Elastic memory drift detection and hot update method and system based on pseudo-label empowerment

Abstract

The application discloses a kind of based on pseudo label empowerment's elastic memory drift detection and hot updating method and system, method includes: obtaining new sample in data stream, to unlabelled sample in new sample, predict its class as pseudo label using current prediction model parameter, simultaneously calculate the confidence of prediction model prediction, and as the original confidence weight of this pseudo label;The new sample of empowerment weight is stored to elastic memory module;Based on the difference between historical data distribution and new received data distribution in elastic memory module, determine drift detection index;When drift detection index exceeds the drift threshold of adaptive change, determine that concept drift occurs;When detecting concept drift, trigger online prediction model update;After training convergence, seamlessly replace the prediction model of current online operation with updated prediction model.The application can efficiently utilize unlabelled data and timely detect and adapt to concept drift model update.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and machine learning technology, specifically to a method and system for elastic memory drift detection and hot update based on pseudo-label weighting. Background Technology

[0002] As artificial intelligence systems are increasingly used in online services and the Internet of Things (IoT), models need to continuously adapt to changes in data distribution over long periods of operation. This phenomenon of data distribution changing over time is commonly known as concept drift. Concept drift can cause existing models to degrade in performance on new data because the models were initially trained on data with different distributions. Traditional solutions include periodically retraining the model with new data or redeploying the model when drift is detected. However, these methods typically rely on large amounts of manually labeled data as a training basis, and acquiring and labeling such massive amounts of streaming data is both expensive and time-consuming. In fact, in high-speed data streams, it is impossible for humans to label every sample in a timely manner, and the amount of actually available labeled data is often very limited.

[0003] Some existing online learning and incremental learning techniques attempt to update models with limited labeling, such as semi-supervised learning methods using unlabeled data. However, directly training the model using predictions of unlabeled samples (i.e., pseudo-labels) can introduce erroneous information, and indiscriminate use of pseudo-labels can further degrade model performance. Additionally, some methods use fixed-size historical data buffers (memory) to prevent the model from forgetting previously learned knowledge, but fixed-size memory is ill-suited to adapting to constantly changing data scales and concept drift frequencies: too small a capacity leads to the overwriting and forgetting of important historical data, while too large a capacity increases storage and computational costs and may introduce stale data that interferes with the current model. Patent application CN113033643A provides a concept drift detection method, system, and electronic device based on weighted sampling. This patent primarily addresses the current methods for solving data drift, which cannot intelligently detect the degree of drift at each time point and update the model accordingly, thus increasing unnecessary retraining and update costs.

[0004] In summary, there is an urgent need for a model update method that can efficiently utilize unlabeled data and promptly detect and adapt to concept drift under limited manual annotation conditions. An ideal method should achieve continuous model learning without interrupting system services: it should be able to extract effective information from unlabeled data through pseudo-label weighting strategies, dynamically save necessary historical data through elastic memory to prevent forgetting, and possess a drift detection mechanism to identify significant changes in data distribution and trigger hot model updates, thereby maintaining high accuracy and stability of the model in new environments. Summary of the Invention

[0005] The technical problem to be solved by this invention is: in a data stream environment, how to make full use of unlabeled data to improve the continuous learning ability of the model when there is only a very small amount of manually labeled data, automatically detect significant changes in data distribution and quickly update the model, so as to ensure that the model performance remains stable and reliable over time.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A method for elastic memory drift detection and hot update based on pseudo-label weighting includes:

[0008] New samples are acquired from the data stream. For unlabeled samples in the new samples, their categories are predicted using the current prediction model parameters and used as pseudo-labels. At the same time, the confidence level of the prediction model is calculated and used as the original confidence weight of the pseudo-label.

[0009] The new samples with assigned weights are stored in the elastic memory module; when the number of stored samples reaches the capacity, the sample with the lowest contribution to the current prediction model or the earliest stored sample is removed.

[0010] Based on the difference between the distribution of historical data and the distribution of newly received data in the elastic memory module, a drift detection index is determined; when the drift detection index exceeds the adaptively changing drift threshold, a concept drift is determined to have occurred.

[0011] When concept drift is detected, the online prediction model is updated; the historical data in the elastic memory module is merged with the current new data into a training set, the prediction model is incrementally trained, and the prediction model parameters are updated by minimizing the overall loss function.

