Data analysis apparatus, method, and program

Abstract

To perform proper data analysis on the basis of the preference of a user.SOLUTION: A data analysis apparatus according to one embodiment includes: a setting unit for setting, in a mixture regression model which represents the distribution of preferences of a group of users and outputs data based on preferences of the users, data indicating the preference of a specific user of the group; and an estimation unit for estimating, by Bayesian estimation, parameters of a regression model representing the distribution of the preference of the specific user in the mixture regression model, and data in accordance with the preference of the specific user based on the regression model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] FIELD Embodiments of the present invention relate to a data analysis device, method, and program. [Background technology]

[0002] In many current web services, such as online shopping, video streaming, and social networking services (SNS), personalized recommendation systems that match the preferences of individual users are widely used. These systems utilize data collected in the process of users using the service, such as product purchases, video playback, and posting rating histories. For example, a model of a user's personal preferences, which indicates the types of items (such as products, videos, or accounts) that the user likes, is estimated from the data, and recommendations are made by presenting the most suitable items based on this personal model.

[0003] Although the application of personalized recommendation systems is expanding, in areas where it is difficult or impossible to create a system for semi-automatically collecting personal data, it is not possible to collect enough data to accurately estimate a model of user preferences. For this reason, it is not easy to realize a personalized recommendation system. For example, in areas related to personal subjectivity, such as estimating emotions or likeability based on facial expressions, writing, or voice, or estimating the aroma, taste, or deliciousness of food and drink, it is difficult to collect data on how users feel about a certain object, making it difficult to create a recommendation system at a practical level.

[0004] Based on the above concerns, an approach called Bayesian Optimization, which discovers the optimal target through trial and error through interaction with the user, is now being used in areas related to individual subjectivity (see, for example, Non-Patent Document 1). However, in problems that require many variables to be adjusted, such as problems that require optimizing facial expressions, sentences, or voices, the number of interactions with the user required to find the optimal target, for example, the optimal solution, i.e., the number of trials, increases, which places a heavy burden on the user.

[0005] For example, Non-Patent Document 2 explains that in order to reduce the burden on users in Bayesian optimization, similar clusters with similar preferences, i.e., groups of users, are found from the interaction history of other users by applying the k-means method, a clustering algorithm, and the number of trials required to find a "good" solution is reduced by using a model of users belonging to this cluster. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] Keito Shiroshita and Masashi Komori, Characteristics of Unpleasant Color Schemes in Living Things, Proceedings of the JSAI2020 Annual Conference, 2020. [Non-patent document 2] Kotaro Kaihatsu, Masaki Matsubara, Hiroyoshi Ito, and Atsuyuki Morishima, Efficient Estimation of Individual and Collective Preference Distributions Based on Gaussian Process Regression, Forum on Data Engineering and Information Management (DEIM), 2022. Summary of the Invention [Problem to be solved by the invention]

[0007] However, in the k-means method, the models and data of other users than the user himself are only used when determining search candidates for a solution, and are not used to estimate a model that represents the user's preferences.

[0008] The present invention has been made in light of the above circumstances, and its purpose is to provide a data analysis device, method, and program that enable appropriate data analysis in line with user preferences. [Means for solving the problem]

[0009] A data analysis device according to one embodiment of the present invention includes a setting unit that sets data indicating the preferences of a specific user belonging to a group of users into a mixed regression model that represents the distribution of preferences of the group of users and outputs data in line with the preferences of the users, and an estimation unit that uses Bayesian estimation to estimate parameters of a regression model that represents the distribution of preferences of the specific user in the mixed regression model, and data in line with the preferences of the specific user based on the regression model.

