A multi-method fusion interpretation method based on feature importance of evidence fusion

Abstract

This invention discloses a multi-method fusion interpretation method for feature importance based on evidence fusion, belonging to the field of machine learning technology. The main steps include: first, converting the importance scores of various interpretation methods for feature outputs into normalized ranking scores and mapping them to propensity probabilities in a binary proposition space; then, calculating the reliability weights of each method by evaluating their stability and fidelity, and discounting the initial evidence accordingly; further, fusing multi-source evidence using D-S evidence theory, and adaptively selecting combination rules based on the degree of conflict; finally, outputting the fusion importance score, uncertainty interval, and interpretation conflict degree for each feature, achieving reliable fusion and uncertainty quantification of the interpretation conclusion. This invention, by fusing evidence from multiple interpretation methods and introducing reliability assessment and conflict adaptation mechanisms, can provide a more consistent and robust interpretation of feature importance with quantifiable uncertainty intervals and conflict diagnostic information.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and specifically to a multi-method fusion interpretation method for feature importance based on evidence fusion. Background Technology

[0002] In the field of machine learning model interpretability research, especially for high-risk decision-making applications such as predicting the severity of traffic accidents, understanding and interpreting the contribution of each input feature to the prediction result—that is, feature importance—is crucial for model mechanism analysis, risk attribution, and subsequent governance policy formulation. Currently, researchers have proposed various methods for interpreting feature importance, explaining the model's decision-making logic from different principles. For example, the scrambled feature importance method observes the degree of model performance degradation by randomly shuffling the value of a feature; the model structure-based feature importance method directly utilizes intrinsic structural information such as gain or coverage from tree models; the Shapley value decomposition method fairly allocates predicted values ​​based on cooperative game theory; the feature removal and re-evaluation method re-evaluates the model after removing features; and the locally interpretable surrogate model method approximates the interpretation by fitting a simple model near a single prediction point. These methods, each with its own focus, collectively constitute a toolset for model interpretability analysis.

[0003] However, in practical applications, especially in scenarios involving complex correlations, high dimensionality, and significant feature collinearity in traffic accident data, the aforementioned interpretation methods often yield inconsistent or even contradictory rankings of feature importance. This inconsistency stems from differences in the underlying theoretical assumptions of each method, varying sensitivities to correlations between features, and differences in their respective focuses on global or local interpretations. Existing technologies typically employ intuitive methods such as simple arithmetic averaging or majority voting to fuse the outputs of different methods in an attempt to obtain a "comprehensive" conclusion. However, such simple fusion strategies have obvious drawbacks: First, they cannot quantify the credibility or uncertainty of the final fusion conclusion, and decision-makers have no way of knowing the reliability of the ranking; second, these strategies treat the results of all interpretation methods equally, failing to consider the differences in reliability and stability of different methods on specific datasets or models, and some methods that may produce unreliable conclusions due to sampling fluctuations or improper handling of collinearity are given equal weight; third, when the evidence given by different methods is highly conflicting, forced fusion may produce counterintuitive results, and existing methods lack detection and diagnosis mechanisms for such conflicts, making it impossible to identify which features have significant disagreements in their importance judgments, and even more difficult to link such disagreements to potential mechanisms such as feature collinearity, interaction effects, or scene heterogeneity in the data. Summary of the Invention

