Data processing method and related device
Through the fusion interpretation technology of multiple interpretable submodels, the problem of poor interpretability of machine learning models' output results is solved, and the interpretation accuracy and credibility of prediction results are improved.
Patent Information
- Application Number
- CN202310666176.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-06
- Publication Date
- 2024-12-06
AI Technical Summary
As the complexity of machine learning models increases, existing interpretation methods cannot accurately interpret the prediction results of the model output, resulting in a lack of interpretability and credibility in the prediction results.
通过多个可解释性子模型分别进行解释,并对得到的待定解释结果进行融合,得到目标解释结果,以更准确地解释机器学习模型的预测结果。
Improve the credibility of interpretation accuracy and predicted results, and achieve a more comprehensive and accurate interpretation of the output results of the machine learning model.
Smart Images

Figure CN119089193A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and in particular to a data processing method and related devices. Background Art
[0002] With the rapid development of artificial intelligence (AI) technology, machine learning models are widely used in various fields, such as e-commerce, Internet, finance, supervision, justice, medical care and education. Specifically, the corresponding machine learning model can be trained based on machine learning, and then the trained machine learning model can be applied in the prediction scenario, specifically, the machine learning model can be used to output the prediction results, and then the prediction results can be used for subsequent related applications.
[0003] As machine learning models become more and more widespread, users are no longer satisfied with the effects of machine learning models (i.e., output prediction results), but hope to have a better understanding of the reasons behind the effects of machine learning models. Explanatory models are used in related technologies for explanation, which mainly refers to the ability to provide explanations for prediction results during the machine learning process, so that users can better understand why the machine learning model outputs such prediction results.
[0004] However, with the continuous development of AI technology, machine learning models are becoming more and more complex. The explanation methods used in related technologies cannot be accurately explained, which makes the prediction results output by the machine learning model lack interpretability, thereby reducing the credibility of the prediction results. Summary of the invention
[0005] In order to solve the above technical problems, the present application provides a data processing method and related devices, which are first interpreted separately through multiple interpretable sub-models and then fused. The target weights obtained by the fusion can more accurately interpret the prediction results output by the machine learning model, thereby improving the accuracy of the interpretation and helping to improve the credibility of the prediction results.
[0006] The embodiments of the present application disclose the following technical solutions:
[0007] On the one hand, an embodiment of the present application provides a data processing method, the method comprising:
[0008] Obtain feature data of multiple dimensions of the object to be predicted;
[0009] Using each of the multiple interpretable sub-models to analyze the feature data of the multiple dimensions, respectively, to obtain multiple pending interpretation results; each of the multiple pending interpretation results is used to characterize the overall weight of the feature data of the multiple dimensions for the prediction result, respectively, and the prediction result is a result obtained by predicting the object to be predicted through a machine learning model based on the feature data of the multiple dimensions;
[0010] The multiple pending explanation results are fused to obtain a target explanation result; the target explanation result is used to characterize the target weights of the feature data of the multiple dimensions for the prediction results respectively.
[0011] In another aspect, an embodiment of the present application provides a data processing device, the device comprising an acquisition unit, an analysis unit and a fusion unit:
[0012] The acquisition unit is used to acquire feature data of multiple dimensions of the object to be predicted;
[0013] The analysis unit is used to analyze the feature data of the multiple dimensions using each of the multiple interpretable sub-models to obtain multiple pending interpretation results; each of the multiple pending interpretation results is used to characterize the overall weight of the feature data of the multiple dimensions for the prediction result, and the prediction result is a result obtained by predicting the object to be predicted through a machine learning model based on the feature data of the multiple dimensions;
[0014] The fusion unit is used to fuse the multiple pending explanation results to obtain a target explanation result; the target explanation result is used to characterize the target weights of the feature data of the multiple dimensions for the prediction results respectively.
[0015] In a possible implementation, the multiple interpretability sub-models include a target interpretability sub-model, and the analysis unit is further configured to:
[0016] The target interpretability sub-model is used to analyze the feature data of the multiple dimensions to obtain an object interpretation result; the object interpretation result is used to characterize the individual weights of the feature data of the multiple dimensions of the object to be predicted for the prediction results, and the individual weights are associated with the object to be predicted.
[0017] In a possible implementation, the number of the target interpretability sub-models is multiple, and the analysis unit is further used for:
[0018] Using each target interpretability sub-model in the plurality of target interpretability sub-models to analyze the feature data of the plurality of dimensions respectively, to obtain a plurality of object interpretation results;
[0019] The fusion unit is also used for:
[0020] A plurality of the object interpretation results are fused to obtain a target object interpretation result; the target object interpretation result is used to characterize the target individual weights of the prediction results for the feature data of multiple dimensions of the object to be predicted.
[0021] In a possible implementation, the analysis unit is further configured to:
[0022] Analyzing the prediction result by using the target weights of the feature data of the multiple dimensions represented by the target interpretation result to obtain a first explanatory result of the prediction result at a feature granularity;
[0023] The prediction result is analyzed by using the individual weights of the feature data of the multiple dimensions represented by the object interpretation result to obtain a second explanatory result of the prediction result at an individual granularity.
[0024] In a possible implementation manner, the fusion unit is further configured to:
[0025] Obtaining model weights corresponding to the multiple interpretability sub-models respectively;
[0026] The multiple pending explanation results are weightedly fused based on the model weights corresponding to the multiple interpretability sub-models to obtain the target explanation result.
[0027] In a possible implementation manner, the object to be predicted belongs to a scene to be predicted, and the fusion unit is further used for:
[0028] Obtain a weighted test sample of the scenario to be predicted; the weighted test sample includes test feature data of multiple dimensions of a test object, the weighted test sample has a test sample label, and the test sample label is used to indicate the overall standard weight of the test feature data of the multiple dimensions for a first sample prediction result, respectively, and the first sample prediction result is obtained by predicting the test object through the machine learning model;
[0029] Analyze the test feature data of the multiple dimensions respectively using each of the multiple interpretable sub-models to obtain multiple test interpretation results; the test interpretation results are used to characterize the overall test weights of the test feature data of the multiple dimensions for the first sample prediction result respectively;
[0030] For each test interpretation result, the model weight of the interpretable sub-model corresponding to the test interpretation result is determined according to the difference between the test interpretation result and the test sample label, and the model weights corresponding to the multiple interpretable sub-models are obtained; the difference is negatively correlated with the model weight.
[0031] In a possible implementation manner, the object to be predicted belongs to a scene to be predicted, and the fusion unit is further used for:
[0032] Acquiring prior information of the scene to be predicted;
[0033] The prior information is used to determine model weights corresponding to the multiple interpretable sub-models.
[0034] In a possible implementation, the plurality of pending interpretation results correspond to an i-th analysis, i being an integer greater than 1, and the fusion unit is further configured to:
[0035] According to the multiple pending explanation results obtained in the (i-1)th analysis, the model weights respectively corresponding to the multiple interpretable sub-models used in the (i-1)th analysis are adjusted to obtain the adjusted model weights respectively corresponding to the multiple interpretable sub-models;
[0036] The adjusted model weights respectively corresponding to the multiple interpretable sub-models used in the i-1th analysis are determined as the model weights respectively corresponding to the multiple interpretable sub-models used in the i-th analysis.
[0037] In a possible implementation, the multiple interpretability sub-models are trained in the following manner:
[0038] Obtaining a training sample set of a scene to be predicted to which the object to be predicted belongs; the training sample set includes a plurality of training samples, each training sample includes training feature data of multiple dimensions of the training sample object, each training sample has a training sample label, and the training sample label is used to indicate the overall standard weight of the training feature data of the multiple dimensions for a second sample prediction result, respectively, and the second sample prediction result is obtained by predicting the training sample object through the machine learning model;
[0039] For each of the multiple interpretable sub-models, the multiple training samples are used to perform model training on the interpretable sub-model until the difference between the training explanation result output by the interpretable sub-model and the training sample label meets the training end condition; the training explanation result is used to characterize the overall training weight of the training feature data of the multiple dimensions for the second sample prediction result respectively.
[0040] In a possible implementation manner, the device further includes a dividing unit and a training unit:
[0041] The division unit is used to divide the training sample set into a plurality of training sample subsets; the number of the training sample subsets is the same as the number of the interpretability sub-models;
[0042] The training unit is used to perform model training for each interpretable sub-model among the multiple interpretable sub-models using training samples included in a training sample subset among the multiple training sample subsets; one training sample subset is used for model training of one interpretable sub-model, and different training sample subsets are used for model training of different interpretable sub-models.
[0043] In a possible implementation, the scene to be predicted to which the object to be predicted belongs is an object classification scene, the machine learning model is a classification model, the prediction result is a classification result, the feature data of the multiple dimensions include historical consumption data, browsing data, and search data, and the device further includes a classification unit:
[0044] The classification unit is used to classify the object to be predicted using the classification model according to the historical consumption data, the browsing data and the search data to obtain the classification result; the classification result is used to indicate whether the object to be predicted is a target object;
[0045] The analysis unit is further used to analyze the classification result using the target weights of the historical consumption data, the browsing data and the search data represented by the target interpretation result for the classification result, so as to obtain an explanatory result corresponding to the classification result.
[0046] In a possible implementation manner, if the classification result indicates that the object to be predicted is the target object, the device further includes a determination unit and a recommendation unit:
[0047] The determining unit is used to select data with the largest target weight from the historical consumption data, the browsing data and the search data to determine as target data;
[0048] The determining unit is further configured to determine the information to be recommended that matches the target data among the information to be recommended associated with the object classification scenario as the recommended information;
[0049] The recommendation unit is used to recommend the recommendation information to the object to be predicted.
[0050] On the other hand, an embodiment of the present application provides a computer device, the computer device comprising a processor and a memory:
[0051] The memory is used to store a computer program and transmit the computer program to the processor;
[0052] The processor is configured to execute the method described in any one of the preceding aspects according to instructions in the computer program.
