Information processing device, display method, and program

Abstract

The present invention provides a display device, a display method, and a program capable of suitably displaying information regarding the path leading to the final state. [Solution] The information processing device 1X mainly comprises a path calculation means 16X and a display control means 17X. The path calculation means 16X determines a path from first information representing the feature quantities of a first event and second information representing the feature quantities of a second event different from the first event, such that the evaluation value, which evaluates the complexity of the path from the first information to the second information, decreases. The display control means 17X displays information about the determined path on a display device.

Description

[Technical Field]

[0001] This disclosure relates to the technical fields of information processing devices, display methods, and programs. [Background technology]

[0002] In various fields, machine learning models are used to make predictions on input data representing specific cases. Related to this, machine learning models are also used to generate suggested modifications to the input data that alter the prediction. For example, Patent Document 1 describes generating suggested modifications to the input data that can obtain the desired prediction. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] International Publication No. 2022 / 034685 [Overview of the project] [Problems that the invention aims to solve]

[0004] When presenting only information showing the improved final state as a suggested modification of input data to obtain the desired prediction, there is a problem in that it is difficult for the user to plan the path to the final state when there is a large discrepancy from the current state.

[0005] One of the purposes of this disclosure is to provide an information processing device, a display method, and a program that can assist users in easily planning the path to the final state, in light of the aforementioned issues. [Means for solving the problem]

[0006] One aspect of an information processing device is: Path calculation means for determining the path from the first information representing the feature amount of the first event to the second information representing the feature amount of the second event different from the first event so that the evaluation value obtained by evaluating the complexity of the path decreases; Display control means for displaying information regarding the determined path on a display device; An information processing apparatus having the above.

[0007] One aspect of the display method is as follows. A computer determines the path from the first information representing the feature amount of the first event to the second information representing the feature amount of the second event different from the first event so that the evaluation value obtained by evaluating the complexity of the path decreases, and displays information regarding the determined path on a display device. This is the display method.

[0008] One aspect of the program is as follows. A program that causes a computer to execute a process of determining the path from the first information representing the feature amount of the first event to the second information representing the feature amount of the second event different from the first event so that the evaluation value obtained by evaluating the complexity of the path decreases, and displaying information regarding the determined path on a display device.

Advantages of the Invention

[0009] In an example of the effect according to the present disclosure, it is possible to assist the user in easily establishing the route from the first event to the second event.

Brief Description of the Drawings

[0010] [Figure 1] Shows the configuration of the information processing system. [Figure 2] A diagram schematically showing the relationship between input examples, corrected examples, and paths. [Figure 3] Shows the hardware configuration of the information processing apparatus. [Figure 4] This is an example of a processor's functional block. [Figure 5] This is an example of a flowchart executed by an information processing device. [Figure 6] This is an example of a graph showing the transition of the Lagrangian function and features from the input example to the modified example. [Figure 7] This is an example of a block diagram for the functional input section when performing path determination based on the Riemannian metric. [Figure 8] This is an example of a block diagram for the functional input section used when determining the functions necessary for designing a Riemannian metric. [Figure 9] This is an example of a graph showing the prediction results and feature changes from the input example to the modified example, based on a prediction model. [Figure 10] This is an example of a graph that shows information about the path from the input example to the corrected example. [Figure 11] This is the first example of displaying the route calculation result. [Figure 12] This is a second example of how the route calculation results are displayed. [Figure 13] This is a third example of how the route calculation results are displayed. [Figure 14] This shows the configuration of the information processing system. [Figure 15] This is a functional block diagram of an information processing device. [Figure 16] This is an example of a flowchart illustrating the processing steps of an information processing device. [Figure 17] This is a functional block diagram of an information processing device. [Figure 18] This is an example of a flowchart illustrating the processing steps of an information processing device. [Modes for carrying out the invention]

[0011] The following describes embodiments of the information processing device, display method, and program with reference to the drawings.

[0012] <First Embodiment> (1) System Configuration Figure 1 shows the configuration of the information processing system 100. The information processing system 100 makes predictions for cases containing multiple features (i.e., characteristics), calculates paths to cases modified to improve the predictions, and displays information about the calculated paths. The information processing system 100 mainly comprises an information processing device 1, an input device 2, and a display device 3. Hereafter, the cases used for prediction will be called "input cases," and the cases modified to improve the predictions will also be called "modified cases."

[0013] The information processing device 1 calculates the path from the input example to the modified example based on the information supplied from the input device 2, and displays information about the calculated path on the display device 3. The information processing device 1 communicates data with the input device 2 and the display device 3 via a communication network, or via direct wireless or wired communication.

[0014] Input device 2 is an interface that accepts user input, which is an external input, and includes, for example, a touch panel, buttons, a keyboard, or a voice input device. Input device 2 supplies information generated based on the user input to information processing device 1.

[0015] The display device 3 is, for example, a display, a projector, etc., and performs a predetermined display based on the display information supplied from the information processing device 1.

[0016] The configuration of the information processing system 100 shown in Figure 1 is an example, and various modifications may be made to this configuration. For example, the input device 2 and the display device 3 may be configured as a single unit. In this case, the input device 2 and the display device 3 may be configured as a tablet terminal integrated with the information processing device 1. The information processing device 1 may also be connected to or have a built-in sound output device, such as a speaker, to output information by sound. Furthermore, the information processing device 1 may be composed of multiple devices. In this case, the multiple devices constituting the information processing device 1 exchange information among themselves that is necessary to execute pre-assigned processes.

