Information processor, information processing method, and program

Abstract

To realize an information processor capable of reflecting the intension of a user in parameters among multiple models.SOLUTION: The information processor includes: a first acquisition unit for acquiring inference data, trained parameters, and feature information associated with a target model or the trained parameters; a presentation unit for presenting the feature information to a user; a second acquisition unit for acquiring intention information indicating the intention of the user regarding a weight of the target model or the trained parameters; and a prediction unit for deriving a prediction result by applying, to the inference data, the trained parameters and a weight determined by the intention information.SELECTED DRAWING: Figure 1

Description

[Technical Field]

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

[0002] A technique using multiple machine learning models (also called models or prediction models) is known. As an example, Patent Document 1 describes a data prediction device in which a user inputs prediction model parameters for each of multiple prediction models. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2006-85645 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the above-mentioned data prediction device does not have a configuration for allowing a user to input parameters between multiple models, and therefore has a problem in that the above-mentioned data prediction device cannot reflect the user's intentions regarding parameters between multiple models.

[0005] The present disclosure has been made in view of the above-mentioned problems, and an exemplary purpose thereof is to provide a technique for reflecting a user's intentions regarding parameters between a plurality of models. [Means for solving the problem]

[0006] An information processing device according to an exemplary aspect of the present disclosure includes a first acquisition means for acquiring data for inference, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters, a presentation means for presenting the feature information to a user, a second acquisition means for acquiring intention information indicating the user's intention regarding the weights of the target models or the learned parameters, and a prediction means for deriving a prediction result by applying at least any of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the data for inference.

[0007] An information processing method according to an exemplary aspect of the present disclosure includes a first acquisition process in which at least one processor acquires inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation process in which the feature information is presented to a user; a second acquisition process in which intention information indicating the user's intention regarding the target models or weights of the learned parameters; and a prediction process in which at least some of the learned parameters for each of the plurality of target models and weights determined by the intention information are applied to the inference data to derive a prediction result.

[0008] The information processing device according to each aspect of the present invention may be realized by a computer, in which case a program that causes the computer to operate as each part (software element) of the information processing device to realize the information processing device also falls within the scope of the present invention. [Effects of the Invention]

[0009] According to an exemplary aspect of the present disclosure, an exemplary effect is achieved in that a technique can be provided that can reflect a user's intentions in parameters between multiple models. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a block diagram illustrating a configuration of an information processing device according to the present disclosure. [Figure 2] FIG. 1 is a flow diagram showing the flow of an information processing method according to the present disclosure. [Figure 3] 1 is a block diagram illustrating a configuration of an information processing device according to the present disclosure. [Figure 4] FIG. 10 is a diagram schematically illustrating the outputs of each endpoint model in the LPV model according to the present disclosure and the interior division ratio parameters by which each output is multiplied. [Figure 5] FIG. 10 is a diagram illustrating an example of a processing flow in an information processing device according to the present disclosure. [Figure 6] 10 shows an example of a graph displayed by an output unit via an input / output unit according to the present disclosure. [Figure 7] FIG. 10 is a diagram illustrating an example of a processing flow in an information processing device according to the present disclosure. [Figure 8] 10 shows another example of a graph displayed by the output unit via the input / output unit according to the present disclosure. [Figure 9] 10 shows yet another example of a graph displayed by the output unit via the input / output unit according to the present disclosure. [Figure 10] 1 is a block diagram illustrating a configuration of an information processing device according to the present disclosure. [Figure 11] 1 is a block diagram illustrating a configuration of an information processing device according to the present disclosure. [Figure 12] FIG. 10 is a diagram illustrating an example of a processing flow in an information processing device according to the present disclosure. [Figure 13] FIG. 1 is a block diagram illustrating a configuration of a computer that functions as an information processing device according to the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0011] The following are examples of embodiments of the present invention. However, the present invention is not limited to the exemplary embodiments shown below, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means employed in the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, embodiments obtained by appropriately omitting some of the technical means employed in the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, the effects mentioned in the exemplary embodiments shown below are examples of effects expected in the exemplary embodiments, and do not define the scope of the present invention. In other words, embodiments that do not exhibit the effects mentioned in the exemplary embodiments shown below may also be included in the scope of the present invention.

[0012] First Exemplary Embodiment A first exemplary embodiment, which is one example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is the basic form of each exemplary embodiment described later. Note that the scope of application of each technical means employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technical means employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure to the extent that no particular technical obstacles arise. Furthermore, each technical means shown in the drawings referred to in describing this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure to the extent that no particular technical obstacles arise.

[0013] (Configuration of information processing device 1) The configuration of the information processing device 1 will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the information processing device 1. As shown in Fig. 1, the information processing device 1 includes a first acquisition unit 21, a presentation unit 22, a second acquisition unit 23, and a prediction unit 24. In this exemplary embodiment, the first acquisition unit 21, the presentation unit 22, the second acquisition unit 23, and the prediction unit 24 respectively realize a first acquisition means, a presentation means, a second acquisition means, and a prediction means.

[0014] (First acquisition unit 21) The first acquisition unit 21 acquires inference data, learned parameters for each of a plurality of target models, and feature information associated with the target model or the learned parameters. The first acquisition unit 21 supplies the acquired inference data and learned parameters to the prediction unit 24. The first acquisition unit 21 also supplies the acquired feature information to the presentation unit 22.

[0015] (Presentation part 22) The presenting unit 22 presents the feature information acquired by the first acquiring unit 21 to the user.

[0016] (Second acquisition unit 23) The second acquisition unit 23 acquires intention information indicating the user's intention regarding the target model or the weights of the learned parameters. The second acquisition unit 23 supplies the acquired intention information to the prediction unit 24.

[0017] (Prediction Section 24) The prediction unit 24 derives a prediction result by applying at least one of the learned parameters for each of the multiple target models and a weight determined by the intention information acquired by the second acquisition unit 23 to the inference data acquired by the first acquisition unit 21.

[0018] (Effects of information processing device 1) As described above, the information processing device 1 includes a first acquisition unit 21 that acquires inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters, a presentation unit 22 that presents the feature information acquired by the first acquisition unit 21 to a user, a second acquisition unit 23 that acquires intention information indicating the user's intention regarding the weights of the target models or the learned parameters, and a prediction unit 24 that derives a prediction result by applying at least one of the learned parameters for each of the plurality of target models and the weights determined by the intention information acquired by the second acquisition unit 23 to the inference data acquired by the first acquisition unit 21. Therefore, the information processing device 1 has the effect of being able to reflect the user's intention in the parameters between a plurality of models.

[0019] (Flow of information processing method S1) The flow of the information processing method S1 will be described with reference to Fig. 2. Fig. 2 is a flow diagram showing the flow of the information processing method S1. As shown in Fig. 2, the information processing method S1 includes a first acquisition process S21, a presentation process S22, a second acquisition process S23, and a prediction process S24.

[0020] (First acquisition process S21) In the first acquisition process S21, the first acquisition unit 21 acquires inference data, learned parameters for each of a plurality of target models, and feature information associated with the target model or the learned parameters. The first acquisition unit 21 supplies the acquired inference data and learned parameters to the prediction unit 24. The first acquisition unit 21 also supplies the acquired feature information to the presentation unit 22.

[0021] (Presentation process S22) In the presentation process S22, the presentation unit 22 presents the feature information acquired by the first acquisition unit 21 to the user.

[0022] (Second acquisition process S23) In the second acquisition process S23, the second acquisition unit 23 acquires intention information indicating the user's intention regarding the target model or the weights of the learned parameters. The second acquisition unit 23 supplies the acquired intention information to the prediction unit 24.

[0023] (Prediction process S24) In the prediction process S24, the prediction unit 24 derives a prediction result by applying at least one of the learned parameters for each of the multiple target models and a weight determined by the intention information acquired by the second acquisition unit 23 to the inference data acquired by the first acquisition unit 21.

[0024] (Effect of information processing method S1) As described above, the information processing method S1 includes a first acquisition process S21 in which the first acquisition unit 21 acquires data for inference, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters, a presentation process S22 in which the presentation unit 22 presents the feature information acquired by the first acquisition unit 21 to the user, a second acquisition process S23 in which the second acquisition unit 23 acquires intention information indicating the user's intention regarding the target models or weights of the learned parameters, and a prediction process S24 in which the prediction unit 24 derives a prediction result by applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information acquired by the second acquisition unit 23 to the data for inference acquired by the first acquisition unit 21. Therefore, the information processing method S1 can achieve the same effects as the information processing device 1 described above.

[0025] Second Exemplary Embodiment A second exemplary embodiment, which is one example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same functions as those described in the above exemplary embodiment will be assigned the same reference numerals, and their description will be omitted as appropriate. The scope of application of each technical means employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technical means employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise. Furthermore, each technical means shown in each drawing referenced to explain this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise.

[0026] The positioning of the algorithm of the processing by the information processing device 1A according to this exemplary embodiment will be described. The present inventors have been studying a Linear Parameter-Varying (LPV) model as a modeling of a system with fluctuations. In the LPV model, as an example, an internal state quantity (internal state variable) x k , and the output state quantity (output state variable) y k is updated and calculated using the following formulas (1A) and (1B).

[0027]

number

[0028] where A (i) , B (i) is a matrix representing each state-space model (also called each end point model) that is distinguished from each other by index i, and μ (i) k is a parameter that defines the internal division ratio (weight) of each model. μ (i) k These are called internal ratio parameters, weight parameters, or scheduling parameters. kare input variables, and C and D are x k and u k k is an index assigned to each state variable, and is, for example, time.

