Generative distribution alignment unsupervised concept drift detection method
By employing a generative distribution-aligned unsupervised concept drift detection method, which utilizes a dual-generator architecture and dynamic weighting mechanism, the difficulty in distinguishing between real and virtual drifts in existing technologies is solved, and accurate detection is achieved in complex data flow environments.
Patent Information
- Application Number
- CN202511718151.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-21
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies cannot effectively distinguish between real and virtual drift and lack robustness in complex data stream environments, resulting in high false alarm rates and frequent missed detections.
An unsupervised concept drift detection method with generative distribution alignment is adopted. By constructing a dual generator architecture with a reference generator and an online generator, combined with a dynamic weighting mechanism and a downstream classifier, the drift score is quantified to identify concept drift.
It achieves accurate detection of concept drift, especially showing significant advantages in distinguishing between real drift and virtual drift, ensuring detection reliability in complex data flow environments.
Smart Images
Figure CN121580108A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of drift detection technology, and in particular to an unsupervised conceptual drift detection method based on generative distribution alignment. Background Technology
[0002] In real-time applications, such as financial risk control and IoT monitoring, data is generated in the form of continuous data streams. The core challenge is that the statistical patterns behind this data subtly change over time, causing previously trained predictive models to gradually become ineffective. This phenomenon is known as concept drift. For example, a credit card fraud pattern may evolve into a new pattern after a period of time, making older anti-fraud models unable to identify the new types of fraudulent behavior.
[0003] Concept drift is mainly divided into two types:
[0004] Reality drift refers to a change in the inherent patterns of things. For example, the underlying patterns of fraudulent behavior may change, rendering the original rules of judgment inapplicable.
[0005] Virtual drift: This refers to a change in the surface characteristics of data, but the underlying patterns remain the same. For example, the spending of normal users may generally increase, but the fundamental rules for identifying them as normal users have not changed.
[0006] To address the issue of concept drift, existing detection technologies are mainly divided into two categories:
[0007] Supervised detection methods: These methods rely on real-time, accurate label feedback, inferring concept drift by monitoring significant changes in model prediction performance metrics (such as classification error rate, F1 score, etc.). Although these methods are statistically reliable, their application is severely limited in real-world data stream environments due to limitations such as delays, high costs, or unavailability in label acquisition.
[0008] Unsupervised detection methods: These methods do not rely on label information and have wider applicability. Based on their theoretical foundation, they can be mainly divided into the following two categories:
[0009] Methods based on changes in data distribution: These methods assume that concept drift causes changes in the source distribution P(X). Detection is achieved by comparing the distribution differences of data features within different time windows. Typical techniques include spatial partitioning methods, histogram modeling methods, and methods based on statistical distance metrics (such as Wasserstein distance, maximum mean difference, etc.).
[0010] Model behavior analysis-based methods assume that concept drift increases uncertainty in model decisions. Drift is detected indirectly by monitoring changes in model output metrics, primarily including methods based on prediction confidence and methods based on prediction consistency.
[0011] The aforementioned shortcomings of existing detection technologies:
[0012] 1. Insufficient ability to distinguish drift types: Existing methods cannot effectively distinguish between real drift and virtual drift. Methods based on changes in data distribution will respond to changes in the distribution P(X) from any source, including virtual drifts that do not affect the decision boundary, leading to a significant increase in false alarm rate, causing unnecessary model updates and wasted resources.
[0013] 2. Limitations of the detection hypothesis: The hypothesis of "increased decision uncertainty" upon which model-based behavioral analysis methods rely does not always hold true in real-world scenarios. Especially in cases of sudden drift, the model may make high-confidence erroneous predictions for samples beyond its cognitive range, leading to serious missed detections.
[0014] 3. Poor adaptability to complex data environments: Real-world data streams are often characterized by high dimensionality, non-stationarity, and noise, making existing methods unstable in terms of distribution estimation and uncertainty measurement, and difficult to achieve an effective balance between detection sensitivity and robustness. Summary of the Invention
[0015] To address the problems of existing technologies failing to effectively distinguish between real and virtual drift, and lacking robustness in complex data flow environments, the present invention aims to provide a generative distributed alignment-based unsupervised concept drift detection method. This method achieves accurate detection of concept drift, exhibiting significant advantages, particularly in distinguishing between real and virtual drift, and ensuring detection reliability in complex data flow environments.
