Model degradation early warning and explainability correction device based on meta-learning and sliding window calibration

Abstract

The application belongs to the technical field of automatic decision system, and particularly relates to a model degradation forward-looking early warning and explainability correction device based on meta-learning and sliding window calibration; comprising: a multi-source data acquisition and sample management module which acquires input sample data, and divides a sample pool after preprocessing; a model health degree evaluation and forward-looking early warning module which generates an error measurement probability, a health score and an early warning threshold according to calculation, and judges whether to trigger early warning according to the health score and the early warning threshold; a concept drift detection and adaptive explanation calibration module which performs drift detection on a meta model when it is judged that early warning is triggered, triggers explanation calibration when it is confirmed that drift occurs, and judges whether to trigger explanation calibration again; an early warning notification and decision support module which pushes early warning notification information to operation and maintenance personnel and returns operation instructions; an update history failure case library is fed back to the model health degree evaluation and forward-looking early warning module; the application provides a model operation and maintenance support with forward-looking early warning capability and continuous explainability.

Description

Technical Field

[0001] This invention belongs to the field of automated decision-making system technology, specifically relating to a model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration. Background Technology

[0002] In automated decision-making systems such as financial credit approval, cybersecurity threat interception, industrial production line quality inspection, and equipment health management, one or more evaluation models (such as classifiers and scorecards) are typically deployed to make binary decisions (pass / reject, normal / abnormal) on input samples. To ensure the effectiveness and reliability of the models in actual operation, the operations and maintenance team needs to continuously monitor the model's performance while meeting the interpretability requirements of regulatory or auditing levels for the model's decision logic. Currently, the main methods for monitoring model performance and maintaining interpretability include periodic model validation based on delayed labels, concept drift detection methods based on unlabeled data, and model interpretability techniques based on methods such as SHAP.

[0003] While the above methods provide their own solutions for delayed label verification, unlabeled drift detection, and model interpretability, in real dynamic environments, the shortcomings of these methods are not isolated but rather superimposed. This makes it difficult for operations and maintenance personnel to perceive the true health status of the model in a timely manner during the label delay period, and also makes it difficult to accurately trace the cause after the model fails.

[0004] (1) The problem of delayed perception of model degradation under delayed labeling conditions. In scenarios where there is a delay in the acquisition of real labels, existing methods face a dilemma. If we wait for the labels to mature before evaluation, although the conclusion is reliable, the perception lag is several weeks or even months, often missing the best time for intervention. If we use a label-free drift detection method, although the response is timely, changes in feature distribution do not necessarily equate to a decline in model performance, so it is difficult to avoid a high false positive rate, and it is also impossible to quantify the extent to which feature changes substantially affect the model performance. Therefore, there is a lack of an intermediate solution that can predict the trend of model performance degradation in advance with sufficient accuracy, leaving maintenance personnel almost unaware of whether the model is failing during the label delay window.

[0005] (2) The problem of interpretability failure caused by concept drift. Mainstream interpretability methods such as SHAP implicitly assume that the data distribution is static and invariant, that is, the background distribution used to calculate the contribution of features is fixed. However, in real dynamic environments, concept drift, that is, the change in the mapping relationship between input features and prediction targets, often occurs frequently. When concept drift occurs, the SHAP value of the same feature for the same type of samples will be systematically shifted, but the interpretation results generated based on the old data distribution still follow the original background distribution. Therefore, the interpretation results are prone to disconnect from the actual decision logic of the model, and the feature importance analysis is also distorted. In scenarios with rigid requirements for model interpretability, the consequences of this problem are particularly prominent. If compliance personnel conduct audits based on outdated interpretations, they may draw conclusions that contradict the facts.

[0006] Furthermore, the two issues mentioned above are often intrinsically linked. Model performance degradation is often caused by concept drift, and once concept drift occurs, interpretability methods based on the old distribution also fail simultaneously. This leaves operations personnel without reliable analytical tools even if they wish to trace the root cause of model degradation afterward. Existing methods treat these two issues separately, and a unified solution has not yet been formed. Summary of the Invention

[0007] To address the technical shortcomings of existing automated decision-making systems, namely the inability to detect model degradation in a timely manner due to delays in acquiring real labels, and the failure of model interpretability methods and the inability to adaptively calibrate in environments with dynamically changing data distributions, this invention proposes a model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration. The device includes: a multi-source data acquisition and sample management module, a model health assessment and prospective early warning module, a concept drift detection and adaptive interpretation calibration module, and an early warning notification and decision support module.