[0029] Figure 4 shows the output of each end model in the LPV model (1st SS model to 5th SS model in Figure 4) and the internal division ratio parameter μ (i) k As shown in Fig. 4, the output of multiple endpoint models at the k-th step is (A (i) x k +B (i) u k ) (i=1~5) For each of the internal division ratio parameters μ (i) k (i=1~5) is multiplied, and x at the k+1th step is k+1 is calculated.

[0030] While such an LPV model is suitable for modeling systems with fluctuations, the internal ratio parameter μ (i) k However, there was a problem that it was difficult to apply to systems where the value of is unknown.

[0031] The present inventors The above internal division ratio parameter μ (i) k The hidden variable (posterior probability) z k Treat it as - Applying hidden variable model learning methods in machine learning, The above internal division ratio parameter μ (i) k the hidden variable z k Calculate as the expected value of By doing so, the inventors have found that it is possible to realize the learning of an LPV model even if the interior ratio parameter μ (i)k The L2PV model (Latent Linear Parameter-Varying model) is defined by the following equations (2A) to (2C), which introduce the following as hidden variables:

[0032]

number

[0033] The following equations (3A) to (3E) are used:

[0034]

number

[0035] By rewriting it into the regression model form (L2PV regression model) specified by (i) k This gave rise to the idea of ​​using this as a learning subject.

[0036] Each process performed by the information processing device 1A described below is based on the above-mentioned formulation and is a process based on the unique viewpoint of the inventor.

[0037] (Configuration of information processing device 1A) The configuration of the information processing device 1A will be described with reference to Fig. 3. Fig. 3 is a block diagram showing the configuration of the information processing device 1A. As shown in Fig. 3, the information processing device 1A includes a control unit 10A, a storage unit 15A, a communication unit 16A, and an input / output unit 17A.

[0038] (Storage section 15A) First, various data (information) stored in the storage unit 15A will be described. Data referenced by the control unit 10A is stored in the storage unit 15A. Examples of the storage unit 15A include, but are not limited to, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination thereof.

[0039] Examples of data stored in the memory unit 15A include, but are not limited to, target data TD, ratio parameter RP, regression coefficient RC, distribution information DI, learning result LR (learned parameter LP), inference data PD, type information CI, intention information WI, candidate data group CDG, and prediction result PR, as shown in Fig. 3. The ratio parameter RP and the regression coefficient RC are also referred to as learned parameters LP. In other words, the learned parameters LP include the ratio parameter RP, and the learned parameters LP include the regression coefficient RC.

[0040] The target data TD is data (learning data) used in the learning process (regression coefficient calculation process, ratio parameter calculation process, and covariance calculation process) in the information processing device 1A. k ) and state variables (~y k ) is expressed as the following equation (4). k ,~x k ,y k ,~y k are sometimes called features. Also, the state variable x k ,~x k are called explanatory variables, and the state variable y k ,~y k The objective variable is sometimes called a target variable. Furthermore, when the objective variable is the target of derivation, the objective variable is sometimes called a predicted value. These specific names do not limit the contents described in this specification.

[0041]

number

[0042] The ratio parameter RP is a parameter that defines the relative weights of a plurality of state space models in an LPV model, and is also referred to as a scheduling parameter. The ratio parameter RP may be an internal ratio parameter or an external ratio parameter. In this exemplary embodiment, in order to describe the case where the ratio parameter RP is an internal ratio parameter, the ratio parameter RP is also referred to as an internal ratio parameter RP. The ratio parameter RP is also referred to as a weight parameter RP or a scheduling parameter RP. As an example, the ratio parameter RP is given by the following equation (5) corresponding to each of m models (model 1 to model m):

[0043]

number

[0044] Here, k is an index similar to the index assigned to each state variable described above, and N represents the dimension of each state variable (the number of samples of each state variable). Also, the index (i) relating to the model is not explicitly stated in the above expression. This is because the ratio parameter RP is divided into an internal ratio parameter vector consisting of components corresponding to models 1 to m for each k. μ k =(μ k (1) , μ k (2) , , μ k (m) ) In this way, the ratio parameter RP may be expressed as an internal ratio parameter vector or a ratio parameter matrix.

[0045] Also, the internal ratio parameter μ for a certain model j k (j) is the N-dimensional target data x k It can also be expressed as the components of an N-dimensional vector having components corresponding to each of the j-th internal division ratio parameters μ k (j)is the N-dimensional target data x k (k=1 to N) (j) , μ2 (j) , , μ N (j) ) are the components of an N-dimensional vector with

[0046] The regression coefficient RC is a coefficient in the L2PV regression model. The regression coefficient RC is expressed as the following equation (6).

[0047]

number

[0048] The distribution information DI includes a covariance matrix Φ of the prior distribution of the latent variables, a covariance parameter η of the prior distribution of the latent variables, and a covariance parameter Ψ of the posterior distribution of the latent variables.

[0049] Hidden variable z k The prior distribution p(z k ) is expressed as the following equation (7).

[0050]

number

[0051] Hidden variable z k In other words, the covariance parameter η of the prior distribution of k |z k ,~x k ,W,η) is the covariance parameter η.

[0052]

number

[0053] Hidden variable z k The posterior distribution p(z k |~y k ,~x k ,W,η) is expressed as the following equation (9).

[0054]

number

[0055] In the above formula, the calligraphy font N on the right side represents a normal distribution. However, this does not mean that the example of the distribution in this exemplary embodiment is limited to a normal distribution. As an example, the hidden variable z k The Dirichlet distribution may be used as the posterior distribution of .

[0056] As will be described later, in the processing by the information processing device 1A, the hidden variable z k The posterior distribution p(z k |~y k ,~x k , W, η) is expressed under the constraints (constraints) of the following equation (10).

[0057]

number

[0058] Therefore, the hidden variable z k Even when a normal distribution is used as the posterior distribution of , a suitable calculation can be performed.

[0059] The learning result LR is data output by the output unit 15, which will be described later. The learning result LR includes the calculated ratio parameter RP and the calculated regression coefficient RC. The learning result LR is also referred to as the learned parameter LP.

[0060] The inference data PD is an internal state quantity input to the L2PV regression model. The inference data PD includes, for example, one or more feature quantities. The feature quantities included in the inference data PD may be referred to as internal state quantities or explanatory variables, but these names do not limit this exemplary embodiment. The L2PV regression model makes predictions by using the inference data PD as input and applying the calculated regression coefficients RC to each of multiple target models.

[0061] The type information CI is information associated with a target model or a trained parameter LP for the target model, and indicates what type of target data TD the target model or trained parameter LP was trained with. In other words, the type information CI is feature information indicating the features of the target model or trained parameter LP.

[0062] An example of type information CI is: What kind of target data TD was used to train (or will be trained) each of the multiple target models? or -What kind of target data TD was each of the multiple learned parameters LP learned (or will be learned) from? In the following, the type information CI is What kind of target data TD was used to train (or will be trained) each of the multiple target models? This will be explained using information showing the following as an example. -What kind of target data TD was each of the multiple learned parameters LP learned (or will be learned) from? The same is true for .

[0063] For example, the type information CI associated with each of the target models a, b, and c is as follows: Target model a: Situation A (e.g., sunny weather), Situation B (e.g., cloudy weather) Target model b: Situation C (e.g., rainy weather) Target model c: Situations A, B, and C Here, the first line in the above example indicates that the target model a has been trained (will be trained) using the target data TD acquired in situation A (e.g., sunny weather) and the target data TD acquired in situation B (e.g., cloudy weather). The same applies to the other lines.

[0064] By referring to the type information CI, the information processing device 1A can identify what situation each target model is associated with, in other words, what situation each target model is preferably used in. Furthermore, the information processing device 1A can also present to the user what situation each target model is preferably used in.

[0065] For example, the type information CI associated with each of the target models a, b, and c includes the following: Target model a: Target data TD1, TD2 Target model b: Target data TD3 Target model c: Target data TD1, TD2, TD3 Here, the first line in the above example indicates that the target model a has been trained (will be trained) using the target data TD1 and TD2. The same applies to the other lines.

[0066] In other words, the type information CI includes: Information indicating whether each of the multiple target models has used (will use) each of the multiple target data TD as training data For example, the type information CI associated with each of the target models a, b, and c includes the following: Target model a: Target data TD1 (○), TD2 (○), TD3 (×) Target model b: Target data TD1(×), TD2(×), TD3(○) Target model c: Target data TD1(○), TD2(○), TD3(○) Here, ○ indicates that it has been used (will be used) for learning, and × indicates that it has not been used (will not be used) for learning.

[0067] In this case, the type information CI may be configured to include information about the circumstances under which each piece of target data TD was acquired. Target data TD1 (e.g., sales volume): Situation A (e.g., sunny weather) Target data TD2 (e.g., sales volume): Situation B (e.g., cloudy weather) Target data TD3 (e.g., sales volume): Situation C (e.g., the weather is rainy) It may contain information such as:

[0068] In addition, when the type information CI includes a correspondence relationship between each piece of target data TD and the situation in which the target data TD was acquired, the type information CI may be expressed as including a relationship between each target model and information related to the target model. In other words, when the type information CI includes a correspondence relationship between each piece of target data and the situation in which the target data was acquired, the type information CI can also be expressed as correspondence relationship information CI regarding the correspondence relationship between the target data TD and a plurality of target models.

