Search support device and search support method

Abstract

To provide a search support device and search support method for performing a search related to a parameter representing an influence degree of a feature quantity at high speed and with high accuracy.SOLUTION: A search support device, in a learning phase s100, calculates at least one or more pieces of SHAP data indicating an influence degree of each feature quantity in a learned model on output data output from the learned model, and generates compressed SHAP data, which is data obtained by compressing the SHAP data, for each SHAP data and stores the compressed SHAP data. The search support device, in the learning phase s100, calculates verification target SHAP data which is SHAP data for the output data output from the learned model by inputting input data to the learned model, calculates a similarity between each of pieces of calculated compressed SHAP data and the calculated verification target SHAP data, and identifies the compressed SHAP data in which the similarity with the verification target SHAP data satisfies a predetermined condition.SELECTED DRAWING: Figure 9

Description

Technical Field

[0001] The present invention relates to a search support device and a search support method.

Background Art

[0002] In the field of machine learning, the use of explainable AI (XAI: eXxplanable Artificial Intelligence) is progressing. XAI is an AI that not only outputs data by an AI model (trained model), but also enables humans to understand the process of the AI leading to the output. Output, but also an AI that enables humans to understand the process of the AI leading to the output.

[0003] In XAI, a Shapley value, which indicates the degree of influence of each feature amount on the output data, is used. As a method of utilizing the Shapley value, for example, for certain data output by a user using AI, the degree of influence of the feature amount derived from the Shapley value (hereinafter referred to as the SHAP value (SHAP: SHapley Additive exPlanations)) Similar to the SHAP value of other past data is searched, and the user interprets the output data of himself / herself. By searching for similar past SHAP values, the user interprets the output data of himself / herself.

[0004] From such a background, Patent Document 1 discloses a method for similar search of SHAP values. Further, Patent Document 2 discloses, as related art, a method for accelerating the pattern search of feature vectors.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

[0006] However, SHAP values, which represent the influence of features in AI, inherently possess a variety of data characteristics due to their nature as AI-based data that can be used for various purposes. Furthermore, the sheer volume of this data can be enormous, making it difficult to achieve both speed and accuracy in searching for SHAP values.

[0007] This invention has been made in view of the current situation, and its purpose is to provide a search support device and a search support method that can perform searches related to parameters representing the degree of influence of feature quantities at high speed and with high accuracy. [Means for solving the problem]

[0008] One aspect of the present invention for solving the above problems is a process comprising a processor and memory, wherein the processor calculates at least one SHAP data set, which is data indicating the degree of influence of each feature in the trained model on the output data output from the trained model, Based on the history of each SHAP data calculated above, Among the features related to the aforementioned SHAP data Identify the features to be deleted. This search support device performs the following processes: generating compressed SHAP data; generating the compressed SHAP data for each SHAP data and storing it in the memory; inputting input data to the trained model to calculate the SHAP data to be verified, which is the SHAP data for the output data output from the trained model; and calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be verified, and identifying the compressed SHAP data whose similarity to the SHAP data to be verified satisfies predetermined conditions. Another aspect of the present invention for solving the above problems is a processor and memory, comprising A search support deviceThe processor performs a process of calculating at least one SHAP data, which is data indicating the degree of influence of each feature in the trained model on the output data output from the trained model. Based on the information regarding the hardware provided by the search support device, Among the features related to the aforementioned SHAP data Identify the features to be deleted. This search support device performs the following processes: generating compressed SHAP data; generating the compressed SHAP data for each SHAP data and storing it in the memory; inputting input data to the trained model to calculate the SHAP data to be verified, which is the SHAP data for the output data output from the trained model; and calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be verified, and identifying the compressed SHAP data whose similarity to the SHAP data to be verified satisfies predetermined conditions. Furthermore, another aspect of the present invention for solving the above problems is a processor and memory, wherein the processor performs a process of calculating at least one SHAP data, which is data indicating the degree of influence of each feature in the trained model on the output data output from the trained model, From the user, Among the features related to the aforementioned SHAP data It accepts the specification of features to be excluded from compression, and by deleting the other features while retaining the data of the specified features to be excluded from compression. This search support device performs the following steps: generating compressed SHAP data; generating the compressed SHAP data for each SHAP data and storing it in the memory; inputting input data to the trained model to calculate the SHAP data to be validated, which is the SHAP data for the output data output from the trained model; and calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be validated, and identifying the compressed SHAP data whose similarity to the SHAP data to be validated satisfies predetermined conditions. [Effects of the Invention]

[0009] According to the present invention, it is possible to perform searches related to parameters that represent the degree of influence of feature quantities quickly and accurately. Configurations, effects, etc. other than those described above will be clarified by the description of the following embodiments.

