be input by the user to the information processing device 100, or may be received from another information processing device.

[0112] The behavior history storage unit 122 stores behavior history data. The behavior history data indicates multiple selection actions by multiple people. The behavior history data associates values ​​of personal attributes with selection results. The model storage unit 123 stores a trained prediction model. The prediction model includes a discrete selection model for each of multiple classes and a class estimation model 152.

[0113] The model generation unit 124 trains a prediction model by machine learning using the option data 131 stored in the option storage unit 121 and the behavior history data stored in the behavior history storage unit 122. The model generation unit 124 stores the trained prediction model in the model storage unit 123. The model generation unit 124 may display the trained prediction model on the display device 111, or may transmit it to another information processing device.

[0114] The prediction unit 125 reads out the trained prediction model from the model storage unit 123. The prediction unit 125 also receives option data and personal data including values ​​of one or more personal attributes. This option data may be the same as the option data 131 used by the model generation unit 124, or may be new option data with a corrected cost. The personal data may be data of a known person included in the behavior history data used by the model generation unit 124, or may be data of a new person not included in the behavior history data.

[0115] The prediction unit 125 inputs the option data and the values ​​of personal attributes of each person included in the personal data into a prediction model to calculate a probability distribution predicting the selection behavior of each person. When predicting the selection behavior of a group, the prediction unit 125 averages the probability distributions of multiple people included in the group to calculate an overall probability distribution. The prediction unit 125 outputs the prediction result. The prediction unit 125 may store the prediction result in a non-volatile storage, may display it on the display device 111, or may transmit it to another information processing device.

[0116] FIG. 10 is a diagram illustrating an example of a behavior history table. The behavior history table 126 is stored in the behavior history storage unit 122. The behavior history table 126 stores a plurality of records, each of which includes a record ID, age, gender, and selection result. One record corresponds to one selection behavior. The behavior history table 126 stores a mixture of records relating to different people. The behavior history table 126 may store two or more records relating to the same person but with different decision-making times.

[0117] The record ID identifies the record. Age and gender are examples of personal attributes. The behavior history table 126 may include personal attributes other than age and gender. The selection result indicates the option selected by the person from multiple options. For example, the selection result of each record is one of bus, train, and taxi.

[0118] FIG. 11 is a flowchart showing an example of a procedure for generating a model. (S10) The model generation unit 124 defines a plurality of classes corresponding to a plurality of options indicated by the option data 131. The model generation unit 124 initializes parameters included in the discrete choice model of each of the plurality of classes. In addition, the model generation unit 124 initializes parameters included in the class estimation model 152.

[0119] (S11) The model generation unit 124 identifies a cost corresponding to a corresponding option for each of a plurality of classes as a reference point. The model generation unit 124 converts the cost of each of the plurality of options into a relative cost using the reference point. Furthermore, the model generation unit 124 converts the relative cost of each of the plurality of options into a relative gain, the larger the value, the more advantageous it is.

[0120] (S12) The model generation unit 124 calculates a probability distribution indicating the selection probability of the multiple options for each of the multiple classes using a discrete choice model. In the discrete choice model, the model generation unit 124 inputs the relative gain of each option into a utility function to calculate the utility, and inputs the utilities of the multiple options into a softmax function to calculate the probability distribution.

[0121] (S13) The model generation unit 124 inputs the value of the individual attribute to the class estimation model 152 for each record included in the behavior history table 126, and calculates the class probability. (S14) The model generation unit 124 synthesizes the probability distributions of the multiple classes calculated in step S12, for each record included in the behavior history table 126, using the class probabilities calculated in step S13 as weights.

[0122] (S15) The model generation unit 124 extracts the selection probability of the correct answer option from the probability distribution synthesized in step S14 for each record included in the behavior history table 126. The model generation unit 124 calculates the logarithm of the extracted selection probabilities for all records included in the behavior history table 126, to calculate the log likelihood.

[0123] (S16) The model generation unit 124 updates the parameter values ​​included in the discrete choice model of each class and the parameter values ​​included in the class estimation model 152 so as to increase the log likelihood. For example, the model generation unit 124 changes the values ​​of each of the multiple parameters by a small amount to calculate the gradient of the log likelihood, and updates the parameter values ​​based on the gradient.

[0124] (S17) The model generation unit 124 determines whether a stopping condition is satisfied. For example, the model generation unit 124 determines whether the number of iterations of steps S12 to S16 exceeds a threshold. Also, for example, the model generation unit 124 determines whether the parameter values ​​have converged, for example, by determining whether the update amounts of all parameters are less than a threshold. If the stopping condition is satisfied, the process proceeds to step S18. If the stopping condition is not satisfied, the process proceeds to step S12.

[0125] (S18) The model generation unit 124 outputs a prediction model including a discrete choice model for each of the multiple classes and the class estimation model 152. FIG. 12 is a flowchart illustrating an example of a procedure for behavior prediction.

[0126] (S20) The prediction unit 125 acquires option data indicating the costs of a plurality of options and personal data indicating the values ​​of one or more personal attributes. (S21) The prediction unit 125 identifies a cost corresponding to a corresponding option for each of a plurality of classes as a reference point. However, the prediction unit 125 may use the reference point at the time of model generation. The model generation unit 124 converts the cost of each of the plurality of options into a relative cost using the reference point. Furthermore, the model generation unit 124 converts the relative cost of each of the plurality of options into a relative gain, which is more advantageous as the value increases.