[0004] To construct a fusion interpretation framework that quantifies uncertainty, assesses method reliability, and intelligently handles conflicting evidence, this invention proposes a feature importance-based multi-method fusion interpretation method, comprising the following steps: S1: Obtain the original importance scores of each feature output in the prediction model from multiple interpretation methods, and convert each original importance score into a normalized importance score based on ranking; S2: For each feature, construct a binary proposition space containing important propositions, unimportant propositions, and sets of ignorance, and map the normalized importance score of the feature to the probability of each interpretation method’s emphasis on the feature. S3: Calculate the reliability weight of each interpretation method based on its stability index under repeated experiments and its fidelity index to the model prediction performance. S4: Based on the important tendency probability corresponding to each interpretation method, construct the initial evidence quality assignment for each feature in the binary proposition space for each interpretation method, and discount the initial evidence quality assignment using reliability weights. S5: For each feature, the evidence quality allocation after discounting multiple interpretation methods is fused and calculated based on DS evidence theory. During the fusion process, the combination rule is adaptively selected according to the degree of conflict between each evidence source to obtain the fused evidence quality allocation for that feature. S6: Based on the quality allocation of the fused evidence, output the fusion importance score, importance uncertainty interval, and degree of conflict of interpretation methods for each feature.

[0005] This invention integrates evidence from multiple interpretation methods and introduces a reliability assessment and conflict adaptation mechanism to provide a more consistent and robust interpretation of feature importance, with quantifiable uncertainty intervals and conflict diagnostic information.

[0006] Furthermore, the multiple interpretation methods are any combination of the following methods: The methods include scrambling feature importance, model structure-based feature importance, Shapley value decomposition, feature deletion and re-evaluation, and locally interpretable surrogate model methods.

[0007] Furthermore, in step S3, the stability index is obtained by: obtaining the feature importance ranking sequence output by each interpretation method through repeated experiments, and calculating the average consistency among these ranking sequences.

[0008] Furthermore, the fidelity index is obtained by: ranking the importance of features output by each interpretation method, deleting important features sequentially using both scrambling and masking strategies, calculating the rate of model performance degradation, and averaging the results under the two strategies.

[0009] Furthermore, in step S3, the reliability weight is obtained by weighted summation of the stability index and the fidelity index.

[0010] Furthermore, in step S4, the initial evidence quality allocation is discounted using reliability weights. Specifically, this includes reducing the quality allocated to important and unimportant propositions according to the reliability weights, and then allocating the reduced quality, along with the quality reflecting the unreliability of the method, to the set of ignorance.

[0011] Furthermore, in step S5, the adaptive selection of the combination rule specifically means: when the conflict degree is lower than a preset threshold, the Dempster combination rule is adopted; when the conflict degree reaches or exceeds the preset threshold, a conservative combination rule that assigns conflict quality to the unknown set is adopted.

[0012] Furthermore, in step S6, the fusion importance score is obtained by adding half of the quality of the fused evidence assigned to the important proposition to the quality of the evidence assigned to the set of ignorance; the importance uncertainty interval is the interval between confidence and likelihood, where confidence is the quality assigned to the important proposition and likelihood is the sum of the quality assigned to the important proposition and the set of ignorance.

[0013] Furthermore, it also includes the following steps: S7: Merge multiple semantically related original features into a feature group. Repeat steps S2 to S6 at the feature group level to obtain the fusion importance score, importance uncertainty interval, and conflict degree of interpretation methods at the feature group level.

[0014] Furthermore, the feature group consists of multiple binary features generated by one-hot encoding of the same categorical variable.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes a feature importance multi-method fusion interpretation method based on evidence fusion, which maps the output of different interpretation methods to a unified binary proposition space and explicitly models the set of ignorance, so that the fusion result not only provides feature importance ranking, but also outputs uncertainty quantification index represented by confidence degree and likelihood degree interval, thereby giving the interpretation conclusion an assessable credibility label. (2) By designing a reliability evaluation mechanism based on reproducible experiments, the stability of each interpretation method under multiple runs and its fidelity to the model prediction performance are comprehensively considered, and the evidence is discounted accordingly to reduce the interpretation noise caused by the fluctuation or infidelity of the method itself, thereby improving the robustness of the fusion conclusion. (3) Calculate and utilize the conflict degree index during the evidence fusion process to achieve adaptive fusion rules: fully integrate consistent information when the conflict is low, adopt a conservative strategy to avoid counterintuitive conclusions when the conflict is high, and the output conflict degree itself can be used as a diagnostic tool to identify possible interpretation discrepancies caused by feature collinearity, interaction effects or scene heterogeneity. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a multi-method fusion interpretation method based on the feature importance of evidence fusion. Detailed Implementation