[0053] On the other hand, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program. When the computer program is executed by a computer device, the computer device executes the method described in any of the above aspects.
[0054] On the other hand, an embodiment of the present application provides a computer program product, including a computer program, which, when executed on a computer device, enables the computer device to execute the method described in any one of the aforementioned aspects.
[0055] It can be seen from the above technical solution that, first, feature data of multiple dimensions of the object to be predicted can be obtained. The object to be predicted can refer to the object that needs to be predicted, and the feature data of multiple dimensions can represent the situation of the object to be predicted from different dimensions. Then, each of the multiple interpretable sub-models can be used to analyze the feature data of multiple dimensions to obtain multiple pending interpretation results. Each of the multiple pending interpretation results can be used to characterize the overall weight of the feature data of multiple dimensions for the prediction result, wherein the prediction result can be the result obtained by predicting the object to be predicted based on the feature data of multiple dimensions through a machine learning model. Finally, multiple pending interpretation results can be fused to obtain a target interpretation result, and the target interpretation result can be used to characterize the target weight of the feature data of multiple dimensions for the prediction result. Based on this, the explanations are first performed separately through multiple interpretable sub-models, and then fused. Usually, different interpretable sub-models focus on different levels of feature data when interpreting, so a more comprehensive explanation can be achieved based on fusion. Therefore, compared with the pending explanation results of a single interpretable sub-model, the target explanation results can more accurately characterize the weights of feature data of multiple dimensions for the prediction results. In other words, the target weight is more accurate than the overall weight. Accordingly, the target weight can more accurately explain the prediction results output by the machine learning model, thereby improving the accuracy of the explanation and helping to improve the credibility of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technical members in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0057] Figure 1 A schematic diagram of an application scenario of a data processing method provided in an embodiment of the present application;
[0058] Figure 2 A flowchart of a data processing method provided in an embodiment of the present application;
[0059] Figure 3 A schematic diagram of the structure of an integrated explanatory model provided in an embodiment of the present application;
[0060] Figure 4 A structural diagram of a data processing device provided in an embodiment of the present application;
[0061] Figure 5 A structural diagram of a terminal provided in an embodiment of the present application;
[0062] Figure 6 A structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION
[0063] The embodiments of the present application are described below in conjunction with the accompanying drawings.
[0064] As machine learning models are widely used in various fields, when using the prediction results output by machine learning models, users hope to understand why such prediction results are output. In order to facilitate the understanding of the prediction results, an interpretable model is usually used for interpretation, which mainly refers to the ability to provide an explanation of the prediction results in the process of machine learning, so that users can understand why the machine learning model outputs such prediction results. Among them, the interpretable model is a machine learning model with explanatory ability, which can be obtained by training the initial model with a designed model structure by constructing training samples. Typical interpretable models can include regression models, decision tree models, rule-based learning models, etc. For example, the regression model can be a Least Absolute Shrinkage and Selection Operator (LASSO) model, and the decision tree model can be an eXtreme Gradient Boosting (XGBoost) model. In addition, a typical interpretable model can also be, for example, a Shapley Additive exPlanations (SHAP) model.
[0065] Based on this, by using the interpretable model to explain the reasons for obtaining this prediction result and ensuring the credibility of the prediction result, it can also be considered to improve the transparency and interpretability of the machine learning model, based on which trust in the machine learning model is established. Among them, interpretability can refer to the ability to provide explanations for the internal operations of the machine learning model and the output prediction results during the machine learning process, thereby assisting in understanding why the machine learning model outputs such a prediction result. In practical applications, the interpretability of prediction results has become a prerequisite for evaluating whether the prediction results can ultimately be adopted, especially in high-risk fields (such as medical care) with high credibility and security requirements, as well as finance (such as banking, insurance) and regulatory fields with significant economic impacts.
[0066] In the related art, when explanation is needed, an interpretable model is directly used for explanation. However, in practical applications, different interpretable models use different explanation logics when explaining. The explanation logic can characterize the process of explanation by the interpretable model. The difference in explanation logic makes different interpretable models focus on different levels of data when explaining. Correspondingly, different interpretable models may get different explanations for the same prediction result. For this reason, the method of using a single interpretable model for explanation in the related art is relatively limited to explaining a certain level of data, and the explanation obtained is not comprehensive enough. Moreover, with the continuous development of AI technology, machine learning models are becoming more and more complex. The explanation method in the related art makes it impossible to explain comprehensively, resulting in low explanation accuracy, which makes the prediction results output by the machine learning model lack of interpretability, thereby reducing the credibility of the prediction results.
[0067] To this end, an embodiment of the present application provides a data processing method, which first uses multiple interpretable sub-models to interpret separately to obtain multiple pending interpretation results, and then fuses the multiple pending interpretation results to obtain a target interpretation result, and the target interpretation result can be used to interpret the prediction results output by the machine learning model. That is, first use multiple interpretable sub-models to interpret from different levels of the data, and then achieve a more comprehensive explanation through fusion. Based on this, compared with the relevant technology, the use of this application can achieve a more comprehensive explanation, so that the prediction results output by the machine learning model can be explained more accurately, the accuracy of the explanation can be improved, and it is conducive to improving the credibility of the prediction results.
[0068] It should be noted that in the specific implementation of the present application, user information and other related data may be involved in the process of data processing. When the above embodiments of the present application are applied to specific products or technologies, the user's separate consent or separate permission is required, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.
[0069] This application can be applied to the field of artificial intelligence, which is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that the machines have the functions of perception, reasoning and decision-making. For example, the automatic execution of the data processing method of this application can be realized based on artificial intelligence technology.
[0070] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, and smart transportation. The embodiments of this application mainly involve machine learning and other directions, and specifically can be to use machine learning methods to perform model training to obtain the machine learning model, interpretable sub-model, etc. used in this application.
[0071] The data processing method provided in the embodiment of the present application can be implemented by a computer device, which can be a terminal or a server, wherein the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminals include but are not limited to smart phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, etc. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application is not limited here. The embodiment of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, audio and video, assisted driving, etc. The embodiment of the present application can be specifically applied to various prediction scenarios corresponding to various fields, such as classification scenarios, recommendation scenarios and other prediction scenarios.
[0072] In practical applications, for prediction scenarios, a trained machine learning model can be used to predict and obtain prediction results. Generally, the object to be predicted can be called an object to be predicted. For example, the object to be predicted can be a user, an enterprise, etc. In order to better understand the prediction results, the data processing method provided in this application can be used to interpret them.
[0073] In order to better understand this application, Figure 1 To explain this application, Figure 1 The following is a schematic diagram of an application scenario of the data processing method provided in the embodiment of the present application. It should be noted that: Figure 1 The objects to be predicted shown are only examples and are not intended to be limiting. Figure 1 In the illustrated scenario, the server 100 is used as an example of the aforementioned computer device for illustration:
[0074] First, the server 100 may obtain feature data of multiple dimensions of the object to be predicted. The object to be predicted may refer to an object that needs to be predicted. The feature data of multiple dimensions may represent the situation of the object to be predicted from different dimensions. For example, the feature data of multiple dimensions may include feature data of three dimensions A, B, and C. And, the prediction process for the object to be predicted is illustrated by the dotted line, that is, according to the feature data of multiple dimensions, the object to be predicted is predicted by the machine learning model to obtain the prediction result.
[0075] Next, the server 100 can use each of the multiple interpretability sub-models to analyze the feature data of multiple dimensions to obtain multiple pending interpretation results. Figure 1 The multiple interpretable sub-models shown in may include interpretable sub-model 1, interpretable sub-model 2, ..., interpretable sub-model M, where M is an integer greater than 1, and accordingly, pending interpretation result 1, pending interpretation result 2, ..., pending interpretation result M can be obtained. Each of the multiple pending interpretation results can be used to characterize the overall weight of the feature data of multiple dimensions for the prediction result, and the overall weight can be used to characterize the importance of the feature data for obtaining the prediction result. Taking the feature data of the three dimensions A, B, and C mentioned above as an example, the pending interpretation result 1 can be the importance of the feature data of the three dimensions A, B, and C determined by the interpretable sub-model 1 for the prediction result, and the pending interpretation result M can be the importance of the feature data of the three dimensions A, B, and C determined by the interpretable sub-model M for the prediction result.
[0076] Finally, the server 100 can fuse multiple pending interpretation results to obtain a target interpretation result. For example, the server 100 can fuse pending interpretation result 1, pending interpretation result 2, ..., pending interpretation result M to obtain a target interpretation result. Similarly, the target interpretation result can be used to characterize the target weights of feature data of multiple dimensions for the prediction results, respectively. The difference is that the target interpretation result is obtained based on the fusion of M pending interpretation results. In practical applications, the interpretability sub-model 1, the interpretability sub-model 2, ..., and the interpretability sub-model M focus on different levels of feature data of the three dimensions of A, B, and C when interpreting. Therefore, it can be considered that the pending interpretation result 1, the pending interpretation result 2, ..., and the pending interpretation result M characterize the importance of A, B, and C to the prediction results from different levels, so a more comprehensive explanation can be achieved based on fusion. Subsequently, the target interpretation result can be used directly for interpretation, specifically for Figure 1 The dotted line part is explained to facilitate understanding of why the machine learning model outputs such a prediction result based on the feature data of the three dimensions A, B, and C.
[0077] It can be seen that in the data processing method provided by the present application, when explanation is needed, it is first explained separately by multiple interpretable sub-models, and then fused to achieve a more comprehensive explanation. Compared with the pending explanation results of a single interpretable sub-model, the target explanation result can more accurately characterize the weights of the feature data of multiple dimensions for the prediction results, that is, the target weight is more accurate than the overall weight. Accordingly, the target weight can be used to more accurately explain the prediction results output by the machine learning model, thereby improving the accuracy of the explanation and helping to improve the credibility of the prediction results.
[0078] Figure 2 This is a flowchart of a data processing method provided in an embodiment of the present application, which is described by taking a server as an example of the aforementioned computer device. The method includes S201-S203:
[0079] S201: Acquire feature data of multiple dimensions of the object to be predicted.
