Attribution analysis method and device based on feature data and electronic equipment

Abstract

Embodiments of the present disclosure disclose a feature data-based attribution analysis method, device and electronic equipment. A specific implementation of the method includes: generating training data according to feature data belonging to in-module features in a swap-in data set and a swap-out data set, wherein the in-module features are in-module features of an original decision model that appears a swap-in swap-out phenomenon; training a preset classification model through the training data, and determining index data of each in-module feature of the original decision model based on the trained preset classification model; selecting a set number of in-module features from each in-module feature of the original decision model as target features according to the determined index data; and analyzing key features that cause the original decision model to produce swap-in swap-out data based on the target features, the swap-in data set and the swap-out data set. The implementation is related to model bottom layer analysis technology, and reveals key features that cause swap-in swap-out by analyzing in-module features of the original decision model.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of model underlying analysis technology, and specifically to attribution analysis methods, apparatus and electronic devices based on feature data. Background Technology

[0002] With the development of artificial intelligence technology, various machine learning and deep learning models are being increasingly widely applied in various production environments and playing an increasingly important role in various business scenarios. However, in many business scenarios, the replacement of old and new strategy systems (or models) can lead to a "swap-in" or "swap-out" phenomenon. Swap-in typically refers to data that passes the new strategy but fails under the old strategy; swap-out typically refers to data that fails under the new strategy but passes under the old strategy. Attribution analysis of this swap-in / swap-out phenomenon is crucial for explaining model decisions and analyzing model performance.

[0003] However, the inventors discovered that current related technologies are basically just statistics on the number of swaps in and out, lacking necessary and accurate attribution analysis, that is, what characteristics mainly cause swaps in and out.

[0004] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure provide attribution analysis methods, devices, electronic devices, computer-readable media, and computer program products based on feature data to solve one or more of the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide an attribution analysis method based on feature data, including: generating training data based on feature data belonging to the input features in the input dataset and the output dataset, wherein the input features are the input features of the original decision model in which the input and output phenomena occur; training a preset classification model using the training data, and determining the index data of each input feature of the original decision model based on the trained preset classification model, wherein the index data is used to characterize the importance of the features; selecting a set number of input features from each input feature of the original decision model based on the determined index data as target features; and analyzing the key features that cause the original decision model to generate input and output data based on the target features, the input dataset, and the output dataset.

[0008] In some embodiments, based on the target features, the input dataset, and the output dataset, the key features that cause the original decision model to generate input and output data are analyzed, including: for each target feature, determining the probability distribution density of the feature data belonging to the target feature in the input and output datasets; determining the population stability index value of the input and output datasets under the target feature based on the probability distribution density of the feature data of each target feature; and determining the target features corresponding to the population stability index values ​​within a preset range as key features with large distribution differences based on the population stability index values ​​of the input and output datasets under each target feature.

[0009] In some embodiments, based on target features, the input dataset, and the output dataset, the key features that cause the original decision model to generate input and output data are analyzed, and the method further includes: for each target feature in the target features, determining a first preset difference between the input dataset and the output dataset under that target feature, wherein the first preset difference is used to characterize the degree of distinction between the input and output data; and selecting a certain number of target features from the target features according to the first preset difference as key features for distinguishing the input and output data.

[0010] In some embodiments, the method further includes: for each key feature, determining the numerical range of the key feature based on feature data belonging to the key feature in the input dataset and output dataset, dividing the numerical range into multiple numerical groups of the key feature at equal intervals; for each numerical group, determining a second preset difference between the input dataset and the output dataset located in the numerical group; and selecting a numerical group from the multiple numerical groups based on the second preset difference as the key feature to distinguish the key numerical range of the input and output data.

[0011] In some embodiments, the preset classification model is a distributed gradient enhancement library XGB tree model, and based on the preset classification model that has been trained, the indicator data of each input feature of the original decision model are determined, including: obtaining the relevant indicator coefficients in the XGB tree model to obtain the importance of each input feature of the original decision model.

[0012] In some embodiments, for each target feature among the target features, determining a first preset difference between the input dataset and the output dataset under that target feature includes: for each target feature among the target features, determining the total input prediction score of the input dataset under that target feature based on the index coefficient of the target feature and the predicted score of each input data in the input dataset by the XGB tree model; determining the total output prediction score of the output dataset under that target feature based on the index coefficient of the target feature and the predicted score of each output data in the output dataset by the XGB tree model; and using the difference between the total input prediction score and the total output prediction score of each target feature as the first preset difference.

[0013] In some embodiments, determining the index data of each input feature of the original decision model based on the trained preset classification model further includes: determining the contribution value of each input feature to the prediction result of the preset classification model by using a machine learning model interpretation algorithm according to the trained preset classification model.

[0014] In some embodiments, for each target feature among the target features, determining a first preset difference between the input dataset and the output dataset under that target feature further includes: in response to determining that the input dataset and the output dataset have the same amount of data, for each target feature among the target features, determining the average contribution of the target feature on the input dataset as the input contribution average, and determining the average contribution of the target feature on the output dataset as the output contribution average; and using the difference between the input contribution average and the output contribution average of each target feature as the first preset difference.

[0015] In some embodiments, training data is generated based on feature data belonging to the input model features in the input dataset and the output dataset, including: determining the risk indicator features at each classification node in the original decision model as the input model features of the original decision model, wherein the original decision model is used for predicting user risk; using user attribute data belonging to the input model features in the input dataset and the output dataset as training samples, and labeling the training samples according to the dataset to which the training samples belong to generate training data; wherein the input dataset stores the input user attribute data, and the output dataset stores the output user attribute data.