[0016] To achieve the above objectives, the present invention provides the following solution:
[0017] An unsupervised concept drift detection method based on generative distribution alignment includes:
[0018] Historical data is acquired, and a reference generator and a discriminator are trained based on the historical data. During the training process, the objective function of the reference generator is set to obtain the optimal reference generator and the optimal discriminator.
[0019] Initialize the online generator and online discriminator, and inherit the parameters of the optimal reference generator and the optimal discriminator, respectively;
[0020] Real data is acquired, and an inherited online generator is trained using the real data. The trained online generator is then used to identify concept drift. During training, the historical data and the real data are input into an auxiliary classifier to obtain similarity weights between the distributions of the real and historical data. These weights are used to dynamically adjust the update direction of the inherited online generator. Simultaneously, the samples generated by the online generator using the real data are used to train the online discriminator to evaluate the distribution differences between the generated samples and the real data, providing a learning signal for the online generator.
[0021] Optionally, the loss function for training the discriminator based on the historical data includes:
[0022] ;
[0023] in, Let the loss function of the discriminator be , Let be a random noise vector in the latent space. , The gradient penalty coefficient is... For samples from the real data distribution, For fake historical data generated using random noise z, The mean, The discriminator assigns scores to the data. For authentic historical data, To calculate the gradient of the input data X, These are sample points randomly sampled from real historical data and generated data.
[0024] Optionally, setting the objective function of the reference generator includes:
[0025] Using the historical data, the objective function of the reference generator is set to generate the sample with the highest discriminator score.
[0026] ;
[0027] in, The objective function for the reference generator.
[0028] Optionally, obtaining the similarity weights includes:
[0029] ;
[0030] in, For similarity weights, For the sigmoid function, To generate the classifier in the model, The new high of discrimination for input sample X is in the form of logits.
[0031] Optionally, dynamically adjusting the update direction of the inherited online generator includes:
[0032] ;
[0033] in, Let the loss function be that of the online generator. , The average score of the samples generated by the two generators on the discriminator.
[0034] Optionally, the loss function for training the online discriminator includes:
[0035] ;
[0036] in, Let the loss function be that of the online discriminator. , The average score of the samples generated by the two generators on the discriminator.
[0037] Optionally, the method further includes:
[0038] Obtain random vectors from the latent space, and input the random vectors into the optimal reference generator and the trained online generator respectively to generate corresponding sample pairs;
[0039] The sample pairs are input into a downstream classifier to obtain prediction results, which are used to calculate the drift score.
[0040] The drift score is subjected to an exponentially weighted average. If the result of the exponentially weighted average exceeds a preset threshold, it proves that a drift has occurred, triggering the model update mechanism and re-initializing the detection cycle.
[0041] Optionally, calculating the drift score includes:
[0042] ;
[0043] in, For drift score, For indicator functions, For batch size, for, This is used by the downstream classifier to classify the samples generated by the two generators.
[0044] Optionally, performing an exponentially weighted average on the drift score includes:
[0045] ;
[0046] in, This is the result of an exponentially weighted average. For smoothing coefficients, This represents the sequence number of the time window.
[0047] The beneficial effects of this invention are as follows:
[0048] A novel detection paradigm based on generative distribution alignment: Unlike traditional methods that directly compare data distributions or monitor model outputs, this invention constructs generative approximations of historical and current distributions, samples from a shared latent space, and quantifies drift by utilizing the inconsistency in predictions by downstream classifiers on generated sample pairs. This improvement shifts the focus of detection from changes in surface features to a fundamental shift in the data generation mechanism.
[0049] Design of a dual-generator architecture and dynamic weighting mechanism: In response to the non-stationary nature of data streams, this invention adopts a dual-model architecture of a reference generator and an online generator. The dynamic weighting mechanism automatically evaluates the similarity of distributions, balancing the preservation of historical knowledge with the learning of new distributions, thus resolving the contradiction between the integrity and timeliness of representation in traditional windowing methods.
[0050] Establishing a drift scoring system based on latent consistency: This invention proposes to calculate the predicted difference of sample pairs generated from homologous latent points by fixing a classifier to obtain a drift score, so that concept drift is mainly manifested as a change in the mapping relationship from the latent space to the input space, thereby enabling more accurate identification of real concept drift.