[0010] A data analysis method according to one aspect of the present invention is a method performed by a data analysis device, and includes: a setting unit of the data analysis device setting data indicating the preferences of a specific user belonging to a group of users in a mixed regression model, which represents the distribution of preferences of the group of users and outputs data in line with the preferences of the users; and an estimation unit of the data analysis device estimating, by Bayesian estimation, parameters of a regression model representing the distribution of preferences of the specific user in the mixed regression model, and data in line with the preferences of the specific user based on the regression model. [Effects of the Invention]

[0011] According to the present invention, it is possible to perform appropriate data analysis in line with the user's preferences. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram showing an application example of a data optimization device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a flowchart showing an example of a procedure of processing operations of a data optimization device according to an embodiment of the present invention. [Figure 3] FIG. 3 is a block diagram showing an example of the hardware configuration of a data optimization device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, an embodiment of the present invention will be described. In this embodiment, a Bayesian optimization method using a mixed regression model will be described. Using this mixed regression model, data and users are clustered based on the response trends shown in the interaction history of all users, and a regression model representing each user's preferences is estimated. This makes it possible to perform clustering and regression simultaneously. This makes it possible to estimate a user model with high accuracy and find the optimal target with fewer trials than existing methods.

[0014] In addition, the above optimization method uses the Dirichlet Process (DP), which makes it possible to handle a virtually infinite number of clusters, and can also handle cases where the number of clusters changes or increases as the number of users or trials increases.

[0015] Furthermore, in ordinary Bayesian optimization, a Gaussian process (GP) is often used as a regression model, but the method of this embodiment uses a model that is linear with respect to parameters (sometimes referred to as a linear model). This is because when using the above GP, the amount of calculation required for the inverse matrix calculation during learning is the cube of the number of data points in the normal method, and the amount of calculation required for the matrix calculation during prediction is the square of the number of data points in the normal method. As the number of users or the number of trials increases, the amount of calculation becomes non-negligible. Even if the regression model is the linear model, if Gaussian basis functions are used, predicted values ​​can be constructed using a linear sum of Gaussian kernels, similar to the GP. If random Fourier features are used, it is also possible to theoretically approximate a linear model to the GP. Random Fourier features are disclosed, for example, in "Ali Rahimi and Benjamin Recht, Random features for large-scale kernel machines, Advances in neural information processing systems, 20, 2007."

[0016] By using this method, it is expected that we can estimate the user model with higher accuracy and find a "good" solution with fewer trials than existing technologies.

[0017] (DP construction using the Chinese Restaurant Process (CRP)) The stick-breaking process or CRP is used to construct the DP. These processes are disclosed, for example, in "Samuel J Gershman and David M Blei, A Tutorial on Bayesian Nonparametric Models, Journal of Mathematical Psychology, 56(1):1-12, 2012." In this embodiment, we will explain the configuration of DP using CRP. CRP is a discrete-time stochastic process, and is explained as a model that expresses how customers who sequentially visit a Chinese restaurant with an infinite number of round tables stochastically select a table to sit at, thereby dividing the customers into groups with a number of tables equal to or less than the number of visiting customers.

[0018] In the infinite mixture model, CRP is used to directly model the generation of the latent variable Z without using the mixture ratio π. The number of clusters belonging to N pieces of data, the latent variable representing this cluster, and the total number of data belonging to cluster k are expressed as follows (1) to (3).

[0019]

number

[0020] Data and clusters correspond one-to-one to customers and tables in CRP. When the latent variable Z follows CRP, the generation probability is given by the probability distribution shown in (4) below.

[0021]

number

[0022] Also, from this equation, when a latent variable other than the i-th latent variable is given, as shown in (5) below, z i It can also be shown that the probability distribution that follows is expressed as follows (6), where α0 is a concentration parameter.

[0023]

number

[0024] The above (6) can be expressed using the following (7) and (8).

[0025]

number

[0026] However, the above (6) can be expressed as (9) below.

[0027]

number

[0028] (regression model) In this embodiment, a regression model shown in the following (11) having the values ​​shown in the following (10) as parameters is used.

[0029]

number

[0030] An example of a regression model is a linear regression model, which is shown in (12) below.

[0031]

number

[0032] Here, the value shown in the following (13) in the above (12) is a basis function, and as this basis function, a polynomial basis function, a Gaussian basis function, a sigmoid basis function, or the like is used.

[0033]

number

[0034] In this case, by introducing the concept of the random Fourier transform, it is possible to approximately express Gaussian process regression (GPR). Although a model such as deep learning can also be used as the regression function, in this embodiment, a parametric linear regression model is adopted to express Gaussian process regression, which is often used in Bayesian optimization.