[0016] In some embodiments, the method further includes: monitoring target user data based on the analyzed key features, wherein the target user data is user attribute data belonging to the key features; and adjusting and training the original decision model using the target user data in response to determining that the target user data has shifted.

[0017] Secondly, some embodiments of this disclosure provide an attribution analysis apparatus based on feature data, comprising: a training data generation unit configured to generate training data based on feature data belonging to input features in the input dataset and the output dataset, wherein the input features are the input features of the original decision model in which input and output phenomena occur; a model training unit configured to train a preset classification model using the training data, and to determine index data of each input feature of the original decision model based on the trained preset classification model, wherein the index data is used to characterize the importance of the features; a target feature selection unit configured to select a predetermined number of input features from each input feature of the original decision model based on the determined index data, as target features; and a key feature analysis unit configured to analyze the key features that cause the original decision model to generate input and output data based on the target features, the input dataset, and the output dataset.

[0018] In some embodiments, the key feature analysis unit includes a first analysis subunit, configured to, for each target feature among the target features, determine the probability distribution density of feature data belonging to the target feature in the incoming and outgoing datasets; determine the population stability index value of the incoming and outgoing datasets under the target feature based on the probability distribution density of the feature data of each target feature; and determine the target features corresponding to the population stability index values ​​within a preset range as key features with large distribution differences based on the population stability index values ​​of the incoming and outgoing datasets under each target feature.

[0019] In some embodiments, the key feature analysis unit further includes a second analysis subunit, configured to determine, for each target feature among the target features, a first preset difference between the swap-in dataset and the swap-out dataset under that target feature, wherein the first preset difference is used to characterize the degree of distinction between the swap-in and swap-out data; and to select a certain number of target features from the target features according to the first preset difference as key features for distinguishing the swap-in and swap-out data.

[0020] In some embodiments, the attribution analysis apparatus further includes a numerical range analysis unit, configured to, for each key feature among the key features, determine the numerical range of the key feature based on feature data belonging to the key feature in the input dataset and output dataset, divide the numerical range into equal intervals to obtain multiple numerical groups of the key feature; for each numerical group among the multiple numerical groups, determine a second preset difference between the input dataset and the output dataset located in the numerical group; and select a numerical group from the multiple numerical groups based on the second preset difference as the key numerical range for distinguishing the input and output data.

[0021] In some embodiments, the preset classification model is a distributed gradient enhancement library XGB tree model, and the model training unit is further configured to obtain the relevant index coefficients in the XGB tree model to obtain the importance of each input feature of the original decision model.

[0022] In some embodiments, the second analysis subunit is further configured to, for each target feature among the target features, determine the total predicted score of the input dataset under the target feature by the XGB tree model for each input data in the input dataset based on the index coefficient of the target feature; determine the total predicted score of the output dataset under the target feature by the XGB tree model for each output data in the output dataset based on the index coefficient of the target feature; and use the difference between the total predicted score of the input and the total predicted score of the output for each target feature as a first preset difference.

[0023] In some embodiments, the model training unit is further configured to determine the contribution value of each input feature to the prediction result of the preset classification model by using a machine learning model interpretation algorithm based on the preset classification model that has been trained.

[0024] In some embodiments, the second analysis subunit is further configured to, in response to determining that the amount of data in the input dataset and the output dataset are the same, for each target feature among the target features, determine the average contribution of the target feature on the input dataset as the input contribution average, and determine the average contribution of the target feature on the output dataset as the output contribution average; and take the difference between the input contribution average and the output contribution average of each target feature as a first preset difference.

[0025] In some embodiments, the training data generation unit is further configured to determine the risk indicator features at each classification node in the original decision model as the input features of the original decision model, wherein the original decision model is used for predicting user risk; to use user attribute data belonging to the input features in the input dataset and the output dataset as training samples, and to label the training samples according to the dataset to which the training samples belong to generate training data; wherein the input dataset stores the input user attribute data, and the output dataset stores the output user attribute data.

[0026] In some embodiments, the attribution analysis apparatus further includes a monitoring unit configured to monitor target user data based on the analyzed key features, wherein the target user data is user attribute data belonging to the key features; and in response to determining that the target user data has shifted, to adjust and train the original decision model using the target user data.

[0027] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the attribution analysis method based on feature data as described in any of the implementations of the first aspect above.

[0028] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the attribution analysis method based on feature data as described in any of the implementations of the first aspect above.

[0029] Fifthly, some embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the attribution analysis method based on feature data as described in any of the implementations of the first aspect above.

[0030] The various embodiments of this disclosure have the following beneficial effects: The attribution analysis method based on feature data in some embodiments of this disclosure can analyze the feature data of the input and output datasets on the input model features, thereby explaining which features play an important role in causing the input and output. Specifically, training data can be generated based on the feature data belonging to the input model features in the input and output datasets, thereby training a preset classification model. Since only the input and output data are considered, using only these data for model training allows the model to fully learn and regress the input and output data, reducing or avoiding the influence of other data.