Brief Description of Drawings

[0010] [Figure 1] It is a diagram showing an example of the hardware configuration included in the search support device according to this embodiment and an example of the functions of the search support device 1. [Figure 2] It is a diagram showing an example of the SHAP matrix of this embodiment. [Figure 3] It is a diagram showing an example of the compressed SHAP matrix. [Figure 4] It is a diagram showing an example of the adoption mandatory items. [Figure 5] It is a diagram showing an example of the SHAP global statistics. [Figure 6] It is a diagram showing an example of the aggregated information. [Figure 7] It is a diagram showing an example of the hardware information. [Figure 8] It is a diagram showing an example of the system constraints. [Figure 9] It is a diagram for explaining the outline of the processing performed by the search support device. [Figure 10] It is a flowchart for explaining the outline of the learning phase. [Figure 11] It is a flowchart for explaining the details of the threshold determination process. [Figure 12] It is a diagram showing an example of the corrected SHAP matrix generated by the threshold determination process. [Figure 13] It is a flowchart for explaining the details of the compressed matrix creation process. [Figure 14] It is a flowchart for explaining an example of the inference phase. [Figure 15] It is a flowchart for explaining the details of the similarity calculation process. [Figure 16] It is a diagram for explaining an example of the processing in the similarity calculation process. [Figure 17] It is a diagram showing an example of the SHAP importance related information input screen. [Figure 18] It is a diagram showing an example of the compressed SHAP matrix confirmation screen. [Figure 19] This figure shows an example of a screen displaying similar records. [Modes for carrying out the invention]

[0011] The search support device and search support method according to this embodiment will be described with reference to the drawings.

[0012] Figure 1 shows an example of the hardware configuration and functions of the search support device 1 according to this embodiment.

[0013] Search support device 1 includes a CPU (Central Processing Unit) and a DSP (Digital Signal Processor). The system comprises a processor 11 such as a GPU (Graphics Processing Unit) or FPGA (Field-Programmable Gate Array), memory 12 such as ROM (Read Only Memory) or RAM (Random Access Memory), storage 13 such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), communication device 14 consisting of a NIC (Network Interface Card), wireless communication module, USB (Universal Serial Interface) module, or serial communication module, input device 15 consisting of a mouse or keyboard, and output device 16 consisting of a liquid crystal display or organic EL (Electro-Luminescence) display.

[0014] The search support device 1 comprises the following functional units: an AI model generation unit 101, a SHAP matrix calculation unit 103, a SHAP importance estimation unit 105, a compressed SHAP matrix generation unit 107, an AI model inference unit 109, a compressed SHAP matrix similarity calculation unit 111, a similar record extraction unit 113, and an input / output unit 115.

[0015] The AI ​​model generation unit 101 performs machine learning using training data to generate a trained model. The AI ​​model generation unit 101 creates multiple types of trained models, each with the same input data type but different output data types. In this embodiment, the trained models may be referred to as AI (Artificial Intelligence).

[0016] The trained model in this embodiment takes attribute information about a patient's health (e.g., age, gender, test data, etc.) as input data and outputs (predicts) the patient's future health status (e.g., risk of disease, risk of requiring long-term care) as a predicted value. Each trained model will output the patient's health status as a predicted value at a different future point in time. Note that the input and output data of such trained models are merely examples and are not intended to limit the scope of the present invention.

[0017] The SHAP matrix calculation unit 103 calculates the degree of influence of each feature that affected the output value when each trained model outputs an output value, based on the SHAP (SHapley Additive exPlanations) algorithm. This degree of influence is based on the SHapley Value. This set of influence values ​​(hereinafter referred to as SHAP values) is stored as the SHAP matrix 300 (hereinafter also referred to as the SHAP matrix), which will be described later.

[0018] The SHAP importance estimation unit 105 estimates the importance of each SHAP value in the SHAP matrix.

[0019] The compressed SHAP matrix generation unit 107 creates compressed SHAP data (compressed SHAP matrix 400, described later) which is data obtained by compressing the SHAP matrix based on the estimation results from the SHAP importance estimation unit 105.

[0020] The AI ​​model inference unit 109 inputs the input data specified by the user into each trained model and outputs a predicted value for each model. The output values ​​are stored in the inference data 600.

[0021] The compressed SHAP matrix similarity calculation unit 111 calculates the similarity between each previously created compressed SHAP matrix and the compressed SHAP matrix to the output value output by the AI ​​model inference unit 109.

[0022] The similarity record extraction unit 113 extracts information such as feature quantities that correspond to the previously created compressed SHAP matrix with the highest similarity.

[0023] The input / output unit 115 outputs various information to the output device 1 6 The system displays the information on its screen, or accepts user input via the input device 15. The input / output unit 115 displays, for example, the SHAP importance-related information input screen 1100, the compressed SHAP matrix confirmation screen 1200, and the similar record display screen 1300.

[0024] The SHAP importance-related information input screen 1100 is a screen that accepts input from the user for parameters to create a compressed SHAP matrix. The compressed SHAP matrix confirmation screen 1200 is a screen that displays the SHAP matrix and the compressed SHAP matrix that has been created. The similar record display screen 1300 is a screen that displays information such as the features extracted by the similar record extraction unit 113.

[0025] Next, the search support device 1 stores the following data: training data 200, SHAP matrix 300, compressed SHAP matrix 400, required items to be adopted 500, inference data 600, lineage 700, SHAP global statistics 800, hardware information 900, and system constraints 1000.

[0026] The training data 200 is the input data used to generate the trained model. The training data 200 contains one or more features (data items), their values, and label data (output data).