[0035] (data) In this embodiment, the value shown in the following (14) is a set of workers (users), and the set of feedback from the workers is a value shown in the following (15).

[0036]

number

[0037] In this embodiment, the set of tuples of the features of the item presented to the worker shown in (16) below and the evaluation value shown in (17) below, which is the evaluation value obtained at the time of this presentation, is defined as the value shown in (18) below, and the set of feedback from all workers is defined as the value shown in (19) below.

[0038]

number

[0039] Here, n in (18) above w is the number of feedbacks from worker w. In this embodiment, the preference distribution of the set of feedbacks and the set of workers is expressed as a mixture of regression models. The set of workers is a cloud. In this embodiment, the entire set of presented item features is set to the value shown in the following (20), and the entire set of feedback is set to the value shown in the following (21).

[0040]

number

[0041] (Data generation process) In this embodiment, the preference distribution of each worker is expressed by a mixture model of regression models. This mixture model is referred to as a mixture of regression models (MoRMs). This MoRMs has (A1) K regression models, (A2) parameters of the mixture ratio with the regression models owned by each worker, and (A3) parameters of each regression model. The parameters of the mixture ratio of the regression models owned by each worker are shown as in (22) below, and the parameters of each regression model are shown as in (23) below.

[0042]

number

[0043] Here, the value shown in the following (24) in the above (22) is taken as a K-dimensional probability vector, and the value shown in the following (25) in the above (23) is taken as a parameter of the k-th regression model.

[0044]

number

[0045] The process of generating MoRMs has the following three steps. First, the parameters shown in (26) below are generated from the prior distribution shown in (27) below, where p(π) in (27) is a Dirichlet distribution shown in (28) below, and p(Θ) in (27) above is a Gaussian distribution shown in (29) below.

[0046]

number

[0047] Here, α0 is the concentration parameter as described above, and λ0 is the accuracy, which are hyperparameters. The 0 in (29) above can be expressed as follows (30):

[0048]

number

[0049] Next, for each worker i=1,…,n w For the reaction, the hidden variable shown in (31) below is set. z wi ={z wi1 ,…,z wiK} …(31) Regarding the above (31), if the i-th sample belongs to the k-th model, then "z wik = 1" and for other k' ≠ k, "z wik´ =0".

[0050] The hidden variables are generated from the multinomial distribution shown in (32) below.

[0051]

number

[0052] Finally, the attribute x of each item shown in (16) above wi is generated from a uniform distribution u(x), and the score y wi Let be generated from a Gaussian distribution with a regression model, as shown in (33) below.

[0053]

number

[0054] This generation step is assumed to be generated for all workers w shown in (14) above, and the random variables generated by this generation step are assumed to be collectively expressed as (34) below.

[0055]

number

[0056] Summarizing the data generation process described above, the joint probability of each random variable is expressed as follows (35).

[0057]

number

[0058] The likelihood function is obtained by integrating and eliminating the hidden variable Z, and is expressed as follows (36).

[0059]

number

[0060] Here, the following (37) holds.

[0061]

number

[0062] When expressing the above mixture model as an infinite mixture model, the mixture rate Π is not expressed explicitly, and the hidden variable z wi The probability distribution for is expressed as CRP. That is, for a value expressed as in (38) below, the CRP is expressed as in (39) below, and the joint probability distribution for all random variables is expressed as in (40) below.

[0063]

number

[0064] (Estimation algorithm) In this embodiment, we will sequentially explain (1-1) a method of estimating all parameters for a finite mixture model by Gibbs sampling, (1-2) a method of estimating only hidden variables for a finite mixture model by collapsed Gibbs sampling, (1-3) a method of estimating a random variable Π for a finite mixture model by marginalizing Gibbs sampling, (2-1) a method of estimating a random variable Π for an infinite mixture model by marginalizing sampling, and (2-2) a method of estimating only hidden variables for an infinite mixture model by collapsed Gibbs sampling.

[0065] (1-1) A method for estimating all parameters for a finite mixture model using Gibbs sampling In this section, we derive the sampling formula for each parameter.