[0031] Next, based on the trained pre-defined classification model, the indicator data of each input feature of the original decision model can be determined. Referring to this indicator data, the target features with higher importance among these input features can be identified. These target features are also the prediction results of the pre-defined classification model, i.e., the features predicted as input or output data that contribute significantly. Thus, based on the target features, the input dataset, and the output dataset, the key features that cause the original decision model to generate input / output data can be analyzed, i.e., which features play a crucial role in input / output. The method disclosed herein can fully explore the inherent relationships between the input features in the decision model, especially the features that have a significant impact on the input / output phenomenon, as these features often affect the model's prediction results. This allows for targeted adjustment and monitoring of the model's input features, thereby reducing input / output phenomena (i.e., the amount of input / output data) and improving the reliability and accuracy of the decision model. Attached Figure Description

[0032] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0033] Figure 1 This is a flowchart of some embodiments of the attribution analysis method based on feature data disclosed herein;

[0034] Figure 2 These are flowcharts of some other embodiments of the attribution analysis method based on feature data disclosed herein;

[0035] Figure 3A This is a schematic diagram of feature importance ranking obtained based on the XGB tree model;

[0036] Figure 3B This is a schematic diagram of feature importance ranking based on SHAP;

[0037] Figure 3C This is a schematic diagram illustrating the distribution differences of a single feature across the input and output datasets.

[0038] Figure 3D This is a schematic diagram of the PSI ranking of the input and output datasets under various features;

[0039] Figure 3E This is a diagram illustrating the difference between the average SHAP values ​​of the input and output datasets under various features.

[0040] Figure 3F From Figure 3E A schematic diagram showing the key features that have made significant contributions;

[0041] Figure 3G This is a schematic diagram illustrating the contribution of key features to distinguishing between the incoming and outgoing datasets under each numerical group;

[0042] Figure 4 This is a schematic diagram of the structure of some embodiments of the feature-based attribution analysis device disclosed herein;

[0043] Figure 5 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0044] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0045] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0046] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0047] Furthermore, the terms “a” and “a plurality” used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as “one or more”.

[0048] Figure 1 A flow 100 is shown, illustrating some embodiments of a feature-based attribution analysis method according to this disclosure. The method may include the following steps:

[0049] Step 101: Generate training data based on the feature data belonging to the input model features in the input dataset and the output dataset.

[0050] In some embodiments, the execution entity (e.g., an analysis server) of the feature-based attribution analysis method disclosed herein can communicate with other electronic devices (such as user terminals, databases, etc.) via wired or wireless connections. Here, the user can send data from the input and output datasets to the execution entity via a terminal. Alternatively, the execution entity can retrieve the input and output datasets from a database.

[0051] Here, the executing entity can identify the feature data belonging to the input model features in both the input and output datasets. This feature data can then be used as training samples. Furthermore, the training samples can be labeled according to their respective datasets. For example, the input dataset might be labeled with 1, and the output dataset with 0, thus obtaining the training data. Optionally, the executing entity can also select a certain proportion of the feature data from these datasets as training samples, and use the remaining proportion as test samples.

[0052] It should be noted that this embodiment only considers features related to the original decision model. That is, it only examines the impact of the input features of the original decision model on input / output changes. Therefore, the input features mentioned above are usually the input features of the original decision model where input / output changes occur, because only these features can cause such changes. Input features are usually inherent properties of the model. After model training is complete, the input features of the model are already determined. The executing entity can determine the attribute features at each classification node in the original decision model as the input features of the original decision model. It can be understood that in a tree model, each branch node typically performs binary classification on the data at that node based on the values ​​of one or more feature parameters. The features analyzed by these branch nodes are the input features of the model.

[0053] In some application scenarios, if the original decision model is used for user risk prediction, the risk indicator features at each classification node in the original decision model can be determined as the input features of the original decision model. In this case, the input dataset typically stores the input user attribute data, while the output dataset typically stores the output user attribute data. Here, the executing entity can use the user attribute data belonging to the input features from both datasets as training samples.

[0054] In some embodiments, if the input features of the new and old models are different, the input features that are common to both models can be selected as the input features mentioned above.

[0055] Step 102: Train the preset classification model using training data, and determine the index data of each input feature of the original decision model based on the trained preset classification model.

[0056] In some embodiments, the executing entity can train a preset classification model using the training data generated in step 101. The preset classification model can be any model capable of classifying and predicting input and output data, such as a classification model that outputs Boolean values ​​or a classification decision model that outputs categories, such as a binary tree model.

[0057] In some embodiments, the preset classification model can be an XGB tree model from a distributed gradient boosting library. XGB, or XGBoost (eXtremeGradientBoosting), is a variant of the GBDT algorithm, a supervised algorithm, a type of boost algorithm, and also an ensemble algorithm. In this case, the XGB tree model can be trained using training samples. Furthermore, the loss function value of the XGB tree model can be determined based on the prediction results and corresponding sample labels of the training samples, thereby adjusting the model parameters until training is complete.

[0058] In some embodiments, the executing entity can also determine the indicator data of each input feature of the original decision model based on the trained preset classification model. The indicator data can be used to characterize the importance of the features. For example, for the XGB tree model described above, the executing entity can obtain the relevant indicator coefficients in the XGB tree model to obtain the importance of each input feature of the original decision model. That is, the obtained relevant indicator coefficients can be used as indicator data. These relevant indicator coefficients can be, for example, the total gain coefficient; the larger the number, the greater the classification contribution. This allows the ranking of the importance of each input feature of the original decision model to be obtained, such as... Figure 3A As shown.