[0069] The intention information WI is information indicating the intention of the user regarding the weights of the target model or the learned parameters LP. The format of the intention information WI is not particularly limited as long as it is information indicating the intention of the user.

[0070] As an example, the intention information WI may be information indicating which model among multiple target models is to be given more importance than other models (or which model is to be given less importance than other models). As another example, the intention information WI may be a numerical value related to the weight of the target model or the learned parameter LP. An example of a process using the intention information WI will be described later.

[0071] The candidate data group CDG includes one or more candidate data CD. Each candidate data CD may be, for example, Information about the type of candidate data CD One or more features Here, the information on the type of the candidate data CD may be, for example, information indicating the circumstances under which each candidate data CD was acquired. For example, when the one or more feature quantities include "day" as an explanatory variable, the information on the type may be, Candidate data CD1 (e.g., daily sales volume): Situation A (e.g., sunny weather) Candidate data CD2 (e.g., daily sales volume): Situation B (e.g., cloudy weather) Candidate data CD3 (e.g., daily sales volume): Situation C (e.g., the weather is rainy) The information may be associated with each candidate data CD as follows.

[0072] In addition, information about the above types is available at Information indicating whether a certain candidate data is similar to other candidate data or data for inference Here, "a certain piece of data is similar to another piece of data" refers to, for example, a case where the type of the certain piece of data is similar to the type of the other piece of data. For example, if candidate data 1 is data on the number of sales acquired in situation A (e.g., sunny weather) as in the above example, candidate data or data for inference that is data on the number of sales in situation A' (e.g., sunny weather) similar to situation A is data similar to candidate data 1.

[0073] The predicted result PR is a predicted result obtained by the L2PV regression model. Examples of the predicted result PR will be described later.

[0074] (Communication unit 16A) The communication unit 16A is an interface for transmitting and receiving data via a network. Examples of the communication unit 16A include, but are not limited to, communication chips for various communication standards such as Ethernet (registered trademark), Wi-Fi (Wireless Fidelity (registered trademark), and wireless communication standards for mobile data communication networks, and USB-compliant connectors.

[0075] (Input / output section 17A) The input / output unit 17A is an interface that receives input of data and outputs data. Examples of the input / output unit 17A include, but are not limited to, a microphone, a camera, a gaze input device, a keyboard, a touchpad, a speaker, and a liquid crystal display.

[0076] (Control unit 10A) The control unit 10A controls each component included in the information processing device 1A. As shown in FIG. 3, the control unit 10A includes an acquisition unit 11 (21, 23), a regression coefficient calculation unit 12, a covariance calculation unit 13, a ratio parameter calculation unit 14, an output unit 15 (22), an initial value determination unit 16, a convergence determination unit 17, and a prediction unit 24. In this exemplary embodiment, the acquisition unit 11 (21, 23) serves as a first acquisition unit and a second acquisition unit. In this exemplary embodiment, the output unit 15 (22) and the prediction unit 24 serve as a presentation unit and a prediction unit, respectively. Specific examples of the processing performed by each unit will be described later with reference to different drawings.

[0077] The acquisition unit 11 (21, 23) acquires data via the communication unit 16A or the input / output unit 17A. Examples of data acquired by the acquisition unit 11 (21, 23) include inference data PD, learned parameters LP (ratio parameters RP and regression coefficients RC), and type information CI. Another example of data acquired by the acquisition unit 11 (21, 23) includes intention information WI. Yet another example of data acquired by the acquisition unit 11 (21, 23) is hidden variables z k The acquiring unit 11 (21, 23) stores the acquired data in the storage unit 15A.

[0078] The regression coefficient calculation unit 12 calculates a regression coefficient RC for each of the plurality of target models by referring to a ratio parameter RP that defines the ratio of the plurality of target models and the target data TD (regression coefficient calculation process). As one example, the ratio parameter RP referred to by the regression coefficient calculation unit 12 is the initial value of the ratio parameter RP determined by an initial value determination unit 16, which will be described later. As another example, the ratio parameter RP referred to by the regression coefficient calculation unit 12 is the ratio parameter RP calculated by a ratio parameter calculation unit 14, which will be described later. The regression coefficient calculation unit 12 stores the calculated regression coefficient RC in the storage unit 15A.

[0079] The covariance calculation unit 13 calculates the target data TD, the ratio parameter RP, the regression coefficient RC, and the latent variable z k By referring to the covariance matrix Φ of the prior distribution of the hidden variable z k The covariance parameter η of the prior distribution of and the hidden variable z k The covariance calculation unit 13 calculates the covariance matrix Ψ of the posterior distribution of the calculated hidden variable z k The covariance parameter η of the prior distribution of and the hidden variable z k and the covariance matrix Ψ of the posterior distribution of the above are stored in the storage unit 15A as distribution information DI.

[0080] The ratio parameter calculation unit 14 calculates the ratio parameter RP by referring to the target data TD and the regression coefficient RC (ratio parameter calculation process). k The ratio parameter calculation unit 14 further refers to the covariance matrix Ψ of the posterior distribution of the ratio parameter RP to calculate the ratio parameter RP. The ratio parameter calculation unit 14 stores the calculated ratio parameter RP in the storage unit 15A.

[0081] Furthermore, the ratio parameter calculation unit 14 refers to the correspondence information CI and calculates the ratio parameter RP under a constraint condition according to the correspondence. As an example, the ratio parameter calculation unit 14 calculates the ratio parameter RP under a constraint condition on the sum of the ratio parameters RP according to the correspondence.

[0082] Furthermore, the ratio parameter calculation unit 14 calculates the ratio parameter RP using an update equation that includes a normalization term according to the correspondence relationship.

[0083] The output unit 15 (22) presents the type information CI to the user. As an example, the output unit 15 (22) outputs an image including the type information CI to the input / output unit 17A.

[0084] The output unit 15 (22) stores the regression coefficients RC and the ratio parameters RP in association with the correspondence. As an example, the output unit 15 associates the regression coefficients RC and the ratio parameters RP with correspondence information CI and supplies them to the storage unit 15A.

[0085] Furthermore, the output unit 15 (22) outputs the regression coefficient RC and the ratio parameter RP (learning result LR) calculated by the ratio parameter calculation unit 14. As one example, the output unit 15 outputs an image including the regression coefficient RC and the ratio parameter RP calculated by the ratio parameter calculation unit 14 to the input / output unit 17A. In this configuration, the output unit 15 may display graphs defined by the regression coefficients RC for at least two of the multiple target models in a manner that allows them to be distinguished from one another. As another example, the output unit 15 outputs the regression coefficient RC and the ratio parameter RP calculated by the ratio parameter calculation unit 14 when the convergence determination unit 17 described later determines that the calculation related to the ratio parameter RP has converged.

[0086] The initial value determination unit 16 determines the initial value of the ratio parameter RP to be referred to by the regression coefficient calculation unit 12. The initial value determination unit 16 stores the determined initial value of the ratio parameter RP in the storage unit 15A.

[0087] The convergence determination unit 17 determines whether or not the calculation regarding the ratio parameter RP has converged. The convergence determination unit 17 supplies the determination result to the output unit 15.

[0088] The prediction unit 24 derives a prediction result PR by applying, to the inference data PD, the regression coefficients RC for each of the plurality of target models and the ratio parameters RP calculated by the ratio parameter calculation unit 14, which are determined according to the type of the inference data PD. The prediction unit 24 stores the derived prediction result PR in the storage unit 15A.

[0089] In addition, the prediction unit 24 derives a prediction result by applying at least one of the learned parameters LP (ratio parameter RP and regression coefficient RC) for each of the multiple target models and a weight determined by the intention information WI to the inference data PD.

[0090] (Example 1 of processing flow in information processing device 1A) 5 is a diagram showing an example of a processing flow in the information processing device 1A according to this exemplary embodiment. Note that the processing example described below can also be regarded as a variational Bayes EM algorithm, but this does not limit this exemplary embodiment. Furthermore, the processing example described below can be regarded as processing for updating each parameter so as to maximize a variational lower bound (VLB) J obtained by the following equation (11).

[0091]

number

[0092] The processing example described below can also be expressed as an algorithm for solving a maximum likelihood problem defined by the model likelihood p in the following equation (12).

[0093]

number

[0094] (Step S11: Acquisition process) In step S11, the acquisition unit 11 (21, 23) acquires target data TD. Here, as described above, the target data TD is data used for the learning process in the information processing device 1A. Details of the target data TD have been explained, so explanation will be omitted here.

[0095] In step S11, the acquisition unit 11 (21, 23) further acquires a parameter m indicating the number of models among the plurality of target models. Here, the number of models m is determined by the internal division ratio parameter vector μ k (i) It can also be expressed as the number of

[0096] Also, in step S11, the acquisition unit 11 (21, 23) acquires correspondence relationship information CI relating to correspondence relationships between the target data TD and a plurality of target models.

[0097] In step S11, the acquisition unit 11 (21, 23) acquires the hidden variable z k As an example, the acquisition unit 11 acquires information about the prior distribution of the hidden variable z k The prior distribution p(z k ) covariance matrix Φ. k The prior distribution p(z k ) may further be obtained.

[0098] (Step S16: Initial value determination process) Next, in step S16, the initial value determination unit 16 determines the initial value of a ratio parameter RP to be referenced in the regression coefficient calculation process S12, which will be described later. As an example, the initial value determination unit 16 determines the initial value of the ratio parameter RP as a random value. By the initial value determination unit 16 determining the initial value of the ratio parameter RP in this way, the regression coefficient RC can be suitably calculated in the regression coefficient calculation process S12, which will be described later. The details of the ratio parameter RP have been explained above, so a detailed explanation will be omitted here.