[0080] In practical applications, the object to be predicted can be called the object to be predicted, and the trained machine learning model can be used to predict the object to be predicted to obtain the prediction result. In order to better understand the prediction result, it can be explained. When an explanation is required, the server can first obtain the feature data of multiple dimensions of the object to be predicted, and the feature data of multiple dimensions can represent the situation of the object to be predicted from different dimensions. Similarly, the feature data of multiple dimensions is also the basis for predicting the object to be predicted, that is, when predicting, the object to be predicted can be predicted by the machine learning model based on the feature data of multiple dimensions to obtain the prediction result.
[0081] It should be noted that this application does not impose any restrictions on the method of obtaining feature data of multiple dimensions. It is understandable that the quality of feature data may affect the effect of the machine learning model in obtaining the prediction results, and also affect the subsequent analysis process. In practical applications, in order to improve the quality of feature data, the acquired raw data can also be preprocessed before machine learning, and the raw data after preprocessing can be used as feature data. Specifically, preprocessing can include steps such as data cleaning, data missing value filling, and data standardization. Through preprocessing, the feature data can be effective data for machine learning and meet the standards of machine learning, thereby improving data quality.
[0082] It should also be noted that this application does not impose any limitations on machine learning models. Specifically, the role of the machine learning model is to predict the object to be predicted to obtain a prediction result. In practical applications, the machine learning model can be associated with the scene to be predicted to which the object to be predicted belongs. For example, when the scene to be predicted is a classification scene, the machine learning model can be a classification model, and when the scene to be predicted is a recommendation scene, the machine learning model can be a recommendation model. Generally, the machine learning model can be obtained by training the initial model using the training samples in the scene to be predicted. In practical applications, corresponding training samples can be constructed according to the actual scene to be predicted, and model training can be performed to obtain the corresponding machine learning model.
[0083] S202: Utilize each of the multiple interpretable sub-models to analyze feature data of multiple dimensions respectively, and obtain multiple pending interpretation results.
[0084] Since the feature data of multiple dimensions is the basis for prediction, that is, the basis for the machine learning model to output such a prediction result, when interpreting, the feature data of multiple dimensions can be analyzed to assist in understanding why such a prediction result is output. In specific implementation, the server can use each of the multiple interpretable sub-models to analyze the feature data of multiple dimensions respectively to obtain multiple pending interpretation results. Among them, each of the multiple pending interpretation results can be used to characterize the overall weight of the feature data of multiple dimensions for the prediction result respectively, and the prediction result can be the result obtained by predicting the prediction object through the machine learning model based on the feature data of multiple dimensions. In practical applications, different interpretable sub-models focus on different levels of feature data when interpreting, so it can be considered that multiple pending interpretation results are the overall weights of the feature data of multiple dimensions for the prediction result respectively from different levels. Based on this, it can be explained separately through multiple interpretable sub-models first, so as to obtain multiple pending interpretation results at different levels.
[0085] It should be noted that this application does not impose any limitation on multiple interpretable sub-models. In practical applications, multiple interpretable sub-models may refer to machine learning models with interpretative capabilities. Usually, multiple interpretable sub-models may also be direct reuse of the aforementioned interpretable models with interpretative capabilities, wherein the aforementioned interpretable models with interpretative capabilities may specifically refer to machine learning models with interpretative capabilities after completing model training. In this way, the direct reuse method is conducive to improving interpretation efficiency and reducing interpretation costs. In addition, multiple interpretable sub-models may also be obtained by model training using a training sample set of the scene to be predicted described by the object to be predicted, so that multiple interpretable sub-models can have good interpretation accuracy in the scene to be predicted, which is conducive to ensuring the accuracy of the interpretation of the prediction results. Specifically, multiple interpretable sub-models may be trained in the following manner:
[0086] First, the server obtains a training sample set of a scene to be predicted to which the object to be predicted belongs. The training sample set may include multiple training samples, each of which may include training feature data of multiple dimensions of the training sample object, and the training feature data of multiple dimensions may represent the situation of the sample object from different dimensions. Each training sample has a training sample label, and the training sample label may be used to represent the overall standard weight of the training feature data of multiple dimensions for the second sample prediction result, and the second sample prediction result may be obtained by predicting the training sample object through a machine learning model. Then, the server may perform model training on the interpretable sub-model using multiple training samples for each interpretable sub-model in the multiple interpretable sub-models. During the model training process, the interpretable sub-model may output a training explanation result, and the training explanation result may be used to characterize the overall training weight of the training feature data of multiple dimensions for the second sample prediction result. Until the difference between the training explanation result output by the interpretable sub-model and the training sample label meets the end training condition, indicating that the interpretable sub-model has the ability to explain and has good explanation accuracy, the model training may be ended at this time.
[0087] Based on this, by using the training sample set of the scene to be predicted to which the object to be predicted belongs to perform model training, multiple interpretable sub-models can have good interpretation accuracy in the scene to be predicted, which is conducive to ensuring the accuracy of the interpretation of the prediction results. It can be understood that on the basis of direct reuse, combined with the method of using the training sample set of the scene to be predicted to which the object to be predicted belongs to perform model training, it can be beneficial to shorten the model training process, and at the same time, it can ensure the prediction accuracy of multiple interpretable sub-models in the scene to be predicted.
[0088] In practical applications, in order to ensure that the interpretable sub-model has good interpretation accuracy, model evaluation can also be performed on each interpretable sub-model after model training. Specifically, a test sample set for model evaluation can be constructed. For example, the original training sample set can be divided into a sample subset for model training and a sample subset for model evaluation. The sample subset for model evaluation constitutes the test sample set. Then, the test sample set can be used to test the effect of each interpretable sub-model that has completed model training, so as to verify and evaluate the interpretation accuracy of each interpretable sub-model, so as to ensure that the interpretable sub-model has good interpretation accuracy.
[0089] This application does not make any limitation on how to use multiple training samples to train the interpretability sub-model. For ease of understanding, the embodiments of this application provide the following methods as examples:
[0090] In order to improve the utilization rate of training samples, in a possible implementation, each training sample in the multiple training samples can be used to perform model training on the interpretability sub-model respectively. Based on this, each training sample can be used to train the models of multiple interpretability sub-models, and the utilization rate of the training samples is the highest.
[0091] In addition, in order to improve the training efficiency, in another possible implementation, the training sample set can also be divided into multiple training sample subsets, and then one training sample subset is used for model training of an interpretable sub-model to achieve parallel training, which is conducive to improving the training efficiency. In the specific implementation, the server can first divide the training sample set into multiple training sample subsets, and the number of training sample subsets is the same as the number of interpretable sub-models. Accordingly, when performing model training, for each interpretable sub-model in multiple interpretable sub-models, the server can use the training samples included in one of the multiple training sample subsets to perform model training. Specifically, one training sample subset can be used for model training of one interpretable sub-model, and different training sample subsets are used for model training of different interpretable sub-models. Based on this, by dividing the training sample set into the same number of training sample subsets as the number of interpretable sub-models, parallel training is achieved, thereby improving the training efficiency.
[0092] This application does not make any limitation on how to divide the training sample set into multiple training sample subsets. In practical applications, the training sample set can be divided into multiple training sample subsets by K-Fold Cross Validation (K-Fold), where K can be used to represent the number of training sample subsets. By dividing in the K-Fold manner, for a certain interpretable sub-model, the training sample subset used to train other interpretable sub-models can also be used as the test sample set after the interpretable sub-model completes training to achieve cross-validation. Based on this, especially when the number of training samples is small, not only can the model training be completed, but the effect of the model training can also be tested and verified based on cross-validation to ensure that the interpretable sub-model has good interpretation accuracy.
[0093] It should also be noted that this application does not impose any restrictions on how to use each interpretability sub-model to analyze feature data of multiple dimensions. In practical applications, different interpretability sub-models may correspond to different analysis methods. For ease of understanding, the embodiments of this application will take multiple interpretability sub-models including the aforementioned LASSO model, XGBoost model, and SHAP model as examples for explanation as follows:
[0094] For the LASSO model, feature data of multiple dimensions can be analyzed based on feature selection methods. Feature selection refers to the process of feature screening and extraction of input feature data during machine learning. The feature selection method can screen out feature data that has a greater impact on the prediction results. Correspondingly, the overall weight of the feature data of each dimension can be determined based on the degree of influence on the prediction results. In specific analysis, the LASSO model can compress the feature data, and when performing regression calculations, it can be based on L1 norm regularization (L1Regularization), and add an L1 norm regularized result, so that the L1 norm regularization result of some feature data becomes zero, thereby achieving feature selection. Usually, the algorithm expression of the LASSO model can be expressed by the following formula:
[0095]
[0096] In the above formula, n can represent the number of dimensions of feature data, n is an integer greater than 1, i can be an integer greater than or equal to 1 and less than or equal to n, which can represent that the dimension of feature data is the i-th dimension, L i It can represent the overall weight corresponding to the feature data of the i-th dimension, x iCan represent the feature vector corresponding to the feature data of the i-th dimension, T can represent the transpose, w can represent the weight vector of the overall weight of the feature data, and λ can represent the regularization coefficient of L1 norm regularization.
[0097] Based on this, the LASSO model can be used to analyze the influence of feature data of each dimension on the prediction results through regression analysis combined with L1 norm regularization, so as to determine the corresponding overall weight. Accordingly, the aforementioned pending explanation results can be obtained to improve the accuracy of the machine learning model.
[0098] For the XGBoost model, the feature data of multiple dimensions can be analyzed based on the explanatory weight calculation method, so as to calculate the influence of the feature data of each dimension on the prediction result. Among them, the explanatory weight can refer to the overall weight generated by a specific algorithm in the machine learning process that can explain the influence of the feature data of each dimension on the prediction result. In practical applications, the XGBoost model can be analyzed in an iterative manner using K decision trees, and usually, K can be an integer greater than 1. When analyzing, for any feature data, the influence of the feature data on the prediction result can be calculated by the distribution difference of the feature data in the K decision trees, so as to determine the overall weight corresponding to the feature data. Based on this, the XGBoost model can be used to analyze the influence of the feature data of each dimension on the prediction result in an iterative manner using K decision trees, so as to determine the corresponding overall weight, and the aforementioned pending explanation result can be obtained accordingly, thereby reducing the complexity of the machine learning model.