[0017] The information processing system 100 is suitably used, for example, in fields such as loan application review, marketing, health checkups, and robotics, to show proposed modifications (including improvement proposals) to the state of a target, such as a customer, along with the path of the state leading up to the modification proposal. In other words, the information processing system 100 is suitably used for recourse or counterfactual explanations.

[0018] Specifically, in the loan application example, it is used to determine the path to correcting the state so that the machine learning model outputs a low probability of bankruptcy. In the marketing example, it is used to determine the path to correcting the state so that the purchase amount increases. In the health checkup example, it is used to determine the path to correcting the state so that the health status improves. In the robotics example, it is used to determine the path to correcting the state so that the robot's trajectory improves.

[0019] Furthermore, generally speaking, it is difficult to plan a path when the goal is too large, and easier when the goal is broken down into smaller parts. Therefore, there is a need to know the path to the revised case. For example, if the goal is to reduce the probability of bankruptcy to 10%, setting the first goal to reduce the probability of bankruptcy to 50% or less, the second goal to reduce it to 30% or less, and the final goal to reduce it to 10% or less makes it easier for the user to plan a path. In such cases, the information processing system 100 can appropriately support the approach of achieving goals through the aforementioned method of setting smaller goals by calculating the path and presenting information about the path to the user.

[0020] Figure 2 is a schematic diagram showing the relationship between the input example, the modified example, and the path proposed by the information processing device 1. In the example in Figure 2, "d" (where d is an integer greater than or equal to 1) types of features are used, and the state of the input example is represented by d types of features "(x1,...,x d The state of the corrected example is represented by d types of features "(y1,...,y dIt is represented by the feature quantity (x1, ..., x) that represents the state of the input case. d ) from which the features (y1, ..., y) represent the state of the corrected case. d The path "γ" in the feature space leading to ) is determined. Details of the method for determining path γ will be described later. Note that the input example is an example of the first event, and the feature quantities of the input example "(x1,...,x d ) is an example of the first piece of information. The modified example is an example of the second event, and the feature of the modified example "(y1,...,y d )" is an example of second-party information.

[0021] (2) Hardware configuration Figure 3 shows the hardware configuration of the information processing device 1. The information processing device 1 includes a processor 11, memory 12, and interface 13 as hardware components. The processor 11, memory 12, and interface 13 are connected via a data bus 19.

[0022] The processor 11 executes predetermined processes by running programs stored in memory 12. The processor 11 is a processor such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), or TPU (Tensor Processing Unit). The processor 11 may be composed of multiple processors. The processor 11 is an example of a computer.

[0023] Memory 12 is composed of various volatile and non-volatile memories, such as RAM (Random Access Memory) and ROM (Read Only Memory). Furthermore, memory 12 stores programs for the information processing device 1 to perform various operations.

[0024] Furthermore, memory 12 stores information necessary for the information processing device 1 to calculate routes and control the display of route-related information, other than the program. For example, memory 12 stores model information for a prediction model that makes predictions from input examples. The prediction model is a machine learning model that outputs prediction data when example data is input. Therefore, the prediction model is generated by machine learning using training data consisting of pre-prepared example data and result data. For example, the example data is data representing multiple types of features that represent the business's management status, such as income, debt, personnel costs, and fixed costs (i.e., feature data), and the result data is data that represents the results of the business, such as whether or not the business went bankrupt, or the calculated probability of bankruptcy. By performing machine learning using such training data, a prediction model is generated that outputs prediction data, which is the probability of bankruptcy, from example data consisting of multiple feature data that represent the business situation. Note that the prediction model is not limited to predicting the probability of bankruptcy, but may also predict any indicator that quantitatively represents the good or bad in various fields such as loan screening, marketing, health checkups, and robotics.

[0025] Memory 12 is used as working memory and may temporarily store information obtained from an external device. In this case, the external device may be a storage device such as a hard disk connected to or built into the information processing device 1, or it may be a storage medium such as flash memory. The external device may also be a server device that communicates data with the information processing device 1. In this case, the external device may consist of multiple server devices. Furthermore, the program executed by the information processing device 1 may be stored in a storage medium other than memory 12.

[0026] Interface 13 is an interface for electrically connecting the information processing device 1 with other devices. These interfaces may be wireless interfaces such as network adapters for wirelessly transmitting and receiving data with other devices, or they may be hardware interfaces for connecting with other devices via cables, etc.

[0027] The hardware configuration of the information processing device 1 is not limited to the configuration shown in Figure 3. For example, the information processing device 1 may include at least one of the input device 2 or the display device 3. Furthermore, the information processing device 1 may be connected to or have a built-in sound output device such as a speaker.

[0028] (3) Processing Overview Figure 4 shows an example of the functional blocks of the processor 11 of the information processing device 1. Functionally, the processor 11 includes an input unit 15, a path calculation unit 16, and a display control unit 17.

[0029] The input unit 15 generates information to be input to the path calculation unit 16 based on information supplied from the input device 2 or information pre-stored in the memory 12. The input unit 15 includes a start point input unit 50, an end point input unit 51, a functional input unit 52, and a feature space input unit 53. The start point input unit 50 acquires input example data, which is data relating to the start point of the path (i.e., input example), and supplies the input example data to the path calculation unit 16. The end point input unit 51 acquires modified example data, which is data relating to the end point of the path (i.e., modified example), and supplies the modified example data to the path calculation unit 16. The functional input unit 52 acquires information specifying the functional (i.e., action functional) used to determine the path by minimizing the functional, and supplies the acquired information to the path calculation unit 16. The feature space input unit 53 acquires information specifying the manifold (figure) that defines the feature space, which is the space of d types of feature quantities, and supplies the acquired information to the path calculation unit 16. Note that the feature space input section 53 is not a mandatory component; for example, it may not be present if the feature space is assumed to be a Euclidean space.