[0066] (Z sampling) In this embodiment, the joint probability distribution is expressed as follows (41).

[0067]

number

[0068] According to Bayes' theorem, the calculation result of the posterior probability distribution (also called the posterior distribution) of Z in the above (41) is expressed as the following (42).

[0069]

number

[0070] Here, the following (43) holds.

[0071]

number

[0072] (Theta sampling) If we extract only the random variables that depend on the parameter Θ shown in (41) above and take the logarithm of the simultaneous probability, we get the following (44).

[0073]

number

[0074]

number

[0075] Here, the following (46) and (47) hold.

[0076]

number

[0077] (Sampling of Pi) The posterior probability distribution of the random variable Π shown in (41) above is expressed as follows (48).

[0078]

number

[0079] Here, the following (49) holds.

[0080]

number

[0081] (1-2) A method for estimating only the hidden variables for finite mixture models using collapsed Gibbs sampling In this section, we derive the collapsed Gibbs sampling for the hidden variable Z shown in (41) above, and explain how to learn a model by sampling only this hidden variable Z. In this method, the parameter Θ and the random variable Π shown in (41) above are integrally eliminated. Z wik The sampling probability for is given by (50) and (51) below.

[0082]

number

[0083] Here, the following (52) and (53) hold.

[0084]

number

[0085] In the above (50), Z wik If we focus on terms that depend only on , the following (54) holds.

[0086]

number

[0087] Here, for the above (54), the following (55) holds, leading to the following (56).

[0088]

number

[0089] To sum up, Z wik The sampling probability for is expressed as follows (57):

[0090]

number

[0091] (1-3) A method for estimating parameters and variables for finite mixture models using collapsed Gibbs sampling with marginalization of only the random variable Π In this section, we explain how to marginalize the random variable Π shown in (41) above and estimate only the parameter Θ and variable Z shown in (41) above by sampling. In this method, the sampling probability in (1-1) above is used for sampling the parameter Θ shown in (41) above. Using the results derived in (1-2) above, Z wik The sampling probability for is expressed as follows (58):

[0092]

number

[0093] (2-1) Estimating variable Z in an infinite mixture model using collapsed Gibbs sampling In this section, we explain how to estimate the model using collapsed Gibbs sampling for an infinite mixture model. In this method, the CRP is calculated as shown in (59) below in the derivation result in (1-2) above.

[0094]

number

[0095] Also, Z wik The sampling probability of is expressed as follows (60).

[0096]

number

[0097] Here, the following (61) holds.

[0098]

number

[0099] (2-2) A method for estimating the model using collapsed Gibbs sampling without marginalizing the parameter Θ in an infinite mixture model In this section, we propose a method for estimating the model using collapsed Gibbs sampling without marginalizing the parameter Θ shown in (41) above in the infinite mixture model. In this method, the parameter Θ shown in (41) above is not marginalized, so the parameter θ corresponding to the existing clusters is k is sampled using the method (1-1) above, and new parameters are sampled for new clusters. wik The sampling probability of is expressed as follows (62).

[0100]

number

[0101] In this embodiment, the estimation algorithm (2-2) above is representatively described as "Algorithm 1" in (63) below.

[0102]

number

[0103] (Construction of predictive distribution) Next, we will show how to construct a predictive distribution. According to the Bayesian estimation method, a predictive distribution can be constructed by substituting multiple parameters sampled from the posterior distribution into the model parameters and averaging these predicted values. For example, if we use the method shown in (1-1) above to sample S parameters and the latent variables shown in (64) below, then we can use (37) above to find the unknown input x of worker w. * The model output y for * is expressed as follows (65).

[0104]

number

[0105] Therefore, by taking the mean over the posterior distribution, the output y * The predictive distribution of is constructed as follows (66).

[0106]

number

[0107] This predictive distribution can be used to predict unknown inputs and calculate the acquisition function described below.

[0108] In addition, a method combining point estimation and Bayesian estimation can be considered as a less computationally intensive method that avoids the averaging operation of S models required in the above-mentioned conventional Bayesian estimation method. In this method, from the parameters sampled for S models, the one parameter with the largest value of the joint probability or likelihood function of the complete data, i.e., the value expressed by (35) or (37) above, is used. This parameter is expressed as shown in (67) below.