[0012] After training converges, the updated prediction model will seamlessly replace the currently running prediction model.

[0013] Technical Effects: This invention makes full use of unlabeled data: Through a pseudo-label weighting strategy, it can effectively utilize massive amounts of unlabeled data for training with only a small amount of manually labeled data. The prediction model learns fully from high-confidence pseudo-labeled samples while treating low-confidence samples cautiously, reducing the interference of erroneous pseudo-labels on the model.

[0014] Flexible memory prevents forgetting and balances new and old knowledge: The flexible memory module can dynamically adjust its storage capacity and content as needed, ensuring that the predictive model remembers important historical samples while also keeping up with the latest data. Compared to a fixed-size buffer, the flexible memory of this invention can expand its storage to include new concept data when concepts change, and shrink its storage to reduce interference from outdated information when concepts are stable. This mechanism effectively alleviates the problem of catastrophic forgetting, enabling the model to quickly recover its recognition ability when encountering recurring old patterns.

[0015] Real-time drift detection prevents sudden drops in predictive model performance: The drift detection mechanism provided in this invention can promptly detect significant changes in data distribution, providing early warning and triggering updates before a substantial decrease in predictive model accuracy. Compared to passively waiting for predictive model performance to deteriorate before retraining, proactive drift detection can minimize the magnitude and duration of performance degradation. This ensures the performance stability of the model during long-term online operation and improves system reliability.

[0016] Hot Model Updates for Seamless Continuous Service: This invention utilizes hot update technology to update model parameters online without taking the model offline or interrupting service. Compared to traditional methods that require offline batch retraining and redeployment, hot updates significantly reduce the latency for the model to adapt to new environments. In one test, the method of this invention began adjusting the model almost immediately after detecting concept drift, restoring the model to a high-precision state in a very short time (e.g., within seconds), while traditional retraining methods often require several hours of offline training. Therefore, this invention greatly improves the system's response speed to environmental changes.

[0017] In this embodiment, the contribution of the current prediction model is obtained using the following formula:

[0018] ;

[0019] In the formula, For sample-based Sensitivity to the loss of the current prediction model For gradient, These are the parameters of the current prediction model. The loss of the current prediction model, This is the output of the current prediction model. For the sample The tag, It is a norm.

[0020] In this embodiment, the difference between the historical data distribution and the newly received data distribution is calculated using Kullback–Leibler divergence, Jensen–Shannon divergence, or a combination of both, to determine the drift detection index.

[0021] In this embodiment, the drift detection index is determined by the Kullback–Leibler divergence, using the following formula:

[0022] ;

[0023] In the formula, Drift detection metrics determined for Kullback–Leibler divergence For historical data distribution, For the distribution of newly received data, For sample index.

[0024] In this embodiment, the drift detection index is determined by the Jensen-Shannon divergence, using the following formula:

[0025] ;

[0026] ;

[0027] In the formula, Drift detection index determined for Jensen–Shannon divergence. For historical data distribution, For the distribution of newly received data, for and The differences between them for and The average distribution.

[0028] In this embodiment, an adaptively changing drift threshold is dynamically updated based on the average error of the current prediction model output of the edge node, and is calculated using the following formula:

[0029] ;

[0030] In the formula, For adaptively changing drift threshold, As the baseline drift threshold, This is an operator that takes the minimum value among multiple input values. It is an empirical constant. This represents the average error output by the current prediction model based on the edge nodes.

[0031] In this embodiment, the expression for minimizing the overall loss function is:

[0032] ;

[0033] In the formula, To minimize the overall loss function, This represents the number of labeled samples. For prediction model parameters, The number of pseudo-labeled samples. For the first A labeled sample The output of the time-ahead prediction model For the first The label of each sample, The tradeoff coefficient for semi-supervised loss, For pseudo-label samples Corresponding to the confidence weights after discounting, For the first A pseudo-label sample, For the first Pseudo-labels for each sample This represents the loss of the current prediction model.

[0034] In this embodiment, without changing the online service, a parameter importance penalty term is introduced into the minimization of the overall loss function to obtain the optimized minimization of the overall loss function, which is expressed by the following formula:

[0035] = ;

[0036] ;

[0037] In the formula, The optimized function minimizes the overall loss function. The initial coefficient for the penalty term, The attenuation constant of the penalty coefficient, It is a natural constant. For time, For prediction model parameters Importance coefficient, For the first The updated prediction model parameters Predictive model parameters after the previous training stage The value of .