[0008] The multi-source data acquisition and sample management module is used to collect each input sample data in the automated decision-making system and preprocess it, and then divide the preprocessed input sample data into sample pools.

[0009] The model health assessment and prospective early warning module is used to calculate the error probability, health score and early warning threshold based on the input sample data, and to determine whether to trigger an early warning based on the health score and early warning threshold.

[0010] The concept drift detection and adaptive interpretation calibration module is used to detect drift in the meta-model when a warning is triggered; and to trigger interpretation calibration and determine whether to trigger interpretation calibration again when drift is confirmed.

[0011] The early warning notification and decision support module includes an information aggregation and push module, an operation feedback receiving module, and a historical record archiving and feedback module. The information aggregation and push module pushes early warning notification information to operation and maintenance personnel, and the operation feedback receiving module receives operation instructions returned by operation and maintenance personnel. The historical record archiving and feedback module updates the historical failure case library and feeds the historical failure case library back to the model health assessment and forward-looking early warning module.

[0012] Preferably, the multi-source data acquisition and sample management module includes a data acquisition module, a preprocessing module, and a sample pool management module;

[0013] The data acquisition module is used to collect data for each input sample in the automated decision-making system, including the feature vector, model prediction score, and prediction confidence of each input sample.

[0014] The preprocessing module is used to preprocess the input sample data, including missing value imputation, outlier removal, format standardization, and unit unification.

[0015] The sample pool management module is used to divide the preprocessed input sample data into the observation period sample pool, the base period sample pool, and the mature sample pool according to the time window.

[0016] Furthermore, the model health assessment and forward-looking early warning module includes a meta-model inference module, a multi-dimensional surrogate index calculation module, and an adaptive threshold determination module;

[0017] The meta-model inference module maintains a pre-trained meta-model, which takes the prediction score, prediction confidence, and statistical characteristics of the sample feature vector in the recent distribution of the main evaluation model as input, and outputs the probability of the main evaluation model generating errors on the sample; and calculates the average probability of generating errors.

[0018] The multidimensional proxy index calculation module calculates the confidence shift based on the predicted confidence distribution of samples in the observation period sample pool and the base period sample pool; it calculates the feature shift step by step based on the feature vectors of samples in the observation period sample pool and the base period sample pool; and it calculates the health score of the main evaluation model based on the average error probability, confidence shift, and feature shift.

[0019] The adaptive threshold determination module updates the warning threshold based on the health score of each case in the historical failure case library; if the health score is less than the warning threshold for k consecutive days, it is determined that a warning is triggered; otherwise, no warning is triggered.

[0020] Furthermore, the formula for calculating the confidence shift is:

[0021]

[0022] in, Indicates confidence level shift. Denotes KL divergence, This represents the confidence distribution of samples in the observation period's sample pool. This represents the confidence distribution of samples in the base period sample pool. Indicates the number of samples in the observation period. The percentage of samples within each confidence interval Indicates the first sample in the base period sample pool The percentage of samples within each confidence interval Indicates the number of bins in the confidence interval;

[0023] The formula for calculating the feature offset is:

[0024]

[0025] in, Indicates feature offset, Represents the total number of features. Representation of features Importance weights in the model, As an indicator of group stability, Indicates the first The distribution of each feature in the sample pool during the current observation period. Indicates the first The distribution of each feature in the current base period sample pool.

[0026] Furthermore, the metamodel in the metamodel inference module is incrementally updated based on the new labeled samples after the true labels of the samples are obtained and confirmed.

[0027] Furthermore, the formula for updating the early warning threshold is:

[0028]

[0029] in, This indicates the updated warning threshold. Indicates the number of failure cases. Indicates the first The case was in the first [number] before failure. The corresponding health score for each day This represents the preset margin coefficient. This indicates taking the median. This is the set of time points within the warning window before failure.

[0030] Preferably, the concept drift detection and adaptive interpretation calibration module includes a concept drift detection module, an interpretability background distribution recalibration module, and an interpretation stability monitoring module;

[0031] The concept drift detection module performs drift detection on the meta-model and obtains the detection result of whether the model has drifted;

[0032] The interpretability background distribution recalibration module triggers interpretive calibration when the model detection result indicates drift. It calculates the contribution of each feature of the sample within the sliding window centered on the current observation period to the model prediction result, which is the interpretive result.

[0033] The explanation stability monitoring module uses the sum of the contributions of each feature to the model prediction results as the explanation vector, calculates the similarity between the explanation vectors of the current day and the previous day and the rank correlation coefficient of the feature importance ranking, and determines whether to trigger explanation calibration again.