[0109]

number

[0110] Using the above parameters, we can obtain the approximate posterior distribution expressed as (69) below, which approximates the posterior distribution of the parameters, i.e., the distribution expressed as (68) below.

[0111]

number

[0112] This can be considered as a method of constructing a posterior distribution by performing point estimation for the parameter Π and the latent variable Z. Note that the term expressed as (70) below in (69) above is the normal distribution of (45) above.

[0113] If an approximate predictive distribution is constructed using the above-mentioned approximate posterior distribution, the following equation (71) is obtained, which represents the approximate predictive distribution of worker w.

[0114]

number

[0115] Here, the following (72) and (73) hold.

[0116]

number

[0117] By using the approximate predictive distribution described above, not only can we avoid manipulating the average of sampled values ​​using the Bayesian estimation method, but we can also analytically calculate the integral, as in Bayesian linear regression, and express a distribution in which the variance of the predicted value changes depending on the variance of the posterior distribution of the parameters.

[0118] Furthermore, although we have shown here how to construct a predictive distribution when using the above-mentioned method (1-1) to estimate the model and parameters, a predictive distribution can also be constructed in the same way when using a different model or estimation method.

[0119] (Bayesian optimization) This section describes a Bayesian optimization method using the predictive distribution of worker preferences. Based on the predictive distribution, an acquisition function is designed to more quickly find the point that maximizes the user's evaluation value, and the next data point to query the worker is determined based on the function value. In this embodiment, two Bayesian optimization strategies, the UCB (Upper Confidence Bound) strategy and the ICD (Inter Cluster Discrepancy) strategy, are described.

[0120] (UCB Policy) In Bayesian optimization, a method for determining query data based on a strategy called the UCB strategy is often adopted. The UCB strategy calculates the magnitude of the expected value of the evaluation value and the magnitude of the uncertainty of the evaluation value of the data point at the time of the query, and selects data based on the criterion that the larger the sum of these, the more preferable it is for the query.

[0121] In this embodiment, we propose a new acquisition function for the UCB policy by calculating the expected value and variance of the proposed model. The acquisition function in the UCB policy is given as follows (74).

[0122]

number

[0123] Here, the following (75) and (76) hold.

[0124]

number

[0125] The α in (74) above is a hyperparameter. The larger this α, the more data with high uncertainty will be selected.

[0126] (ICD Policy) The ICD strategy prioritizes the selection of points where the difference between the probability density functions of the regression models corresponding to each cluster is large. This strategy allows us to quickly determine which cluster each worker is likely to belong to. Furthermore, by utilizing the expected value of the regression model of the cluster obtained from previous responses, we can expect to be able to select data with high evaluation values ​​more quickly.

[0127] In this embodiment, we adopt the Jensen-Shannon (JS) divergence to calculate the difference between the probability distributions, because the JS divergence is symmetric for two probability distributions. The acquisition function in the ICD strategy is given as follows (77):

[0128]

number

[0129] Here, the terms in (77) above are expressed as in (78) below, leading to (79) below.

[0130]

number

[0131] Here, the terms in (77) above correspond to the JS divergence of probability distributions p and q as shown in (80) below, and are given by (81) below.

[0132]

number

[0133] The KL divergence of the predictive distribution of the regression model corresponding to the kth and lth clusters shown in (78) above is calculated as follows (82):

[0134]

number

[0135] Therefore, the JS divergence is calculated as follows (83).

[0136]

number

[0137] (Overall flow of Bayesian optimization) The overall flow of Bayesian optimization is given as "Algorithm 2" as shown in (84) below.

[0138]

number

[0139] In this process, the next data to query for each worker is determined based on the acquisition function, and the pair of data and the evaluation value at the time of determination is added to the dataset. After repeating the above process a certain number of times, the model is updated.