[0027] The SHAP matrix 300 is data that stores multiple SHAP values. The SHAP matrix consists of rows representing "cases" set for each run of the trained model (output of output values), and columns representing the feature values ​​related to the trained model in that case.

[0028] (SHAP matrix) Figure 2 shows an example of the SHAP matrix 300 of this embodiment. This SHAP matrix 300 has rows 301 representing each case and columns 302 representing the values ​​of each feature in each case. The value of each feature indicates the degree of influence on the output data output from the trained model. The value of the feature is, for example, any value between 0 and 1.

[0029] Next, the compressed SHAP matrix 400 shown in Figure 1 is compressed data obtained by deleting some of the feature information from the SHAP matrix.

[0030] (Compressed SHAP matrix) Figure 3 shows an example of a compressed SHAP matrix. This compressed SHAP matrix 400 consists of one or more rows of data, and each row consists of three data items: the case ID 401, the feature ID 402 which is the identifier of one of the feature items in that case, and the feature value 403 for that item.

[0031] Next, the 500 mandatory features shown in Figure 1 are data that stores the mandatory features, which are features that must always be present in the compressed SHAP matrix. The 500 mandatory features are set for each user project.

[0032] (Required item for recruitment) Figure 4 shows an example of the required items 500 for adoption. These required items 500 include a project ID 501 in which the ID of a project set by the user to achieve a predetermined business objective using a trained model is set; a field 502 in which the business field to which the project belongs is set; a customer 503 in which information about the target audience of the project (e.g., the name of the customer) is set; a KPI 504 in which an evaluation indicator (in this embodiment, a KPI) representing the objective to be achieved is set (corresponding to the type of output value of the trained model); and a request data source 505 in which the required items for adoption in the trained model related to the project are set. The data content of the required items 500 is set in advance by the user, for example.

[0033] Next, the inference data 600 shown in Figure 1 is the data of each output value (predicted value) obtained by inputting the input data (user-specified input data and training data) into the trained model.

[0034] Lineage 700 stores information (for example, threshold information, which will be described later) related to cases where the prediction was valid among the cases in which an output value (predicted value) was obtained by inputting input data into each trained model.

[0035] SHAP Global Statistics 800 is a dataset that accumulates the execution results (prediction results) of trained models.

[0036] (SHAP Global Statistics) Figure 5 shows an example of SHAP Global Statistics 800. This SHAP Global Statistics 800 has the following data items: Test ID 801, which contains the ID of the prediction made; Model ID 802, which contains the ID of the trained model used for the prediction; Feature ID 803, which contains information about the subject (patient, etc.) of the prediction; KPI 804, which contains evaluation indicators related to the prediction (corresponding to the type of output value of the trained model); and Test Results 805, which contains data related to the prediction. Test Results 805 contains the output value of the trained model for the subject and one or more SHAP values ​​corresponding to that output value.

[0037] In this embodiment, aggregated information 850, which is obtained by aggregating the contents of SHAP global statistics 800, is used.

[0038] Figure 6 shows an example of aggregated information 850. Aggregated information 850 has the following data items: KPI 851 where the evaluation index (type of output value) is set, feature name 852 where the name of the feature related to that evaluation index is set, minimum value 853 where the minimum value of that feature is set, average value 854 where the average value of that feature is set, and maximum value 855 where the maximum value of that feature is set.

[0039] Next, the hardware information 900 shown in Figure 1 is data relating to the hardware state of the search support device 1. The system constraints 1000 are data relating to the hardware constraints of the search support device 1 when creating the compressed SHAP matrix, which will be described later. The system constraints 1000 are created based on the hardware information 900.

[0040] (Hardware information) Figure 7 shows an example of hardware information 900. Hardware information 900 has the following data items: time 901, which sets the time (timing) when the data was acquired; CPU usage rate 902, which sets the usage rate of the CPU 11 of the search support device 1 at that time; memory usage rate 903, which sets the available amount of memory 12 of the search support device 1 at that time; and storage usage rate 904, which sets the available amount of storage 13 of the search support device 1 at that time. Hardware information 900 is updated as needed by a predetermined hardware monitoring program.

[0041] (System constraints) Figure 8 shows an example of a system constraint 1000. This system constraint 1000 has the following data items: number of data 1001 which sets the conditions of the SHAP matrix that will be the basis of the compressed SHAP matrix (in this embodiment, the length of the columns of the SHAP matrix); required CPU usage 1002 which sets the CPU usage required to create the compressed SHAP matrix from the SHAP matrix with those conditions; required memory amount 1003 which sets the amount of memory required to create the compressed SHAP matrix from the SHAP matrix with those conditions; required storage capacity 1004 which sets the storage capacity required to create the compressed SHAP matrix from the SHAP matrix with those conditions; required time 1005 which sets the estimated time required to create the compressed SHAP matrix from the SHAP matrix with those conditions; and compression ratio 1006 of the compressed SHAP matrix achieved under those conditions (compression ratio relative to the original SHAP matrix).