[0038] In this embodiment, the discounted confidence weight is represented by the following formula:

[0039] ;

[0040] ;

[0041] ;

[0042] In the formula, The confidence weight after discounting. The original confidence weights, For activation function, Predicting the category to which a sample belongs in the predictive model The probability, These are two predicted probability distributions obtained for the same unlabeled sample under different perturbations. The total number of categories, To predict the entropy value of the probability distribution, The degree of difference between the two predicted distributions. , They are respectively , The weight.

[0043] This invention also provides a system for elastic memory drift detection and hot update based on pseudo-label weighting, comprising:

[0044] The pseudo-label weighting module is used to acquire new samples in the data stream. For unlabeled samples in the new samples, it uses the current prediction model parameters to predict their category as pseudo-labels. At the same time, it calculates the confidence level of the prediction model and uses it as the original confidence weight of the pseudo-label.

[0045] The storage module is used to store new samples with assigned weights into the elastic memory module; when the number of stored samples reaches the capacity, the sample with the lowest contribution to the current prediction model or the earliest stored sample is removed.

[0046] The drift detection module is used to determine the drift detection index based on the difference between the historical data distribution and the newly received data distribution in the elastic memory module; when the drift detection index exceeds the adaptively changing drift threshold, it is determined that concept drift has occurred.

[0047] The update module is used to trigger an online prediction model update when concept drift is detected; it merges historical data in the elastic memory module with the current new data into a training set, incrementally trains the prediction model, and applies the minimization of the overall loss function to update the prediction model parameters.

[0048] The replacement module is used to seamlessly replace the currently running prediction model with the updated prediction model after training convergence. Attached Figure Description

[0049] Figure 1 This invention provides an elastic memory drift detection and hot update method based on pseudo-label weighting.

[0050] Figure 2 This is a block diagram of an elastic memory drift detection and hot update system based on pseudo-label weighting, according to an embodiment of the present invention. Detailed Implementation

[0051] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0052] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0053] Please see Figure 1 As shown, this invention provides a method for elastic memory drift detection and hot update based on pseudo-label weighting, comprising:

[0054] S10: Obtain new samples from the data stream. For unlabeled samples in the new samples, predict their categories using the current prediction model parameters as pseudo-labels. At the same time, calculate the confidence level of the prediction model and use it as the original confidence weight of the pseudo-label.

[0055] In this embodiment, high-confidence predictions are assigned a larger weight, while low-confidence predictions are assigned a smaller weight, thereby reducing the adverse impact of uncertain pseudo-labels on the prediction model during subsequent training. For new samples... If the current prediction model parameters are The original confidence weight is defined as follows:

[0056] ;

[0057] In the formula, The original confidence weights, To predict model parameters Below, the prediction model output This represents the probability value of the true category. This represents the true category of the actual sample.

[0058] Furthermore, the original confidence weights are adjusted based on information entropy and inconsistency. Apply a discount and obtain the confidence weight after the discount. Specifically, it can be obtained through the following formula:

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula, The confidence weight after discounting. The original confidence weights, For activation function, Predicting the category to which a sample belongs in the predictive model The probability, These are two predicted probability distributions obtained for the same unlabeled sample under different perturbations. The total number of categories, To predict the entropy value of the probability distribution, The degree of difference between the two predicted distributions. , They are respectively , The weights. The specific activation function is the Sigmoid function.

[0063] S20: The new samples with assigned weights are stored in the elastic memory module; when the number of stored samples reaches the capacity, the sample with the lowest contribution to the current prediction model or the earliest stored sample is removed.

[0064] In this embodiment, the new sample will be assigned weights. Data is stored in a flexible memory module. When a new concept is detected, the storage size of the flexible memory module can be appropriately expanded to accommodate the new data distribution. The capacity of the flexible memory module adapts to each time step. The contribution to the current prediction model is obtained using the following formula:

[0065] ;

[0066] In the formula, For sample-based Sensitivity to the loss of the current prediction model For gradient, These are the parameters of the current prediction model. The current prediction model loss is... This is the output of the current prediction model. For the sample The tag, It is a norm.