[0059] Optionally, based on the trained pre-defined classification model, the executing entity can also employ a machine learning model interpretation algorithm (SHAP, SHapley Additive exPlanations) to determine the contribution value of each input feature to the prediction result of the pre-defined classification model, such as... Figure 3B As shown. In other words, a defined contribution value can also be used as metric data. SHAP is typically used as a unified method to interpret the output of any machine learning model. The core concept is the Shapley value. The Shapley value can be used to answer the question: In a team project, how much does each member contribute? Note that contribution here is not only positive but also negative. Applied to machine learning models, it can explain how much each sample and each feature contributes to the corresponding predicted value.

[0060] Step 103: Based on the determined indicator data, select a set number of input features from the input features of the original decision model as target features.

[0061] It should be noted that the above metrics provide a ranking of feature importance based on their contribution to distinguishing between the input and output datasets. This feature importance ranking focuses on the feature data of the two datasets themselves, without considering the model's decision-making process; therefore, it is only a preliminary analysis of data differentiation. Subsequent analyses should focus on the top N feature data.

[0062] In some embodiments, the executing entity may select a predetermined number of input features from the input features of the original decision model based on the indicator data determined in step 102, as target features. For example, the executing entity may select a predetermined number of input features as target features in descending order of indicator data values ​​(i.e., importance). This predetermined number can be set according to actual needs, such as 10, 20, or 30.

[0063] It is understandable that, as described above, some input features have a positive effect on distinguishing between input and output data, i.e., they increase the degree of difference between the two. Conversely, some input features have a negative effect on distinguishing between input and output data, i.e., they decrease or weaken the degree of difference between the two. Therefore, in some embodiments, the executing entity can first determine the input features that have a positive effect among the input features of the original decision model based on the indicator data. Then, for these positively effective input features, the executing entity can continue to select a set number of target features from them, in descending order of indicator data values.

[0064] Step 104: Based on the target features, the input dataset, and the output dataset, analyze the key features that cause the original decision model to generate input and output data.

[0065] In some embodiments, based on the target features selected in step 103, and the input and output datasets, the executing entity can analyze the key features that cause the original decision model to generate the input and output data. As an example, the executing entity can further select fewer than the aforementioned set number of target features as key features from the target features, such as... Figure 3A and 3B The top two examples shown are zijie_final and ali_final. For instance, for each target feature, the executing entity can also count the amount of data belonging to that target feature in the input and output datasets. Then, key features can be selected from these target features in descending order of data volume, such as the top five target features.

[0066] In some embodiments, the executing entity can further analyze the target features based on the input and output datasets to extract key features. Specifically, for each target feature, the probability distribution density of the feature data belonging to that target feature in the input and output datasets can first be determined. For example... Figure 3C As shown, this diagram illustrates the distribution differences of a single feature, ali_final, across sets A and B. Set A represents the incoming dataset; set B represents the outgoing dataset; Density represents the probability distribution density; and the horizontal axis represents the numerical values ​​of the feature data. Let P be the probability that a continuous random variable falls within a certain interval, and Δx > 0 be the length of the interval. The ratio of P / Δx is called the "average probability distribution density" of the random variable within that interval. If the limit of this ratio exists as the interval length Δx → 0, then this limit is called the probability distribution density of the random variable at point x, or simply the distribution density.

[0067] Next, the executing entity can determine the Population Stability Index (PSI) of the incoming and outgoing datasets under the target feature based on the probability distribution density of the feature data for each target feature. PSI typically measures data stability by comparing the fluctuations of historical and current sample distributions. It is generally used to quantitatively assess the distribution differences between the observed (actual, equivalent to the incoming dataset) and the development (expect, equivalent to the outgoing dataset) samples. This method usually involves first dividing the two samples into n groups according to certain rules (equal width, equal frequency), and then comparing the similarity of their distributions. That is: PSI = SUM[(observed sample percentage - development sample percentage) * ln(observed sample percentage / development sample percentage)]. This yields... Figure 3D The diagram shows the PSI ranking of the input and output datasets under each feature. As can be seen from the diagram, some features in the input and output datasets have significantly different distributions, while others have relatively small differences. When distinguishing between the input and output datasets, it is usually necessary to focus on the features with larger distribution differences.

[0068] Furthermore, based on the population stability index (PSI) values ​​of the incoming and outgoing datasets for each target feature, the executing entity can identify the target features corresponding to PSI values ​​within a preset range as key features with significant distributional differences. According to PSI empirical values: PSI values ​​between 0.25 and 1.0 typically indicate features with large distributional differences, such as... Figure 3D Left-side characteristics; PSI values ​​between 0.1 and 0.25 typically indicate moderate distributional differences, such as... Figure 3D Intermediate characteristics; PSI values ​​between 0 and 0.1 typically indicate characteristics with small distributional differences, such as... Figure 3D The right-hand side features. Here, the executing entity can set the preset range to 0.25-1.0. In this case, the executing entity can identify target features with PSI values ​​between 0.25 and 1.0 as key features.

[0069] As described above, the attribution analysis method based on feature data in some embodiments of this disclosure can analyze the feature data of the input and output datasets on the input model features, thereby explaining which features play an important role in causing the input and output. Specifically, training data can be generated based on the feature data belonging to the input model features in the input and output datasets, thereby training a preset classification model. Since only the input and output data are considered, using only these data for model training allows the model to fully learn and regress the input and output data, reducing or avoiding the influence of other data.