[0034] Furthermore, the concept drift detection module performs drift detection on the meta-model, including:

[0035] The PSI values ​​of the features calculated from the observation period sample pool and the base period sample pool are used to determine if more than half of the features have a PSI value greater than a preset threshold. If so, it is determined that the model has drifted;

[0036] Alternatively, a linear trend indicator can be calculated based on the model's error rate on mature samples within the most recent L time windows; if the linear trend indicator is greater than 0 and the cumulative increase in the error rate exceeds a preset threshold, it is determined that the model has drifted.

[0037] Furthermore, the formula for calculating the contribution of each feature to the model's prediction results is as follows:

[0038]

[0039]

[0040] in, This represents the contribution value of the j-th feature calculated based on the base period distribution. For the set of all features, Features that are not included in F A subset, Only in subsets The model predictions calculated above, For the feature subset Add features The model predictions after that, This indicates that the result is calculated based on the sliding window sample distribution. The contribution value of each feature, This represents the sample distribution within the sliding window. Represents the samples within the sliding window The contribution value of the j-th feature.

[0041] Furthermore, the concept drift detection and adaptive interpretation calibration module also maintains a lightweight online alternative model that tracks changes in the meta-model decision logic in real time, serving as an auxiliary trigger signal for interpretation calibration.

[0042] The beneficial effects of this invention are as follows:

[0043] (1) This invention achieves forward-looking model degradation early warning under label delay conditions, effectively shortening the perception delay. Through meta-modeling and back-distillation techniques, this invention can predict the performance degradation trend of the main evaluation model in advance without waiting for real labels. Traditional methods typically require waiting for labels to mature before model failure can be detected, while this invention can significantly advance the warning point, providing the operations team with valuable early intervention time, thus changing the passive situation where perception lags behind failure occurrence in existing methods. Simultaneously, through multi-dimensional proxy indicators and adaptive threshold mechanisms, this invention can effectively reduce the frequency of false positive warnings and reduce unnecessary model retraining costs.

[0044] (2) This invention achieves dynamic adaptive calibration of interpretability in a concept drift environment. Through a sliding window dynamic recalibration mechanism, this invention enables the background distribution of the interpretability method to be automatically updated as the data distribution changes, thereby ensuring that the directional stability and ranking stability of the model interpretation remain at an acceptable level in dynamic environments. Traditional static interpretability methods experience a significant decrease in interpretability after concept drift occurs, even producing misleading analytical results. This invention effectively solves this problem through interpretability stability monitoring and automatic recalibration mechanisms, meeting the needs of scenarios with rigid requirements for model interpretability, such as financial regulation and medical auditing.

[0045] (3) The linkage between the early warning mechanism and the interpretation calibration mechanism of this invention improves the pertinence and efficiency of operation and maintenance. This invention integrates model degradation early warning and interpretability calibration into a unified framework. After the early warning mechanism identifies the occurrence of concept drift and assesses its impact on model performance, the adaptive interpretation mechanism immediately provides feature-level drift attribution analysis, enabling operation and maintenance personnel not only to know the potential problems of the model, but also to understand the specific manifestations and possible causes of the problems. In addition, this invention is not limited to specific business areas and can be applied to various automated decision-making scenarios centered on evaluating models. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the structure of the model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration in this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] To achieve proactive early warning of performance degradation in evaluation models within automated decision-making systems and interpretable adaptive calibration in contexts of concept drift, this invention proposes a model degradation proactive early warning and interpretable correction device based on meta-learning and sliding window calibration. This invention comprises two complementary core mechanisms. The first mechanism involves constructing a meta-model to predict the probability of prediction errors by the main evaluation model on samples without real labels. This meta-model, trained using backdistillation techniques, can detect model performance degradation trends in advance without real labels, thus significantly advancing the early warning point. Simultaneously, it distinguishes different degrees of drift through multi-dimensional surrogate indicators and adaptive thresholds, reducing the possibility of false positive warnings. The second mechanism involves automatically performing a sliding window dynamic recalibration of the interpretability background distribution when concept drift occurs. This ensures that the interpretation of feature contributions always reflects the true decision-making logic under the current data distribution. Simultaneously, a monitoring index for interpretability stability is set, automatically triggering recalibration when significant changes occur in the interpretation. A lightweight online alternative model is introduced to incrementally track the evolution of the main model's decision logic, serving as an auxiliary trigger signal for interpretation calibration. This allows for feedforward correction of the interpretation results when the main model's logic shifts, ensuring consistency between interpretability and model decisions. The two mechanisms are interconnected. The early warning mechanism continuously monitors the model's degradation risk during label delay. Once significant concept drift is detected and its impact on model performance is confirmed, the adaptive interpretation mechanism is activated, providing feature-level drift attribution analysis. The device specifically includes the following:

[0049] like Figure 1 As shown, the device proposed in this invention includes a multi-source data acquisition and sample management module, a model health assessment and prospective early warning module, a concept drift detection and adaptive interpretation calibration module, and an early warning notification and decision support module. The following provides a detailed description of each model:

[0050] (1) Multi-source data acquisition and sample management module

[0051] The multi-source data acquisition and sample management module is the data foundation of the device. It is used to collect and preprocess each input sample data in the automated decision-making system and divide the preprocessed input sample data into sample pools. This module includes a data acquisition module, a preprocessing module, and a sample pool management module.

[0052] The data acquisition module obtains raw data, i.e., the data of each input sample, through two methods: real-time streaming access and batch synchronization. This includes the feature vector, model prediction score and prediction confidence of each input sample, as well as the decision results returned synchronously by the business system and the delayed real labels.

[0053] The preprocessing module is used to preprocess the input sample data to ensure that data from different sources and at different times are comparable; preprocessing includes missing value imputation, outlier removal, format standardization, and unit unification.

[0054] The sample pool management module is used to divide the preprocessed input sample data into three pools based on time windows: the observation period sample pool, the base period sample pool, and the mature sample pool. Specifically, samples arriving within a set number of days before the current date are assigned to the observation period sample pool, samples from earlier historical periods are assigned to the base period sample pool, and samples with valid labels and sufficient performance periods (e.g., 1-3 months for short-term loans, 6-12 months for mortgages, and 3-6 months for industrial equipment failure predictions) are assigned to the mature sample pool. Each sample pool updates over time and manages the timestamp alignment of the samples to ensure correct association between feature data and subsequently acquired valid labels, avoiding evaluation bias caused by time mismatches.

[0055] (2) Model health assessment and forward-looking early warning module

[0056] The model health assessment and prospective early warning module calculates error probability, health score, and early warning threshold based on input sample data, and determines whether to trigger an early warning based on the health score and early warning threshold. This module includes a meta-model inference module, a multi-dimensional surrogate indicator calculation module, and an adaptive threshold determination module.

[0057] This module's function is to quantitatively assess and make a forward-looking judgment on the performance degradation risk of the main evaluation model during the current observation period, even before real labels are obtained. The core processing idea of ​​this module is to directly predict the probability of the main evaluation model making errors on unlabeled samples using a meta-model, i.e., generating an error probability estimate. This is supplemented by two dimensions: predicted confidence distribution shift and feature marginal distribution shift. These three factors are then combined into a comprehensive health score. Based on this, by comparing the health score with a warning threshold dynamically calibrated based on historical failure cases, it is determined whether there is a significant performance degradation risk and an warning is triggered. Specifically:

[0058] The metamodel inference module maintains a pre-trained metamodel. This model takes the prediction score and prediction confidence of the main evaluation model on the sample, as well as the statistical characteristics of the sample's feature vector in the recent distribution, as inputs. Its output is the probability that the main evaluation model might make a wrong prediction on that sample, i.e., the probability of making a wrong prediction. Statistical characteristics, such as the degree of deviation of the feature value from the base period mean, and the density of the feature in the nearest neighbor.

[0059] In some preferred embodiments of the present invention, the main evaluation model may be a network security threat interception model or an industrial production line quality inspection model.

[0060] The average probability of generating error is calculated based on the probability of generating error across all samples in the observation period sample pool using the main evaluation model. This probability is the baseline error probability estimate for the current time point:

[0061]

[0062] in The total number of samples during the observation period. For the meta-model to the first The probability of error in the master model predicted for each sample.

[0063] In some preferred embodiments of the present invention, the meta-model is trained using a reverse distillation technique, specifically for error prediction modeling of out-of-time samples, enabling it to capture error patterns of the main model under conditions of distribution variation.

[0064] The multidimensional proxy indicator calculation module calculates the confidence shift based on the predicted confidence distribution of samples in the observation period sample pool and the base period sample pool; it calculates the feature shift step by step based on the feature vectors of samples in the observation period sample pool and the base period sample pool; and it calculates the health score of the main evaluation model based on the average error probability, confidence shift, and feature shift. Specifically:

[0065] The multidimensional surrogate index calculation section is responsible for calculating two-dimensional auxiliary indices. The first dimension is the KL divergence shift of the predicted confidence distribution, which is the KL divergence of the model output confidence distribution compared to the base period. It reflects the overall change in the model's certainty regarding the current sample. The calculation formula is as follows:

[0066]

[0067] in, Indicates confidence level shift. Denotes KL divergence, This represents the confidence distribution of samples in the observation period's sample pool. This represents the confidence distribution of samples in the base period sample pool. Indicates the number of samples in the observation period. The percentage of samples within each confidence interval Indicates the first sample in the base period sample pool The percentage of samples within each confidence interval This indicates the number of bins in the confidence interval.