[0042] In this embodiment, the search support device 1 creates system constraints 1000 based on hardware information 900. For example, the search support device 1 uses a predetermined algorithm (regression analysis, machine learning, etc.) to calculate the relationship between the length of the compressed SHAP matrix, the hardware configuration, the creation time, and the compression ratio, based on each compressed SHAP matrix created in the past, the hardware information 900 at the time of its creation, the time required to create the compressed SHAP matrix, and the compression ratio of the compressed SHAP matrix. The calculated relationship is then set for each record in the system constraints 1000. In addition, the user may perform a SHAP matrix compression test using the search support device 1 in advance, and The result may be entered into system constraint 1000.

[0043] In this embodiment, the number of data points 901 is assumed to be the length of the columns of the SHAP matrix, but other conditions such as the row length may also be set. Furthermore, the method for creating the system constraint 1000 and the data items described herein are examples, and the present invention does not particularly limit the method for creating the constraint or the data items.

[0044] The functions of each functional unit of the search support device 1 described above are realized by the processor 11 reading and executing a program stored in the memory 12 or storage 13. The program can also be distributed, for example, by recording it on a recording medium. Furthermore, the search support device 1 may be implemented in whole or in part using virtual information processing resources provided using virtualization technology, process space isolation technology, etc., such as a virtual server provided by a cloud system. Also, all or part of the functions provided by the search support device 1 may be implemented, for example, by services provided by a cloud system via an API (Application Programming Interface), etc. Next, we will explain the processes performed by the search support device 1.

[0045] Figure 9 is a diagram illustrating the overview of the processing performed by the search support device 1.

[0046] First, the search support device 1 creates a trained model using the training data 200, and also creates a SHAP matrix and a compressed SHAP matrix corresponding to the training data 200 (corresponding to the data output by the trained model) (training phase s100). In this case, the search support device 1 outputs multiple training models, each with different types of data. Done Create a model.

[0047] Meanwhile, the search support device 1 obtains an output value by inputting the input data that the user is currently trying to infer into a pre-trained model selected by the user from among multiple pre-trained models created in the learning phase s100 (hereinafter referred to as "this pre-trained model"). The search support device 1 creates a SHAP matrix and a compressed SHAP matrix corresponding to that output value. The search support device 1 searches for a compressed SHAP matrix created in the learning phase s100 that is similar to the compressed SHAP matrix it created, and displays the search result on the screen (this concludes the inference phase s200). The learning phase s100 and the inference phase s200 will be described below.

[0048] <Learning Phase> Figure 10 is a flowchart illustrating the overview of the learning phase s100. First, the AI ​​model generation unit 101 creates a trained model (AI) (s110). For example, the AI ​​model generation unit 101 performs machine learning using the dataset (data of multiple items) for each case and the corresponding label data (output data) as training data, thereby creating multiple trained models that output different types of data.

[0049] This pre-trained model is created, for example, by performing machine learning using deep learning. In this embodiment, this pre-trained model is a neural network having an input layer into which a dataset is input, one or more hidden layers that extract and output features from the dataset, and an output layer that outputs predetermined output values ​​from the features. The neural network in the pre-trained model may be, for example, a CNN (Convolutional Neural Network), an SVM (Support Vector Machine), a Bayesian network, or a regression tree.

[0050] Next, the SHAP matrix calculation unit 103 creates a SHAP matrix for each feature corresponding to the output values ​​output during the process of creating the trained model created in s110 (s130). The matrix is ​​created, for example, by calculating the marginal contribution of each feature using marginalization.

[0051] Next, the SHAP importance estimation unit 105 estimates the importance of each feature in the SHAP matrix created in s130, determines a threshold to be used for data compression, and then executes a threshold determination process s150, which modifies the SHAP matrix based on this threshold. Details of the threshold determination process s150 will be described later.

[0052] The SHAP importance estimation unit 105 estimates the importance of each data item (feature) in the modified SHAP matrix created in the threshold determination process s150 by calling the compression matrix creation process s170, and then creates a compressed SHAP matrix. Details of the compression matrix creation process s170 will be described later. This concludes the learning phase s100. Next, we will explain the details of the threshold determination process s150 and the compression matrix creation process s170.

[0053] <Threshold determination process> Figure 11 is a flowchart illustrating the details of the threshold determination process s150. The SHAP importance estimation unit 105 determines thresholds for the feature values ​​that serve as the basis for compression when creating a compressed SHAP matrix based on the SHAP matrix created in s130 (s151, s153).

[0054] Specifically, the SHAP importance estimation unit 105 first calculates a provisional standard by analyzing the frequency of occurrence (density distribution) of the values ​​of each feature in each SHAP matrix created in s130 (s151).

[0055] Specifically, the SHAP importance estimation unit 105 refers to the SHAP global statistics 800 or aggregated information 850 to identify the values ​​(or ranges of values) of each feature in each record of the SHAP matrix and their frequency of occurrence (density), and sets the value of the feature with a particularly low frequency of occurrence as a provisional threshold. As a result, the SHAP importance estimation unit 105 classifies the data into a set of data where the feature values ​​are greater than the threshold and a set of data where the feature values ​​are less than the threshold, and sets the boundary between these two data sets as a provisional threshold (i.e., identifies the trough between the two peaks in terms of frequency of occurrence). For example, the SHAP importance estimation unit 105 sets the value of the feature with the lowest density as a provisional threshold.