[0017] The following will describe in detail the feature importance multi-method fusion interpretation method based on evidence fusion according to the present invention, with reference to the accompanying drawings and specific embodiments. Those skilled in the art will understand that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. The present invention can be applied to any scenario that utilizes machine learning models for prediction and requires understanding feature contribution. Figure 1 As shown, the feature importance multi-method fusion interpretation method based on evidence fusion provided by this invention can be summarized in the following sequential steps: S1: Obtain the original importance scores of each feature output in the prediction model from multiple interpretation methods, and convert each original importance score into a normalized importance score based on ranking; S2: For each feature, construct a binary proposition space containing important propositions, unimportant propositions, and sets of ignorance, and map the normalized importance score of the feature to the probability of each interpretation method’s emphasis on the feature. S3: Calculate the reliability weight of each interpretation method based on its stability index under repeated experiments and its fidelity index to the model prediction performance. S4: Based on the important tendency probability corresponding to each interpretation method, construct the initial evidence quality assignment for each feature in the binary proposition space for each interpretation method, and discount the initial evidence quality assignment using reliability weights. S5: For each feature, the evidence quality allocation after discounting multiple interpretation methods is fused and calculated based on DS evidence theory. During the fusion process, the combination rule is adaptively selected according to the degree of conflict between each evidence source to obtain the fused evidence quality allocation for that feature. S6: Based on the quality allocation of the fused evidence, output the fusion importance score, importance uncertainty interval, and conflict degree of the interpretation method for each feature; S7: Merge multiple semantically related original features into a feature group. Repeat steps S2 to S6 at the feature group level to obtain the fusion importance score, importance uncertainty interval, and conflict degree of interpretation methods at the feature group level.

[0018] The implementation environment of this embodiment includes a pre-trained traffic accident severity prediction model and a dataset containing historical accident records. The prediction model can be any machine learning model capable of outputting predicted probabilities, such as gradient boosting decision trees, random forests, or neural networks. The dataset contains several samples, each sample consisting of... The system consists of several feature variables (such as road conditions, weather, and time) and a target variable representing the severity of the accident. The core objective of this invention is to address this pre-trained prediction model. For the use of The importance of each feature is explained through a robust, reliable, and diagnostic multi-method fusion approach.

[0019] The entire method begins by obtaining raw feature importance assessments from multiple interpretation methods. In this embodiment, the invention simultaneously employs five widely used interpretation methods with different principles, including a scrambled feature importance method, a model structure-based feature importance method, a Shapley value decomposition method, a feature deletion and re-evaluation method, and a locally interpretable surrogate model method. For a given prediction model... and dataset Run each of the five methods separately.

[0020] Specifically, the scrambling feature importance method calculates its importance by randomly shuffling the values ​​of a feature across all samples and observing the decrease in the overall predictive performance of the model (such as accuracy and log loss); the model structure-based feature importance method (e.g., for tree models) calculates it by statistically analyzing the total reduction in impurity (gain) or the number of samples covered when a feature splits across all tree nodes; the Shapley value decomposition method, based on cooperative game theory, calculates the marginal contribution of each feature to the prediction result of a single sample and then obtains global importance by averaging the absolute values ​​of all samples; the feature deletion re-evaluation method re-evaluates the model performance by setting all values ​​of a feature to zero or deleting it, measuring importance by the degree of performance degradation; and the locally interpretable surrogate model method fits a simple interpretable model (such as a linear model) within the local neighborhood of the target model's prediction results, uses the coefficients of this simple model to approximate the feature importance of the original model in this local area, and obtains a global perspective by aggregating on a representative sample set. Each method... For each feature Output a raw importance score Due to differences in calculation principles and dimensions, these scores have different numerical ranges and scales, and direct comparison or fusion will introduce bias.