[0099] (Step S12: Regression coefficient calculation process) Subsequently, in step S12, the regression coefficient calculation unit 12 calculates the ratio parameter (internal division ratio parameter vector) RP expressed as the following equation (13):

[0100]

number

[0101] and the target data TD expressed as the following equation (14):

[0102]

number

[0103] and a regression coefficient RC for each of the plurality of target models, which is expressed as the following equation (15) with reference to

[0104]

number

[0105] As an example, the regression coefficient calculation unit 12 calculates a regression coefficient RC expressed as the following equation (17) by referring to the ratio parameter RP and the target data TD using the following equation (16).

[0106]

number

[0107]

number

[0108] Here, the asterisk on the shoulder of W indicates the updated value, and in the calculation formula, the operation symbol with a cross in a circle represents the Kronecker product. T represents transposition. Also, Ψ k is the hidden variable z k represents the covariance parameter of the posterior distribution of

[0109] The regression coefficient calculation unit 12 may store the calculated regression coefficient RC (learned parameter LP) and the target data TD (or type information CI) in association with each other in the storage unit 15A.

[0110] (Step S13: Covariance calculation process) Subsequently, in step S13, the covariance calculation unit 13 calculates the target data TD expressed as the following equation (18):

[0111]

number

[0112] and the ratio parameter (internal division ratio parameter vector) RP expressed as the following equation (19):

[0113]

number

[0114] and the regression coefficient RC expressed as the following equation (20):

[0115]

number

[0116] and the hidden variable z k The hidden variables z k The covariance parameter η of the prior distribution of the hidden variable z k The covariance matrix of the posterior distribution of {Ψ k} k=1 N As an example, the covariance calculation unit 13 calculates the following equation (21):

[0117]

number

[0118] (where Λ k is given by the following equation (22)

[0119]

number

[0120] By the hidden variable z k The covariance matrix of the posterior distribution of {Ψ k} k=1 N Furthermore, the covariance calculation unit 13 calculates the following equation (23):

[0121]

number

[0122] By the hidden variable z k Calculate the covariance parameter η of the prior distribution of , where N is the number of samples for each state variable as described above, and r is the covariance parameter η of the prior distribution of . k , and as an example, r=1.

[0123] (Step S14: Ratio parameter calculation process) Subsequently, in step S14, the ratio parameter calculation unit 14 calculates the target data TD expressed as the following equation (24):

[0124]

number

[0125] and the regression coefficient RC expressed as the following equation (25):

[0126]

number

[0127] and the hidden variable z k The covariance matrix of the posterior distribution of {Ψ k}k=1 N With reference to the above, the ratio parameter (internal division ratio parameter vector) RP expressed as the following equation (26) is

[0128]

number

[0129] As an example, the ratio parameter calculation unit 14 calculates (updates) the following equation (27):

[0130]

number

[0131] By μ k The ratio parameter calculation unit 14 calculates (updates) the ratio parameter RP by executing the process of calculating the following equation (28) under constraint conditions according to the correspondence indicated by the correspondence information CI.

[0132]

number

[0133] By μ k By performing the process of calculating the ratio parameter (internal division ratio parameter vector) RP expressed by the following equation (30), under the constraints (constraints) expressed by the following equations (29A) to (29C),

[0134]

number

[0135]

number

[0136] Here, the first equation of the constraint condition (29A) can be expressed by explicitly specifying the index (i) related to the model as follows: Σ i=1 m μ k (i) = 1 In other words, the first equation of the constraint condition equation (29A) indicates that the sum of the ratio parameter RP across the indexes related to the model is 1. Furthermore, the second equation of the constraint condition equation (29A) indicates that the value of the ratio parameter RP is equal to or greater than 0. In this way, by calculating the ratio parameter RP under the constraint condition equation (29A), the ratio parameter calculation unit 14 can suitably calculate the ratio parameter RP even when, for example, a normal distribution is adopted as the posterior distribution of the latent variable.

[0137] In addition, in the above constraint equation (29B), M i denotes the set of target data TD corresponding to the i-th endpoint model. In other words, M i indicates a set of target data TD that the correspondence relationship information CI indicates will be used as training data for the i-th end point model. Therefore, the above constraint condition equation (29B) indicates that the ratio parameter μ calculated using the j-th target data TD from the set of target data TD that indicates will be used as training data for the i-th end point model. j The sum of these ratio parameters RP is 1. That is, the ratio parameter calculation unit 14 calculates the ratio parameters under the constraint condition according to the correspondence relationship, which is a constraint condition on the sum of the ratio parameters RP, by using the above constraint condition formula (29B). On the other hand, the above constraint condition formula (29C) calculates the ratio parameters μ in the end point model that does not use the j-th target data TD as training data. h indicates that ∇ ...

[0138] For example, consider a model that predicts sales of drinking water in a store. As an example, the endpoint models are (1) Model 1, which predicts sales when a measure is implemented, (2) Model 2, which predicts sales when the content volume is changed, (3) Model 3, which predicts sales when the appearance is changed, (4) Model 4, which predicts sales when the external environment changes, and Model 5, which predicts sales under all of the situations (1) to (5).

[0139] As an example of this case, if the target data TD is data on sales when a measure is implemented, the correspondence relationship information CI indicates that the target data TD is used as learning data for Model 1 and Model 5.

[0140] As described above, the ratio parameter calculation unit 14 may calculate the ratio parameter RP using an update equation including a normalization term according to the correspondence relationship. For example, the ratio parameter calculation unit 14 calculates the ratio parameter RP using the following equations (31A) and (31B).

[0141]

number

[0142] Of the equations (31A), the following equation (32) is a normalization term according to the correspondence relationship.

[0143]

number

[0144] In equation (31A), γ k is a coefficient indicating the strength of regularization. That is, equation (32) is the ratio parameter μ calculated using the k-th target data TD. k is the ratio parameter μ calculated using target data TD similar to the k-th target data TD k’This indicates that the value of the ratio parameter RP is close to the value of the target data TD. With this configuration, when two pieces of target data TD are similar to each other and the ratio parameter RP is calculated using each of the two pieces of target data TD, the ratio parameter calculation unit 14 calculates a ratio parameter RP that is close in value. Therefore, the ratio parameter calculation unit 14 can calculate a ratio parameter RP with improved interpretability.

[0145] The ratio parameter calculation unit 14 may associate the calculated ratio parameter RP (learned parameter LP) with the target data TD (or type information CI) and store them in the storage unit 15 A. Furthermore, the ratio parameter calculation unit 14 may associate the calculated ratio parameter RP with a regression coefficient RC, and then store them in the storage unit 15 A in association with the target data TD (or type information CI).

[0146] (Step S17: Convergence determination process) Subsequently, in step S17, the convergence determination unit 17 determines whether or not the series of processes in the above-mentioned steps S12, S13, and S14 have converged. This may be expressed as determining whether or not the above-mentioned variational Bayes EM algorithm has converged, or as determining whether or not the calculation related to the ratio parameter RP in step S14 has converged. As an example, the convergence determination unit 17 determines whether or not the variational lower bound (VLB) J obtained by the following formula (33),

[0147]

number

[0148] and determines that the series of processes in steps S12, S13, and S14 described above have converged if the change in the variation lower limit is equal to or less than a predetermined threshold. For example, in the nth iteration of the series of processes including steps S12, S13, and S14 described above, the convergence determination unit 17 compares the (n-1)th variation lower limit with the nth variation lower limit in the convergence determination process, and determines that the series of processes in steps S12, S13, and S14 described above have converged if the absolute value of the difference between them is equal to or less than a predetermined threshold.

[0149] (Step S15: Output process) If the convergence determination unit 17 determines in step S17 that the calculation for the ratio parameter RP has "converged," then in step S15 the output unit 15 (22) outputs the regression coefficient RC calculated by the regression coefficient calculation unit 12 in step S12 and the ratio parameter RP (learning result LR) calculated by the ratio parameter calculation unit 14 in step S14. As an example, the output unit 15 (22) associates the regression coefficient RC and the ratio parameter RP with the type of target data TD and stores them in the storage unit 25A. The stored target data TD is then referenced as candidate data CD in the similarity calculation process described below. In this way, if the convergence determination unit 17 determines that the calculation for the ratio parameter RP has "converged," the output unit 15 (22) outputs the learning result LR, thereby making it possible to output a suitable learning result LR.

[0150] In step S15, the output unit 15 (22) outputs the hidden variable z calculated by the covariance calculation unit 13 in step S13. k The covariance parameter η of the prior distribution of the hidden variable z k With this configuration, the output unit 15 (22) may further output the covariance matrix Ψ of the posterior distribution of the hidden variable z k The covariance parameter η of the prior distribution of and the hidden variable z k The covariance matrix Ψ of the posterior distribution of σ can be presented to the user.

[0151] In step S15, the output unit 15 (22) may be configured to display graphs defined by the regression coefficients RC for at least two of the plurality of target models in a manner that allows them to be distinguished from one another.

[0152] 6 shows an example of a graph that the output unit 15 (22) displays via the input / output unit 17A in this step. In the example shown in FIG. 6, the output unit 15 (22) displays the regression coefficients RC calculated for each of the multiple target models, which are expressed as the following equation (34), in the regression coefficient calculation process in step S12.