[0099] In addition, the XGBoost model can also use K decision trees to analyze feature data of multiple dimensions to achieve the purpose of classification. Specifically, the algorithm expression of the XGBoost model to achieve classification can be expressed by the following formula:
[0100]
[0101] In the above formula, X j It can represent the classification results for j objects to be predicted. In the binary classification, X j The value of can be 0 or 1, K can represent the number of decision trees in the XGBoost model, k can be an integer greater than or equal to 1 and less than or equal to K, which can indicate that the decision tree is the kth decision tree, f k It can represent the prediction function of the kth decision tree, x j It can represent the feature vector corresponding to the feature data of the jth object to be predicted. Taking this application as an example, x j It may include feature vectors corresponding to the feature data of the aforementioned multiple dimensions.
[0102] For the SHAP model, the feature data of multiple dimensions can be analyzed based on the explanatory weight method. The core idea is to calculate the marginal contribution of the feature data of each dimension to the prediction results. At the same time, the interaction between the feature data of each dimension can be considered, and the influence of the feature data of each dimension on the prediction results can be analyzed. Finally, the overall weight of the feature data of each dimension on the prediction results can be calculated. In practical applications, the SHAP model can express the overall weight of the feature data by calculating the SHAP value of the feature data of each dimension. The specific algorithm expression for calculating the SHAP value of the feature data by the SHAP model can be expressed as follows:
[0103]
[0104] In the above formula, n can represent the number of dimensions of feature data, n is an integer greater than 1, i can be an integer greater than or equal to 1 and less than or equal to n, which can represent that the dimension of feature data is the i-th dimension, It can represent the SHAP value of the feature data of the i-th dimension. S can represent a feature data subset. S is included in the feature data set composed of feature data of n dimensions. S does not contain the feature data of the i-th dimension. S It can represent the feature vector corresponding to the feature data contained in S, x S∪{i} It can represent the feature data contained in S and the feature vector corresponding to the feature data of the i-th dimension.
[0105] Based on this, the SHAP model can be used to analyze the influence of feature data of each dimension on the prediction results by calculating the SHAP value of the feature data, thereby determining the corresponding overall weight. The aforementioned pending explanation results can be obtained accordingly, which can explain and evaluate the machine learning model and improve the reliability of the machine learning model.
[0106] S203: Fusing multiple pending interpretation results to obtain a target interpretation result.
[0107] After obtaining multiple pending explanation results, the server can fuse the multiple pending explanation results to obtain a target explanation result. Similarly, the target explanation result can be used to characterize the target weights of feature data of multiple dimensions for the prediction results respectively. The difference is that the target explanation result is the result of fusing multiple pending explanation results, and multiple pending explanation results at different levels are integrated based on the fusion to achieve a more comprehensive explanation. For example, still taking the aforementioned multiple interpretable sub-models including the LASSO model, the XGBoost model and the SHAP model as an example, the analysis results at three different levels based on feature selection, explanatory weights and the interaction between feature data of each dimension can be integrated through fusion to achieve a more comprehensive explanation.
[0108] After obtaining the target interpretation result, the target interpretation result can be used to assist in understanding why the machine learning model outputs such a prediction result. It is understandable that the target interpretation result represents the target weights of the feature data of multiple dimensions for the prediction results, that is, the importance of the prediction results respectively. The larger the target weight, the more important it is, and the more contribution it makes to the prediction results. This is the process of using the target interpretation result to explain the prediction result. It should be noted that this application does not make any restrictions on how to use the target interpretation result to explain the prediction result. For ease of understanding, the embodiments of this application provide the following methods as examples:
[0109] In practical applications, the prediction results output by the machine learning model can be user-oriented, so that users can perform subsequent related applications based on the prediction results. Correspondingly, the target interpretation results can also be user-oriented. Therefore, in one possible implementation, the target interpretation results can be directly displayed to the user. Then, the user can independently analyze and understand why such a prediction result is output based on the target weights of the feature data of multiple dimensions represented by the target interpretation results. In this way, the target interpretation results are directly displayed to the user, which is convenient for assisting the user to understand the prediction results.
[0110] In addition, in order to facilitate users to understand the prediction results more intuitively, in another possible implementation, after obtaining the target explanation result, the prediction result can be analyzed by first using the target weights of the feature data of multiple dimensions represented by the target explanation result to obtain the explanatory result corresponding to the prediction result, which can be used to indicate why such a prediction result is output. In this way, the explanatory result can be displayed to the user, so that the user can understand the prediction result more intuitively.
[0111] From the above, it can be seen that the interpretation method provided by the present application can specifically be a method of integrating multiple pending interpretation results obtained by multiple interpretable sub-models to obtain a target interpretation result. This method can also be called an integrated method of integrating the interpretation results of multiple interpretable sub-models, which is beneficial to improving the accuracy of interpretation. Corresponding to the integration method provided in the embodiment of the present application, the embodiment of the present application also provides an integrated explanatory model, which can be used to interpret the final target interpretation result. Among them, the integrated explanatory model can include multiple interpretable sub-models and a fusion layer in terms of model structure. The fusion layer can be used to fuse multiple pending interpretation results to obtain the target interpretation result.
[0112] For ease of understanding, still taking the aforementioned number of multiple interpretable sub-models as M as an example, the present application embodiment provides a structural schematic diagram of an integrated interpretable model, which can be seen in Figure 3 shown. Figure 3 A structural schematic diagram of an integrated explanatory model is shown. Specifically, the integrated explanatory model may include an interpretable sub-model 1, an interpretable sub-model 2, ..., an interpretable sub-model M, and a fusion layer. When using the integrated explanatory model for interpretation, the feature data of multiple dimensions of the object to be predicted can be input into the integrated explanatory model. First, the interpretable sub-model 1, the interpretable sub-model 2, ..., and the interpretable sub-model M are used to analyze the feature data of multiple dimensions respectively to obtain the pending interpretation result 1, the pending interpretation result 2, ..., and the pending interpretation result M. Then, the pending interpretation result 1, the pending interpretation result 2, ..., and the pending interpretation result M are input into the fusion layer, and the fusion layer is used to fuse the pending interpretation result 1, the pending interpretation result 2, ..., and the pending interpretation result M to obtain the target interpretation result. Accordingly, the fused target interpretation result can be used as the output of the integrated explanatory model for subsequent interpretation of the prediction result.
[0113] Regarding the above embodiments, it should be noted that the present application does not impose any limitation on how to merge multiple pending interpretation results. For ease of understanding, the present application provides the following methods as examples:
[0114] In practical applications, when interpreting the same prediction result, the interpretation accuracy of different interpretable sub-models may be different, resulting in differences in the accuracy of multiple pending interpretation results. In order to improve the interpretation accuracy, in a possible implementation method, corresponding model weights can be determined for multiple interpretable sub-models. The model weights of the interpretable sub-models can be used to reflect the interpretation accuracy of the interpretable sub-models. The interpretable sub-models with higher model weights can be considered to have higher interpretation accuracy, and the corresponding pending interpretation results have higher accuracy. Then, when fusion is performed, a weighted fusion method can be adopted, which is conducive to improving the interpretation accuracy. In the specific implementation, the server can first obtain the model weights corresponding to the multiple interpretable sub-models respectively. The model weights of the interpretable sub-models can be used to reflect the interpretation accuracy of the interpretable sub-models. The interpretable sub-models with higher model weights can be considered to have higher interpretation accuracy. Then, the server can perform weighted fusion on the multiple pending interpretation results based on the model weights corresponding to the multiple interpretable sub-models to obtain the target interpretation result. Based on this, a weighted fusion method is adopted for multiple pending interpretation results, so that the more accurate pending interpretation results are more important to the target interpretation results, which is conducive to improving the interpretation accuracy of the target interpretation results.
[0115] In practical applications, the weighted fusion process can be achieved through the following formula:
[0116]
[0117] In the above formula, y MAIME can refer to the target explanation result, M can refer to the number of interpretable sub-models, and w t It can refer to the model weight of the tth interpretable sub-model among the M interpretable sub-models, y t It may refer to the pending explanation result obtained by the t-th interpretable sub-model. Wherein, M is an integer greater than 1, and t is an integer greater than or equal to 1 and less than or equal to M.
[0118] In practical applications, the weighted fusion method can be used to integrate the model aggregation for interpretable model ensemble (MAIME). Specifically, the MAIME method can be used to weightedly integrate multiple pending explanation results obtained by multiple interpretable sub-models to obtain a comprehensive target explanation result, which can then be used to explain the prediction result. By integrating the prediction results, a more comprehensive target explanation result is used to explain the prediction results, so that sufficient explanations can be provided to facilitate understanding of the prediction results.
[0119] It should be noted that this application does not limit the method of obtaining the model weights corresponding to multiple interpretable sub-models. For ease of understanding, the following method is provided as an example in the embodiment of this application:
[0120] It is understandable that the same interpretability sub-model may have different interpretation accuracy for different prediction scenarios. Therefore, in practical applications, the prediction scenario to which the object to be predicted belongs can be called the prediction scenario, and then the model weights of each interpretability sub-model can be flexibly determined according to the prediction scenario, so that the model weights can more accurately reflect the interpretation accuracy of each interpretability sub-model in the current prediction scenario.