[0021] Therefore, the next crucial step is to transform these heterogeneous original scores to a comparable scale. This invention adopts a robust strategy that abandons the use of the absolute values ​​of the original scores, instead relying on the relative ranking relationships between their features. For each interpretation method... Based on its output raw importance score For all Sort the features to obtain each feature Ranking under this method Rank 1 indicates the most important. Then, this ranking is normalized to... Intervals, defined by the formula: , Calculate the normalized sort score The higher the score, the more important the feature is from the perspective of this method. Through this transformation, the present invention unifies the different outputs of the five methods into an importance measure with the same value range and semantics (1 for most important, 0 for least important).

[0022] After obtaining standardized scores, this invention introduces Dempster-Shafer evidence theory as a fusion framework. For this purpose, it is necessary to evaluate each feature... This invention constructs a proposition space for evidence processing. It defines a binary proposition space. ,in Indicates the "characteristics" of a proposition It is important. Indicates the "characteristics" of a proposition "It's unimportant." Furthermore, the "ignorance" characteristic of DS theory... It was also introduced to represent features Whether it is important is "completely uncertain". Next, each method needs to be considered. The given normalized score Transform the proposition into a method for applying this approach. The degree of support, i.e., the "probability of a significant tendency". This transformation is achieved through a smooth S-shaped function, for example: , , Among them, parameters It is a soft threshold used to control the proportion of features considered "important," for example, it can be set to all features. The median; parameter The steepness of the control function affects the sensitivity of the probability to small changes in the ranking. Thus, It is a value between 0 and 1, which intuitively reflects the method. Characteristics Important probability tendency, This naturally corresponds to the tendency to consider it unimportant.

[0023] However, not all explanatory methods are equally reliable. Some methods may be highly sensitive to data perturbations, leading to unstable results; others may have insufficient fidelity because the importance of their definitions does not fully align with the actual behavior of the model that depends on that feature. Indiscriminately fusing these inconsistent pieces of evidence can contaminate the final conclusion. Therefore, this invention designs a reproducible, experiment-driven reliability assessment module for each explanatory method. Assign a reliability weight This reliability is a weighted composite of two core indicators: stability and fidelity.

[0024] Stability metrics are used in quantitative methods The consistency of output results under slight data perturbations. This is calculated using repeated experimental techniques such as K-fold cross-validation or bootstrap sampling. A slightly different subset of data, and run the method on each subset. ,get A sequence of features ranked by importance. Then, this... Kendall's harmonic coefficients between each pair of sorted sequences Finally, take all pairs. average .because The range of values ​​is Normalize it linearly to From the interval, we obtain the stability index: . The closer the value is to 1, the more stable the method is.

[0025] Loyalty metrics are used as a measurement method. Whether the identified "important features" are truly crucial to the model's predictions. The specific evaluation method is: according to the methodology... The features are ranked from most important to least important. The most important features are then removed (or scrambled, masked) sequentially, and the performance degradation curve of the model on the test set is observed. The faster and more significant the performance degradation, the more critical the identified features are, and the higher their fidelity. To eliminate the bias of a single removal strategy, this embodiment uses both "scrambling removal" and "masking removal" (e.g., padding with the mean) strategies for testing, and the normalized average of the area under the performance degradation curves (or the early slope of degradation) of both strategies is taken as the fidelity index. .

[0026] Ultimately, the method Reliability weight We obtain this by weighted summation of stability and fidelity: , in, It is an adjustable hyperparameter used to balance the emphasis on stability and fidelity, and its optimal value can usually be determined on the validation set through grid search.

[0027] Possessing significant probabilities and method reliability Then, the basic probability assignment function in DS theory can be constructed and adjusted in two steps.