[0153]

number

[0154] Among them, the regression coefficient W of Model 1 (1) The graph L1 defined by the regression coefficient W of Model 2 (2) and graph L2 defined by the above are displayed so as to be distinguishable from each other. In this way, according to the information processing device 1A according to this exemplary embodiment, a plurality of models are used and the ratio parameter RP of each model can be determined by learning, so that an output result having a range (as an example, an output result having a range defined by the above graph L1 and graph L2) can be generated.

[0155] 6, the output unit 15 (22) may refer to the correspondence relationship information CI and display what models the graph L1 and the graph L2 correspond to. In FIG. 6, the output unit 15 displays that the graph L1 is a model corresponding to the situation A, and the graph L2 is a model corresponding to the situation B.

[0156] (Example of processing executed by the prediction unit 24) The prediction unit 24 derives multiple prediction results PR by applying, to the inference data PD, the regression coefficient RC for each of the multiple target models and the ratio parameter RP determined according to the type of the inference data PD, among the ratio parameters RP calculated by the ratio parameter calculation unit 14. As an example, when the inference data PD includes a feature amount (explanatory variable) x0, the prediction unit 24 derives the prediction results PR by applying, to the feature amount x0, the regression coefficient RC for each of the multiple target models and the ratio parameter RP determined according to the type of the inference data PD.

[0157] In addition, the ratio parameter determined according to the type of inference data PD refers, as an example, to a ratio parameter learned (updated) by the above-mentioned learning process using learning data (target data) associated with a type that is the same as or similar to the type of the inference data PD, but this does not limit this exemplary embodiment.

[0158] For example, assume that each of the multiple target data TD is sales of drinking water at multiple stores. Also, a situation in which a campaign for the drinking water is being conducted is defined as situation A, and a situation in which the volume of the drinking water has been increased is defined as situation B. Examples of situations include, but are not limited to, a situation in which the appearance of the drinking water container has been changed, a situation in which a new store has opened near the store, a situation in which it is raining, and a situation in which all of these situations are included.

[0159] For example, when predicting the sales of a drink at a certain store at a certain date and time x0, where a campaign for the drink is being carried out and the volume of the drink has been increased, the prediction unit 24 calculates the regression coefficient W of model 1 corresponding to situation A at the certain date and time x0. (1) The predicted result P1 is derived by applying the regression coefficient W of Model 2 corresponding to situation B at a certain date and time x0. (2) The prediction unit 24 derives a prediction result P2 by applying the ratio parameter μ (1) and the ratio parameter μ of Model 2 (2) The prediction result RP is derived by applying

[0160] In this way, the information processing device 1A according to this exemplary embodiment can generate a prediction result PR according to the type of inference data PD. Therefore, the information processing device 1A according to this exemplary embodiment can present a prediction result PR that has favorable explainability to the user.

[0161] Furthermore, the output unit 15 (22) may display graphs defined by the regression coefficients R C of each of the multiple models. The output unit 15 (22) may display a graph L1 defined by the regression coefficients R C of model 1 corresponding to situation A, a graph L2 defined by the regression coefficients R C of model 2 corresponding to situation B, and a graph L3 defined by the regression coefficients R C of model 3 corresponding to situation C in a manner that allows them to be distinguished from one another. Situation C may include at least one of situations A and B, or may be a situation different from situations A and B. According to the information processing device 1A according to this exemplary embodiment, even with the above-mentioned configuration, it is possible to present to the user a prediction result PR that has favorable explainability.

[0162] (Example 2 of processing flow in information processing device 1A) 7 is a diagram showing another example of the flow of processing in the information processing device 1A according to this exemplary embodiment. Note that, in the processing example described below, a case where similar data is used is described, but this does not limit this exemplary embodiment.

[0163] (Step S21) In step S21, the acquisition unit 11 (21, 23) acquires inference data PD and a type PDC of the inference data PD. Here, the example shown in Fig. 7 illustrates a data structure in which a type PDC is included as part of the inference data PD, but this does not limit the present exemplary embodiment. Furthermore, the inference data PD acquired in this step includes one or more feature PDFs.

[0164] In step S21, the acquisition unit 11 (21, 23) acquires learned parameters LP including the ratio parameter RP and the regression coefficient RC. Furthermore, the acquisition unit 11 (21, 23) acquires type information CI associated with the learned parameters LP.

[0165] In the following, a case will be described in which the acquisition unit 11 (21, 23) acquires the learned parameter LP1 and the type information CI1 associated with the learned parameter LP1, and the learned parameter LP2 and the type information CI2 associated with the learned parameter LP2 in step S21. Also, a case will be described in which the type information CI1 indicates that the model corresponds to situation A, and the type information CI2 indicates that the model corresponds to situation B.

[0166] (Step S31_1) In step S31_1, the acquisition unit 11 (21, 23) executes a similarity calculation process with reference to the inference data PD and the candidate data group CDG. In the example shown in FIG. 7, in this step, the following data are used as the candidate data group CDG: Candidate data CD1 whose type is C1 and contains one or more features F1 Candidate data CD2 whose type is C2 and includes one or more feature values ​​F2 is referenced.

[0167] In the similarity calculation process, the acquisition unit 11 (21, 23) calculates the similarity between the inference data PD and each candidate data CD included in the candidate data group CDG as follows: The difference between the index indicating the type of the inference data PD and the index indicating the type of each candidate data CD included in the candidate data group CDG As an example, if the type is related to weather, the indicator indicating the type is calculated as follows: Sunny: 5, Sunny: 4, Cloudy: 3, Light rain: 2, Heavy rain: 1 The type of the inference data PD is "clear (index: 5)" and the types of the candidate data CD1 and CD2 are as follows: Candidate data CD1 type C1: Sunny (index: 4) Candidate data CD1 type C2: Heavy rain (index: 1) If so, the acquisition unit 11 (21, 23) The similarity between the inference data PD and the candidate data CD1 is calculated as 5-4=1. The similarity between the inference data PD and the candidate data CD2 is calculated as 5-1=4.

[0168] (Step S31_2) In step S31_2, the acquisition unit 11 (21, 23) executes a process of comparing the similarity calculated in step S31_1 with a similarity threshold. Here, the similarity threshold may be stored in the storage unit 15A, for example, but this does not limit the present exemplary embodiment. The acquisition unit 11 (21, 23) refers to the result of the similarity determination process and selects (identifies, determines) similar data that is similar to the inference data PD acquired in step S21 from the multiple candidate data CD included in the candidate data group CDG.

[0169] For example, when the threshold value of the similarity is 2 and the inference data PD, the candidate data CD1, and the candidate data CD2 are as described above, the acquisition unit 11 (21, 23) Determine that the similarity between the inference data PD and the candidate data CD1 (1) is equal to or less than the threshold value of similarity (2); Based on the above determination, the candidate data CD1 is determined to be similar to the inference data PD.

[0170] Similarly, the acquisition unit 11 (21, 23) The similarity between the inference data PD and the candidate data CD2, which is 4, is determined to be greater than the similarity threshold value, which is 2; Based on the above determination, it is determined that the candidate data CD2 is not similar to the inference data PD.

[0171] Then, based on the above determination results, the acquisition unit 11 (21, 23) identifies the candidate data CD1 as similar data that is similar to the inference data PD.

[0172] Moreover, the acquisition unit 11 (21, 23) searches and acquires the ratio parameter RP and the regression coefficient RC associated with the similar data from, for example, the learning result LR stored in the storage unit 15A.

[0173] (Step S22_1) In step S22_1, the output unit 15 (22) presents to the user the type information CI acquired by the acquisition unit 11 (21, 23) in step S21. An example of an image presented by the output unit 15 (22) in step S22_1 will be described with reference to Fig. 8. Fig. 8 shows another example of a graph displayed by the output unit 15 (22) via the input / output unit 17A in this step.

[0174] As shown in FIG. 8, the output unit 15 (22) presents a graph L1 defined by the regression coefficient RC1 of the model corresponding to the situation A indicated by the type information CI1, and a graph L2 defined by the regression coefficient RC2 of the model corresponding to the situation B indicated by the type information CI2.

[0175] 8, the output unit 15 (22) presents a prediction result P1, which is the result obtained by applying the regression coefficient RC1 of the model corresponding to situation A to the feature x0, which is the value of the feature PDF of the inference data PD, and a prediction result P2, which is the result obtained by applying the regression coefficient RC2 of the model corresponding to situation B. With this configuration, the information processing device 1A can present the prediction results of the model for the inference data PD to the user.

[0176] 8, the output unit 15 (22) presents a prediction result PP, which is a result obtained by applying the regression coefficient RC and the ratio parameter RP to the feature quantity x0, which is the value of the feature quantity PDF of the inference data PD. Here, the output unit 15 (22) presents, as the prediction result PP, a prediction result PP obtained by applying the ratio parameter RP of the candidate data CD1, which is similar data to the above-mentioned inference data PD. The prediction result PP in FIG. 8 is obtained by applying γ (1) and γ, the value of the ratio parameter RP of the model corresponding to the graph L2 in the candidate data CD1. (2) This is the result obtained by applying the above. With this configuration, the information processing device 1A can present to the user the prediction result obtained by applying the ratio parameter RP to the inference data PD.

[0177] Furthermore, the output unit 15 (22) may display an interface for acquiring the intention information WI. As an example, the output unit 15 (22) displays an interface "Use ratio of L1: XX" that accepts an operation to change the weight of the model or learned parameter LP1 corresponding to situation A. Similarly, the output unit 15 (22) displays an interface "Use ratio of L2: YY" that accepts an operation to change the weight of the model or learned parameter LP2 corresponding to situation B.