[0051] This invention achieves accurate detection of concept drift, and shows a significant advantage in distinguishing between real drift and virtual drift, ensuring detection reliability in complex data flow environments. Attached Figure Description
[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0053] Figure 1 This is a flowchart of an unsupervised concept drift detection method based on generative distribution alignment according to an embodiment of the present invention;
[0054] Figure 2 This is a schematic diagram of online data stream distribution fitting according to an embodiment of the present invention;
[0055] Figure 3 This is a conceptual drift diagram of an embodiment of the present invention;
[0056] Figure 4 This is a schematic diagram illustrating the impact of the training ratio on online learning in an embodiment of the present invention;
[0057] Figure 5 is a schematic diagram of the algorithm detection effect of the embodiment of the present invention; (a) is a line graph showing the performance score of training and prediction using the first dataset 2D4C-Sudden on four classification models: random forest, SVM, KNN, and neural network, and recording the performance score of each time window; (b) is a line graph showing the performance score of training and prediction using the second dataset 2D4C-Gradual on four classification models: random forest, SVM, KNN, and neural network, and recording the performance score of each time window; (c) is a line graph showing the performance score of training and prediction using the third dataset 2D4C-Incremental on four classification models: random forest, SVM, KNN, and neural network, and recording the performance score of each time window. Line graphs of the data: (d) is a line graph showing the performance scores of four classification models (random forest, SVM, KNN, and neural network) trained and predicted using the fourth dataset Result (Sudden), with each time window recorded; (e) is a line graph showing the performance scores of four classification models (random forest, SVM, KNN, and neural network) trained and predicted using the fifth dataset Result (Gradual), with each time window recorded; (f) is a line graph showing the performance scores of four classification models (random forest, SVM, KNN, and neural network) trained and predicted using the sixth dataset Result (Incremental), with each time window recorded. Detailed Implementation
[0058] 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.
[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0060] like Figure 1As shown in the figure, this embodiment discloses an unsupervised concept drift detection method based on generative distribution alignment, including: acquiring historical data; training a reference generator and a discriminator based on the historical data; during the training process, setting the objective function of the reference generator to obtain the optimal reference generator and the optimal discriminator; initializing an online generator and an online discriminator, and inheriting the parameters of the optimal reference generator and the optimal discriminator respectively; acquiring real data; training the inherited online generator with the real data; and using the trained online generator to identify concept drift; during the training process, inputting the historical data and the real data into an auxiliary classifier to obtain the similarity weights between the distributions of the real and historical data, which are used to dynamically adjust the update direction of the inherited online generator; and simultaneously using the samples generated by the online generator with the real data to train the online discriminator to evaluate the distribution difference between the generated samples and the real data, providing a learning signal for the online generator.
[0061] This embodiment discloses an unsupervised concept drift detection method based on generative distribution alignment, including:
[0062] Step 1: The overall process begins with the historical distribution modeling stage. A generative adversarial network framework is used to train a reference generator on historical data, enabling it to learn the distribution characteristics of historical data and fix the parameters to establish a benchmark reference system for subsequent detection.
[0063] Step 2: Online Distribution Learning and Dynamic Adaptation Stage. First, the online generator is initialized and inherits the parameters of the reference generator to ensure knowledge continuity. Then, the similarity weights of the new and old data distributions are calculated through the auxiliary classifier. Based on these weight values, the update direction of the online generator is dynamically adjusted. When the weight is high, the emphasis is on preserving historical features, and when the weight is low, the emphasis is on redistribution adaptation, thus achieving a balance between model stability and adaptability.
[0064] Step 3: In the drift detection and quantization stage, the system samples from the latent space and inputs them into two generators to generate sample pairs. The drift score is obtained by calculating the proportion of inconsistencies in predictions through the downstream classifier. This score reflects the degree of influence of changes in data distribution on the model's decision.
[0065] Step 4: Finally, in the decision and update phase, the drift score is smoothed over time. When it exceeds a preset threshold, concept drift is determined to have occurred, triggering the model update mechanism and restarting the detection cycle. The entire technical solution achieves accurate identification of concept drift through generative distribution alignment and prediction consistency checks.