[0140] Next, an outline of the operation according to this embodiment will be described. Figure 1 is a diagram showing an application example of a data optimization device according to one embodiment of the present invention. As shown in FIG. 1, a data optimization device 100, which is a data analysis device according to one embodiment of the present invention, includes a preset parameter processing unit 10, a Bayesian optimization execution unit 20, an estimation / optimization result processing unit 30, a recording unit 40, and an input / output unit 50. The Bayesian optimization execution unit 20 includes a mixed regression model parameter estimation unit 21 and a mixed regression model acquisition function optimization unit 22 . The recording unit 40 includes a preset parameter recording unit 41 , a data recording unit 42 , and a parameter recording unit 43 .

[0141] FIG. 2 is a flowchart showing an example of a procedure of processing operations of a data optimization device according to an embodiment of the present invention. (B1) The preset parameter processing unit 10 inputs the hyperparameters of the mixed regression model, the hyperparameters of the acquisition function, the function form and features of the model, etc. as preset parameters, and stores them in the preset parameter recording unit 41 of the recording unit 40 (S1).

[0142] (B2) The Bayesian optimization execution unit 20 inputs the preset parameters from the preset parameter recording unit 41, and collects data through interactions with workers, i.e., users, in the framework of Bayesian optimization, i.e., the method shown in the above-mentioned "Algorithm 2." The mixed regression model parameter estimation unit 21 estimates the model parameters, and the mixed regression model acquisition function optimization unit 22 estimates the optimal point (S2). The preset parameter processing unit 10 stores the acquired data in the data recording unit 42 of the recording unit 40, and stores the estimated model parameters and optimum points in the parameter recording unit 43 of the recording unit 40.

[0143] (B3) The estimation and optimization result processing unit 30 outputs the model parameters and optimal points stored in the parameter recording unit 43 to the input / output unit 50. The input / output unit 50 receives the model parameters and optimal points from the estimation and optimization result processing unit 30 and outputs them to the external device 200 (S3).

[0144] In the above embodiment, the number of interactions with each worker, i.e., the user, is the same, but the number of interactions may be different for each worker. Also, the timing for updating the parameters of the mixed regression model may be such that some or all of the parameters are updated each time data is obtained from each worker.

[0145] This embodiment can be applied regardless of the type of data collected through interactions with workers. For example, any data may be used, such as evaluation values ​​for products, videos, or accounts, values ​​representing the likes or dislikes of emotions from facial expressions, sentences, or voices, and evaluation values ​​of the aroma, taste, or deliciousness of food and drink.

[0146] FIG. 3 is a block diagram showing an example of the hardware configuration of a data optimization device according to an embodiment of the present invention. 3, the data optimization device 100 according to the embodiment is configured, for example, by a server computer or a personal computer, and has a hardware processor 111A such as a CPU (Central Processing Unit). A program memory 111B, a data memory 112, an input / output interface 113, and a communication interface 114 are connected to the hardware processor 111A via a bus 115.

[0147] The communication interface 114 includes, for example, one or more wireless communication interface units, and enables transmission and reception of information to and from a communication network. As the wireless interface, for example, an interface that adopts a low-power wireless data communication standard such as a wireless LAN (Local Area Network) is used.

[0148] The input / output interface 113 is connected to an external device 200 that is attached to the data optimization device 100 and includes an input device and an output device used by a user or the like. The input / output interface 113 takes in operation data input by a user or the like through an input device such as a keyboard, a touch panel, a touchpad, a mouse, etc., and outputs output data to an output device including a display device using a liquid crystal or an organic EL (Electro Luminescence) display, etc. The input and output devices may be devices built into the data optimization device 100, or may be input and output devices of other information terminals that can communicate with the data optimization device 100 via a network.

[0149] The program memory 111B is a non-transitory tangible storage medium that is a combination of a non-volatile memory that can be written to and read from at any time, such as a hard disk drive (HDD) or a solid state drive (SSD), and a non-volatile memory such as a read only memory (ROM), and stores programs necessary to execute various control processes, etc., according to one embodiment.

[0150] The data memory 112 is a tangible storage medium that is, for example, a combination of the above-mentioned nonvolatile memory and a volatile memory such as RAM (Random Access Memory), and is used to store various data acquired and created during various processing steps.