[0178] (Step S23) In step S23, the acquisition unit 11 (21, 23) acquires intention information WI. As an example, when the output unit 15 (22) presents the image shown in FIG. 8 to the user, the acquisition unit 11 (21, 23) accepts an operation to select the interface "Use ratio of L1: XX" and further accepts an input of a numerical value. The acquisition unit 11 (21, 23) acquires the accepted numerical value as intention information WI indicating the user's intention regarding the model corresponding to the graph L1 or the weight of the learned parameter LP of the model. The same applies when the acquisition unit 11 (21, 23) accepts an operation to select the interface "Use ratio of L2: YY".

[0179] Another example will be described with reference to Fig. 9. Fig. 9 shows yet another example of a graph that the output unit 15 (22) displays via the input / output unit 17A in step S22_1.

[0180] The output unit 15 (22) may display a slider for acquiring the intention information WI, as shown in FIG. 9. In this case, the acquisition unit 11 (21, 23) accepts an operation on the slider "proportion of use of L1" or the slider "proportion of use of L2." As an example, when an operation to slide the slider "proportion of use of L1" toward "high" or "low" is accepted, a numerical value corresponding to the amount of sliding is acquired as the intention information WI indicating the user's intention regarding the model corresponding to the graph L1 or the weight of the learned parameter LP of the model. The same applies when an operation on the slider "proportion of use of L2" is accepted.

[0181] The acquisition unit 11 (21, 23) may receive the intention information WI every time the inference data PD changes, or may use the same intention information WI as last time even if the inference data PD changes. The acquisition unit 11 (21, 23) may be configured to receive the intention information WI every predetermined period (for example, every day or every week). The acquisition unit 11 (21, 23) may be configured so that the user sets when the acquisition unit 11 (21, 23) will acquire the intention information WI (every time the inference data changes, or every predetermined period).

[0182] (Step S24) In step S24, the prediction unit 24 derives a prediction result by applying at least one of the learned parameters LP for each of the plurality of target models and a weight determined by the intention information WI to the inference data PD.

[0183] As an example, the prediction unit 24 derives the prediction result PR by applying at least one of the learned parameters LP for each of the multiple target models and a weight determined by the ratio parameter RP and the intention information WI to the inference data PD.

[0184] As an example of this configuration, when the intention information WI acquired by the acquisition unit 11 (21, 23) indicates a numerical value, the prediction unit 24 changes the value of the ratio parameter RP to the numerical value indicated by the intention information WI.

[0185] For example, in the image shown in Fig. 8, it is assumed that the acquisition unit 11 (21, 23) receives an operation to select the interface "Use ratio of L1: XX" and further receives an input of a numerical value "0.4". In this case, the prediction unit 24 calculates γ (1) Change the value to "0.4".

[0186] Similarly, in the image shown in Fig. 8, it is assumed that the acquisition unit 11 (21, 23) receives an operation to select the interface "Use ratio of L2: YY" and further receives an input of a numerical value "0.7". In this case, the prediction unit 24 calculates γ (2) Change the value to "0.7".

[0187] Then, the prediction unit 24 applies the value of the ratio parameter RP of the model corresponding to graph L1, which is "0.4," and the value of the ratio parameter RP of the model corresponding to graph L2, which is "0.7," to the prediction results P1 and P2, to derive the prediction result PP.

[0188] Here, the numerical range of the ratio parameter RP acquired by the acquisition unit 11 (21, 23) is not particularly limited. In the above example, the ratio parameter "0.4" of the model corresponding to graph L1 and the ratio parameter "0.7" of the model corresponding to graph L2 are summed to "1.1", which is an unnormalized ratio parameter (weight parameter), and such a case is also included in this exemplary embodiment. In this way, if the sum of the ratio parameters does not become 1, the prediction unit 24, when displaying the prediction result, Calculate the normalized weight parameters (0.63 and 0.37 in the above example) by dividing each ratio parameter (0.7 and 0.4 in the above example) by the sum of the ratio parameters, The normalized weight parameters may be used as an internal division ratio to derive the prediction result PP.

[0189] Furthermore, if the numerical value of the ratio parameter is negative, the prediction unit 24 may be configured to apply the numerical value as an external ratio to derive the prediction result PP. That is, the prediction result PP may be above graph L1 or below graph L2, rather than between graph L1 and graph L2.

[0190] As another example of this configuration, when the intention information WI acquired by the acquisition unit 11 (21, 23) indicates an emphasized model, the prediction unit 24 changes the value of the ratio parameter RP in a predetermined manner. For example, assume that the acquisition unit 11 (21, 23) acquires intention information WI indicating that an emphasis is placed on a model defined by a graph L1.

[0191] In this case, as an example of a predetermined method, the prediction unit 24 increases the value of the ratio parameter RP of the model defined by the graph L1 by 10%. For example, if the value of the ratio parameter RP of the model defined by the graph L1 is "0.3", the prediction unit 24 increases the value of the ratio parameter RP by 10% to "0.33".

[0192] Furthermore, the prediction unit 24 may lower the value of the ratio parameter RP of a model other than the emphasized model (the model defined by graph L2 in FIG. 8) by 10%. For example, if the value of the ratio parameter RP of the model defined by graph L2 is "0.7", the prediction unit 24 changes the value of the ratio parameter RP to "0.63", which is a 10% decrease.

[0193] Furthermore, the prediction unit 24 may be configured to determine to what extent the user's intention is reflected. For example, assume that the prediction unit 24 determines to reflect 50% of the user's intention. In this case, if the acquisition unit 11 (21, 23) acquires intention information WI indicating that the value of the ratio parameter RP of a model with a value of "0.3" of the ratio parameter RP is to be changed to "0.4," the prediction unit 24 reflects 50% of the user's intention and changes the value of the ratio parameter RP to "0.35."

[0194] In this way, the information processing device 2A derives the prediction result PR by applying the learned parameters LP of the target model and the weights determined by the ratio parameters RP and the intention information WI to the inference data PD. Therefore, the information processing device 2A can derive the prediction result RP that reflects the intention of the user.

[0195] (Step S22_2) In step S22_2, the output unit 15 (22) presents the prediction result PR derived in step S24 to the user. As an example, the output unit 15 (22) may change γ in the above-mentioned FIG. 8 or FIG. 9 to the value of the ratio parameter RP changed in step S24 and present the result.

[0196] Also, as in the example above, if the ratio parameter of the model corresponding to graph L1 is "0.4" and the ratio parameter of the model corresponding to graph L2 is "0.7", the values ​​of the ratio parameter RP to be presented may be "0.4" and "0.7" indicated by the intention information WI, or may be "0.63" and "0.37" after normalization.

[0197] (Effects of information processing device 1A) As described above, in the information processing device 1A, the prediction result RP is derived by applying the learned parameters LP and weights determined by the intention information WI indicating the user's intention regarding the target model or the weights of the learned parameters LP to the inference data PD.

[0198] For example, consider a model for predicting sales of drinking water in a store. In this case, when the image shown in Fig. 8 is displayed in step S22_1 described above, if the user thinks that the sales at a certain date and time x0 are close to situation A, the user can increase the usage rate of model L1 defined by L1.

[0199] With this configuration, the information processing device 1A can reflect the user's intentions in the parameters (weights) between a plurality of models.

[0200] Third Exemplary Embodiment A third exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same functions as those described in the above exemplary embodiment will be assigned the same reference numerals, and their description will be omitted as appropriate. The scope of application of each technical means employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technical means employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise. Furthermore, each technical means shown in each drawing referenced to explain this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise.

[0201] (Configuration of information processing device 2A) The configuration of the information processing device 2A will be described with reference to Fig. 10. Fig. 10 is a block diagram showing the configuration of the information processing device 2A. As shown in Fig. 10, the information processing device 2A includes a control unit 20A, a storage unit 25A, a communication unit 26A, and an input / output unit 27A.

[0202] The memory unit 25A stores data referenced by the control unit 20A, similar to the memory unit 15A described above. Examples of data stored in the memory unit 25A include, but are not limited to, the learning result LR, inference data PD, type information CI, intention information WI, and prediction result PR, as shown in Fig. 10. The learning result LR, inference data PD, type information CI, intention information WI, and prediction result PR are as described above.

[0203] The communication unit 26A is an interface that transmits and receives data via a network, similar to the above-described communication unit 16A.

[0204] The input / output unit 27A is an interface that receives input of data and outputs data, similar to the above-described input / output unit 17A.

[0205] (Control unit 20A) The control unit 20A controls each component included in the information processing device 2A. As shown in Fig. 10, the control unit 20A also includes a first acquisition unit 21, an output unit 22, a second acquisition unit 23, and a prediction unit 24. In this exemplary embodiment, the first acquisition unit 21, the output unit 22, the second acquisition unit 23, and the prediction unit 24 respectively realize a first acquisition means, a presentation means, a second acquisition means, and a prediction means.

[0206] The first acquisition unit 21 acquires inference data PD, a learning result LR (learned parameters LP) for each of a plurality of target models, and type information CI associated with the target model or the learning result LR. The first acquisition unit 21 stores the acquired inference data PD, learning result LR, and type information CI in the storage unit 25A.

[0207] The output unit 22 presents the type information CI to the user. Examples of images presented by the output unit 22 are as shown in FIGS.

[0208] The second acquisition unit 23 acquires the intention information WI. The second acquisition unit 23 stores the acquired intention information WI in the storage unit 25A.