[0068] The second dimension is the weighted PSI offset of the marginal distribution of each feature, which summarizes the magnitude of the distribution change of each feature in the current observation period relative to the base period. It is used to quantify the degree of change in the input data itself. The calculation formula is as follows:

[0069]

[0070] in, Indicates feature offset, Represents the total number of features. Representation of features Importance weights in the model, As an indicator of group stability, Indicates the first The distribution of each feature in the sample pool during the current observation period. Indicates the first The distribution of each feature in the current base period sample pool.

[0071] The meta-model output is combined with the aforementioned two-dimensional auxiliary indicators according to preset weights to output the health score of the main evaluation model. The formula for calculating the health score is as follows:

[0072]

[0073] in, This represents the health score of the meta-model. This represents the average probability of generating an error in the test. This indicates the confidence shift after normalization. This represents the feature offset after normalization. , , These are the preset first, second, and third weighting coefficients, and their sum is 1; Health Score The value ranges from 0 to 100, and a higher score indicates that the model is more likely to remain effective at the current point in time.

[0074] In some preferred embodiments of the present invention, the metamodel in the metamodel inference module is incrementally updated based on new labeled samples after the true labels of the samples are acquired and confirmed, thereby continuously improving its ability to predict error patterns of the main model. The accumulated cases from earlier periods continuously optimize the warning thresholds and metamodel parameters for subsequent cycles, thus gradually improving the warning accuracy of the entire device over time without requiring repeated manual adjustments.

[0075] The adaptive threshold determination module updates the warning threshold based on the health score of each case in the historical failure case database; if the health score remains below the warning threshold for k consecutive days, a warning is triggered; otherwise, no warning is triggered. Specifically:

[0076] The adaptive threshold determination module dynamically calibrates the early warning threshold based on a historical failure case library. The system maintains a set of historical failure cases, recording the evolution trajectory of health scores in the days leading up to each confirmed failure by the main assessment model. Typical score intervals within the pre-failure early warning window are extracted from these cases, and the current threshold is dynamically calibrated accordingly. Daily health scores are received and compared with the early warning threshold. When the health score falls below the threshold for several consecutive days, a forward-looking early warning is triggered, and the warning level is output. The historical failure case library upon which this module relies is periodically updated by the historical record archive (historical case library) within the early warning notification and decision support module, allowing the threshold to be gradually optimized as actual usage scenarios change.

[0077] Suppose there are a total of The failure case, the first The case was in the first [number] before failure. The corresponding health score for the day is The warning threshold The update method is as follows:

[0078]

[0079] in, This indicates the updated warning threshold. Indicates the number of failure cases. Indicates the first The case was in the first [number] before failure. The health score corresponding to the day, the preferred one. This is obtained by statistically analyzing the average lead time before the prediction error probability of the meta-model begins to rise continuously before a significant conceptual drift occurs in the model's history. This represents the preset margin coefficient. This indicates taking the median. This is the set of time points within the warning window before failure.

[0080] Calculate current health score daily and the warning threshold The comparison and judgment logic is as follows:

[0081]

[0082] in The threshold is the number of consecutive days. If the above conditions are met, the model is deemed to have a significant risk of performance degradation, triggering a forward-looking warning and recording the warning level. If the score fluctuates but does not consistently fall below the threshold, it is considered a benign fluctuation and no warning is triggered.

[0083] This invention proposes a proactive early warning method for unlabeled model degradation based on meta-learning and backdistillation. This method constructs a meta-model to predict the prediction error probability of the main evaluation model on samples without real labels, and uses backdistillation technology to train the meta-model specifically for error prediction on out-of-time samples. Simultaneously, it combines multi-dimensional surrogate indicators such as the distribution change of model prediction confidence and the divergence of feature marginal distributions to construct a label-free model health score. Then, it dynamically calibrates the early warning threshold based on historical failure cases, thereby proactively detecting model performance degradation trends even when real labels are missing.