[0151] A data optimization device 100 according to one embodiment of the present invention can be configured as a data analysis device having software processing function units, including a preset parameter processing unit 10, a Bayesian optimization execution unit 20, an estimation / optimization result processing unit 30, and an input / output unit 50, as shown in FIG. 3.

[0152] Each information storage unit and recording unit 40 used as a working memory by each unit of the data optimization device 100 can be configured by using the data memory 112 shown in Fig. 3. However, these configured storage areas are not essential components within the data optimization device 100, and may be areas provided in an external storage medium such as a USB (Universal Serial Bus) memory, or a storage device such as a database server located in the cloud.

[0153] The above processing function units can be realized by having the hardware processor 111A read and execute a program stored in the program memory 111B. Note that some or all of these processing function units may be realized in various other forms, including integrated circuits such as an application specific integrated circuit (ASIC) or a field-programmable gate array (FPGA).

[0154] The present invention is not limited to the above-described embodiments, and various modifications can be made in the implementation stage without departing from the spirit of the invention. Furthermore, the embodiments may be implemented in appropriate combinations, in which case the combined effects can be obtained. Furthermore, the above-described embodiments include various inventions, and various inventions can be extracted by combining selected elements from the disclosed elements. For example, if the problem can be solved and the desired effect can be obtained even if some elements are deleted from all elements shown in the embodiments, the configuration from which these elements are deleted can be extracted as an invention.

[0155] The methods described in each embodiment can be stored as a program (software means) that can be executed by a computer on a recording medium such as a magnetic disk (e.g., a floppy disk, a hard disk, etc.), an optical disk (e.g., a CD-ROM, a DVD, an MO, etc.), or a semiconductor memory (e.g., a ROM, a RAM, a flash memory, etc.), or can be distributed by transmitting it via a communication medium. The program stored on the medium also includes a configuration program that configures the software means (including not only an execution program but also tables and data structures) that the computer executes. The computer that realizes this device reads the program stored on the recording medium and, in some cases, configures the software means using the configuration program, and executes the above-mentioned processing by controlling the operation of this software means. The term "recording medium" as used herein is not limited to a storage medium for distribution, but also includes a storage medium such as a magnetic disk or semiconductor memory installed inside the computer or in a device connected via a network. [Explanation of symbols]

[0156] 100...Data optimization device 200...External device 10...Pre-setting parameter processing section 20...Bayesian optimization execution unit 21...Mixed regression model parameter estimation unit 22...Mixed regression model acquisition function optimization unit 30...Estimation and optimization result processing section 40...Recording section 41...Pre-setting parameter recording section 42...Data recording unit 43...Parameter recording section 50...Input / output section

Claims

1. a setting unit that sets data indicating the preferences of a specific user belonging to the group into a mixed regression model that represents a distribution of preferences of the group of users and outputs data in line with the preferences of the users; an estimation unit that estimates, by Bayesian estimation, parameters of a regression model that represents a distribution of preferences of the specific user in the mixed regression model, and data that is in line with the preferences of the specific user based on the regression model; A data analysis device comprising:

2. The estimation unit Calculating an acquisition function related to the estimation based on an expected value of an evaluation value for the data to be estimated and a magnitude of uncertainty of the evaluation value; Inferring data according to the specific user's preferences based on the acquisition function; The data analysis device according to claim 1 .

3. The estimation unit calculating an acquisition function related to the estimation based on the difference in the probability distributions constituting the mixed regression model; Inferring data according to the specific user's preferences based on the acquisition function; The data analysis device according to claim 1 .

4. A method performed by a data analysis device, comprising: setting data indicating the preferences of a specific user belonging to the group in a mixed regression model, which represents a distribution of preferences of a group of users and outputs data in line with the preferences of the users, by a setting unit of the data analysis device; an estimation unit of the data analysis device estimates, by Bayesian estimation, parameters of a regression model that represents a distribution of preferences of the specific user in the mixed regression model, and data that is in line with the preferences of the specific user based on the regression model; A data analysis method comprising:

5. A data analysis program that causes a processor to function as each unit of the data analysis device according to any one of claims 1 to 3.

Images