[0209] The prediction unit 24 derives the prediction result PR by applying at least one of the learning results LR for each of the plurality of target models and a weight determined by the intention information WI to the inference data PD. The process by which the prediction unit 24 derives the prediction result PR is as described above.

[0210] (Effects of information processing device 2A) As described above, in the information processing device 2A, the prediction result PR is derived by applying at least one of the learning results LR for each of the plurality of target models and the weight determined by the intention information WI.

[0211] With this configuration, the information processing device 2A can reflect the user's intentions in the parameters (weights) between a plurality of models.

[0212] Fourth Exemplary Embodiment A fourth exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same functions as those described in the above exemplary embodiment will be assigned the same reference numerals, and their description will be omitted as appropriate. The scope of application of each technical means employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technical means employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise. Furthermore, each technical means shown in each drawing referenced to explain this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise.

[0213] (Configuration of information processing device 1B) The configuration of the information processing device 1B will be described with reference to Fig. 11. Fig. 11 is a block diagram showing the configuration of the information processing device 1B. As shown in Fig. 11, the information processing device 1B includes a control unit 10B, a storage unit 15B, a communication unit 16A, and an input / output unit 17A. The communication unit 16A and the input / output unit 17A have been described above, and therefore description thereof will be omitted.

[0214] The memory unit 15B stores data referenced by the control unit 10B, similar to the memory unit 15A described above. Examples of data stored in the memory unit 15B include, but are not limited to, target data TD, ratio parameter RP, regression coefficient RC, learning result LR, inference data PD, type information CI, intention information WI, and prediction result PR, as shown in Fig. 11. The target data TD, ratio parameter RP, regression coefficient RC, learning result LR, inference data PD, type information CI, intention information WI, and prediction result PR have been described above, and therefore further description will be omitted.

[0215] (control unit 10B) The control unit 10B controls each component included in the information processing device 1B, similar to the control unit 10A described above. As shown in FIG. 11 , the control unit 10B includes an acquisition unit 11 (21, 23), a regression coefficient calculation unit 12, a ratio parameter calculation unit 14, an output unit 15 (22), an initial value determination unit 16, a convergence determination unit 17, and a prediction unit 24. In this exemplary embodiment, the acquisition unit 11 (21, 23) serves as a first acquisition unit and a second acquisition unit. In this exemplary embodiment, the output unit 15 (22) and the prediction unit 24 serve as an output unit and a prediction unit, respectively. The acquisition unit 11 (21, 23), the output unit 15 (22), the initial value determination unit 16, the convergence determination unit 17, and the prediction unit 24 have been described above, and therefore further description thereof will be omitted.

[0216] The regression coefficient calculation unit 12 calculates a regression coefficient RC for each of the plurality of target models by referring to the ratio parameter RP that defines the ratio of the plurality of target models and the target data TD. In this exemplary embodiment, the regression coefficient calculation unit 12 calculates the regression coefficient RC for each of the plurality of target models by referring to the hidden variable z kThe covariance parameter η of the prior distribution of and the hidden variable z k The regression coefficients RC are calculated using the least squares method without updating the covariance matrices Φ and Φ of the prior distributions.

[0217] The ratio parameter calculation unit 14 calculates the ratio parameter RP by referring to the target data TD and the regression coefficient RC. In this exemplary embodiment, the ratio parameter RP is not limited to an internal ratio parameter, and may be an external ratio parameter.

[0218] (Example of processing flow in information processing device 1B) 12 is a diagram showing an example of the flow of processing in the information processing device 1B according to this exemplary embodiment. The acquisition processing S11, the initial value determination processing S16, and the output processing S15 are as described above, and therefore descriptions thereof will be omitted.

[0219] (Step S12: Regression coefficient calculation process) In step S12, the regression coefficient calculation unit 12 calculates the ratio parameter RP expressed as the following equation (35):

[0220]

number

[0221] and the target data TD expressed as the following equation (36):

[0222]

number

[0223] and the regression coefficient RC for each of the plurality of target models, which is expressed as the following equation (37):

[0224]

number

[0225] As an example, the regression coefficient calculation unit 12 calculates the regression coefficient W that minimizes the value of the evaluation function of the following equation (38) with the value of the ratio parameter RP and the value of the target data TD set as fixed values. (i) Calculate the value of

[0226]

number

[0227] That is, the regression coefficient calculation unit 12 calculates the regression coefficient W using the least squares method.

[0228] (Step S14: Ratio parameter calculation process) Subsequently, in step S14, the ratio parameter calculation unit 14 calculates the target data TD expressed as the following equation (39):

[0229]

number

[0230] and the regression coefficient RC expressed as the following equation (38):

[0231]

number

[0232] and the ratio parameter RP expressed as the following equation (41)

[0233]

number

[0234] As described above, the ratio parameter calculation unit 14 calculates (updates) the hidden variables z k The covariance parameter η of the prior distribution of and the hidden variable z k The ratio parameter RP is calculated without updating the covariance matrix Φ of the prior distribution of . -1The ratio parameter RP is calculated by the following equation (42) with ρ = 0:

[0235]

number

[0236] Furthermore, similarly to the above-described exemplary embodiment, the ratio parameter calculation unit 14 calculates the ratio parameter RP under the constraints expressed as the following equations (43A) and (43B).

[0237]

number

[0238] (Step S17: Convergence determination process) Subsequently, in step S17, the convergence determination unit 17 determines whether or not the series of processes in steps S12, S13, and S14 described above has converged. In step S17, the convergence determination unit 17, for example, determines the variational lower bound (VLB) J obtained by the following equation (44), as in the exemplary embodiment:

[0239]

number

[0240] If the change in the variation lower limit is equal to or less than a predetermined threshold, it is determined that the series of processes in steps S12, S13, and S14 have converged.

[0241] As another example, the convergence determination unit 17 determines whether the value of the ratio parameter RP has converged. For example, in the nth convergence determination process of repeating a series of processes including the above-mentioned steps S12, S13, and S14, the convergence determination unit 17 compares the value of the ratio parameter RP at the (n-1)th time with the value of the ratio parameter RP at the nth time, and determines that the series of processes at the above-mentioned steps S12, S13, and S14 has converged if the absolute value of the difference between them is equal to or smaller than a predetermined threshold.

[0242] (Effects of information processing device 1B) As described above, the information processing device 1B calculates the regression coefficient RC using the least squares method. With this configuration, the information processing device 1B also reflects the user's intention in the parameters between a plurality of models, just like the information processing device 1A.

[0243] [Software implementation example] Some or all of the functions of the information processing devices 1, 1A, 1B, and 2A (hereinafter also referred to as "each of the above devices") may be realized by hardware such as an integrated circuit (IC chip), or by software.

[0244] In the latter case, each of the above devices is realized by, for example, a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in Figure 13. Figure 13 is a block diagram showing the hardware configuration of computer C that functions as each of the above devices.

[0245] The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program P for causing the computer C to operate as each of the above-mentioned devices. In the computer C, the processor C1 reads and executes the program P from the memory C2, thereby realizing the functions of each of the above-mentioned devices.

[0246] The processor C1 may be, for example, a central processing unit (CPU), a graphic processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a tensor processing unit (TPU), a quantum processor, a microcontroller, or a combination thereof. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination thereof.

[0247] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may also include a communication interface for transmitting and receiving data to and from other devices. The computer C may also include an input / output interface for connecting input / output devices such as a keyboard, mouse, display, and printer.

[0248] Furthermore, the program P can be recorded on a non-transitory tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can acquire the program P via such a recording medium M. The program P can also be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also acquire the program P via such a transmission medium.

[0249] [Appendix A] This disclosure includes the techniques described in the following appendices. However, the present invention is not limited to the techniques described in the following appendices, and various modifications are possible within the scope of the claims.

[0250] (Appendix A1) a first acquisition means for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation means for presenting the feature information to a user; a second acquisition means for acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction means for applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data to derive a prediction result; An information processing device comprising:

[0251] (Appendix A2) The learned parameters include regression coefficients, The presentation means further presents a result obtained by applying the regression coefficient to the inference data. 10. The information processing device according to claim 1,

[0252] (Appendix A3) the learned parameters include a ratio parameter that defines a ratio of the plurality of target models; The presentation means further presents a result obtained by applying the regression coefficients and the ratio parameters to the inference data. 10. The information processing device according to claim 9, wherein the information processing device is a

[0253] (Appendix A4) The prediction means performs the following on the inference data: At least one of learned parameters for each of the plurality of target models; a weight determined by the ratio parameter and the intention information; The prediction result is derived by applying 10. The information processing device according to claim 9, wherein the information processing device is a

[0254] (Appendix A5) The regression coefficients and the ratio parameters are a regression coefficient calculation process that calculates a regression coefficient for each of a plurality of target models by referring to the ratio parameters and the learning data; and A ratio parameter calculation process for calculating the ratio parameters by referring to the learning data and the regression coefficients. are the regression coefficients and ratio parameters learned by the learning process including 10. The information processing device according to claim 9, wherein the information processing device is a device for processing information.

[0255] (Appendix A6) The learning process for the regression coefficients and the ratio parameters is a covariance calculation process for calculating covariance parameters of the prior distribution of the latent variables and a covariance matrix of the posterior distribution of the latent variables by referring to the training data, the ratio parameters, the regression coefficients, and information on the prior distribution of the latent variables; Further comprising: The ratio parameter calculation process is Calculate the ratio parameter by further referring to the covariance matrix of the posterior distribution of the hidden variables. 10. The information processing device according to claim 9, wherein the information processing device is a

[0256] [Appendix B] This disclosure includes the techniques described in the following appendices. However, the present invention is not limited to the techniques described in the following appendices, and various modifications are possible within the scope of the claims.