[0030] The path calculation unit 16 calculates a path based on the information supplied from the input unit 15. In this case, the path calculation unit 16 minimizes the action functional based on the input example data, the modified example data, and the information specifying the action functional and feature space, and determines the path that minimizes the action functional. The path calculation unit 16 supplies the path calculation result to the display control unit 17.

[0031] The display control unit 17 generates display information based on the route calculation result generated by the route calculation unit 16, and controls the display of the display device 3 by supplying the generated display information to the display device 3. The specific processing of the display control unit 17 will be described later with reference to the display example.

[0032] The input unit 15, route calculation unit 16, and display control unit 17 described in Figure 4 can be implemented, for example, by the processor 11 executing a program. Alternatively, the necessary programs may be recorded on any non-volatile storage medium and installed as needed to implement each component. At least a portion of these components may be implemented not only by software programs, but also by a combination of hardware, firmware, and software. Furthermore, at least a portion of these components may be implemented using a user-programmable integrated circuit, such as an FPGA (Field-Programmable Gate Array) or a microcontroller. In this case, the program composed of the above components may be implemented using this integrated circuit. At least a portion of each component may also be implemented by an ASSP (Application Specific Standard Produce), ASIC (Application Specific Integrated Circuit), or quantum processor (quantum computer control chip). Thus, each component may be implemented by various hardware. The same applies to other embodiments described later. Furthermore, each of these components may be implemented by the collaboration of multiple computers, for example, using cloud computing technology.

[0033] FIG. 5 shows an example of a flowchart executed by the information processing apparatus 1. First, the input unit 15 of the information processing apparatus 1 acquires input case data, corrected case data, and information specifying an action functional and a feature space (step S11). In this case, the input unit 15 may acquire this information by receiving user input by the input device 2, or may read out this information from the memory 12 if this information is stored in the memory 12 in advance as initial information. Next, the path calculation unit 16 of the information processing apparatus 1 calculates a path that minimizes the specified action functional (step S12). Then, the display control unit 17 of the information processing apparatus 1 displays information regarding the calculated path on the display device 3 (step S13).

[0034] (4) Details of the route determination method Next, details of the path determination by the path calculation unit 16 will be described.

[0035] (4-1) Path determination based on the Lagrange function The path calculation unit 16 calculates a path γ that minimizes the action functional "I(γ)" shown in the following equation (1) using the Lagrange function "L". Note that the value of the action functional I(γ) corresponds to "an evaluation value obtained by evaluating the complexity of the path from the first information to the second information".

[0036]

Equation

[0037] Here, "γ" represents a path on the feature space of feature quantities. "M" is a manifold or a figure representing the feature space of feature quantities, for example, a Euclidean space. Also, "t" represents a mediating variable, and the mediating variable corresponding to the input case is set to "t = 0", and the mediating variable corresponding to the corrected case is set to "t = T". Note that hereinafter, unless otherwise specified, "T = 1". Also, "X (0) " represents the set of starting points of the path γ, and "X (0) ⊂ M" holds. "X (T) " represents the set of ending points of the path γ, and "X (T) ⊂ M" holds.

[0038] Thus, the action functional "I(γ)" can be expressed as the integral of the Lagrangian function L(t,γ,γ′) with t, γ, and the first derivative of γ, γ′ (dot notation is used in each equation). Note that the Lagrangian function L may be designed to take not only the first derivative of γ, but also the second or higher derivative of γ as an argument. The specific Lagrangian function L is determined by the functional input section 52.

[0039] The Lagrangian function L may be obtained through machine learning using deep learning with a neural network. Specific examples of machine learning the Lagrangian function using a neural network architecture are disclosed, for example, in the following literature. “LAGRANGIAN NEURAL NETWORKS” arXiv:2003.04630v2 [cs.LG] 30 Jul 2020

[0040] The path calculation unit 16 may solve the optimization problem shown in equation (1) by reducing it to a nonlinear programming method or by reducing it to a differential dynamic programming method. Examples of the nonlinear programming methods mentioned above include the prim-dual interior-point method, gradient method, genetic algorithm, local search, and successive improvement method. These algorithms can be implemented in, for example, ALGLIB, NLopt, SciPy, IPOPT, and CasADi. As an example of the differential dynamic programming method mentioned above, the method described in the following literature may be used. Z. Xie, C K. Liu, and K. Hauser, Differential dynamic programming with nonlinear constraints, 2017 IEEE international conference on robotics and automation (ICRA), 2017, pp. 695‐702.

[0041] To approximately solve equation (1), the path calculation unit 16 approximates the integral of the action functional with a point cloud using a point selection method. In this case, the path calculation unit 16 solves the following equation (2), which is a transformation of equation (1).

[0042]

number

[0043] The selected point may be the leftmost endpoint of each of the N intervals obtained by dividing the range of the parameter t "0 ≤ t ≤ 1" into intervals, or it may be the midpoint. The path calculation unit 16 may also approximate using Simpson's rule. The path calculation unit 16 then determines the path γ by solving equation (2) using the nonlinear programming or differential dynamic programming described above.