[0070] Next, based on the trained pre-defined classification model, the indicator data of each input feature of the original decision model can be determined. Referring to this indicator data, the target features with higher importance among these input features can be identified. These target features are also the prediction results of the pre-defined classification model, i.e., the features predicted as input or output data that contribute significantly. Thus, based on the target features, the input dataset, and the output dataset, the key features that cause the original decision model to generate input / output data can be analyzed, i.e., which features play a crucial role in input / output. The method disclosed herein can fully explore the inherent relationships between the input features in the decision model, especially the features that have a significant impact on the input / output phenomenon, as these features often affect the model's prediction results. This allows for targeted adjustment and monitoring of the model's input features, thereby reducing input / output phenomena (i.e., the amount of input / output data) and improving the reliability and accuracy of the decision model.

[0071] In some application scenarios, the executing entity can also use the following analysis method to determine the key features that distinguish between the swapped-in and swapped-out data. For each target feature, a first preset difference between the swapped-in and swapped-out datasets under that target feature can be determined. This first preset difference can characterize the degree of distinction between the swapped-in and swapped-out data and can be set according to the actual situation. Next, a certain number of target features can be selected from the target features based on the first preset difference as the key features for distinguishing between the swapped-in and swapped-out data. This certain number is not limited.

[0072] In some embodiments, the first preset difference can be the difference in the total predicted scores of the XGB tree models of the input and output datasets. Specifically, for each target feature, the executing entity can determine the total predicted score of the input dataset under that target feature based on the index coefficient of that target feature and the predicted scores of each input data in the input dataset by the XGB tree model. It is understood that after inputting each input data into the XGB tree model, a predicted score for that input data is obtained. Multiplying the predicted score of each input data by the index coefficient of each target feature yields the specific predicted score for each input data under that target feature. Thus, the sum of the feature predicted scores of all input data in the input dataset can be the total predicted score of the input dataset under each target feature.

[0073] Similarly, for each target feature, based on its index coefficient and the predicted scores of each exit data point in the exit dataset by the XGB tree model, the executing entity can determine the total predicted score for the exit dataset under that target feature. Then, the difference between the total predicted score for the entry and the total predicted score for the exit under each target feature can be used as the first preset difference.

[0074] Optionally, the first preset difference can also be the difference between the average contributions of each target feature in the input and output datasets. Specifically, assuming the input and output datasets have the same amount of data, for each target feature, the executing entity can determine the average contribution of that target feature in the input dataset as the input contribution average; and can determine the average contribution of that target feature in the output dataset as the output contribution average. Then, the difference between the input contribution average and the output contribution average for each target feature can be used as the first preset difference.

[0075] In other words, when the amount of data in the input and output datasets is the same, we can also use the difference between the average SHAP values ​​of each feature on the input and output datasets to measure the contribution of each feature to the splitting of the input and output datasets. We calculate the difference between the average SHAP value of each feature on dataset A (input) and the average SHAP value on dataset B (output), and then sort them, as follows: Figure 3E As shown. Figure 3F That is, from Figure 3E Key features that ranked highly and made significant contributions were selected. Figure 3E and 3F It can be seen that some features have a positive effect on the splitting of the input and output datasets, even if the model predicts a larger split between the input and output datasets, while other features have the opposite effect.

[0076] Understandably, by analyzing the difference in the model's predicted total score for each target feature between the two datasets, or the difference in the SHAP average of each target feature with respect to the two datasets, the key features that cause the original decision model to generate input and output data can be further selected, thereby improving the accuracy of the analysis and selection results.

[0077] See also Figure 2 This illustrates a flow 200 of another embodiment of the feature-based attribution analysis method according to this disclosure. The method may include the following steps:

[0078] Step 201: For each key feature, determine the numerical range of the key feature based on the feature data belonging to the key feature in the input and output datasets.

[0079] In some embodiments, for the identified key features, the impact of these key features on the input and output data of the original decision model under different numerical ranges can be further analyzed. Specifically, the execution subject of the feature data-based attribution analysis method of this disclosure, for each key feature, first determines the numerical range of the key feature based on the feature data belonging to the key feature in the input and output datasets. As an example, the execution subject can compare the numerical ranges of the key feature in the input and output datasets respectively. Thus, the numerical range with a relatively larger numerical range can be taken as the numerical range of the key feature. Optionally, since the analysis mainly distinguishes the data that contributes significantly to the input and output data, the numerical range with a relatively smaller numerical range, i.e., the numerical range that exists in both datasets, can also be taken as the numerical range of the key feature. This can reduce the amount of data to be processed and improve the analysis efficiency.

[0080] Step 202: Divide the numerical range into equal intervals to obtain multiple numerical groups of the key feature.

[0081] In some embodiments, the executing entity can divide the numerical range of a key feature into equally spaced intervals to obtain multiple numerical groups for that key feature. Here, the number of equally spaced intervals can be adjusted according to the numerical range of each feature, and is not limited thereto. Figure 3G As shown, the key feature ali_final can be divided into 10 numerical groups with an equal interval of 9.1.

[0082] Step 203: For each of the multiple sets of values, determine a second preset difference between the incoming and outgoing datasets within that set of values.

[0083] In some embodiments, for each of a plurality of numerical groups, the executing entity may determine a second preset difference between the incoming and outgoing datasets within that numerical group. This second preset difference can also be used to characterize the degree of distinction between the incoming and outgoing data, and may be the same as or different from the first preset difference. As an example, the second preset difference may be the difference in the model's predicted total score for the incoming and outgoing datasets within each numerical group under this key feature.