[0257] (Appendix B1) At least one processor a first acquisition process for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation process of presenting the feature information to a user; a second acquisition process of acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction process for deriving a prediction result by applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data; An information processing method comprising:

[0258] (Appendix B2) The learned parameters include regression coefficients, In the presentation process, the at least one processor further presents a result obtained by applying the regression coefficients to the inference data. 1. The information processing method described in Appendix B1.

[0259] (Appendix B3) the learned parameters include a ratio parameter that defines a ratio of the plurality of target models; In the presentation process, the at least one processor further presents a result obtained by applying the regression coefficients and the ratio parameters to the inference data. 1. The information processing method described in Appendix B2.

[0260] (Appendix B4) In the prediction process, the at least one processor performs the following on the inference data: At least one of learned parameters for each of the plurality of target models; a weight determined by the ratio parameter and the intention information; The prediction result is derived by applying The information processing method described in Appendix B3.

[0261] (Appendix B5) The regression coefficients and the ratio parameters are a regression coefficient calculation process in which the at least one processor calculates a regression coefficient for each of a plurality of target models by referring to the ratio parameters and learning data; and a ratio parameter calculation process in which the at least one processor calculates the ratio parameter by referring to the training data and the regression coefficients; are the regression coefficients and ratio parameters learned by the learning process including An information processing method as described in Appendix B3 or B4.

[0262] (Appendix B6) The learning process for the regression coefficients and the ratio parameters is a covariance calculation process in which the at least one processor calculates covariance parameters of the prior distribution of the latent variables and a covariance matrix of the posterior distribution of the latent variables by referring to the training data, the ratio parameters, the regression coefficients, and information on the prior distribution of the latent variables; Further comprising: In the ratio parameter calculation process, the at least one processor: Calculate the ratio parameter by further referring to the covariance matrix of the posterior distribution of the hidden variables. The information processing method described in Appendix B5.

[0263] [Appendix C] This disclosure includes the techniques described in the following appendices. However, the present invention is not limited to the techniques described in the following appendices, and various modifications are possible within the scope of the claims.

[0264] (Appendix C1) A program that causes a computer to function as an information processing device, The computer a first acquisition means for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation means for presenting the feature information to a user; a second acquisition means for acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction means for applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data to derive a prediction result; An information processing program that functions as a

[0265] (Appendix C2) The learned parameters include regression coefficients, The presentation means further presents a result obtained by applying the regression coefficient to the inference data. An information processing program as described in Appendix C1.

[0266] (Appendix C3) the learned parameters include a ratio parameter that defines a ratio of the plurality of target models; The presentation means further presents a result obtained by applying the regression coefficients and the ratio parameters to the inference data. An information processing program as described in Appendix C2.

[0267] (Appendix C4) The prediction means performs the following on the inference data: At least one of learned parameters for each of the plurality of target models; a weight determined by the ratio parameter and the intention information; The prediction result is derived by applying An information processing program as described in Appendix C3.

[0268] (Appendix C5) The regression coefficients and the ratio parameters are a regression coefficient calculation process that calculates a regression coefficient for each of a plurality of target models by referring to the ratio parameters and the learning data; and A ratio parameter calculation process for calculating the ratio parameters by referring to the learning data and the regression coefficients. are the regression coefficients and ratio parameters learned by the learning process including An information processing program according to Appendix C3 or C4.

[0269] (Appendix C6) The learning process for the regression coefficients and the ratio parameters is a covariance calculation process for calculating covariance parameters of the prior distribution of the latent variables and a covariance matrix of the posterior distribution of the latent variables by referring to the training data, the ratio parameters, the regression coefficients, and information on the prior distribution of the latent variables; Further comprising: The ratio parameter calculation process is Calculate the ratio parameter by further referring to the covariance matrix of the posterior distribution of the hidden variables. An information processing program as described in Appendix C5.

[0270] [Appendix D] This disclosure includes the techniques described in the following appendices. However, the present invention is not limited to the techniques described in the following appendices, and various modifications are possible within the scope of the claims.

[0271] (Appendix D1) at least one processor, a first acquisition process for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation process of presenting the feature information to a user; a second acquisition process of acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction process for deriving a prediction result by applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data; An information processing device that executes the above.

[0272] The information processing device may further include a memory, and the memory may store a program for causing the at least one processor to execute each of the processes.

[0273] (Appendix D2) The learned parameters include regression coefficients, In the presentation process, the at least one processor further presents a result obtained by applying the regression coefficients to the inference data. 10. The information processing device according to claim 9, wherein the information processing device is an information processing device according to claim 1.

[0274] (Appendix D3) the learned parameters include a ratio parameter that defines a ratio of the plurality of target models; In the presentation process, the at least one processor further presents a result obtained by applying the regression coefficients and the ratio parameters to the inference data. 10. The information processing device according to claim 9, wherein the information processing device is an information processing device according to claim 1, wherein

[0275] (Appendix D4) In the prediction process, the at least one processor performs the following on the inference data: At least one of learned parameters for each of the plurality of target models; a weight determined by the ratio parameter and the intention information; The prediction result is derived by applying 10. The information processing device according to claim 9, wherein the information processing device is an information processing device according to claim 1, wherein

[0276] (Appendix D5) The regression coefficients and the ratio parameters are a regression coefficient calculation process that calculates a regression coefficient for each of a plurality of target models by referring to the ratio parameters and the learning data; and A ratio parameter calculation process for calculating the ratio parameters by referring to the learning data and the regression coefficients. are the regression coefficients and ratio parameters learned by the learning process executed by the at least one processor. An information processing device according to appendix D3 or D4.

[0277] (Appendix D6) The learning process for the regression coefficients and the ratio parameters is the at least one processor: a covariance calculation process for calculating covariance parameters of the prior distribution of the latent variables and a covariance matrix of the posterior distribution of the latent variables by referring to the training data, the ratio parameters, the regression coefficients, and information on the prior distribution of the latent variables; Further execute In the ratio parameter calculation process, the at least one processor: Calculate the ratio parameter by further referring to the covariance matrix of the posterior distribution of the hidden variables. 10. The information processing device according to claim 9, wherein said information processing device is a

[0278] [Appendix E] This disclosure includes the techniques described in the following appendices. However, the present invention is not limited to the techniques described in the following appendices, and various modifications are possible within the scope of the claims.

[0279] (Appendix E1) A non-transitory recording medium on which a program that causes a computer to function as an information processing device is recorded, The program causes the computer to: a first acquisition process for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation process of presenting the feature information to a user; a second acquisition process of acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction process for deriving a prediction result by applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data; A non-transitory recording medium on which a program for executing a program is recorded. [Explanation of symbols]

[0280] 1, 1A, 1B, 2A Information processing equipment 11 Acquisition Department 12 Regression coefficient calculation section 13 Covariance calculation part 14 Ratio parameter calculation section 15, 22 Output section 16 Initial value determination section 17 Convergence judgment section 18, 24 Prediction Section 21 First Acquisition Section 22 Presentation section 23 Second Acquisition Section TD Target Data RP ratio parameter RC regression coefficient DI distribution information LR learning results LP trained parameters PD inference data CI Type Information WI intention information CD candidate data CDG candidate data set PR prediction results

Claims

1. a first acquisition means for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation means for presenting the feature information to a user; a second acquisition means for acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction means for applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data to derive a prediction result; An information processing device comprising:

2. The learned parameters include regression coefficients, The presentation means further presents a result obtained by applying the regression coefficient to the inference data. The information processing device according to claim 1 .

3. the learned parameters include a ratio parameter that defines a ratio of the plurality of target models; The presentation means further presents a result obtained by applying the regression coefficients and the ratio parameters to the inference data. The information processing device according to claim 2 .

4. The prediction means performs the following on the inference data: At least one of learned parameters for each of the plurality of target models; a weight determined by the ratio parameter and the intention information; The prediction result is derived by applying The information processing device according to claim 3 .

5. The regression coefficients and the ratio parameters are a regression coefficient calculation process that calculates a regression coefficient for each of a plurality of target models by referring to the ratio parameters and the learning data; and A ratio parameter calculation process for calculating the ratio parameters by referring to the learning data and the regression coefficients. are the regression coefficients and ratio parameters learned by the learning process including 5. The information processing device according to claim 3 or 4.

6. The learning process for the regression coefficients and the ratio parameters is a covariance calculation process for calculating covariance parameters of the prior distribution of the latent variables and a covariance matrix of the posterior distribution of the latent variables by referring to the training data, the ratio parameters, the regression coefficients, and information on the prior distribution of the latent variables; Further comprising: The ratio parameter calculation process is Calculate the ratio parameter by further referring to the covariance matrix of the posterior distribution of the hidden variables. The information processing device according to claim 5 .

7. At least one processor a first acquisition process for acquiring inference data, learned parameters for each of a plurality of target models, and feature information associated with the target models or the learned parameters; a presentation process of presenting the feature information to a user; a second acquisition process of acquiring intention information indicating the intention of the user regarding the target model or weights of the learned parameters; a prediction process for deriving a prediction result by applying at least one of the learned parameters for each of the plurality of target models and a weight determined by the intention information to the inference data; An information processing method comprising:

8. 2. A program for causing a computer to function as the information processing device according to claim 1, the program causing a computer to function as the first acquisition means, the presentation means, the second acquisition means, and the prediction means.

Images