[0028] The first step is to construct the initial basic probability assignment (BBA). For the interpretation method... and characteristics First, an initial evidence allocation is established without considering the reliability of the method. The method... The full credibility (considered 1 for now) is allocated to the binary proposition space. Specifically, it is allocated to "important" propositions. The quality is That is, the probability of its importance itself: accordingly, the probability of assigning importance to "unimportant" propositions. The quality is At this moment, ignorance is gathered. The quality is 0. This allocation reflects the ideal situation, i.e., the method. When completely reliable, the degree to which the evidence supports the judgment of feature importance. This initial BBA satisfies the condition that the sum of probabilities is 1, i.e. .

[0029] The second step is to introduce method reliability for evidence discounting. Because of the method... The reliability is It is necessary to use The initial evidence above should be discounted. The principle of discounting is: the more reliable the method (…). The closer to 1, the more of the initial evidence is preserved; the less reliable the method ( The smaller the value, the larger the portion of the evidence considered "uncertain." The discount is applied using the following formula: After the discount, the allocation is determined to be "important" for specific propositions. The mass becomes: .

[0030] After the discount, the value allocated to the definitive proposition is deemed "unimportant". The mass becomes: .

[0031] After the discount, the additional ignorance quality resulting from the unreliability of the method is merged with the initial ignorance quality and jointly allocated to the "ignorance" set. : .

[0032] It can be verified that the discounted BPA still satisfies the condition that the sum of all qualities equals 1. Through this discounting operation, the evidentiary quality of methods with low reliability is largely transferred to those representing global uncertainty. In contrast, the influence of highly reliable methods is naturally reduced in subsequent integrations, while the evidence of highly reliable methods is largely preserved.

[0033] The next step is the core process of fusing multiple sources of evidence. For the same feature... It possesses five discounted pieces of evidence (i.e., five BBAs) from five different interpretation methods. The DS evidence theory provides rules for combining multiple BBAs. This invention employs a conflict-adaptive fusion strategy. First, the degree of conflict between any two pieces of evidence is calculated. It is defined as the sum of the products of the masses of all supporting disjoint propositions in two pieces of evidence. For example, method A supports... Quality and Methods B Support The product of their qualities contributes to the conflict level. When the degree of conflict Less than a preset threshold When the conflict level is low, indicating minimal discrepancy between the two methods, this invention employs the standard Dempster combination rule for fusion. This rule effectively merges consistent information and proportionally redistributes conflict quality to critical issues, thereby arriving at a more definitive conclusion under low conflict conditions. However, when the conflict level is high... Reaching or exceeding the threshold This indicates a fundamental disagreement between the two methods in assessing the importance of this feature. Forcing the application of the Dempster rule at this point could lead to counterintuitive, overly certain results. Therefore, this invention switches to a conservative combination rule, such as the Yager rule or the Dubois-Prade rule. Taking one of these as an example, the conservative rule will properly combine the quality of all consistent parts in the two pieces of evidence, while merging all conflicting quality... All assigned to the "ignorant" set This approach reduces the certainty of the conclusions reached. While increasing the quality of evidence, this invention avoids making arbitrary judgments in high-conflict situations, making it more consistent with reality. In this embodiment, the present invention sequentially fuses the evidence from the five methods in pairs, and dynamically selects the combination rule at each step based on the real-time calculated conflict degree, ultimately obtaining a result tailored to the feature. A unified basic probability allocation .

[0034] After the integration is complete, it will be necessary to start from the final BBA. Extract explanatory metrics that are favorable to decision-makers. First, calculate each feature. The importance score of the fusion. This invention uses "betting probability" as the ranking criterion, and its calculation formula is: , The meaning of this formula is that it represents the set of ignorance. The uncertainty quality represented is evenly distributed between the two possibilities of "important" and "unimportant", thus obtaining a point estimate between confidence and likelihood, which is suitable for ranking feature importance.