[0121] In the specific implementation, the object to be predicted belongs to the scene to be predicted, and for the scene to be predicted, there is usually existing prior information. The prior information can refer to the existing explanatory experience, knowledge, etc. in the scene to be predicted, which can be used to reflect the characteristics of the scene to be predicted and the accuracy of the interpretation when the explanatory sub-model is applied to the scene to be predicted. Therefore, in a possible implementation method, the server can first obtain the prior information of the scene to be predicted, and then use the prior information to determine the model weights corresponding to multiple interpretable sub-models. Based on this, the model weights of each interpretable sub-model can be determined using the prior information of the scene to be predicted, and the model weights can be determined simply and conveniently. In this way, when the same interpretable sub-model is applied to different scenes to be predicted, the model weights can be re-determined using the prior information corresponding to the scene to be predicted. It is a way to dynamically adjust the model weight so that it can match the current scene to be predicted, which is conducive to improving the accuracy of the interpretation in the current scene to be predicted.
[0122] Among them, the prior information can be determined based on the domain expert knowledge in the field to which the scene to be predicted belongs, the existing interpretation experience in the scene to be predicted, etc. It can be considered as a priori hypothesis with a certain accuracy and credibility, and can be used as a basis for determining the model weight. Based on this, by combining the prior information in the interpretation process, it is helpful to improve the accuracy of the interpretation.
[0123] In practical applications, the aforementioned multiple interpretable sub-models can be models that have been trained and have explanation capabilities. In the case where the object to be predicted belongs to the scene to be predicted, in order to make the model weight more accurately represent the accuracy of the explanation in the scene to be predicted, in another possible implementation, the weight test samples in the scene to be predicted can be used to test multiple interpretable sub-models to evaluate the accuracy of the explanation of each interpretable sub-model in the scene to be predicted, and then determine the corresponding model weight.
[0124] In a specific implementation, the server may first obtain a weighted test sample of the scene to be predicted, and the weighted test sample may include test feature data of multiple dimensions of the test object. The weighted test sample has a test sample label, and the test sample label can be used to represent the overall standard weight of the test feature data of multiple dimensions for the first sample prediction result, and the first sample prediction result can be obtained by predicting the test object through a machine learning model. Among them, the test feature data of multiple dimensions can represent the situation of the test object from different dimensions, and the first sample prediction result can be obtained by predicting the test object through a machine learning model based on the test feature data of multiple dimensions. Then, the server can use each of the multiple interpretable sub-models to analyze the test feature data of multiple dimensions respectively, and obtain multiple test interpretation results. For each of the multiple test interpretation results, the test interpretation result can be used to characterize the overall test weight of the test feature data of multiple dimensions for the first sample prediction result, that is, it can refer to the importance of the test feature data of multiple dimensions determined by the interpretable sub-model for the first sample prediction result. It is understandable that the difference between the test interpretation result and the test sample label can reflect the interpretation accuracy of the interpretability sub-model. Therefore, for each test interpretation result, the server can determine the model weight of the interpretability sub-model corresponding to the test interpretation result based on the difference between the test interpretation result and the test sample label. Accordingly, the model weights corresponding to multiple interpretability sub-models can be obtained. Among them, the smaller the difference between the test interpretation result and the test sample label, the closer the overall test weight represented by the test interpretation result is to the overall standard weight, and it can be considered that the interpretation accuracy of the test interpretation result is higher. Accordingly, it can be determined that the model weight of the interpretability sub-model that obtains the test interpretation result can be higher. That is, the difference can be negatively correlated with the model weight.
[0125] Based on this, by using the weight test samples of the scene to be predicted to test the interpretation accuracy of each interpretable sub-model, it is further ensured that the model weight can more accurately represent the interpretation accuracy of the interpretable sub-model in the scene to be predicted, which is conducive to improving the interpretation accuracy of the target interpretation result. In this way, the same interpretable sub-model can be re-determined by testing when applied to different scenes to be predicted. This is another way to dynamically adjust the model weight so that it can match the current scene to be predicted, which is conducive to improving the interpretation accuracy in the current scene to be predicted.
[0126] It should be noted that this application does not impose any restrictions on the number of weight test samples for the scene to be predicted. For example, in practical applications, in order to improve efficiency, the number of weight test samples can be set to be relatively small so that the model weights can be determined more quickly. Correspondingly, in order to improve accuracy, the number of weight test samples can be set to be relatively large so that the interpretation accuracy of each interpretability sub-model can be more accurately evaluated.
[0127] Usually, the number of analyses using multiple interpretable sub-models to achieve interpretation can be multiple times, and the multiple analyses can be performed sequentially. It is understandable that the same interpretable sub-model may have different pending interpretation results obtained in different analysis processes. Therefore, for the case of multiple analyses, the embodiment of the present application provides another way to dynamically adjust the model weight. Specifically, the model weight can be adjusted according to the pending interpretation result obtained in the previous analysis, and then the adjusted model weight is used for weighted fusion in the next analysis. Based on this, in the entire analysis process, the model weight can be adjusted according to the specific circumstances of each analysis, which is more flexible. In specific implementation, if the aforementioned multiple pending interpretation results correspond to the i-th analysis, i is an integer greater than 1, and when the multiple pending interpretation results obtained for the i-th analysis need to be weighted fused, the aforementioned acquisition of the model weights corresponding to the multiple interpretable sub-models respectively in specific implementation, the server can adjust the model weights corresponding to the multiple interpretable sub-models used in the i-1 analysis according to the multiple pending interpretation results obtained in the i-1 analysis, and obtain the adjusted model weights corresponding to the multiple interpretable sub-models respectively. Then, the server can determine the adjusted model weights corresponding to the multiple interpretable sub-models used in the i-1th analysis as the model weights corresponding to the multiple interpretable sub-models used in the i-th analysis. Based on this, adjusting the model weights used in the i-1th analysis according to the specific circumstances of the i-1th analysis is conducive to improving the model weights used in the i-th analysis to more accurately reflect the interpretation accuracy of the interpretable sub-models. In this way, the model weights can be dynamically adjusted according to the specific circumstances of each analysis during the entire analysis process, so that the model weights used in the next analysis have higher accuracy, which is conducive to improving the interpretation accuracy.
[0128] In specific implementation, the accuracy of multiple pending interpretation results obtained in the i-1th time can be marked by manual marking, so as to adjust the model weight based on the multiple pending interpretation results obtained in the i-1th time. For example, still taking the number of the aforementioned multiple interpretable sub-models as M as an example, if the accuracy of the Mth pending interpretation result obtained in the i-1th time after manual marking is the highest, and the model weight of the Mth interpretable sub-model in the i-1th analysis is not the largest, then during the adjustment, the model weight of the Mth interpretable sub-model can be adjusted upward.
[0129] In the actual prediction process, when predicting different objects to be predicted belonging to the same scene to be predicted, the dimensions of the feature data based on them are usually the same. Correspondingly, different objects to be predicted belonging to the same scene to be predicted can be considered to have commonalities from the perspective of feature granularity. The commonality reflects the importance of feature data of multiple dimensions determined based on the characteristics of the scene to be predicted for the prediction results. In actual applications, there are also individualities between different objects to be predicted. In order to achieve a finer-grained explanation, on the basis of achieving the explanation at the feature granularity, the explanation at the individual granularity can also be performed. Specifically, a target interpretability sub-model that can be analyzed at the individual granularity can be trained, and the target interpretability sub-model can be used to achieve the explanation at the individual granularity later. In other words, the target interpretability sub-model can achieve the explanation at the feature granularity at the level of focusing on the feature data to obtain the pending explanation result. On this basis, the level of individual granularity can also be focused on to achieve the explanation at the individual granularity.
[0130] Specifically, the aforementioned multiple interpretability sub-models may include a target interpretability sub-model. When performing interpretation, the server may use the target interpretability sub-model to analyze the feature data of multiple dimensions to obtain an object interpretation result. Among them, the object interpretation result can be used to characterize the individual weights of the feature data of multiple dimensions of the object to be predicted for the prediction results, and the individual weights are associated with the object to be predicted. For example, for the scene to be predicted, the feature data A, B, and C of three dimensions, then for the object to be predicted 1, the corresponding feature data are A1, B1, and C1, and for the object to be predicted 2, the corresponding feature data are A2, B2, and C2. From the feature granularity point of view, the pending interpretation results obtained can all be the overall weights at the feature granularity. The overall weight can be, for example, the importance of A, B, and C, that is, the overall weight may not be directly related to the object to be predicted, but is associated with the dimension of the feature data. From the perspective of individual granularity, the object interpretation result obtained can be the individual weight at the individual granularity. For the object 1 to be predicted, the individual weight can be, for example, the importance of B1, A1, C1. For the object 2 to be predicted, the individual weight can be, for example, the importance of A2, B2, C2. That is, the individual weight is associated with the object to be predicted and can reflect individual differences. Based on this, the target interpretability sub-model can be used to interpret from the individual granularity, output the importance of the feature data of each individual object to be predicted, and refine the granularity of interpretability to each object to be predicted, so as to achieve a finer-grained explanation and facilitate a better understanding of the prediction results. In this way, for each individual object to be predicted, a personalized explanation at the individual granularity can be provided on the basis of the feature granularity to achieve a personalized explanation. This feature granularity combined with the individual granularity interpretation method, for those prediction scenarios with high interpretability requirements (such as the aforementioned prediction scenario for financial risk control in the financial field), a finer-grained explanation can better assist in understanding the prediction results to ensure the credibility of the prediction results. In addition, according to the needs and characteristics of the prediction scenario, the required explanation granularity can be flexibly selected to achieve customized explanations, which is convenient for better meeting the needs of different scenarios and achieving differentiated effects.
[0131] It should be noted that the present application does not impose any restrictions on the specific type and number of the target interpretability sub-model. For ease of understanding, the present application embodiment provides the following methods as examples:
[0132] Generally, an interpretable model that can achieve explanation at the individual granularity can be selected as the target interpretable sub-model. For example, the aforementioned XGBoost model can be selected as the target interpretable sub-model. Based on this, by directly selecting an interpretable model with explanation capability at the individual granularity, it is beneficial to improve the explanation efficiency.