[0084] (3) Concept drift detection and adaptive interpretation calibration module

[0085] This module is another core function of the device, responsible for detecting changes in data distribution and dynamically calibrating model interpretability when necessary. The concept drift detection and adaptive interpretation calibration module includes a concept drift detection module, an interpretability background distribution recalibration module, and an interpretation stability monitoring module.

[0086] When the concept drift detection and adaptive interpretation calibration module determines that an early warning has been triggered, it performs drift detection on the meta-model. Upon confirmation of drift, it triggers interpretation calibration and determines whether to trigger it again. Specifically:

[0087] The concept drift detection module performs drift detection on the meta-model and obtains the detection result of whether the model has drifted;

[0088] Concept drift detection is performed on two levels. The first level is covariate drift detection, which monitors the PSI value of each feature in the current observation period relative to the base period. When the changes in most features exceed a preset threshold, it is determined that the input distribution has changed significantly. For each feature... The formula for calculating the PSI value of the current observation period relative to the base period is:

[0089]

[0090] in For the number of equal-frequency bins, and The first The feature in the first The percentage of samples in the current observation period and the base period within each bin. If more than half of the features have a PSI value greater than a preset threshold. If the input distribution changes significantly, it is considered that the input distribution has changed.

[0091] The second level is the detection of changes in conditional distribution. This is achieved by comparing the temporal trend of model prediction errors on recent mature samples to determine whether the mapping relationship between input and output has changed. Let's assume the recent... Within each time window (a full performance period is defined as one time window), the model's error rates on mature samples are as follows: Calculate the linear trend indicator as (number of mature samples with incorrect predictions within the window / number of mature samples within the window):

[0092]

[0093] in and These are the time sequence number and the mean error rate, respectively. If... And the cumulative increase in error rate exceeds the preset threshold, that is If a significant signal is detected in at least one of the two levels, the system considers the mapping relationship to have undergone conceptual drift. The system confirms the occurrence of conceptual drift and records the type and intensity of the drift when at least one significant signal is detected in either level.

[0094] The interpretability background distribution recalibration module triggers interpretive calibration when the model detection result indicates drift. It calculates the contribution of each feature of the samples within a sliding window centered on the current observation period to the model prediction result, which is the interpretive result. Specifically:

[0095] The interpretability background distribution recalibration module is initiated after drift confirmation. For SHAP-based interpretability methods, its calculation relies on a background dataset to estimate the baseline contribution of each feature. The SHAP value is calculated as follows:

[0096]

[0097] in, This represents the contribution value of the j-th feature calculated based on the base period distribution. For the set of all features, A subset of F (excluding features) ), Only in subsets The model predictions calculated above, For the feature subset Add features The model predictions are then updated. Standard SHAP uses all historical data as the background dataset. This invention switches this background dataset to a sliding window sample centered on the current observation period, i.e., switching from base period samples to the latest time window samples. The system maintains a fixed-length time window, which only includes samples arriving within the most recent few days, and slides forward daily as new samples arrive. When concept drift is confirmed, the SHAP background distribution is immediately switched to the sample distribution within this sliding window, and the SHAP values ​​of each feature, i.e., the contribution of each feature to the model prediction results, are recalculated.

[0098]

[0099] in, This indicates that the result is calculated based on the sliding window sample distribution. The contribution value of each feature, This represents the sample distribution within the sliding window. Represents the samples within the sliding window The contribution value of the j-th feature, The sample distribution within the sliding window ensures that the interpretation of feature contributions is always based on the current data distribution, thereby avoiding the continued influence of the old distribution on the interpretation results.

[0100] The stability monitoring module uses the sum of the contributions of each feature to the model's prediction results as the explanatory vector. It calculates the similarity between the explanatory vectors of the current day and the previous day, as well as the rank correlation coefficient of the feature importance ranking, and determines whether to trigger explanatory calibration again. Specifically:

[0101] After background distribution recalibration, the system continuously monitors the stability of the interpretation results. The current interpretation vector is calculated daily. Explanation vectors from the previous day Cosine similarity between :

[0102]

[0103] Calculate the Kendall rank correlation coefficient for ranking the importance of each feature:

[0104]

[0105] in For the total number of features, The number of feature pairs with consistent contribution ranking. The number of feature pairs with inconsistent contribution rankings. If for several consecutive days... Below the threshold or Below the threshold This indicates that the interpretation results are undergoing drastic changes, and the system automatically triggers a recalibration.

[0106] In some preferred embodiments of the present invention, the system maintains a lightweight online substitution model implemented using incremental Ridge regression. This model fits the prediction target of the main evaluation model and updates its parameters daily using new samples.