[0056] Note that the density distribution analysis method described here is an example, and various other determination methods can be employed. Furthermore, the SHAP importance estimation unit 105 may accept input for this threshold value from the user.

[0057] Then, the SHAP importance estimation unit 105 adjusts the provisional threshold calculated in s151 based on the thresholds previously calculated for other types of trained models that were calculated in s110 (s153). For example, if the threshold for other types of trained models recorded in lineage 700 is lower than the threshold calculated in s151, the SHAP importance estimation unit 105 sets the threshold calculated in s151 to a lower value according to the degree of the discrepancy between the two.

[0058] Next, the SHAP importance estimation unit 105 determines the mandatory items to be adopted as data items (features) in the compressed SHAP matrix, which are data items (features) that will always be adopted regardless of the threshold calculated in s151 (s155).

[0059] For example, the SHAP importance estimation unit 105 accepts input from the user for mandatory items to be adopted. Alternatively, for example, the SHAP importance estimation unit 105 may automatically select mandatory items to be adopted based on the history of past mandatory item designations. 105 may obtain the required recruitment items to be adopted from 500 records of required recruitment items that have the same or similar fields, target audiences, or KPIs.

[0060] Furthermore, the SHAP importance estimation unit 105 determines the data compression method when creating the compressed SHAP matrix based on the set system constraints (s 1 57).

[0061] In this embodiment, the SHAP importance estimation unit 105 determines the data compression ratio in creating the compressed SHAP matrix, and specifically determines the proportion of items to be deleted from each feature (column compression).

[0062] For example, the SHAP importance estimation unit 105 accepts input for an upper limit on the creation time of the compressed SHAP matrix. The SHAP importance estimation unit 105 obtains the current hardware state from the hardware information 900 and refers to the system constraints 1000 to determine the current hardware constraints, the input upper limit on the creation time, and the compression ratio of the SHAP matrix corresponding to the SHAP matrix created in s130.

[0063] The method for determining the compression ratio using the system constraint 1000 described here is just one example. For example, the SHAP importance estimation unit 105 may accept a compression ratio specification from the user. Also, although the SHAP importance estimation unit 105 is assumed to perform column compression in the above example, it may also perform row-based compression.

[0064] Then, the SHAP importance estimation unit 105 determines the final threshold (s159) based on the threshold determined in s153, the mandatory items to be included determined in s155, and the compression ratio determined in s157. Specifically, the SHAP importance estimation unit 105 excludes the mandatory items to be included determined in s155 from compression, and further lowers the threshold value determined in s153 as necessary to satisfy the feature compression ratio determined in s157.

[0065] Then, the SHAP importance estimation unit 105 creates a modified SHAP matrix (s161) in which the values ​​of features in each row and column of the SHAP matrix created in s130 that are below the threshold determined in s159 are set to 0. This completes the threshold determination process s150.

[0066] Figure 12 shows an example of a modified SHAP matrix generated by the threshold determination process s150. In this modified SHAP matrix 300, the values ​​of each element of the SHAP matrix created in s130 whose value was less than the threshold are set to 0 (indicated by 303).

[0067] <Processing to create compression matrix> Figure 13 is a flowchart illustrating the details of the compression matrix creation process s170. The compressed SHAP matrix generation unit 107 obtains the modified SHAP matrix created in the threshold determination process s150 (s171).

[0068] The compressed SHAP matrix generation unit 107 selects one row of the modified SHAP matrix obtained in s171 (s173), and for each column value (feature value) of the selected row, it obtains the features whose values ​​are not zero, and the data item names of those features (s175).

[0069] The compressed SHAP matrix generation unit 107 creates a record for one row of the compressed SHAP matrix (s177). Specifically, for example, the compressed SHAP matrix generation unit 107 creates new data or adds it to an existing compressed SHAP matrix, with each record consisting of the case ID (or row number) of the row selected in s171, the data item name obtained in s173, and the value obtained in s173.

[0070] The compressed SHAP matrix generation unit 107 checks whether the currently selected row of the SHAP matrix is ​​the last row (s179). If the currently selected row of the SHAP matrix is ​​the last row (s179: Yes), the created compressed SHAP matrix is ​​stored (s181), and the compressed matrix creation process s170 ends (s183). On the other hand, if the currently selected row of the SHAP matrix is ​​not the last row (s179: No), the compressed SHAP matrix generation unit 107 repeats the process in s173 to select the next row. Next, we will explain the inference phase s200.

[0071] <Inference Phase> Figure 14 is a flowchart illustrating an example of the inference phase s200.

[0072] The inference phase s200 begins after the user has performed inference using the trained model. For example, the AI ​​model inference unit 109 receives a specification from the user for the trained model and the input data (data to be inferred) to be input to this trained model, and outputs output data (predicted values) by inputting the input data to the trained model. The inference phase s200 begins when this output is received.

[0073] First, the AI ​​model inference unit 109 obtains the predicted value (s210).

[0074] The AI ​​model inference unit 109 creates a SHAP matrix corresponding to the predicted values ​​obtained in s210, following the same algorithm as in s130 (s230).