[0044] Figure 6 is an example of a graph showing the transition of the Lagrangian function L and features from the input example to the modified example. In Figure 6, it is assumed that d types of features are used, and the features at the time of the input example are "(x1,...,x d ) and the features at the time of the corrected example are "(y1,...,y d )” Graph G01 shows the transition of the Lagrangian function L value for the path that minimizes the action functional I in equation (1), and graph G02 shows an example of the transition of the Lagrangian function L value for the path that does not minimize the action functional I.

[0045] Here, the action functional shown in graph G01 corresponds to the area enclosed by graph G01 and the horizontal axis in the interval where the parameter t is 0 to T (i.e., the hatched region in Figure 6). The path calculation unit 16 then determines the path γ that minimizes this area. Note that the area enclosed by graph G02 and the horizontal axis in the interval where the parameter t is 0 to T is clearly larger than the area enclosed by graph G01 and the horizontal axis, and does not minimize the action functional I. Therefore, the path specified by graph G02 is not the path γ determined by the path calculation unit 16.

[0046] (4-2) Path determination based on the Riemannian metric In a suitable example of minimizing the action functional I of equation (1), the path calculation unit 16 may calculate the shortest path with respect to the Riemann metric. For example, when the path calculation unit 16 calculates the shortest path based on the length of the Riemann metric, it transforms the Lagrangian function L shown in equation (1) into the following equation (3) using the inner product "g".

[0047]

number

[0048] Here, the Riemannian metric is the inner product g for each point x∈M. x It refers to something that is determined. And the inner product g x When it is a standard dot product, it corresponds to the Euclidean metric, and the length of the path is the Euclidean distance. Also, let "v" be a vector and "A" be a matrix, and the dot product "g x If (v,v)=(v,Av) holds, then the length of path γ is the Mahalanobis distance. This metric will henceforth be called the Mahalanobis metric.

[0049] The Riemannian metric may be obtained by machine learning. In this case, the Fisher information may be determined by machine learning, and the Mahalanobis distance may be determined by machine learning. Furthermore, the Riemannian metric may be obtained by deep learning using a neural network.

[0050] Here, "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (4) holds true.

[0051]

number

[0052] In this case, the Lagrangian function L in equation (1) is expressed by the following equation (5).

[0053]

number

[0054] Note that ||γ||² in equations (4) and (5) represents the L² norm of path γ, but it may also be the Mahalanobis distance. Function "f" represents the prediction model and is a function that takes features as arguments. Function f may be a probability distribution, a probability, a utility function, or an error function. Function f may also be an existing basic model or a function obtained through machine learning. Examples of machine learning for function f include XGBoost, LightGBM, and deep learning using neural networks. Function f does not have to be a smooth function. If function f is not smooth, it may be approximated by a smooth function and optimization may be performed to obtain path γ.

[0055] Figure 7 shows an example of a block diagram of the functional input unit 52 when performing path determination based on the Riemann metric. The functional input unit 52 includes a Riemann metric input unit 522 and a functional setting unit 523. In this case, the Riemann metric input unit 522 obtains information specifying the inner product g by referring to the memory 12 or from the input device 2. The functional setting unit 523 then sets the functional shown in equation (1) based on the information specifying the inner product g obtained by the Riemann metric input unit 522 and equation (3).

[0056] Figure 8 is an example of a block diagram of the functional input unit 52 when determining the function f necessary for designing the Riemann metric. The functional input unit 52 includes a function input unit 521, a Riemann metric design unit 522A, and a functional setting unit 523. In this case, the function input unit 521 obtains information specifying the function f by referring to the memory 12 or from the input device 2. The Riemann metric design unit 522A then obtains information specifying the function f from the function input unit 521 and designs the Riemann metric corresponding to the inner product g based on the information specifying the function f. In this case, for example, the relationship between the function f and the inner product g (for example, "g x (v,v)=f(x) 2||v||2 2 Information indicating the function f is pre-stored in memory 12, etc., and the Riemann metric design unit 522A refers to this information to determine the inner product g from the specified function f. Then, the functional setting unit 523 sets the Lagrangian function L to be used in the minimization problem of the action functional I in equation (1) based on the Riemann metric designed by the Riemann metric design unit 522A.

[0057] The path calculation unit 16 may determine the path based on the energy of the Riemann metric, instead of using equation (3) to determine the path based on the length of the Riemann metric.

[0058] When determining the path based on the energy of the Riemann metric, the path calculation unit 16 converts the Lagrangian function L shown in equation (1) to the following equation (6) using the inner product g.

[0059]

number

[0060] Here, when the integral of the action functional in equation (1) is approximated by a point selection method using a point group, the path calculation unit 16 solves equation (7) below as a minimization problem to determine the path γ based on the energy of the Riemann metric.

[0061]

number

[0062] Also, "g x (v,v)=f(x) 2 ||v||2 2 If we do this, equation (6) is transformed into the following equation (8).

[0063]

number

[0064] Here, the same solution is obtained, under the condition that the function f that determines the inner product g is smooth, whether the path is determined based on the length of the Riemann metric using equation (3) or based on the energy of the Riemann metric using equation (6). On the other hand, in terms of computational complexity, the calculation after discrete approximation by the selection method is complex when determining the path based on the length of the Riemann metric because it involves the calculation of square roots, whereas the calculation of square roots does not occur when determining the path based on the energy of the Riemann metric. Therefore, the computational load tends to be lighter when determining the path based on the energy of the Riemann metric.