[0035] Secondly, to quantify the uncertainty, an importance interval is output. The lower bound of the interval is the "confidence level". This indicates that the evidence clearly supports the characteristics. The minimum level of importance; the upper bound of the interval is the "likelihood". , indicating features The highest degree of potential importance (i.e., no evidence against its importance). Interval width. It directly reflects the total amount of uncertainty. The wider the width, the greater the divergence between different methods or the less reliable the method itself is, and the more cautious the conclusion should be.

[0036] Finally, the total conflict degree calculated during the fusion process will be... It is also output as a key diagnostic indicator. A high degree of conflict. The text explicitly states that different interpretation methods have differing opinions on the importance of this feature. This is often not a random error, but rather may reveal collinearity or interaction effects between this feature and other features, or heterogeneity in its importance across different subsets of data (scenarios). For example, in traffic accident prediction, "whether it is peak hour" may be highly correlated with "road congestion index," and different methods may confuse their importance when attributing them, leading to conflicts.

[0037] To further enhance the robustness and readability of the explanation, this invention also proposes a group feature-level fusion mechanism. In real-world data, a semantic concept is often broken down into multiple technical features. A typical example is that after one-hot encoding, categorical variables generate multiple binary feature columns (e.g., "direction" is encoded as four independent features: "east," "south," "west," and "north"). If these split features are evaluated and fused separately, their importance may be diluted across the columns or "drift" between different methods (e.g., method A considers "east" important, while method B considers "north" important), leading to an ambiguous and unstable judgment of the overall importance of the "direction" concept. To address this issue, this invention introduces a group feature layer before or after the aforementioned feature-level fusion. First, the original feature set is fused according to the business semantics. Divide into several non-overlapping groups Each group Features that semantically belong to the same concept (such as one-hot encoded columns for all "directions") are included. Then, evidence fusion is performed again at the group level. Specifically, for each interpretation method... and each feature group First, based on this method, each feature within the group... After the discount, BBA calculates the value of each feature. The value is then calculated. A weighted average (with weights that can be equal or adjusted based on factors such as feature variance) is then used to obtain the "probability of importance" for that group. At the same time, the quality of ignorance for all features within the group is taken. The maximum value is used as a characterization of the uncertainty of the method at this group level. This is a conservative estimate. Next, following the steps at the feature level, the method reliability is used... and calculated , This method is used to construct a group-level propositional space. The basic probability allocation is then performed. Finally, at the group level, the conflict-adaptive DS fusion rule is applied to the group-level BBAs provided by the five methods to obtain the group-level fusion results. and output group-level results. Sort, Range and Conflict Ultimately, the system outputs two layers of results: one layer is a fusion interpretation at the feature group level, which provides importance ranking and uncertainty, suitable for high-level decision analysis and policy discussions; the other layer is a fusion interpretation at the original feature level, which can provide fine-grained positioning references when it is necessary to delve deeper into the specific driving factors within an important concept (for example, if the "direction" group is found to be important, further investigation is needed to determine whether "eastward" or "northward" plays a major role).

[0038] In summary, this invention effectively solves the problems of inconsistent conclusions, lack of credible quantification, and conflict diagnosis among multiple interpretation methods in traffic accident prediction models by introducing DS evidence theory to construct a unified fusion framework. Its technical effects are mainly reflected in the following aspects: First, this method maps the outputs of different interpretation methods to a unified binary proposition space and explicitly models the set of ignorance. This allows the fusion result to not only provide a ranking of feature importance but also output uncertainty quantification indicators represented by confidence and likelihood intervals, thereby giving the interpretation conclusions an assessable credibility label. Second, by designing a reliability evaluation mechanism based on reproducible experiments, it comprehensively considers the stability of each interpretation method under multiple runs and its faithfulness to the model's predictive performance. Based on this, it discounts the evidence, significantly reducing interpretation noise caused by the method's own fluctuations or infidelity, and improving the robustness of the fusion conclusions. Furthermore, it innovatively calculates and utilizes the conflict degree index during the evidence fusion process to achieve adaptive fusion rules: when the conflict is low, it fully integrates consistent information; when the conflict is high, it adopts a conservative strategy to avoid counterintuitive conclusions. At the same time, the output conflict degree itself can be used as a diagnostic tool to identify interpretation discrepancies that may be caused by feature collinearity, interaction effects, or scene heterogeneity.