[0075] The AI ​​model inference unit 109 creates a compressed SHAP matrix (hereinafter referred to as the SHAP data to be verified) for the SHAP matrix created in s230 by calling the compression matrix creation process s170 with respect to the modified SHAP matrix created in s230 (s250).

[0076] The AI ​​model inference unit 109 executes a similarity calculation process s270, which calculates the similarity between the compressed SHAP matrix created in s250 and the compressed SHAP matrices of each case created in the past. Details of the similarity calculation process s270 will be described later.

[0077] The AI ​​model inference unit 109 identifies past compressed SHAP matrices that yielded high similarity scores among those calculated in the similarity calculation process s270. The AI ​​model inference unit 109 then displays information on the screen about the case corresponding to the identified compressed SHAP matrice (for example, information about the input data input to the trained model). Here, we will explain the details of the similarity calculation process s270.

[0078] <Similarity Calculation Process> Figure 15 is a flowchart illustrating the details of the similarity calculation process s270. The similar record extraction unit 113 obtains the compressed SHAP matrix created in s250 (s271).

[0079] The compressed SHAP matrix similarity calculation unit 111 retrieves one record data row from each row of a previously created compressed SHAP matrix that is the same case as the case related to the compressed SHAP matrix obtained in s271 (hereinafter referred to as "this case," for example, data from the same project) (s272).

[0080] The compressed SHAP matrix similarity calculation unit 111 compares the values ​​of each column (each feature) of the compressed SHAP matrix obtained in s271 with the values ​​of each column (each feature) of the compressed SHAP matrix obtained in s273 (s273).

[0081] For each feature, if its value is set in both compressed SHAP matrices (i.e., if any non-zero value of the feature is set in both compressed SHAP matrices) (s273:Yes), the compressed SHAP matrix similarity calculation unit 111 performs the process in s275 for that feature. On the other hand, if it is detected that the value of the feature (any non-zero value) is not set in either of the compressed SHAP matrices (s273:No), the compressed SHAP matrix similarity calculation unit 111 (temporarily) creates a column for that feature in the compressed SHAP matrix that does not have a value set for that feature, and sets a baseline value (in this case, 0) for the value of that feature (s274). After that, the process in s275 is performed.

[0082] In s275, the compressed SHAP matrix similarity calculation unit 111 calculates the similarity between the compressed SHAP matrix obtained in s271 and the compressed SHAP matrix obtained in s273 with respect to the feature in question.

[0083] Specifically, the compressed SHAP matrix similarity calculation unit 111 sets the similarity so that the closer the value of the feature in the compressed SHAP matrix obtained in s271 is to the value of the feature in the compressed SHAP matrix obtained in s273, the larger the similarity value becomes. For example, the compressed SHAP matrix similarity calculation unit 111 sets the similarity to the reciprocal of the difference between the two values. Note that the similarity calculation method described here is just one example.

[0084] The compressed SHAP matrix similarity calculation unit 111 checks whether the processes s272 to s275 have been performed for all rows related to the case of the compressed SHAP matrix created in the past for the present case (s276). If the processes s272 to s275 have been performed for all rows (s276: Yes), the compressed SHAP matrix similarity calculation unit 111 executes the process s277. If there are rows for which the processes s272 to s275 have not been performed (s276: No), the compressed SHAP matrix similarity calculation unit 111 repeats the processes from s272 onwards for those rows.

[0085] In s277, the compressed SHAP matrix similarity calculation unit 111 stores the similarity calculated so far (s277). Subsequently, the similar record extraction unit 113 identifies compressed SHAP matrices whose similarity satisfies predetermined conditions (for example, compressed SHAP matrices with similarity higher than a predetermined threshold, and compressed SHAP matrices up to a predetermined rank in terms of similarity). The compressed SHAP matrix similarity calculation unit 111 then displays various pieces of information associated with the identified compressed SHAP matrix (for example, the feature quantities of the corresponding SHAP matrix, and information about the input data for the trained model corresponding to that SHAP matrix). This completes the similarity calculation process s270.

[0086] Figure 16 illustrates an example of the process in the similarity calculation process s270. As shown in the figure, the compressed SHAP matrix 400a for case "001" is the row data having the features "F01", "F02", "F08", "F09", and "F10", and the past compressed SHAP matrix 400 is the row data having the features "F01", "F02", "F03", "F07", and "F09". b If such a combination exists, the similar record extraction unit 113 detects the feature quantities "F01", "F02", "F03", "F07", "F08", "F08", "F09", and "F10" that exist in either of the two compressed SHAP matrices 400a and 400b (code 430). Then, the similar record extraction unit 113 detects the feature quantities "F03" and "F07" which have values ​​set in only one of the compressed SHAP matrices. 、 For each of "F08" and "F10", set the value of the other feature to "0" (symbol 440).

[0087] In this way, when comparing values ​​for the same item, setting the value to 0 if one of the compressed SHAP matrices does not have a value can improve the efficiency of the comparison process.