[0066] S1. Specific implementation of distributed modeling:
[0067] S11. This invention employs a generative modeling method to construct a data distribution representation. Its core is learning a compact representation of the data distribution through a generative adversarial network framework. In the historical distribution modeling stage, an initial data window is used to train a reference generator G1 and a discriminator D1. The generator uses a deep neural network architecture to map random noise in the latent space to the data space, generating synthetic samples that conform to the historical data distribution. The discriminator is optimized through a loss function to ensure accurate differentiation between real historical data and generated samples. After training, the parameters of the reference generator G1 are fixed, serving as the distribution benchmark for subsequent drift detection. The specific implementation is as follows:
[0068] S12. Model Architecture and Training Objectives: The system uses the WGAN-div framework to construct a generative model. The optimization objective of this framework is based on Wasserstein distance, combined with a gradient penalty term to ensure training stability. The loss function of the discriminator D1 is defined as:
[0069] ;
[0070] Where z is a random noise vector in the latent space, x is a sample in the real data distribution, and k and p are gradient penalty coefficients. The first and second terms calculate the Wasserstein distance between the generated distribution and the real distribution, and the third term forces the discriminator to satisfy the 1-Lipschitz continuity condition through gradient penalty.
[0071] S13. The optimization objective of generator G is to minimize:
[0072] ;
[0073] By minimizing this objective function, the generator learns to generate samples that give the discriminator the highest score, thereby driving the generation distribution p. g Continuously approximating the true data distribution p data .
[0074] S14. Technical Motivation: In non-stationary data stream environments, data distributions evolve continuously over time, lacking global consistency. Traditional drift detection methods typically employ a sliding window mechanism, retaining only a fixed number of recent samples to estimate the current distribution. This method implicitly assumes that a short-term sequence of recent samples is sufficient to characterize the current complete data distribution. However, under non-stationary conditions, this assumption is difficult to hold. Figure 2 As shown, a local window can only capture a limited segment of the data generation process and often cannot fully reflect its overall structure.
[0075] While increasing the window size can improve distribution coverage, it inevitably introduces additional detection latency, potentially spanning multiple drift stages, thus exacerbating the complexity of distribution modeling. This inherent contradiction between timeliness and representation fidelity constitutes the fundamental limitation of window-based methods in dynamic environments.
[0076] To address this challenge, this invention employs a generative modeling approach to mitigate the structural incompleteness problem. By training a generative model online to approximate the current distribution of the streaming data, more representative distribution estimates can be generated without increasing memory usage or introducing additional latency. This approach effectively alleviates the inherent representational deficiencies of the sliding window method, providing a more stable foundation for downstream alignment and drift detection in rapidly evolving data streams.
[0077] S2, Online Distributed Learning:
[0078] S21. Dual Generator Framework Construction: This invention constructs a dual generator architecture consisting of a reference generator and an online generator for distribution alignment and concept drift detection. One of the generators, mentioned in S1, is used to learn the old data distribution. The online generator G2 and the online discriminator D2 inherit the parameters of G1 and D2 respectively during system initialization.
[0079] S22. Dynamic Weight Calculation and Distribution Alignment: The system achieves accurate distribution alignment through a dynamic weight calculation module. The training objective function of the auxiliary classifier C is:
[0080] ;
[0081] in, For the sigmoid function, N is the batch size.
[0082] Based on the output of classifier C, the formula for calculating the dynamic weight w is:
[0083] ;
[0084] This weight value reflects the alignment between the old and new distributions in real time. When w approaches 1, it indicates a high degree of similarity between the distributions, and when it approaches 0, it indicates a significant difference between the distributions.
[0085] S23. Online generator update mechanism:
[0086] The update process of the online generator G2 employs a composite loss function design, which is one of the core innovations of this invention. Specifically, the optimization objective of G2 is jointly guided by the reference discriminator D1 and the online discriminator D2, and its loss function is defined as:
[0087] ;
[0088] This composite loss function achieves a precise balance between preserving historical distributions and learning new distributions through dynamic weights w. When w is high, it indicates a high similarity between the current data distribution and historical distributions. In this case, the loss function is mainly dominated by the first term, forcing the online generator G2 to maintain consistency with the reference generator G1 and preventing the forgetting of historical distribution knowledge. When w is low, it indicates a significant change in distribution. The loss function then becomes dominated by the second term, driving G2 to converge quickly to the new data distribution.
[0089] Considering the constraint that historical original samples are not stored, the update of the reference discriminator D1 is achieved by comparing the outputs of G2 and G1, and its loss function is:
[0090] ;
[0091] This design ensures that D1 can continuously evaluate the deviation of the online generator G1 from the reference generator G1, while maintaining the Lipschitz continuity of the discriminator through a gradient penalty term. Meanwhile, the online discriminator D2 is trained using newly arrived real data and samples generated by the online generator G2, with the following optimization objective:
[0092] ;
[0093] This design enables D2 to accurately assess the distributional differences between generated samples and real new data samples, providing accurate learning signals for the online generator G2.