[0039] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0040] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0041] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0042] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A multi-method fusion interpretation method based on feature importance in evidence fusion, characterized in that, Including the following steps: S1: Obtain the original importance scores of each feature output in the prediction model from multiple interpretation methods, and convert each original importance score into a normalized importance score based on ranking; S2: For each feature, construct a binary proposition space containing important propositions, unimportant propositions, and sets of ignorance, and map the normalized importance score of the feature to the probability of each interpretation method’s emphasis on the feature. S3: Calculate the reliability weight of each interpretation method based on its stability index under repeated experiments and its fidelity index to the model prediction performance. S4: Based on the important tendency probability corresponding to each interpretation method, construct the initial evidence quality assignment for each feature in the binary proposition space for each interpretation method, and discount the initial evidence quality assignment using reliability weights. S5: For each feature, the evidence quality allocation after discounting multiple interpretation methods is fused and calculated based on DS evidence theory. During the fusion process, the combination rule is adaptively selected according to the degree of conflict between each evidence source to obtain the fused evidence quality allocation for that feature. S6: Based on the quality allocation of the fused evidence, output the fusion importance score, importance uncertainty interval, and conflict degree of the interpretation method for each feature.

2. The feature importance fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, The multiple interpretation methods are any combination of the following methods: The methods include scrambling feature importance, model structure-based feature importance, Shapley value decomposition, feature deletion and re-evaluation, and locally interpretable surrogate model methods.

3. The feature importance fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, In step S3, the stability index is obtained by: obtaining the feature importance ranking sequence output by each interpretation method through repeated experiments, and calculating the average consistency between these ranking sequences.

4. The feature importance fusion interpretation method based on evidence fusion as described in claim 3, characterized in that, The fidelity index is obtained by ranking the importance of features output by each interpretation method, deleting important features sequentially using both scrambling and masking strategies, calculating the rate of model performance degradation, and averaging the results of the two strategies.

5. The feature importance fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, In step S3, the reliability weight is obtained by weighted summation of the stability index and the fidelity index.

6. The feature importance fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, In step S4, the initial evidence quality allocation is discounted using reliability weights. Specifically, this includes reducing the quality allocated to important and unimportant propositions according to the reliability weights, and then allocating the reduced quality and the quality reflecting the unreliability of the method to the set of ignorance.

7. The feature importance fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, In step S5, the adaptive selection of the combination rule is as follows: when the conflict degree is lower than the preset threshold, the Dempster combination rule is adopted; when the conflict degree reaches or exceeds the preset threshold, the conservative combination rule that allocates the conflict quality to the unknown set is adopted.

8. The feature importance multi-method fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, In step S6, the fusion importance score is obtained by adding half of the quality of the fused evidence assigned to the important proposition to the quality of the evidence assigned to the set of ignorance; the importance uncertainty interval is the interval from confidence to likelihood, where confidence is the quality assigned to the important proposition and likelihood is the sum of the quality assigned to the important proposition and the set of ignorance.

9. The feature importance fusion interpretation method based on evidence fusion as described in claim 1, characterized in that, It also includes the following steps: S7: Merge multiple semantically related original features into a feature group. Repeat steps S2 to S6 at the feature group level to obtain the fusion importance score, importance uncertainty interval, and conflict degree of interpretation methods at the feature group level.

10. The feature importance fusion interpretation method based on evidence fusion as described in claim 9, characterized in that, The feature group consists of multiple binary features generated by one-hot encoding of the same categorical variable.

Images