[0088] Here, we will explain the screen displayed by the search support device 1. (SHAP Importance-Related Information Input Screen) Figure 17 shows an example of the SHAP importance-related information input screen 1100. The SHAP importance-related information input screen 1100 includes a project name display field 1110 where the project name is displayed, a target value input field 1120 where the user inputs key performance indicators (KPIs) related to the project, a feature input field 1130 where the user inputs features in the trained model, and a required item input field 1140 where the user inputs required items to be adopted. As shown in the mismatch pattern input field 1150, the SHAP importance estimation unit 105 may also accept the specification of combinations of features to be excluded when creating the compressed SHAP matrix.

[0089] The SHAP importance-related information input screen 1100 is displayed, for example, when the user decides which data to input into the trained model, or when the user inputs the 500 required items to be adopted.

[0090] (Compressed SHAP matrix confirmation screen) Figure 18 shows an example of the compressed SHAP matrix confirmation screen 1200. The compressed SHAP matrix confirmation screen 1200 includes a list display 1210 of the SHAP matrices before compression (matrices of SHAP values) and a list display 1220 of the SHAP matrices after compression (matrices of SHAP values). In addition, the column lengths 1211 (number of features) of the SHAP matrices before compression and 1221 (number of features) of the SHAP matrices after compression are displayed. This allows the user to check how much the SHAP matrices have been compressed.

[0091] The compressed SHAP matrix confirmation screen 1200 is displayed, for example, when a compressed SHAP matrix is ​​created or when input is received from the user.

[0092] (Similar Records Display Screen) Figure 19 shows an example of the Similar Records display screen 1300. The Similar Records display screen 1300 displays the ID 1310 of each case that was determined to have a high degree of similarity, the similarity score 1320 of that case, the attribute information 1330 (data input to the trained model) entered in that case, and the output data 1340 (predicted value) output by the trained model in that case.

[0093] The similar record display screen 1300 is displayed, for example, during the similarity calculation process s270.

[0094] As described above, in the learning phase s100, the search support device 1 of this embodiment calculates each SHAP data for each output data output from the trained model to which the learning data has been input, and generates and stores compressed SHAP data for each SHAP data. Meanwhile, in the inference phase s200, the search support device 1 calculates the SHAP data to be validated that corresponds to the predicted value for the data to be inferred, calculates the similarity between the calculated SHAP data to be validated and each compressed SHAP data, and identifies the compressed SHAP data whose similarity satisfies predetermined conditions.

[0095] In other words, the search support device 1 of this embodiment performs a search for SHAP data by comparing compressed SHAP data with other SHAP data. Thus, the search support device 1 of this embodiment can perform searches related to SHAP data, which are parameters representing the degree of influence of feature quantities, at high speed and with high accuracy.

[0096] Furthermore, the search support device 1 of this embodiment provides a history of each SHAP data (SHAP global Based on statistics (800), compressed SHAP data is generated by identifying the features to be compressed among the features related to the SHAP data.

[0097] Specifically, the search support device 1 of this embodiment identifies a threshold for the degree of influence in the SHAP data based on the history of each SHAP data (SHAP global statistics 800), identifies SHAP data with an influence below the threshold among the values ​​of the SHAP data as data for the feature quantities to be compressed, and generates compressed SHAP data by removing the data for the identified feature quantities.

[0098] This allows us to identify the features that should be compressed and generate compressed SHAP data that is more suitable for accurate searching.

[0099] Furthermore, the search support device 1 of this embodiment generates compressed SHAP data by identifying the feature quantities to be compressed among the feature quantities related to SHAP data, based on hardware information (hardware information 900 and system constraints 1000) provided by the search support device 1.

[0100] Specifically, the search support device 1 of this embodiment determines the compression ratio of SHAP data based on hardware information (hardware information 900 and system constraints 1000) provided by the search support device 1, and generates compressed SHAP data based on the determined compression ratio.

[0101] This allows for the generation of compressed SHAP data suitable for searching, depending on the hardware status of the search support device 1 that performs the search.

[0102] Furthermore, the search support device 1 of this embodiment accepts from the user the specification of features related to SHAP data that should not be compressed, and generates compressed SHAP data based on the specified features that should not be compressed.

[0103] This allows for the retention of essential features in the compressed SHAP data based on user knowledge (domain knowledge), enabling more appropriate searches.

[0104] Furthermore, when the search support device 1 of this embodiment calculates the similarity, if it detects a feature that exists only in either the compressed SHAP data or the SHAP data to be verified, it sets the influence value of the feature in the SHAP data where the feature does not exist to a predetermined reference value (0 in this embodiment), thereby calculating the similarity between the compressed SHAP data and the SHAP data to be verified.

[0105] This makes it easy to compare each feature of the compressed SHAP data with each feature of the SHAP data being validated, and to calculate the similarity.

[0106] Furthermore, the search support device 1 of this embodiment outputs information about the created compressed SHAP data (compressed SHAP matrix confirmation screen 1200). This allows the user to confirm how the SHAP data was compressed.

[0107] Furthermore, the search support device 1 of this embodiment displays a screen that accepts the specification of feature quantities to be excluded from compression (SHAP importance-related information input screen 1100). This allows the user to arbitrarily specify feature quantities to be excluded from compression.