[0084] Optionally, the second preset difference can also be the difference between the sum of the contribution values ​​of the key features to the input and output datasets within each numerical group. This sum of contribution values ​​can be the sum of SHAP values ​​or the sum of SHAP values ​​plus SHAPbase values. It should be noted that for each predicted sample, the model generates a predicted value, and the SHAP value is the numerical value assigned to each feature in that sample. Assume the i-th sample is x. i The j-th feature of the i-th sample is x. i,j The model predicts the value yi for the i-th sample. The baseline of the entire model (usually the mean of the target variable for all samples) is ybas. e Therefore, the SHAP value usually follows the following equation.

[0085] y i =y base +f(x i，1 )+f(x i，2 )+…+f(;

[0086] Wherein, the SHAP base value is y here. base .

[0087] Step 204: Select a numerical group from multiple numerical groups based on the second preset difference as the key feature to distinguish the key numerical range of the swapped-in and swapped-out data.

[0088] In some embodiments, the executing entity may select a numerical group from multiple numerical groups based on the second preset difference determined in step 203, as a key feature to distinguish the key numerical range of the swapped-in and swapped-out data. For example... Figure 3G As shown, within each numerical group, the left side represents the difference in the model's predicted total score, and the right side represents the difference in the sum of the contribution values. From Figure 3G It can be concluded that for the key feature *ali_final*, the difference between the model's total predicted score for the incoming dataset and the model's total predicted score for the outgoing dataset varies across different value ranges. This indicates that the feature's contribution to the splitting of the incoming and outgoing datasets differs depending on its value range. We need to focus on the distribution changes of the numerical ranges that contribute significantly to the splitting of the incoming and outgoing datasets. In other words, we can select the numerical range with the largest second-preset difference as the key numerical range.

[0089] Furthermore, the above figure also shows that the second preset difference calculated by different methods is positively correlated with their ability to distinguish between incoming and outgoing data. In other words, both methods can accurately analyze the key numerical range.

[0090] The attribution analysis method based on feature data in this disclosure further enriches and improves the key features with different numerical ranges, and their impact on the original decision model when data is swapped in or out. In other words, it considers not only the interactions between various features, but also the impact of data distribution on the differences in model prediction scores between the swapped-in and swapped-out datasets. It answers the question of which value ranges of a certain feature play a crucial role in the splitting of the two datasets.

[0091] As described above, this solution first employs a pre-defined classification model (such as XGBoost) to characterize the user profiles of the two datasets being swapped in and out. It then uses a feature importance ranking method based on the domain classifiers of the swapped-in and swapped-out datasets to characterize the relative contribution of each feature to the differences between the two datasets. Furthermore, it characterizes the differences in the distribution of each feature between the two datasets using the PSI metric to represent the magnitude of these differences.

[0092] Secondly, we explore the strength of each feature's influence on the difference in model prediction scores between the two datasets, calculating the difference between the average SHAP values ​​of each feature on the two datasets. This difference reflects the strength and direction of the feature's contribution to the distinction between the two datasets. After ranking, we can determine the strength of each feature's contribution to the distinction between the two datasets, thus answering which features play a crucial role in the in-and-out transition.

[0093] Finally, we explore the strength of each feature's influence on the difference in model prediction scores between the two datasets across different value ranges. Using scatter plots of the SHAP values ​​for each feature, we examine the differences in SHAP value distribution density between the two datasets. Simultaneously, we can observe the impact of different value ranges on the model prediction scores. After grouping the features, we calculate the difference between the sum of the model prediction scores of the two datasets within each group, and the difference between the sum of (SHAP value + SHAP base value) of the two datasets within each group. Using these two strongly correlated differences, we can measure the impact of different value ranges of the feature on the difference in model prediction scores between the two datasets.

[0094] Using the three steps described above, a relatively complete, detailed, and accurate attribution analysis can be performed on swap-in and swap-out data. Furthermore, these three steps constitute the best practice for swap-in and swap-out attribution analysis.

[0095] In some application scenarios, once the key features are identified, target user data can be monitored. This target user data typically consists of user attribute data belonging to the aforementioned key features. Thus, when the target user data shifts, such as deviating from its original distribution range, the original decision model can be retrained using the target user data. This allows the model to learn the shifted data distribution, thereby improving its robustness and predictive accuracy. It is understood that the method disclosed herein is not only applicable to user risk classification and prediction models, helping to reduce the number of users experiencing risk incidents, but also applicable to any classification and decision model.

[0096] Further reference Figure 4 As a response to the above Figure 1 , 2 The present disclosure provides some embodiments of an attribution analysis apparatus based on feature data to implement the method shown. These apparatus embodiments are similar to... Figure 1 , 2 The methods and embodiments shown correspond to those described. This device can be specifically applied to various electronic devices.

[0097] like Figure 4 As shown, the feature-based attribution analysis apparatus 400 in some embodiments may include: a training data generation unit 401, configured to generate training data based on feature data belonging to the input features in the input dataset and the output dataset, wherein the input features are the input features of the original decision model in which the input and output phenomena occur; a model training unit 402, configured to train a preset classification model using the training data, and determine the index data of each input feature of the original decision model based on the trained preset classification model, wherein the index data is used to characterize the importance of the features; a target feature selection unit 403, configured to select a set number of input features from each input feature of the original decision model based on the determined index data, as target features; and a key feature analysis unit 404, configured to analyze the key features that cause the original decision model to generate input and output data based on the target features, the input dataset, and the output dataset.