[0107]

[0108] in For learning rate, For Ridge regression loss function:

[0109]

[0110] The parameter change trend of the alternative model can indirectly reflect the change direction of the decision logic of the main model. When the feature weights of the alternative model drift significantly, it also serves as an auxiliary trigger signal for interpreting calibration.

[0111] This invention proposes an adaptive interpretability calibration and stability monitoring method for concept drift awareness. When concept drift is detected, this method switches the background distribution of the interpretability method from a fixed historical benchmark to a dynamically updated current distribution via a sliding window, achieving piecewise reweighting of the interpretation results. Simultaneously, it calculates the cosine similarity and Kendall rank correlation coefficient of interpretation results from adjacent time windows. When the stability index falls below a preset threshold, recalibration is automatically triggered. Furthermore, a lightweight online alternative model tracks changes in the main model's decision logic in real time, ensuring that the interpretability results always remain consistent with the model's actual decisions. This method, in conjunction with the aforementioned early warning mechanism, provides feature-level attribution information after the early warning mechanism identifies concept drift and confirms its impact on model performance, forming a complete technical process from perception to diagnosis.

[0112] (4) Early warning notification and decision support module

[0113] The early warning notification and decision support module is used to push early warning notification information to operations and maintenance personnel and receive operation instructions from them; it updates the historical failure case library and feeds the historical failure case library back to the model health assessment and proactive early warning module; specifically:

[0114] This module receives various early warning information and calibration results output by the preceding modules, pushes structured notification content to maintenance personnel, and is responsible for the accumulation and feedback of cases to support the adaptive iteration of the device. This module includes an information aggregation and push module, an operation feedback receiving module, and a historical record archiving and feedback module.

[0115] The information aggregation and push module summarizes the time-series trend of model health scores, warning trigger time and level, concept drift detection result summary, and feature importance comparison analysis after adaptive calibration into structured notifications, i.e., warning notification information. These are pushed to operations and maintenance personnel through enterprise communication tools, email, or operations and maintenance workbench. The push strategy is configured differently according to the warning level.

[0116] The operation feedback receiving module is responsible for receiving operation instructions returned by operation and maintenance personnel, such as confirming warnings, starting model retraining, and marking false alarms, and transmitting the confirmed results to the historical record archive.

[0117] The historical record archiving and feedback module maintains a complete history of early warnings and calibrations. Once the true label is finally obtained and confirmed, the health score sequence of the main evaluation model within this evaluation cycle, early warning trigger records, and the results of whether the model has truly failed are integrated into a complete case and stored in the historical failure case library. This case library is periodically fed back to the adaptive threshold determination module for threshold recalibration, and simultaneously fed back to the meta-model inference module for incremental updates of the meta-model. This allows the early warning threshold and the predictive ability of the meta-model to continuously improve over time without the need for repeated manual adjustments.

[0118] In summary, this invention provides automated decision-making systems with model operation and maintenance support that features forward-looking early warning capabilities and continuous interpretability. This invention does not rely on specific business semantics and can be applied to any automated decision-making scenario centered on model evaluation, demonstrating promising application prospects.

[0119] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration, characterized in that, include: The module includes: multi-source data acquisition and sample management, model health assessment and prospective early warning, concept drift detection and adaptive interpretation and calibration, and early warning notification and decision support. The multi-source data acquisition and sample management module is used to acquire each input sample data in the automated decision-making system and preprocess it, and then divide the preprocessed input sample data into sample pools. The model health assessment and prospective early warning module is used to calculate and generate error probability, health score and early warning threshold based on input sample data, and to determine whether to trigger an early warning based on the health score and early warning threshold. The concept drift detection and adaptive interpretation calibration module is used to perform drift detection on the meta-model when it is determined that an early warning has been triggered; and to trigger interpretation calibration and determine whether to trigger interpretation calibration again when it is confirmed that a drift has occurred. The early warning notification and decision support module includes an information aggregation and push module, an operation feedback receiving module, and a historical record archiving and feedback module. The information aggregation and push module pushes early warning notification information to operation and maintenance personnel, and the operation feedback receiving module receives operation instructions returned by operation and maintenance personnel. The historical record archiving and feedback module updates the historical failure case library and feeds the historical failure case library back to the model health assessment and forward-looking early warning module.

2. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 1, characterized in that, The multi-source data acquisition and sample management module includes a data acquisition module, a preprocessing module, and a sample pool management module; The data acquisition module is used to collect data for each input sample in the automated decision-making system, including the feature vector, model prediction score, and prediction confidence of each input sample. The preprocessing module is used to preprocess the input sample data, including missing value imputation, outlier removal, format standardization, and unit unification. The sample pool management module is used to divide the preprocessed input sample data into the observation period sample pool, the base period sample pool, and the mature sample pool according to the time window.

3. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 2, characterized in that, The model health assessment and forward-looking early warning module includes a meta-model inference module, a multi-dimensional proxy indicator calculation module, and an adaptive threshold determination module. The meta-model inference module maintains a pre-trained meta-model, which takes the prediction score and prediction confidence of the main evaluation model on the sample, as well as the statistical characteristics of the sample feature vector in the recent distribution, as inputs, and outputs the probability of the main evaluation model generating errors on the sample. Calculate the average probability of generating errors; The multidimensional proxy index calculation module calculates the confidence shift based on the predicted confidence distribution of samples in the observation period sample pool and the base period sample pool; and calculates the feature shift step by step based on the feature vectors of samples in the observation period sample pool and the base period sample pool. The health score of the main evaluation model is calculated based on the average error probability, confidence bias, and feature bias. The adaptive threshold determination module updates the warning threshold based on the health score of each case in the historical failure case library; if the health score is less than the warning threshold for k consecutive days, it is determined that a warning is triggered; otherwise, no warning is triggered.

4. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 3, characterized in that, The formula for calculating confidence shift is: ； in, Indicates confidence level shift. Denotes KL divergence, This represents the confidence distribution of samples in the observation period's sample pool. This represents the confidence distribution of samples in the base period sample pool. Indicates the number of samples in the observation period. The percentage of samples within each confidence interval Indicates the first sample in the base period sample pool The percentage of samples within each confidence interval Indicates the number of bins in the confidence interval; The formula for calculating the feature offset is: ； in, Indicates feature offset, Represents the total number of features. Representation of features Importance weights in the model, As an indicator of group stability, Indicates the first The distribution of each feature in the sample pool during the current observation period. Indicates the first The distribution of each feature in the current base period sample pool.

5. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 3, characterized in that, The metamodel in the metamodel inference module is incrementally updated based on the new labeled samples after the real labels of the samples are obtained and confirmed.

6. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 3, characterized in that, The formula for updating the warning threshold is: ； in, This indicates the updated warning threshold. Indicates the number of failure cases. Indicates the first The case was in the first [number] before failure. The corresponding health score for each day This represents the preset margin coefficient. This indicates taking the median. This is the set of time points within the warning window before failure.

7. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 1, characterized in that, The concept drift detection and adaptive interpretation calibration module includes a concept drift detection module, an interpretability background distribution recalibration module, and an interpretation stability monitoring module. The concept drift detection module performs drift detection on the meta-model to obtain the detection result of whether the model has drifted; The interpretability background distribution recalibration module triggers interpretation calibration when the model detection result indicates drift, and calculates the contribution value of each feature of the sample within the sliding window centered on the current observation period to the model prediction result, which is the interpretation result. The interpretation stability monitoring module uses the sum of the contributions of each feature to the model prediction results as the interpretation vector, calculates the similarity between the interpretation vectors of the current day and the previous day and the rank correlation coefficient of the feature importance ranking, and determines whether to trigger interpretation calibration again.

8. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 7, characterized in that, The process by which the concept drift detection module performs drift detection on the meta-model includes: The PSI values ​​of the features calculated from the observation period sample pool and the base period sample pool are used to determine if more than half of the features have a PSI value greater than a preset threshold. If so, it is determined that the model has drifted; Alternatively, a linear trend indicator can be calculated based on the model's error rate on mature samples within the most recent L time windows; if the linear trend indicator is greater than 0 and the cumulative increase in the error rate exceeds a preset threshold, it is determined that the model has drifted.

9. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 8, characterized in that, The formula for calculating the contribution of each feature to the model prediction result is as follows: ； ； in, This represents the contribution value of the j-th feature calculated based on the base period distribution. For the set of all features, Features that are not included in F A subset, Only in subsets The model predictions calculated above, For the feature subset Add features The model predictions after that, This indicates that the result is calculated based on the sliding window sample distribution. The contribution value of each feature, This represents the sample distribution within the sliding window. Represents the samples within the sliding window The contribution value of the j-th feature.

10. The model degradation prospective early warning and interpretability correction device based on meta-learning and sliding window calibration according to claim 7, characterized in that, The concept drift detection and adaptive interpretation calibration module also maintains a lightweight online substitution model that tracks changes in the meta-model decision logic in real time, serving as an auxiliary trigger signal for interpretation calibration.

Images