[0108] Furthermore, the search support device 1 of this embodiment displays information about the feature quantities associated with the compressed SHAP data (similar record display screen 1300). This allows the user to verify You can find information related to the target SHAP data, the data to be inferred, etc.

[0109] The present invention is not limited to the embodiments described above, and can be implemented using any components without departing from its spirit. The embodiments and modifications described above are merely examples, and the present invention is not limited to these as long as the features of the invention are not impaired. Furthermore, although various embodiments and modifications have been described above, the present invention is not limited to these. Other embodiments conceivable within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention.

[0110] For example, the configuration of each functional unit described in this embodiment is just one example; for instance, a part of one functional unit may be incorporated into another, or multiple functional units may be configured as a single functional unit. [Explanation of Symbols]

[0111] 1. Search support device, 300 SHAP matrix, 400 compressed SHAP matrix

Claims

1. Equipped with a processor and memory, The aforementioned processor, A process to calculate at least one SHAP data, which is data showing the degree of influence of each feature in the trained model on the output data output from the trained model, Based on the history of each SHAP data calculated above, the process of generating compressed SHAP data is performed by identifying the feature quantities to be deleted from the feature quantities related to the SHAP data, The process of generating the compressed SHAP data for each SHAP data and storing it in the memory, A process to calculate the SHAP data to be validated, which is the SHAP data for the output data output from the trained model, by inputting input data to the trained model, The process involves calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be verified, and identifying the compressed SHAP data whose similarity to the SHAP data to be verified satisfies predetermined conditions. A search assistance device that performs the following actions.

2. A search support device comprising a processor and memory, The aforementioned processor, A process to calculate at least one SHAP data, which is data showing the degree of influence of each feature in the trained model on the output data output from the trained model, Based on information regarding the hardware provided by the search support device, the process involves identifying the feature quantities to be deleted from the feature quantities related to the SHAP data, thereby generating compressed SHAP data. The process of generating the compressed SHAP data for each SHAP data and storing it in the memory, A process to calculate the SHAP data to be validated, which is the SHAP data for the output data output from the trained model, by inputting input data to the trained model, The process involves calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be verified, and identifying the compressed SHAP data whose similarity to the SHAP data to be verified satisfies predetermined conditions. A search assistance device that performs the following actions.

3. Equipped with a processor and memory, The aforementioned processor, A process to calculate at least one SHAP data, which is data showing the degree of influence of each feature in the trained model on the output data output from the trained model, The process involves receiving a user's specification of features to be excluded from compression among the features related to the SHAP data, and generating compressed SHAP data by deleting the other features while retaining the data of the specified features to be excluded from compression. The process of generating the compressed SHAP data for each SHAP data and storing it in the memory, A process to calculate the SHAP data to be validated, which is the SHAP data for the output data output from the trained model, by inputting input data to the trained model, The process involves calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be verified, and identifying the compressed SHAP data whose similarity to the SHAP data to be verified satisfies predetermined conditions. A search assistance device that performs the following actions.

4. The aforementioned processor, Based on the history of each SHAP data calculated above, a threshold for the degree of influence in the SHAP data is identified, data with an influence below the threshold in the SHAP data is identified as the feature data to be deleted, and the compressed SHAP data is generated by removing the identified feature data from the SHAP data. The search support device according to claim 1.

5. The aforementioned processor, Based on information regarding the hardware provided by the search support device, the compression ratio of the SHAP data is determined, and the compressed SHAP data is generated based on the determined compression ratio. The search support device according to claim 2.

6. The aforementioned processor, The compressed SHAP data is generated as data consisting of combinations of the names of each feature and the values ​​of those features. A search support device according to any one of claims 1 to 3.

7. The aforementioned processor, When calculating the similarity, if a feature is detected that exists only in either the compressed SHAP data or the SHAP data to be verified, the influence value of that feature in the SHAP data that does not contain that feature is set to a predetermined standard value, thereby calculating the similarity between the compressed SHAP data and the SHAP data to be verified. A search support device according to any one of claims 1 to 3.

8. The system includes an output device that outputs information about the generated compressed SHAP data. A search support device according to any one of claims 1 to 3.

9. The system includes an output device that displays a screen for receiving user requests to specify the feature quantities to be excluded from compression. The search support device according to claim 3.

10. The system includes an output device that displays information about the feature quantities associated with the identified compressed SHAP data. A search support device according to any one of claims 1 to 3.

11. The information processing device performs a process of calculating at least one SHAP data, which is data indicating the degree of influence of each feature in the trained model on the output data output from the trained model, Based on information about the hardware provided by the aforementioned information processing device, the process of generating compressed SHAP data involves identifying the feature quantities to be deleted from among the feature quantities related to the SHAP data, and The process involves generating the compressed SHAP data for each SHAP data and storing it in memory. A process to calculate the SHAP data to be validated, which is the SHAP data for the output data output from the trained model, by inputting input data to the trained model, The process involves calculating the similarity between each of the generated compressed SHAP data and the calculated SHAP data to be verified, and identifying the compressed SHAP data whose similarity to the SHAP data to be verified satisfies predetermined conditions. A search assistance method that performs this task.

Images