[0065] Figure 9 is an example of a graph showing the prediction results and feature changes from the input example to the modified example using a prediction model. In Figure 9, as an example, d types of features are used, and the features at the time of the input example are "(x1,...,x d ) and the features at the time of the corrected example are "(y1,...,y d )” Graph G03 shows the transition of the prediction results (i.e., the value of function f) by the prediction model for the path γ that minimizes the action functional I in equation (1), and graph G04 shows an example of the transition of the value of function f for a path that does not minimize the action functional I. The horizontal axis of graphs G03 and G04 represents the length of the path based on the Mahalanobis distance. The horizontal axis of graphs G03 and G04 is an example of the first axis corresponding to the length of the path, and the vertical axis of graphs G03 and G04 is an example of the second axis corresponding to the predicted value at a point on the path. Note that the “distance of the path” is not limited to the Mahalanobis distance, but can be any distance determined from the Riemannian metric. The horizontal and vertical axes of graphs G03 and G04 form a two-dimensional coordinate space.

[0066] In the minimization problem of the action functional I shown in equation (1), determining the path γ based on the Riemannian metric is equivalent to determining the path γ that minimizes the area between the graph showing the prediction results from the prediction model in the interval where the parameter t is 0 to T and the horizontal axis. This area is the value obtained by integrating the function f with respect to the length of the path (i.e., ∫ γIt is equal to f(x)ds). And, since the area enclosed by graph G03 and the horizontal axis in the interval where the parameter t is 0 to T (i.e., the area of ​​the hatched region) is minimized, the path calculation unit 16 determines the path γ specified by graph G03. On the other hand, the area enclosed by graph G04 and the horizontal axis in the interval where the parameter t is 0 to T is clearly larger than the area enclosed by graph G03 and the horizontal axis, and the path specified by graph G04 is not the path γ determined by the path calculation unit 16.

[0067] (4-3) Path determination considering sparsity The path calculation unit 16 may set constraints (soft constraints) in the minimization problem of the action functional I to output a sparse path with respect to the change in the feature quantities from the input. In this case, the path calculation unit 16 sets constraints (soft constraints) in the minimization problem shown in equation (1) regarding at least one of the change in the feature quantities or the gradient of the path γ. This makes it possible for the path calculation unit 16 to determine a sparse path γ.

[0068] First, let's explain an example of setting constraints (soft constraints) regarding changes in feature quantities. The path calculation unit 16 adds a regularization term of the L1 norm related to changes in feature quantities to the Lagrangian function L shown in equation (3), as shown in equation (9) below.

[0069]

number

[0070] The regularization term of the L1 norm shown in equation (9) corresponds to a constraint on the change in feature quantities. "λ1" is a parameter that adjusts sparsity. By setting the Lagrangian function in this way, the path calculation unit 16 can find a path that is sparse with respect to the change in feature quantities from the input, while using a path length based on the Riemannian metric.

[0071] In recourse analysis, being spars means that the number of features being changed is small, or the magnitude of the changes in the features is small. For example, when deciding on a modification to reduce the probability of developing a specific disease in a health checkup, changing only two features, weight and cholesterol level, is more spars and therefore more explainable than changing five features, weight, white blood cell count, hemoglobin count, cholesterol level, and uric acid level.

[0072] Note that in equation (9), "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (10) holds true.

[0073]

number

[0074] Furthermore, even when the path calculation unit 16 determines the path based on the energy of the Riemann metric, it may also include a regularization term for the L1 norm related to the change in features. In this case, the path calculation unit 16 solves the minimization problem of the action functional I in equation (1) using the following equation (11), which is obtained by adding a regularization term for the L1 norm related to the change in features to the Lagrangian function L based on equation (6).

[0075]

number

[0076] Also, in equation (11), "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (12) holds true.

[0077]

number

[0078] Next, we will explain an example of setting constraints (soft constraints) on the gradient of path γ. The path calculation unit 16 further adds a regularization term of the L1 norm with respect to the gradient of path γ to the Lagrangian function L shown in equation (3), as shown in equation (13) below.

[0079]

number

[0080] The regularization term of the L1 norm shown in equation (13) corresponds to the constraint on the gradient of path γ. "λ2" is a parameter that adjusts sparsity. By setting the Lagrangian function in this way, the path calculation unit 16 can find a path γ that is sparse with respect to the gradient of path γ, while using the length of the path as the length based on the Riemannian metric.

[0081] Here, when the integral of the action functional I in equation (1) is approximated by a point cloud using the point selection method, the path calculation unit 16 solves equation (14) below as the minimization problem of the action functional I when determining the path based on the energy of the Riemann metric. Note that the parameter λ2 is set to 1 here.

[0082]

number

[0083] In equation (13), "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (15) holds true.

[0084]

number

[0085] Furthermore, even when the path calculation unit 16 determines the path based on the energy of the Riemann metric, it may also include a regularization term for the L1 norm with respect to the gradient of path γ. In this case, the path calculation unit 16 solves the problem in equation (1) using the following equation (16), which is obtained by adding a regularization term for the L1 norm with respect to the gradient of path γ to equation (6).

[0086]

number

[0087] In this case as well, when the path calculation unit 16 solves the minimization problem of the action functional I using energy based on the Riemann metric, it becomes possible to find a path γ that is sparse with respect to the gradient and reduce the computational load.

[0088] Note that in equation (16), "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (17) holds true.

[0089]

number

[0090] Next, we will explain an example in which both constraints on the change in feature quantities (soft constraints) and constraints on the gradient of path γ (soft constraints) are imposed.

[0091] The path calculation unit 16 adds both a regularization term of the L1 norm related to the change in feature quantities and a regularization term of the L1 norm related to the gradient of the path γ to the Lagrangian function L shown in equation (3), as shown in equation (18) below.