[0098] In some embodiments, the key feature analysis unit 404 may include a first analysis subunit (not shown in the figure), configured to, for each target feature among the target features, determine the probability distribution density of feature data belonging to the target feature in the input dataset and the output dataset; determine the population stability index value of the input dataset and the output dataset under the target feature based on the probability distribution density of the feature data of each target feature; and determine the target features corresponding to the population stability index values ​​within a preset range as key features with large distribution differences based on the population stability index values ​​of the input dataset and the output dataset under each target feature.

[0099] In some embodiments, the key feature analysis unit 404 may further include a second analysis subunit (not shown in the figure), configured to determine, for each target feature in the target features, a first preset difference between the swap-in dataset and the swap-out dataset under that target feature, wherein the first preset difference is used to characterize the degree of distinction between the swap-in and swap-out data; and to select a certain number of target features from the target features according to the first preset difference as key features for distinguishing the swap-in and swap-out data.

[0100] In some embodiments, the attribution analysis device 400 may further include a numerical range analysis unit (not shown in the figure), configured to, for each key feature among the key features, determine the numerical range of the key feature based on feature data belonging to the key feature in the input dataset and output dataset, divide the numerical range into multiple numerical groups of the key feature at equal intervals; for each numerical group among the multiple numerical groups, determine a second preset difference between the input dataset and the output dataset located in the numerical group; and select a numerical group from the multiple numerical groups based on the second preset difference as the key numerical range for distinguishing the input and output data.

[0101] In some embodiments, the preset classification model is a distributed gradient enhancement library XGB tree model, and the model training unit 402 can be further configured to obtain the relevant index coefficients in the XGB tree model to obtain the importance of each input feature of the original decision model.

[0102] In some embodiments, the second analysis subunit may be further configured to, for each target feature among the target features, determine the total predicted score of the input dataset under the target feature by the XGB tree model for each input data in the input dataset based on the index coefficient of the target feature; determine the total predicted score of the output dataset under the target feature by the XGB tree model for each output data in the output dataset based on the index coefficient of the target feature; and use the difference between the total predicted score of the input and the total predicted score of the output for each target feature as a first preset difference.

[0103] In some embodiments, the model training unit 402 may be further configured to determine the contribution value of each input feature to the prediction result of the preset classification model by using a machine learning model interpretation algorithm based on the preset classification model that has been trained.

[0104] In some embodiments, the second analysis subunit may be further configured to, in response to determining that the amount of data in the input dataset and the output dataset are the same, for each target feature among the target features, determine the average contribution of the target feature on the input dataset as the input contribution average, and determine the average contribution of the target feature on the output dataset as the output contribution average; and take the difference between the input contribution average and the output contribution average of each target feature as a first preset difference.

[0105] In some embodiments, the training data generation unit 401 may be further configured to determine the risk index features at each classification node in the original decision model as the input features of the original decision model, wherein the original decision model is used for predicting user risk; to use user attribute data belonging to the input features in the input dataset and the output dataset as training samples, and to label the training samples according to the dataset to which the training samples belong to generate training data; wherein the input dataset stores the input user attribute data, and the output dataset stores the output user attribute data.

[0106] In some embodiments, the attribution analysis apparatus 400 may further include a monitoring unit (not shown in the figure) configured to monitor target user data based on the analyzed key features, wherein the target user data is user attribute data belonging to the key features; and in response to determining that the target user data has shifted, to adjust and train the original decision model using the target user data.

[0107] It is understandable that the units described in the attribution analysis device 400 and the reference Figure 1 , 2 The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the attribution analysis device 400 and the units contained therein, and will not be repeated here.

[0108] The following is for reference. Figure 5 It shows a schematic diagram of the structure of an electronic device 500 suitable for implementing some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0109] like Figure 5As shown, the electronic device 500 may include a processing unit 501 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage device 508 into a random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device 500. The processing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0110] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, speakers, vibrators, etc.; storage devices 508 including, for example, disks, memory cards, etc.; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 5 Each box shown can represent a device or multiple devices as needed.

[0111] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 509, or installed from storage device 508, or installed from ROM 502. When the computer program is executed by processing device 501, it performs the functions defined in the methods of some embodiments of this disclosure.

[0112] It should be noted that, in some embodiments of this disclosure, the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0113] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0114] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: generate training data based on feature data belonging to the input features in the input dataset and the output dataset, wherein the input features are the input features of the original decision model exhibiting the input / output phenomenon; train a preset classification model using the training data, and based on the trained preset classification model, determine the indicator data of each input feature of the original decision model, wherein the indicator data is used to characterize the importance of the features; select a predetermined number of input features from the input features of the original decision model as target features based on the determined indicator data; and analyze the key features that cause the original decision model to generate input / output data based on the target features, the input dataset, and the output dataset.

[0115] Furthermore, computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0116] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0117] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a training data generation unit, a model training unit, a target feature selection unit, and a key feature analysis unit. The names of these units do not necessarily limit the specific unit; for example, the training data generation unit may also be described as "a unit that generates training data based on feature data belonging to the input model features in the input and output datasets."

[0118] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0119] Some embodiments of this disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the above-described attribution analysis methods based on feature data.

[0120] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. An attribution analysis method based on feature data, comprising: Training data is generated based on the feature data belonging to the input model features in the input dataset and the output dataset, wherein the input model features are the input model features of the original decision model in which the input and output phenomena occur. The preset classification model is trained using the training data, and the index data of each input feature of the original decision model is determined based on the trained preset classification model, wherein the index data is used to characterize the importance of the features. Based on the determined indicator data, a set number of input features are selected from the input features of the original decision model as target features. Based on the target features, the input dataset, and the output dataset, the key features that cause the original decision model to generate input and output data are analyzed.