[0094] S24. Technological Motivation: In dataflow learning, concept drift manifests as a dynamic change in the distribution of categories in the feature space. Based on whether it affects the decision boundary, concept drift can be divided into real drift and virtual drift, such as... Figure 3 As shown. This distinction is crucial for maintaining model performance, but existing unsupervised detection methods are significantly lacking in this regard.
[0095] Traditional unsupervised methods primarily focus on the temporal changes of the marginal feature distribution P(X), but fail to effectively distinguish between real and virtual drifts. Distribution comparison-based methods respond to any change in the feature distribution, including virtual drifts that do not affect the decision boundary, leading to a significantly higher false alarm rate in complex environments.
[0096] To address this problem, some research has turned to heuristic methods based on decision boundaries. These methods assume that an increase in sample density near the decision boundary indicates the occurrence of concept drift. However, this assumption has significant limitations: virtual drift may only bring samples close to the decision boundary without actually crossing it, triggering false alarms; while in sudden or structural drift, real drift may cause the new distribution to completely avoid the original decision region, without leading to an increase in boundary density, thus causing missed detections.
[0097] In unsupervised environments, the true correspondence between distributions before and after drift cannot be directly observed due to the lack of labels and prior information. This fundamental challenge makes accurately distinguishing drift types extremely difficult.
[0098] To address the aforementioned problems, this invention proposes an innovative distribution alignment scheme based on a dual-generator architecture. The core idea of this scheme is to establish effective alignment between old and new distributions in an unsupervised environment by constructing a structured distribution correspondence. Specifically, the system utilizes a reference generator and an online generator to fit historical and current data distributions respectively, achieving structural alignment of distributions through sharing a latent space. This generative distribution alignment method provides a new technical path to overcome the limitations of traditional unsupervised methods in distinguishing between real and virtual drift.
[0099] It is also worth noting that a key design consideration in the generative distribution alignment framework of this invention is the ratio of training iterations for the generator to those for the discriminator. This hidden parameter has a significant impact on the stability of model training and the quality of distribution fitting.
[0100] From the theoretical perspective of adversarial training, a dynamic balance needs to be maintained between the generator and the discriminator. When the discriminator is undertrained, its gradient signal contains more noise and is directionally unstable, causing oscillations in the generator's update process and making it difficult to converge to the ideal distribution representation. This instability further affects the distribution alignment quality in the latent space, thus disrupting the semantic coherence of the generated samples.
[0101] Conversely, when the discriminator is overtrained, its discriminative power may become too strong, causing the gradient signals received by the generator to become sparse and lack sufficient information. In this case, the generator's update pace will be over-constrained, making it difficult to fully explore the feature space of the data distribution, thus affecting the model's speed of adaptation to new distributions, such as... Figure 4 As shown.
[0102] This invention, through theoretical analysis and experimental verification, reveals that maintaining an appropriate training ratio is crucial for ensuring the stability of the distribution alignment process. A suitable ratio allows the discriminator to provide sufficiently informative and stable gradient guidance, while ensuring the generator has enough flexibility to adapt to distribution changes. This delicate balance ensures that the model's distribution mapping in the latent space maintains semantic coherence, providing a reliable foundation for downstream concept drift detection.
[0103] S3, Online Distributed Learning:
[0104] After completing the distribution alignment, the system enters the drift detection stage. This stage achieves accurate identification of concept drift by quantitatively analyzing the predicted differences between sample pairs.
[0105] Drift score calculation is the core step in the detection process. The system samples N random vectors z from the potential space. i The reference generator G1 and the online generator G2 are input respectively to generate corresponding sample pairs. These sample pairs are then used by the downstream classifier f to obtain prediction results. The drift score d is calculated using the following formula:
[0106] ;
[0107] Here, I(·) is an indicator function that takes a value of 1 when the predictions are inconsistent. This score reflects the degree of consistency in model predictions under the same semantic context.
[0108] Time smoothing is used to improve detection stability. The system performs an exponentially weighted average of the raw drift scores:
[0109] ;
[0110] Here, α is a smoothing coefficient that controls the weighting of the current score and historical scores. This process effectively suppresses interference caused by instantaneous fluctuations.