[0092]

number

[0093] The regularization term of the L1 norm including the parameter λ1 shown in equation (18) corresponds to a constraint on the change in feature quantities, and the regularization term of the L1 norm including the parameter λ2 corresponds to a constraint on the gradient of the path γ. By setting the Lagrangian function in this way, the path calculation unit 16 can find a sparse path with respect to both the change in feature quantities and the gradient of the path γ when solving the minimization problem of the action functional I using length based on the Riemann metric.

[0094] Note that in equation (18), "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (19) holds true.

[0095]

number

[0096] Furthermore, even when the path calculation unit 16 determines the path based on the energy of the Riemann metric, it may also provide both a regularization term for the L1 norm related to the change in feature quantities and a regularization term for the L1 norm related to the gradient of the path γ. In this case, the path calculation unit 16 solves the problem in equation (1) using the following equation (20), which is obtained by adding these L1 norm regularization terms to equation (6).

[0097]

number

[0098] In this case as well, when the path calculation unit 16 solves the problem of minimizing the action functional I using energy based on the Riemann metric, it becomes possible to find a sparse path γ with respect to both the change in feature quantities and the gradient of the path γ, and the computational load can be reduced.

[0099] Note that in equation (20), "g x (v,v)=f(x) 2 ||v||2 2 If we assume that, then the following equation (21) holds true.

[0100]

number

[0101] (5) Display example This section describes the information regarding the route γ that the display control unit 17 displays based on the route calculation result of the route calculation unit 16.

[0102] Figure 10 is an example of a graph representing information about the path from the input example to the modified example. Figure 10 shows an example of calculating the shortest path γ related to the Riemannian metric based on equation (3) or equation (6), and the function f corresponding to the prediction model outputs the percentage of probability. Furthermore, the function f makes predictions with d types of features as arguments, and the features at the time of the input example are "(x1,...,x d ) and the features at the time of the corrected example are "(y1,...,y d )"

[0103] In each graph shown in Figure 10, "T=1" is used, and the parameter t is varied from 0 to 1 on the horizontal axis. The parameter t corresponds to the length of the path. Then, the graphs G1 show the prediction results, with the "model prediction result" which shows the value of the function f on the vertical axis, and feature graphs G2 and G3 show "feature 1" and "feature 2", which represent the two types of features included in the d types of features input to the function f, on the vertical axis.

[0104] Here, the path calculation unit 16 determines the path γ by solving the minimization problem shown in equation (1) based on equation (3) or equation (6). The path γ determined by the path calculation unit 16 represents the values ​​of d types of features, including "feature 1" and "feature 2d," for 0 ≤ t ≤ 1, and corresponds to the information represented by feature graphs G2 and G3. Furthermore, the path calculation unit 16 or the display control unit 17 calculates the value of function f for 0 ≤ t ≤ 1 as the "model prediction result" based on the values ​​of d types of features for 0 ≤ t ≤ 1. The value of function f for 0 ≤ t ≤ 1 calculated in this case corresponds to the information represented by the prediction result graph G1.

[0105] Figure 11 shows a first display example representing the calculation result of path γ displayed on the display device 3 by the display control unit 17. When the path calculation unit 16 calculates the shortest path γ related to the Riemann metric, the display control unit 17 generates display information based on the calculation result of path γ and supplies the generated display information to the display device 3, thereby displaying the display screen shown in Figure 11 on the display device 3. In this case, function f is assumed to output the probability of bankruptcy, and the d types of features are assumed to include at least three types of features: "annual income," "debt," and "health status."

[0106] In this case, the display control unit 17 obtains the values ​​of "annual income," "debt," and "health status" from the route calculation unit 16 as the calculation results of the route γ when the parameter t is between 0 and 1, and displays them as graphs. Here, the display control unit 17 displays graphs for "annual income," "debt," and "health status" that represent the transition from the values ​​in the input example to the values ​​in the modified example.

[0107] The display control unit 17 may, instead of displaying graphs for all d types of features, display graphs for some of the d types of features (i.e., fewer than d types). For example, the display control unit 17 may display graphs for a type of feature specified by the user via the input device 2, or it may display graphs for a predetermined type of feature.

[0108] Furthermore, the display control unit 17 calculates the probability of ruin when the parameter t is between 0 and 1, based on the function f which is a prediction model for the probability of ruin and the features when the parameter t is between 0 and 1, and displays a prediction result graph representing the probability of ruin when the parameter t is between 0 and 1 on the display screen.

[0109] As shown in Figure 11, the display control unit 17 displays information representing the transition from the input example to the corrected example for each feature and prediction model as information related to the determined path γ. This allows the display control unit 17 to present to the user the transition of each feature and prediction model's prediction results up to the target corrected example.

[0110] Figure 12 shows a second display example representing the calculation result of path γ displayed on the display device 3 by the display control unit 17. In this example, the path calculation unit 16 calculates the shortest path γ related to the Riemannian metric by setting constraints on the change of feature quantities according to equation (9) or equation (11), and the display control unit 17 generates display information based on the calculation result of path γ. The display control unit 17 then supplies the generated display information to the display device 3, causing the display screen shown in Figure 12 to be displayed on the display device 3.

[0111] In this case, by setting constraints on the changes in feature quantities according to equation (9) or equation (11), the shortest path γ related to the Riemannian metric is calculated, and the fluctuations in the values ​​of "annual income," "debt," and "health status" are smaller compared to the example in Figure 11. In this way, by setting constraints on the changes in feature quantities according to equation (9) or equation (11), a sparse path can be calculated with respect to the changes in feature quantities from the input, and this path can be presented to the user.