[0127] (S22) The prediction unit 125 calculates a probability distribution indicating the selection probability of the multiple options for each of the multiple classes using a discrete choice model. In the discrete choice model, the prediction unit 125 inputs the relative gain of each option into a utility function to calculate the utility, and inputs the utilities of the multiple options into a softmax function to calculate the probability distribution.

[0128] (S23) The prediction unit 125 inputs the values ​​of the personal attributes for each of the one or more persons into the class estimation model 152 to calculate the class probability. (S24) The prediction unit 125 uses the class probabilities calculated in step S23 as weights for each of the one or more people to synthesize the probability distributions of the multiple classes calculated in step S22.

[0129] (S25) The prediction unit 125 averages the probability distributions of one or more people combined in step S24 to calculate an overall probability distribution. (S26) The prediction unit 125 outputs the calculated overall probability distribution. Note that the prediction unit 125 may output the probability distribution of each individual person.

[0130] As described above, the information processing device 100 according to the second embodiment generates a prediction model for predicting a person's selection behavior using behavior history data collected from the real world. This allows the information processing device 100 to simulate measures such as changing the cost of options in a virtual environment using the generated prediction model.

[0131] Furthermore, the information processing device 100 generates a prediction model that takes into account differences in reference points that are difficult to observe directly, thereby improving prediction accuracy compared to a prediction model that does not take reference points into account or a prediction model that approximates the reference points of all people with the same fixed value.

[0132] Furthermore, the information processing device 100 assumes a class having a cost corresponding to any one of a plurality of options as a reference point, and generates a discrete choice model for each class. Furthermore, the information processing device 100 generates a class estimation model 152 that calculates the probability of belonging to each class from the value of the personal attribute. The prediction model uses the probability of belonging as a weight to combine the prediction results for each class, and calculates the final prediction result. This enables the information processing device 100 to implement a prediction model that takes into account differences in reference points, improving prediction accuracy.

[0133] Furthermore, the information processing device 100 collectively trains the discrete choice model of each class and the class estimation model 152 using the behavior history data. As a result, the discrete choice model and the class estimation model 152 are trained so that the final prediction result approaches the actual selection result. Furthermore, the information processing device 100 distinguishes between the parameters of the discrete choice model of a certain class and the parameters of the discrete choice model of another class and adjusts the values. As a result, the prediction model can appropriately reflect the difference in the reference point in the prediction result. [Explanation of symbols]

[0134] 10. Information processing device 11 Storage section 12 Processing section 13, 16a, 16b Feature data 14 Historical Data 15a, 15b class 17 Predictive Models 18a, 18b, 19 Models

Claims

1. acquiring first feature amount data indicating a plurality of first feature amounts corresponding to a plurality of options, and history data in which values ​​of attributes of a person are associated with selection results of the person for the plurality of options; generating second feature data indicating a plurality of second feature data converted from the plurality of first feature data by using a first feature data corresponding to the reference option for each of a plurality of classes having different reference options designated from among the plurality of options; generating a prediction model, which is a first model for calculating a prediction result of predicting a selection behavior for the plurality of options, including a plurality of first models corresponding to the plurality of classes and a second model for calculating a classification result into the plurality of classes based on a value of the attribute, using the history data and the second feature amount data of each of the plurality of classes; A model generation program that causes a computer to carry out the processing.

2. the classification result includes a plurality of class probabilities corresponding to the plurality of classes; The prediction model combines a plurality of prediction results calculated by the plurality of first models using the plurality of class probabilities. The model generating program according to claim 1.

3. The process of generating the prediction model includes a process of determining a value of a first parameter included in each of the plurality of first models and a value of a second parameter included in the second model using the history data. The model generating program according to claim 1.

4. the process of generating the prediction model includes a process of updating a value of a parameter included in each of the plurality of first models based on the prediction result calculated from the second feature amount data corresponding to the first model and the selection result. The model generating program according to claim 1.

5. the second model is a neural network that calculates a plurality of class probabilities corresponding to the plurality of classes from the values ​​of the attributes; The model generating program according to claim 1.

6. acquiring first feature amount data indicating a plurality of first feature amounts corresponding to a plurality of options, and history data in which values ​​of attributes of a person are associated with selection results of the person for the plurality of options; generating second feature data indicating a plurality of second feature data converted from the plurality of first feature data by using a first feature data corresponding to the reference option for each of a plurality of classes having different reference options designated from among the plurality of options; generating a prediction model, which is a first model for calculating a prediction result of predicting a selection behavior for the plurality of options, including a plurality of first models corresponding to the plurality of classes and a second model for calculating a classification result into the plurality of classes based on a value of the attribute, using the history data and the second feature amount data of each of the plurality of classes; A model generation method in which processing is performed by a computer.

7. a storage unit that stores first feature amount data indicating a plurality of first feature amounts corresponding to a plurality of options, and history data in which values ​​of attributes of a person are associated with selection results of the person for the plurality of options; a processing unit that generates, for each of a plurality of classes having different reference options designated from among the plurality of options, second feature data indicating a plurality of second feature values ​​converted from the plurality of first feature values ​​using first feature values ​​corresponding to the reference option, and calculates a prediction result predicting selection behavior for the plurality of options, the prediction model including a plurality of first models corresponding to the plurality of classes and a second model that calculates a classification result into the plurality of classes based on a value of the attribute, using the history data and the second feature data for each of the plurality of classes; An information processing device having the above configuration.

Images