2. The attribution analysis method based on feature data according to claim 1, wherein, Based on the target features, the input dataset, and the output dataset, the key features that cause the original decision model to generate input and output data are analyzed, including: For each target feature in the target features, determine the probability distribution density of the feature data belonging to that target feature in the input dataset and the output dataset; Based on the probability distribution density of the feature data of each target feature, determine the population stability index value of the incoming dataset and the outgoing dataset under the target feature. Based on the input dataset and the output dataset, the target features corresponding to the population stability index values ​​under each target feature are determined as key features with large distribution differences.

3. The attribution analysis method based on feature data according to claim 1, wherein, The step of analyzing the key features that cause the original decision model to generate the swap-in / switch-out data based on the target features, the swap-in dataset, and the swap-out dataset also includes: For each of the target features, a first preset difference is determined between the input dataset and the output dataset under that target feature, wherein the first preset difference is used to characterize the degree of distinction between the input and output datasets; Based on the first preset difference, a certain number of target features are selected from the target features as key features to distinguish the swapped-in and swapped-out data.

4. The attribution analysis method based on feature data according to claim 3, wherein, The method further includes: For each of the key features, the numerical range of the key feature is determined based on the feature data belonging to the key feature in the input dataset and the output dataset. The numerical range is then divided into multiple numerical groups of the key feature at equal intervals. For each of the plurality of numerical groups, a second preset difference is determined between the incoming dataset and the outgoing dataset located within that numerical group; Based on the second preset difference, a numerical group is selected from the plurality of numerical groups as a key feature to distinguish the key numerical range of the swapped-in and swapped-out data.

5. The attribution analysis method based on feature data according to claim 3, wherein, The preset classification model is the XGB tree model from the distributed gradient enhancement library, and... The indicator data for each input feature of the original decision model, determined based on the pre-trained classification model, includes: Obtain the relevant index coefficients in the XGB tree model to obtain the importance of each input feature of the original decision model.

6. The attribution analysis method based on feature data according to claim 5, wherein, For each of the target features, determining a first preset difference between the incoming dataset and the outgoing dataset under that target feature includes: For each of the target features, based on the index coefficient of the target feature, the XGB tree model predicts the score of each swapped-in data in the swapped-in dataset, and determines the total swapped-in prediction score of the swapped-in dataset under the target feature. Based on the index coefficient of the target feature, the XGB tree model predicts the score of each data point in the data point to be exchanged in the data point, and determines the total predicted score of the data point to be exchanged under the target feature. The difference between the predicted total score for each target feature and the predicted total score for the feature removed is used as the first preset difference.

7. The attribution analysis method based on feature data according to claim 3, wherein, The method for determining the index data of each input feature of the original decision model based on the trained preset classification model further includes: Based on the trained preset classification model, a machine learning model interpretation algorithm is used to determine the contribution value of each input feature to the prediction result of the preset classification model.

8. The attribution analysis method based on feature data according to claim 7, wherein, The step of determining a first preset difference between the input dataset and the output dataset under each of the target features further includes: In response to determining that the amount of data in the input dataset and the output dataset are the same, for each target feature in the target features, the average contribution of the target feature on the input dataset is determined as the input contribution average, and the average contribution of the target feature on the output dataset is determined as the output contribution average. The difference between the average value of the input contribution and the average value of the output contribution for each target feature is used as the first preset difference.

9. The attribution analysis method based on feature data according to claim 1, wherein, The step of generating training data based on feature data belonging to the input model features in the input and output datasets includes: The risk indicator features at each classification node in the original decision model are determined as the input features of the original decision model, wherein the original decision model is used for predicting user risk. The user attribute data belonging to the input model features in the input dataset and the output dataset are used as training samples, and the training samples are labeled according to the dataset to which they belong to generate training data. The swap-in dataset stores the user attribute data of the swapped-in user, and the swap-out dataset stores the user attribute data of the swapped-out user.

10. The attribution analysis method based on feature data according to any one of claims 1-9, wherein, The method further includes: Based on the key features identified in the analysis, the target user data is monitored, wherein the target user data is user attribute data belonging to the key features; In response to determining that the target user data has shifted, the original decision model is adjusted and trained using the target user data.

11. An attribution analysis device based on feature data, comprising: The training data generation unit is configured to generate training data based on the feature data belonging to the input model features in the input dataset and the output dataset, wherein the input model features are the input model features of the original decision model in which the input and output phenomena occur. The model training unit is configured to train a preset classification model using the training data, and based on the trained preset classification model, determine the index data of each input feature of the original decision model, wherein the index data is used to characterize the importance of the features. The target feature selection unit is configured to select a set number of input features from the input features of the original decision model based on the determined index data, as target features. The key feature analysis unit is configured to analyze the key features that cause the original decision model to generate the swap-in and swap-out data based on the target features, the swap-in dataset, and the swap-out dataset.

12. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the attribution analysis method based on feature data as described in any one of claims 1-10.

13. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by the processor, it implements the attribution analysis method based on feature data as described in any one of claims 1-10.

14. A computer program product comprising a computer program that, when executed by a processor, implements the attribution analysis method based on feature data as described in any one of claims 1-10.

Images