[0112] Figure 13 shows a third display example representing the calculation result of path γ displayed on the display device 3 by the display control unit 17. In this example, the path calculation unit 16 calculates the shortest path γ related to the Riemannian metric by setting constraints on the gradient of path γ according to equation (13) or equation (16), and the display control unit 17 generates display information based on the calculation result of path γ. The display control unit 17 then supplies the generated display information to the display device 3, causing the display screen shown in Figure 13 to be displayed on the display device 3.

[0113] In this case, by setting constraints on the gradient of path γ according to equation (13) or equation (16), the shortest path γ with respect to the Riemannian metric is calculated, and the number of features that vary depending on the parameter t is smaller compared to the example in Figure 11. Thus, by setting constraints on the gradient of path γ according to equation (13) or equation (16), a path γ with respect to the gradient can be calculated and presented to the user.

[0114] <Second Embodiment> Figure 14 shows the configuration of the information processing system 100A. The information processing system 100A mainly consists of an information processing device 1A and a terminal device 8. The information processing device 1A and the terminal device 8 communicate data via the network 7.

[0115] The information processing device 1A is one or more devices that function as a server (including a cloud server), and in the first embodiment, it performs processing related to the information processing performed by the information processing device 1. In this case, the information processing device 1A receives input information that the information processing device 1 receives from the input device 2 via the network 7 from the terminal device 8. In addition, the information processing device 1A transmits display information that the information processing device 1 has transmitted to the display device 3 via the network 7 to the terminal device 8.

[0116] Terminal device 8 is a terminal having input, display, and communication functions, and functions as the input device 2 and display device 3 in the first embodiment. Terminal device 8 may be, for example, a personal computer, a tablet terminal, or a PDA (Personal Digital Assistant). Terminal device 8 transmits input information generated based on received user input to information processing device 1A via the network 7. Also, when terminal device 8 receives display information from information processing device 1A, it displays information based on said display information.

[0117] The information processing device 1A according to the second embodiment can suitably perform the input processing and output processing performed by the information processing device 1 in the first embodiment, targeting the user of the terminal device 8.

[0118] <Third Embodiment> Figure 15 is a functional block diagram of the information processing device 1X. The information processing device 1X mainly comprises a route calculation means 16X and a display control means 17X. The information processing device 1X can be the information processing device 1 in the first embodiment or the information processing device 1A in the second embodiment.

[0119] The path calculation means 16X determines a path from first information representing the feature quantities of the first event to second information representing the feature quantities of a second event different from the first event, such that the evaluation value, which is an evaluation of the complexity of the path from first information to second information, decreases. Here, a lower evaluation value is considered preferable. The path calculation means 16X can be the path calculation unit 16 in the first and second embodiments.

[0120] The display control means 17X displays information regarding the determined route on the display device. The display control means 17X can be the display control unit 17 in the first and second embodiments.

[0121] Figure 16 is an example of a flowchart executed by the information processing device 1X. The path calculation means 16X determines a path from first information representing the feature quantities of the first event and second information representing the feature quantities of a second event different from the first event, such that the evaluation value, which evaluates the complexity of the path from the first information to the second information, decreases (step S21). Next, the display control means 17X displays information about the determined path on the display device (step S22).

[0122] The information processing device 1X according to the third embodiment can accurately determine the path from the first information to the second information and display information related to the determined path.

[0123] <Fourth Embodiment> Figure 17 is a functional block diagram of the information processing device 1Y. The information processing device 1Y mainly comprises a route calculation means 16Y and a display control means 17Y. The information processing device 1Y can be the information processing device 1 in the first embodiment or the information processing device 1A in the second embodiment.

[0124] The path calculation means 16Y determines a path from first information representing the feature quantities of a first event to second information representing the feature quantities of a second event different from the first event, such that the evaluated value, which is an evaluation of the energy of the path from the first information to the second information, decreases. The path calculation means 16Y can be the path calculation unit 16 in the first or second embodiment.

[0125] The path calculation means 16Y determines a path from first information representing the feature quantities of a first event to second information representing the feature quantities of a second event different from the first event, such that the evaluated value, which is an evaluation of the energy of the path from the first information to the second information, decreases. The path calculation means 16Y can be the path calculation unit 16 in the first or second embodiment.

[0126] The display control means 17Y displays information regarding the determined route on the display device. The display control means 17Y can be the display control unit 17 in the first and second embodiments.

[0127] Figure 18 is an example of a flowchart executed by the information processing device 1Y. The path calculation means 16Y determines a path from the first information representing the feature quantity of the first event and the second information representing the feature quantity of a second event different from the first event, such that the evaluated value, which is an evaluation of the energy of the path from the first information to the second information, decreases (step S31). The display control means 17Y displays information about the determined path on the display device (step S32).

[0128] The information processing device 1X according to the fourth embodiment can accurately determine the path from the first information to the second information and display information related to the determined path.

[0129] Furthermore, some or all of the above embodiments (including modifications, hereinafter the same) may also be described as follows, but are not limited to these. Moreover, not limited to the devices, methods, and storage media described in the following appendices, some or all of the configurations described in the appendices may also be subordinate to various hardware, software, various recording means (including storage media) for recording software, or systems, without departing from the above embodiments.