[0133] In practical applications, in order to improve the accuracy of interpretation at the individual granularity, the number of target interpretability sub-models can be multiple. In this way, multiple target interpretability sub-models can be used to achieve more comprehensive analysis at the individual granularity, which is conducive to improving the accuracy of interpretation at the individual granularity. In specific implementation, the server can first use each target interpretability sub-model in the multiple target interpretability sub-models to analyze the feature data of multiple dimensions respectively, and obtain multiple object interpretation results. Multiple target interpretability sub-models can focus on different levels, and the corresponding multiple object interpretation results can represent the individual weights of the feature data of multiple dimensions of the object to be predicted from different levels for the prediction results. Then, the server can fuse the multiple object interpretation results to obtain the target object interpretation result. Similarly, the target object interpretation result can be used to represent the target individual weights of the feature data of multiple dimensions of the object to be predicted for the prediction results respectively. The difference is that the target object interpretation result is obtained by fusing multiple object interpretation results. Compared with the object interpretation result obtained by a single target interpretability sub-model, the target object interpretation result can more comprehensively represent the individual weight, that is, the target individual weight has higher accuracy. Correspondingly, the target individual weights can be used to more accurately explain the prediction results of the object to be predicted output by the machine learning model.
[0134] After obtaining the target object interpretation result, the target object interpretation result can be used to assist in understanding the prediction result for the object to be predicted. This application does not make any restrictions on how to use the target object interpretation result to interpret the prediction result. For ease of understanding, in an embodiment of the present application, similar to the method of using the target interpretation result to assist the prediction result, the target object interpretation result can be directly displayed to the user so that the user can independently analyze and understand why such a prediction result is output. It is also possible to first analyze the prediction result using the target object interpretation result to obtain the corresponding explanatory result, and then display the explanatory result to the user so that the user can intuitively understand the prediction result.
[0135] For the object to be predicted, the target interpretation result is obtained by analyzing the feature granularity, and the target object interpretation result is obtained by analyzing the individual granularity. In order to better understand how to use the interpretation result (including the target interpretation result and the target object interpretation result) to interpret the prediction result, the method of first using the interpretation result to analyze the prediction result and then displaying the explanatory result to the user is used as an example for explanation. In specific implementation, the server can analyze the prediction result using the target weights of the feature data of multiple dimensions represented by the target interpretation result to obtain the first explanatory result of the prediction result at the feature granularity, and the first explanatory result can be used to indicate why such a prediction result is output based on the situation analysis of the target weights of the feature data of multiple dimensions. Similarly, the server can analyze the prediction result using the individual weights of the feature data of multiple dimensions represented by the object interpretation result to obtain the second explanatory result of the prediction result at the individual granularity, and the second explanatory result can be used to indicate why such a prediction result is output based on the situation analysis of the target individual weights of the feature data of multiple dimensions. After that, the first explanatory result and the second explanatory result can be displayed to the user so that the user can more intuitively understand the prediction result from the feature granularity and the individual granularity.
[0136] In order to better understand the present application, the present application embodiment will take the scene to be predicted to which the object to be predicted belongs as the object classification scene as an example to further introduce the present application. Specifically, the object classification scene may refer to a prediction scene in which the object to be predicted needs to be classified. Correspondingly, the aforementioned machine learning model may be a classification model for predicting the object to be predicted, and the aforementioned prediction result may be a classification result, and the classification result may be used to indicate the category of the object to be predicted. And in the object classification scene, the aforementioned feature data of multiple dimensions may include historical consumption data, browsing data, and search data. Among them, the historical consumption data may reflect the existing consumption of the object to be predicted, the browsing data may reflect the content browsed by the object to be predicted, and the search data may reflect the content to be predicted. When predicting, the classification model may be used to classify the object to be predicted according to the historical consumption data, browsing data, and search data to obtain the classification result, and the classification result may be used to indicate whether the object to be predicted is the target object, that is, whether the category of the object to be predicted is the target object. When explaining, the historical consumption data, browsing data, and search data represented by the target explanation result may be used to analyze the classification result for the target weight of the classification result, and obtain the explanatory result corresponding to the classification result to assist in understanding the classification result. Based on this, for the object to be predicted, whether the object to be predicted is the target object can be determined through prediction, and why it is determined whether the object to be predicted is the target object can be understood through the explanatory results, so that the classification results can be better understood.
[0137] In practical applications, the target object may refer to the type of object concerned in the object classification scenario, and specifically, it may be to take subsequent related application and other measures for the object to be predicted that is the target object. Generally, if the classification result indicates that the object to be predicted is the target object, then subsequent related application and other measures may be taken. For ease of understanding, the embodiment of the present application provides a method of linking with recommendation as an example of the measures adopted subsequently. Specifically, first, the data with the largest target weight can be screened from the historical consumption data, browsing data and search data to be determined as the target data. The target weight is the largest, indicating that the contribution to determining that the object to be predicted is the target object is the largest, and the target data can be considered to be the data dimension that the object to be predicted is most interested in. Therefore, when linking with the recommendation, the information to be recommended that matches the target data in the information to be recommended associated with the object classification scenario can be determined as the recommended information, and the recommended information can be recommended to the object to be predicted. Based on this, the data dimension that the object to be predicted is most likely to be interested in can be evaluated based on the target weight, and then the recommended information matching it can be recommended for it, which is conducive to improving the recommendation efficiency. At the same time, by linking with the recommendation, it is convenient for the object to be predicted to understand the content that he is most interested in more conveniently, which is conducive to improving the user experience.
[0138] For further understanding, the object classification scenario specifically refers to the scenario of classifying e-commerce users and multiple interpretable sub-models including the aforementioned LASSO model, XGBoost model and SHAP model are used as examples to provide a more detailed description as follows:
[0139] Among them, the object to be predicted may refer to an e-commerce user, specifically, a user of an e-commerce platform, and the e-commerce platform may provide e-commerce users with functions such as commodity search, browsing, and purchase. The purpose of classifying e-commerce users is to predict the purchase potential of e-commerce users, so as to provide more accurate sales forecasts and user demand forecasts for the e-commerce platform. Correspondingly, the target object may refer to the type of e-commerce user with relatively large purchase potential, for example, an e-commerce user with relatively large purchase potential may be referred to as a purchasing user. For e-commerce users, in practical applications, the aforementioned historical consumption data may refer to the historical purchase records of e-commerce users on the e-commerce platform, the browsing data may refer to the browsing records of e-commerce users on the e-commerce platform, and the search data may refer to the search records of e-commerce users on the e-commerce platform. When interpreting, the LASSO model, the XGBoost model, and the SHAP model may be used to analyze the historical purchase records, browsing records, and search records, respectively, to obtain the corresponding three pending interpretation results, each of which may be used to characterize the overall weight of the historical purchase records, browsing records, and search records for the prediction of the purchase potential of the object to be predicted, so as to evaluate the influence of the historical purchase records, browsing records, and search records on the prediction of the purchase potential of the object to be predicted.
[0140] Specifically, the LASSO model can be used based on the feature selection method to evaluate the influence of historical purchase records, browsing records, and search records on the purchase potential prediction of the object to be predicted. Similarly, the XGBoost model can be used based on the explanatory weight calculation method to calculate the influence of historical purchase records, browsing records, and search records on the purchase potential prediction of the object to be predicted. And, the SHAP model can be used based on the explanatory weight method to explain the influence of historical purchase records, browsing records, and search records on the purchase potential prediction of the object to be predicted. In practical applications, the greater the degree of influence, the more important it is for the purchase potential prediction, and the greater the corresponding overall weight in the pending explanation result. Among them, the LASSO model analyzes from the level of feature selection, the XGBoost model analyzes from the level of explanatory weights, and the SHAP model simultaneously considers the interaction between historical purchase records, browsing records, and search records. Therefore, by using the LASSO model, XGBoost model, and SHAP model to analyze separately, multiple levels of explanation can be achieved. Finally, the aforementioned MAIME method can be used for fusion. The three pending explanation results obtained by using the LASSO model, XGBoost model, and SHAP model are integrated to obtain a more comprehensive target explanation result, so as to understand why the classification result that the e-commerce user is a purchasing user or the e-commerce user is not a purchasing user is obtained. The explanation is more accurate, which is conducive to improving the credibility of the classification results, and assisting the e-commerce platform to understand the reasons why the e-commerce user is a purchasing user or not a purchasing user based on the e-commerce user's historical purchase, search, and browsing behavior.
[0141] Correspondingly, if the target weight is the largest in the search record, it means that the e-commerce user is more interested in the products that he / she searched for. Then, the e-commerce user can be recommended products that match the search record, so as to realize recommendation linkage, so that the e-commerce user can browse the products that he / she is interested in, which is conducive to improving user experience and user satisfaction. It is understandable that e-commerce users are more likely to buy the products that they are interested in, so this method is also conducive to increasing the sales of products on the e-commerce platform. Based on this, the e-commerce user's historical purchase, search and browsing behavior can be analyzed to determine whether the e-commerce user is a purchasing user, and when it is determined that the e-commerce user is a purchasing user, the product can be recommended to improve product sales and user satisfaction.
[0142] It can be seen from the above technical solution that, first, feature data of multiple dimensions of the object to be predicted can be obtained. The object to be predicted can refer to the object that needs to be predicted, and the feature data of multiple dimensions can represent the situation of the object to be predicted from different dimensions. Then, each of the multiple interpretable sub-models can be used to analyze the feature data of multiple dimensions to obtain multiple pending interpretation results. Each of the multiple pending interpretation results can be used to characterize the overall weight of the feature data of multiple dimensions for the prediction result, wherein the prediction result can be the result obtained by predicting the object to be predicted based on the feature data of multiple dimensions through a machine learning model. Finally, multiple pending interpretation results can be fused to obtain a target interpretation result, and the target interpretation result can be used to characterize the target weight of the feature data of multiple dimensions for the prediction result. Based on this, the explanations are first performed separately through multiple interpretable sub-models, and then fused. Usually, different interpretable sub-models focus on different levels of feature data when interpreting, so a more comprehensive explanation can be achieved based on fusion. Therefore, compared with the pending explanation results of a single interpretable sub-model, the target explanation results can more accurately characterize the weights of feature data of multiple dimensions for the prediction results. In other words, the target weight is more accurate than the overall weight. Accordingly, the target weight can more accurately explain the prediction results output by the machine learning model, thereby improving the accuracy of the explanation and helping to improve the credibility of the prediction results.