[0111] The decision-making mechanism determines drift based on a preset threshold. During offline verification, the system determines the detection threshold τ by analyzing the drift score distribution across multiple detection windows. During online detection, if the smoothed drift score exceeds the threshold, a drift is considered to have occurred.
[0112] ;
[0113] The model update strategy is initiated after drift is confirmed. The system assigns the parameters of the current online generator G2 to the reference generator G1 and reinitializes the detection period. This mechanism ensures that the baseline distribution can be updated in a timely manner to adapt to the new data environment. The final algorithm pseudocode is shown in Table 1.
[0114] Table 1
[0115]
[0116] The online experimental results on the dataset with concept drift are shown in Figures 5(a)-(f). The broken lines represent the performance values of the downstream model in each small window sequence. The dashed lines in the upper three rows represent drift occurrences, including both virtual and real concept drifts, while the dashed lines in the lower three rows represent actual detections. After a detection occurs, retraining is performed directly using the latest window.
[0117] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. An unsupervised concept drift detection method based on generative distribution alignment, characterized in that, include: Historical data is acquired, and a reference generator and a discriminator are trained based on the historical data. During the training process, the objective function of the reference generator is set to obtain the optimal reference generator and the optimal discriminator. Initialize the online generator and online discriminator, and inherit the parameters of the optimal reference generator and the optimal discriminator, respectively; Acquire real data, train an inherited online generator using the real data, and use the trained online generator to identify concept drift. During training, the historical data and the real data are input into the auxiliary classifier to obtain the similarity weights between the distribution of the real and historical data. These weights are used to dynamically adjust the update direction of the inherited online generator. At the same time, the samples generated by the online generator using the real data are used to train the online discriminator to evaluate the distribution differences between the generated samples and the real data, thus providing a learning signal for the online generator.
2. The unsupervised concept drift detection method with generative distribution alignment according to claim 1, characterized in that, The loss function for training the discriminator based on the historical data includes: ； in, Let the loss function be that of the discriminator. Let be a random noise vector in the latent space. , The gradient penalty coefficient is... For samples from the real data distribution, For fake historical data generated using random noise z, The mean, The discriminator assigns scores to the data. For authentic historical data, To calculate the gradient with respect to the input data X, These are sample points randomly sampled from real historical data and generated data.
3. The unsupervised concept drift detection method with generative distribution alignment according to claim 1, characterized in that, Setting the objective function of the reference generator includes: Using the historical data, the objective function of the reference generator is set to generate the sample with the highest discriminator score. ； in, The objective function for the reference generator.
4. The unsupervised concept drift detection method with generative distribution alignment according to claim 1, characterized in that, Obtaining the similarity weights includes: ； in, For similarity weights, For the sigmoid function, To generate the classifier in the model, The new high of discrimination for input sample X is in the form of logits.
5. The unsupervised concept drift detection method with generative distribution alignment according to claim 1, characterized in that, Dynamically adjusting the update direction of the inherited online generator includes: ； in, Let the loss function be that of the online generator. , The average score of the samples generated by the two generators on the discriminator.
6. The unsupervised concept drift detection method with generative distribution alignment according to claim 1, characterized in that, The loss function for training the online discriminator includes: ； in, The loss function of the online discriminator. , The average score of the samples generated by the two generators on the discriminator.
7. The unsupervised concept drift detection method with generative distribution alignment according to claim 1, characterized in that, The method also includes: Obtain random vectors from the latent space, and input the random vectors into the optimal reference generator and the trained online generator respectively to generate corresponding sample pairs; The sample pairs are input into a downstream classifier to obtain prediction results, which are used to calculate the drift score. The drift score is subjected to an exponentially weighted average. If the result of the exponentially weighted average exceeds a preset threshold, it proves that a drift has occurred, triggering the model update mechanism and re-initializing the detection cycle.
8. The unsupervised concept drift detection method with generative distribution alignment according to claim 7, characterized in that, Calculating the drift score includes: ； in, For drift score, For indicator functions, For batch size, for, This is used by the downstream classifier to classify the samples generated by the two generators.
9. The unsupervised concept drift detection method with generative distribution alignment according to claim 7, characterized in that, The exponentially weighted average of the drift scores includes: ； in, This is the result of an exponentially weighted average. For smoothing coefficients, This represents the sequence number of the time window.