[0130] [Note 1] A path calculation means that determines a path such that an evaluation value, which is an evaluation of the complexity of the path from the first information to the second information, decreases, based on first information representing the features of the first event and second information representing the features of a second event different from the first event. A display control means that displays information regarding the determined route on a display device, An information processing device having [Note 2] The aforementioned evaluation value is the area of ​​the portion between the path and the first axis in the interval on the first axis from the first event to the second event in a coordinate space having a first axis corresponding to the length of the path and a second axis corresponding to the predicted value at a point on the path. The route calculation means is an information processing device according to Appendix 1, which determines the route that minimizes the area. [Note 3] The information processing apparatus according to Appendix 1 or 2, wherein the display control means displays information on the display device indicating the change in the feature quantity based on the path. [Note 4] The information processing apparatus described in Appendix 2, wherein the display control means displays information indicating the path in the coordinate space on the display device. [Note 5] The path calculation means determines the path such that the evaluation value obtained by evaluating the path based on a Lagrangian function that defines an evaluation for determining the path by minimizing a functional decreases, according to the information processing apparatus according to any one of the appendices 1 to 4. [Note 6] The information processing device described in Appendix 5, wherein the path calculation means calculates the path based on the evaluation defined based on the length of the Riemann metric in the feature space. [Note 7] The information processing apparatus according to any one of the appendices 1 to 6, wherein the path calculation means sets constraints on the changes in the feature quantities along the path when calculating the path. [Note 8] The information processing apparatus according to any one of the appendices 1 to 7, wherein the route calculation means sets constraints on the gradient of the route when calculating the route. [Note 9] The information processing apparatus according to any one of the appendices 1 to 8, wherein the path calculation means provides a first constraint condition regarding the change of the feature quantity in the path and a second constraint condition regarding the gradient of the path when calculating the path. [Note 10] The path calculation means determines the path in the feature space when an input is received that defines the feature space which is the space of the feature quantities, as described in any one of the appendices 1 to 9. [Note 11] Based on first information representing the features of the first event and second information representing the features of a second event different from the first event, the path is determined such that the complexity of the path from the first information to the second information is evaluated, and the evaluation value decreases. The information regarding the determined route is displayed on the display device. Display method. [Note 12] Based on first information representing the features of the first event and second information representing the features of a second event different from the first event, the path is determined such that the complexity of the path from the first information to the second information is evaluated, and the evaluation value decreases. A program that causes a computer to perform the process of displaying information about the determined route on a display device. [Note 13] A storage medium containing the program described in Appendix 12.

[0131] In each of the embodiments described above, the program can be stored using various types of non-transitory computer-readable medium and supplied to a computer, such as a processor. Non-transitory computer-readable mediums include various types of tangible storage mediums. Examples of non-transitory computer-readable mediums include magnetic storage mediums (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical storage mediums (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). The program may also be supplied to the computer by various types of transient computer-readable mediums. Examples of transient computer-readable mediums include electrical signals, optical signals, and electromagnetic waves. Transitory computer-readable mediums can supply the program to the computer via wired communication channels such as electric wires and optical fibers, or via wireless communication channels.

[0132] Although the present invention has been described above with reference to embodiments, the present invention is not limited to the above embodiments. Various modifications to the structure and details of the present invention can be made as understood by those skilled in the art within the scope of the present invention. That is, the present invention includes the full disclosure, including the claims, and of course, various modifications and alterations that those skilled in the art could make in accordance with the technical idea. Furthermore, the above-mentioned patent and non-patent disclosures are incorporated herein by reference. [Explanation of Symbols]

[0133] 1, 1A, 1X, 1Y Information Processing Device 2 Input devices 3 Display device 8 Terminal devices 11 processors 12 memory 13 Interfaces 100, 100A Information Processing System

Claims

1. A path calculation means that determines a path such that an evaluation value, which is an evaluation of the complexity of the path from the first information to the second information, decreases, based on first information representing the features of the first event and second information representing the features of a second event different from the first event. A display control means that displays information regarding the determined route on a display device, An information processing device having

2. The aforementioned evaluation value is the area of ​​the portion between the path and the first axis in the interval on the first axis from the first event to the second event in a coordinate space having a first axis corresponding to the length of the path and a second axis corresponding to the predicted value at a point on the path. The information processing apparatus according to claim 1, wherein the path calculation means determines the path that minimizes the area.

3. The information processing apparatus according to claim 1, wherein the display control means displays information on the display device indicating the change in the feature quantity based on the path.

4. The information processing apparatus according to claim 2, wherein the display control means displays information indicating the path in the coordinate space on the display device.

5. The information processing apparatus according to claim 1, wherein the path calculation means determines the path such that the evaluation value obtained by evaluating the path based on a Lagrangian function that defines an evaluation for determining the path by minimizing a functional decreases.

6. The information processing apparatus according to claim 5, wherein the path calculation means calculates the path based on the evaluation defined based on the length of the Riemann metric in the feature space.

7. The information processing apparatus according to claim 1, wherein the path calculation means provides constraints regarding the changes in the feature quantities along the path when calculating the path.

8. The information processing apparatus according to claim 1, wherein the route calculation means provides constraints regarding the gradient of the route when calculating the route.

9. Computers Based on first information representing the features of the first event and second information representing the features of a second event different from the first event, the path is determined such that the complexity of the path from the first information to the second information is evaluated, and the evaluation value decreases. The information regarding the determined route is displayed on the display device. Display method.

10. Based on first information representing the features of the first event and second information representing the features of a second event different from the first event, the path is determined such that the complexity of the path from the first information to the second information is evaluated, and the evaluation value decreases. A program that causes a computer to perform the process of displaying information about the determined route on a display device.

Images