[0143] It should be noted that, based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods.
[0144] based on Figure 2 Corresponding to the data processing method provided in the embodiment, the embodiment of the present application further provides a data processing device 400, wherein the data processing device 400 includes an acquisition unit 401, an analysis unit 402, and a fusion unit 403:
[0145] The acquisition unit 401 is used to acquire feature data of multiple dimensions of the object to be predicted;
[0146] The analysis unit 402 is used to analyze the feature data of the multiple dimensions using each of the multiple interpretable sub-models to obtain multiple pending interpretation results; each of the multiple pending interpretation results is used to characterize the overall weight of the feature data of the multiple dimensions for the prediction result, and the prediction result is a result obtained by predicting the object to be predicted through a machine learning model based on the feature data of the multiple dimensions;
[0147] The fusion unit 403 is used to fuse the multiple pending interpretation results to obtain a target interpretation result; the target interpretation result is used to represent the target weights of the feature data of the multiple dimensions for the prediction results.
[0148] In a possible implementation, the multiple interpretability sub-models include a target interpretability sub-model, and the analysis unit is further configured to:
[0149] The target interpretability sub-model is used to analyze the feature data of the multiple dimensions to obtain an object interpretation result; the object interpretation result is used to characterize the individual weights of the feature data of the multiple dimensions of the object to be predicted for the prediction results, and the individual weights are associated with the object to be predicted.
[0150] In a possible implementation, the number of the target interpretability sub-models is multiple, and the analysis unit is further used for:
[0151] Using each target interpretability sub-model in the plurality of target interpretability sub-models to analyze the feature data of the plurality of dimensions respectively, to obtain a plurality of object interpretation results;
[0152] The fusion unit is also used for:
[0153] A plurality of the object interpretation results are fused to obtain a target object interpretation result; the target object interpretation result is used to characterize the target individual weights of the prediction results for the feature data of multiple dimensions of the object to be predicted.
[0154] In a possible implementation, the analysis unit is further configured to:
[0155] Analyzing the prediction result by using the target weights of the feature data of the multiple dimensions represented by the target interpretation result to obtain a first explanatory result of the prediction result at a feature granularity;
[0156] The prediction result is analyzed by using the individual weights of the feature data of the multiple dimensions represented by the object interpretation result to obtain a second explanatory result of the prediction result at an individual granularity.
[0157] In a possible implementation manner, the fusion unit is further configured to:
[0158] Obtaining model weights corresponding to the multiple interpretability sub-models respectively;
[0159] The multiple pending explanation results are weightedly fused based on the model weights corresponding to the multiple interpretability sub-models to obtain the target explanation result.
[0160] In a possible implementation manner, the object to be predicted belongs to a scene to be predicted, and the fusion unit is further used for:
[0161] Obtain a weighted test sample of the scenario to be predicted; the weighted test sample includes test feature data of multiple dimensions of a test object, the weighted test sample has a test sample label, and the test sample label is used to indicate the overall standard weight of the test feature data of the multiple dimensions for a first sample prediction result, respectively, and the first sample prediction result is obtained by predicting the test object through the machine learning model;
[0162] Analyze the test feature data of the multiple dimensions respectively using each of the multiple interpretable sub-models to obtain multiple test interpretation results; the test interpretation results are used to characterize the overall test weights of the test feature data of the multiple dimensions for the first sample prediction result respectively;
[0163] For each test interpretation result, the model weight of the interpretable sub-model corresponding to the test interpretation result is determined according to the difference between the test interpretation result and the test sample label, and the model weights corresponding to the multiple interpretable sub-models are obtained; the difference is negatively correlated with the model weight.
[0164] In a possible implementation manner, the object to be predicted belongs to a scene to be predicted, and the fusion unit is further used for:
[0165] Acquiring prior information of the scene to be predicted;
[0166] The prior information is used to determine model weights corresponding to the multiple interpretable sub-models.
[0167] In a possible implementation, the plurality of pending interpretation results correspond to an i-th analysis, i being an integer greater than 1, and the fusion unit is further configured to:
[0168] According to the multiple pending explanation results obtained in the (i-1)th analysis, the model weights respectively corresponding to the multiple interpretable sub-models used in the (i-1)th analysis are adjusted to obtain the adjusted model weights respectively corresponding to the multiple interpretable sub-models;
[0169] The adjusted model weights respectively corresponding to the multiple interpretable sub-models used in the i-1th analysis are determined as the model weights respectively corresponding to the multiple interpretable sub-models used in the i-th analysis.
[0170] In a possible implementation, the multiple interpretability sub-models are trained in the following manner:
[0171] Obtaining a training sample set of a scene to be predicted to which the object to be predicted belongs; the training sample set includes a plurality of training samples, each training sample includes training feature data of multiple dimensions of the training sample object, each training sample has a training sample label, and the training sample label is used to indicate the overall standard weight of the training feature data of the multiple dimensions for a second sample prediction result, respectively, and the second sample prediction result is obtained by predicting the training sample object through the machine learning model;
[0172] For each of the multiple interpretable sub-models, the multiple training samples are used to perform model training on the interpretable sub-model until the difference between the training explanation result output by the interpretable sub-model and the training sample label meets the training end condition; the training explanation result is used to characterize the overall training weight of the training feature data of the multiple dimensions for the second sample prediction result respectively.
[0173] In a possible implementation manner, the device further includes a dividing unit and a training unit:
[0174] The division unit is used to divide the training sample set into a plurality of training sample subsets; the number of the training sample subsets is the same as the number of the interpretability sub-models;
[0175] The training unit is used to perform model training for each interpretable sub-model among the multiple interpretable sub-models using training samples included in a training sample subset among the multiple training sample subsets; one training sample subset is used for model training of one interpretable sub-model, and different training sample subsets are used for model training of different interpretable sub-models.
[0176] In a possible implementation, the scene to be predicted to which the object to be predicted belongs is an object classification scene, the machine learning model is a classification model, the prediction result is a classification result, the feature data of the multiple dimensions include historical consumption data, browsing data, and search data, and the device further includes a classification unit:
[0177] The classification unit is used to classify the object to be predicted using the classification model according to the historical consumption data, the browsing data and the search data to obtain the classification result; the classification result is used to indicate whether the object to be predicted is a target object;
[0178] The analysis unit is further used to analyze the classification result using the target weights of the historical consumption data, the browsing data and the search data represented by the target interpretation result for the classification result, so as to obtain an explanatory result corresponding to the classification result.
[0179] In a possible implementation manner, if the classification result indicates that the object to be predicted is the target object, the device further includes a determination unit and a recommendation unit:
[0180] The determining unit is used to select data with the largest target weight from the historical consumption data, the browsing data and the search data to determine as target data;
[0181] The determining unit is further configured to determine the information to be recommended that matches the target data among the information to be recommended associated with the object classification scenario as the recommended information;
[0182] The recommendation unit is used to recommend the recommendation information to the object to be predicted.
[0183] It can be seen from the above technical solution that, first, feature data of multiple dimensions of the object to be predicted can be obtained. The object to be predicted can refer to the object that needs to be predicted, and the feature data of multiple dimensions can represent the situation of the object to be predicted from different dimensions. Then, each of the multiple interpretable sub-models can be used to analyze the feature data of multiple dimensions to obtain multiple pending interpretation results. Each of the multiple pending interpretation results can be used to characterize the overall weight of the feature data of multiple dimensions for the prediction result, wherein the prediction result can be the result obtained by predicting the object to be predicted based on the feature data of multiple dimensions through a machine learning model. Finally, multiple pending interpretation results can be fused to obtain a target interpretation result, and the target interpretation result can be used to characterize the target weight of the feature data of multiple dimensions for the prediction result. Based on this, the explanations are first performed separately through multiple interpretable sub-models, and then fused. Usually, different interpretable sub-models focus on different levels of feature data when interpreting, so a more comprehensive explanation can be achieved based on fusion. Therefore, compared with the pending explanation results of a single interpretable sub-model, the target explanation results can more accurately characterize the weights of feature data of multiple dimensions for the prediction results. In other words, the target weight is more accurate than the overall weight. Accordingly, the target weight can more accurately explain the prediction results output by the machine learning model, thereby improving the accuracy of the explanation and helping to improve the credibility of the prediction results.
[0184] The embodiment of the present application further provides a computer device, which may be a terminal. For example, the terminal is a smart phone:
[0185] Figure 5 The block diagram shows a partial structure of a smart phone provided in an embodiment of the present application. Figure 5The smartphone includes: a radio frequency (RF) circuit 510, a memory 520, an input unit 530, a display unit 540, a sensor 550, an audio circuit 560, a wireless fidelity (WiFi) module 570, a processor 580, and a power supply 590. The input unit 530 may include a touch panel 531 and other input devices 532, the display unit 540 may include a display panel 541, and the audio circuit 560 may include a speaker 561 and a microphone 562. Those skilled in the art will appreciate that Figure 5 The structure of the smartphone shown in the figure does not constitute a limitation of the smartphone, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.
[0186] The memory 520 can be used to store software programs and modules. The processor 580 executes various functional applications and data processing of the smartphone by running the software programs and modules stored in the memory 520. The memory 520 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the smartphone (such as audio data, a phone book, etc.), etc. In addition, the memory 520 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.
[0187] The processor 580 is the control center of the smartphone, which uses various interfaces and lines to connect various parts of the entire smartphone, and executes various functions of the smartphone and processes data by running or executing software programs and / or modules stored in the memory 520, and calling data stored in the memory 520. Optionally, the processor 580 may include one or more processing units; preferably, the processor 580 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 580.
[0188] In this embodiment, the steps performed by the processor 580 in the smartphone may be based on Figure 5 The structure shown is implemented.
[0189] The computer device provided in the embodiment of the present application may also be a server. Figure 6 As shown, Figure 6The structural diagram of the server 600 provided in the embodiment of the present application, the server 600 may have relatively large differences due to different configurations or performances, and may include one or more processors, such as a central processing unit (CPU) 622, and a memory 632, one or more storage media 630 (such as one or more mass storage devices) storing application programs 642 or data 644. Among them, the memory 632 and the storage medium 630 can be short-term storage or permanent storage. The program stored in the storage medium 630 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the server. Furthermore, the central processing unit 622 can be configured to communicate with the storage medium 630 and execute a series of instruction operations in the storage medium 630 on the server 600.
[0190] The server 600 may also include one or more power supplies 626, one or more wired or wireless network interfaces 650, one or more input and output interfaces 658, and / or one or more operating systems 641, such as Windows Server 2000 or Windows Server 2003. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM etc.
[0191] In this embodiment, the CPU 622 in the server 600 may perform the following steps:
[0192] Obtain feature data of multiple dimensions of the object to be predicted;
[0193] Using each of the multiple interpretable sub-models to analyze the feature data of the multiple dimensions, respectively, to obtain multiple pending interpretation results; each of the multiple pending interpretation results is used to characterize the overall weight of the feature data of the multiple dimensions for the prediction result, respectively, and the prediction result is a result obtained by predicting the object to be predicted through a machine learning model based on the feature data of the multiple dimensions;
[0194] The multiple pending explanation results are fused to obtain a target explanation result; the target explanation result is used to characterize the target weights of the feature data of the multiple dimensions for the prediction results respectively.
[0195] According to one aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium is used to store a computer program. When the computer program is executed by a computer device, the computer device executes the data processing method described in the above-mentioned embodiments.
[0196] According to one aspect of the present application, a computer program product is provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in various optional implementations of the above-mentioned embodiments.
[0197] The descriptions of the processes or structures corresponding to the above-mentioned figures have different emphases. For parts that are not described in detail in a certain process or structure, please refer to the relevant descriptions of other processes or structures.
[0198] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0199] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.
[0200] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0201] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0202] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the relevant technology or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store program codes.
[0203] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, ordinary technical members in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A data processing method, characterized in that: The method comprises: Obtain feature data of multiple dimensions of the object to be predicted; Using each of the multiple interpretable sub-models to analyze the feature data of the multiple dimensions, respectively, to obtain multiple pending interpretation results; each of the multiple pending interpretation results is used to characterize the overall weight of the feature data of the multiple dimensions for the prediction result, respectively, and the prediction result is a result obtained by predicting the object to be predicted through a machine learning model based on the feature data of the multiple dimensions; The multiple pending explanation results are fused to obtain a target explanation result; the target explanation result is used to characterize the target weights of the feature data of the multiple dimensions for the prediction results respectively.
2. The method according to claim 1, characterized in that The multiple interpretable sub-models include a target interpretable sub-model, and the method further includes: The target interpretability sub-model is used to analyze the feature data of the multiple dimensions to obtain an object interpretation result; the object interpretation result is used to characterize the individual weights of the feature data of the multiple dimensions of the object to be predicted for the prediction results, and the individual weights are associated with the object to be predicted.
3. The method according to claim 2, characterized in that The number of the target interpretability sub-models is multiple, and the target interpretability sub-models are used to analyze the feature data of the multiple dimensions to obtain the object interpretation result, including: Using each target interpretability sub-model in the plurality of target interpretability sub-models to analyze the feature data of the plurality of dimensions respectively, to obtain a plurality of object interpretation results; The method further comprises: A plurality of the object interpretation results are fused to obtain a target object interpretation result; the target object interpretation result is used to characterize the target individual weights of the prediction results for the feature data of multiple dimensions of the object to be predicted.
4. The method according to claim 2, characterized in that: The method further comprises: Analyzing the prediction result by using the target weights of the feature data of the multiple dimensions represented by the target interpretation result to obtain a first explanatory result of the prediction result at a feature granularity; The prediction result is analyzed by using the individual weights of the feature data of the multiple dimensions represented by the object interpretation result to obtain a second explanatory result of the prediction result at an individual granularity.
5. The method according to claim 1, characterized in that The step of fusing the plurality of pending interpretation results to obtain a target interpretation result includes: Obtaining model weights corresponding to the multiple interpretability sub-models respectively; The multiple pending explanation results are weightedly fused based on the model weights corresponding to the multiple interpretability sub-models to obtain the target explanation result.
6. The method according to claim 5, characterized in that The object to be predicted belongs to a scene to be predicted, and obtaining the model weights corresponding to the multiple interpretability sub-models respectively includes: Obtain a weighted test sample of the scenario to be predicted; the weighted test sample includes test feature data of multiple dimensions of a test object, the weighted test sample has a test sample label, and the test sample label is used to indicate the overall standard weight of the test feature data of the multiple dimensions for a first sample prediction result, respectively, and the first sample prediction result is obtained by predicting the test object through the machine learning model; Analyze the test feature data of the multiple dimensions respectively using each of the multiple interpretable sub-models to obtain multiple test interpretation results; the test interpretation results are used to characterize the overall test weights of the test feature data of the multiple dimensions for the first sample prediction result respectively; For each test interpretation result, the model weight of the interpretable sub-model corresponding to the test interpretation result is determined according to the difference between the test interpretation result and the test sample label, and the model weights corresponding to the multiple interpretable sub-models are obtained; the difference is negatively correlated with the model weight.
7. The method according to claim 5, characterized in that The object to be predicted belongs to a scene to be predicted, and obtaining the model weights corresponding to the multiple interpretability sub-models respectively includes: Acquiring prior information of the scene to be predicted; The prior information is used to determine model weights corresponding to the multiple interpretable sub-models.
8. The method according to claim 5, characterized in that The multiple pending explanation results correspond to the i-th analysis, where i is an integer greater than 1, and the obtaining of model weights corresponding to the multiple interpretability sub-models respectively includes: According to the multiple pending explanation results obtained in the (i-1)th analysis, the model weights respectively corresponding to the multiple interpretable sub-models used in the (i-1)th analysis are adjusted to obtain the adjusted model weights respectively corresponding to the multiple interpretable sub-models; The adjusted model weights respectively corresponding to the multiple interpretable sub-models used in the i-1th analysis are determined as the model weights respectively corresponding to the multiple interpretable sub-models used in the i-th analysis.
9. The method according to any one of claims 1 to 8, characterized in that: The multiple interpretable sub-models are trained in the following way: Obtaining a training sample set of a scene to be predicted to which the object to be predicted belongs; the training sample set includes a plurality of training samples, each training sample includes training feature data of multiple dimensions of the training sample object, each training sample has a training sample label, and the training sample label is used to indicate the overall standard weight of the training feature data of the multiple dimensions for a second sample prediction result, respectively, and the second sample prediction result is obtained by predicting the training sample object through the machine learning model; For each of the multiple interpretable sub-models, perform model training on the interpretable sub-model using the multiple training samples until the difference between the training explanation result output by the interpretable sub-model and the training sample label meets the training end condition; The training explanation result is used to characterize the overall training weights of the training feature data of the multiple dimensions for the second sample prediction result respectively.
10. The method according to claim 9, characterized in that Before performing model training on the interpretability sub-model using the training sample, the method includes: Dividing the training sample sample set into a plurality of training sample subsets; the number of the training sample subsets is the same as the number of the interpretability sub-models; The step of performing model training on each of the multiple interpretable sub-models by using the training sample comprises: For each interpretable sub-model among the multiple interpretable sub-models, model training is performed using the training samples included in a training sample subset among the multiple training sample subsets; one training sample subset is used for model training of one interpretable sub-model, and different training sample subsets are used for model training of different interpretable sub-models.
11. The method according to claim 1, characterized in that: The to-be-predicted scene to which the to-be-predicted object belongs is an object classification scene, the machine learning model is a classification model, the prediction result is a classification result, the feature data of the multiple dimensions include historical consumption data, browsing data, and search data, and the method further includes: According to the historical consumption data, the browsing data and the search data, the object to be predicted is classified using the classification model to obtain the classification result; the classification result is used to indicate whether the object to be predicted is a target object; The classification results are analyzed using the target weights of the historical consumption data, the browsing data and the search data represented by the target explanation results for the classification results, to obtain the explanation results corresponding to the classification results.
12. The method according to claim 11, characterized in that If the classification result indicates that the object to be predicted is the target object, the method further includes: Filtering data with the largest target weight from the historical consumption data, the browsing data, and the search data to determine as target data; Determining the information to be recommended associated with the object classification scenario and matching the target data as recommended information; The recommendation information is recommended to the object to be predicted.
13. A data processing device, characterized in that: The device comprises an acquisition unit, an analysis unit and a fusion unit: The acquisition unit is used to acquire feature data of multiple dimensions of the object to be predicted; The analysis unit is used to analyze the feature data of the multiple dimensions respectively using each of the multiple interpretable sub-models to obtain multiple pending interpretation results; Each of the multiple pending interpretation results is used to characterize the overall weight of the feature data of the multiple dimensions for the prediction result, and the prediction result is the result obtained by predicting the object to be predicted through a machine learning model based on the feature data of the multiple dimensions; The fusion unit is used to fuse the multiple pending interpretation results to obtain a target interpretation result; The target interpretation result is used to characterize the target weights of the feature data of the multiple dimensions for the prediction results respectively.
14. A computer device, characterized in that: The computer device comprises a processor and a memory: The memory is used to store a computer program and transmit the computer program to the processor; The processor is configured to execute the method according to any one of claims 1 to 12 according to the instructions in the computer program.
15. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and when the computer program is executed by a computer device, the computer device executes the method according to any one of claims 1 to 12.
16. A computer program product comprising a computer program, characterized in that When the method is executed on a computer device, the computer device is enabled to execute the method according to any one of claims 1 to 12.