[0067] When the number of stored samples reaches the capacity limit, priority is given to removing them. For lower-level or earliest writers, elastic scaling / shrinkage is implemented to balance representativeness and resource consumption.

[0068] S30, based on the difference between the distribution of historical data and the distribution of newly received data in the elastic memory module, determine the drift detection index; when the drift detection index exceeds the adaptively changing drift threshold, it is determined that concept drift has occurred.

[0069] In this embodiment, the difference between the historical data distribution and the newly received data distribution can be calculated using Kullback–Leibler divergence, Jensen–Shannon divergence, or a combination of both, to determine the drift detection index.

[0070] Specifically, when determining the drift detection index using the Kullback–Leibler divergence, the following formula is applied:

[0071] ;

[0072] In the formula, Drift detection metrics determined for Kullback–Leibler divergence For historical data distribution, For the distribution of newly received data, This is the sample index.

[0073] The drift detection index is determined using the Jensen–Shannon divergence, applying the following formula:

[0074] ;

[0075] ;

[0076] In the formula, Drift detection index determined for Jensen–Shannon divergence. For historical data distribution, For the distribution of newly received data, for and The differences between them for and The average distribution.

[0077] When determining the drift detection index by combining Kullback–Leibler divergence and Jensen–Shannon divergence, the drift detection index determined based on Kullback–Leibler divergence and Jensen–Shannon divergence are calculated separately first, and then the average of the two is taken as the final drift detection index.

[0078] Undoubtedly, if it is determined that no concept drift has occurred, the prediction model is working normally and does not need to be updated.

[0079] In this embodiment, when the drift detection index determined by any of the above methods exceeds the drift threshold... When a concept drift occurs, it is determined that a concept drift has taken place. To balance false positives and false negatives, an adaptively changing drift threshold is dynamically updated based on the average error of the current prediction model output at the edge node. :

[0080] ;

[0081] In the formula, For adaptively changing drift threshold, As the baseline drift threshold, This is an operator that takes the minimum value among multiple input values. Let be an empirical constant, where , This represents the average error based on the current prediction model output from the edge nodes. The edge nodes are computational entities deployed at the data acquisition end or access network side, calculating the current prediction model output error locally.

[0082] S40, when concept drift is detected, the online prediction model is updated; the historical data in the elastic memory module is merged with the current new data into a training set, the prediction model is incrementally trained, and the prediction model parameters are updated by minimizing the overall loss function.

[0083] In this embodiment, upon detecting concept drift, a hot update process for the prediction model is immediately triggered. Representative historical data stored in the elastic memory module, along with the current new data, are used to perform incremental training on the prediction model, adjusting the model parameters in a timely manner to adapt to the new data distribution. The expression for minimizing the overall loss function is:

[0084] ;

[0085] In the formula, To minimize the overall loss function, This represents the number of labeled samples. For prediction model parameters, The number of pseudo-labeled samples. For the first A labeled sample The output of a time-based prediction model, such as a class probability distribution or score, For the first The label of each sample, This is a tradeoff coefficient for the semi-supervised loss, used to adjust the strength of the impact of the pseudo-label loss relative to the labeled loss. For pseudo-label samples Corresponding to the confidence weights after discounting, For the first A pseudo-label sample, For the first Pseudo-labels for each sample The loss for the current prediction model can be applied, for example, by applying cross-entropy loss.

[0086] Furthermore, without altering the online service, a parameter importance penalty term is introduced into the minimization of the overall loss function to obtain the optimized minimization of the overall loss function. An elastic memory constraint term is further included in the minimization of the overall loss function to suppress drastic changes in important historical parameters. The optimized expression for the minimization of the overall loss function is then:

[0087] = ;

[0088] ;

[0089] In the formula, The optimized function minimizes the overall loss function. The initial coefficient for the penalty term, The attenuation constant of the penalty coefficient, It is a natural constant. For time, For prediction model parameters Importance coefficient, For the first The updated prediction model parameters Predictive model parameters after the previous training stage The value of .

[0090] S50 seamlessly replaces the currently running prediction model with the updated prediction model after training convergence.

[0091] In this embodiment, "hot update" emphasizes that model training and service are carried out simultaneously, meaning that external services are not interrupted during the update process. This can be achieved by updating parameters in the background. Subsequently, the online prediction model parameters are seamlessly replaced, or the old and new models are used in parallel for a short period and then switched after consistency verification. To avoid ambiguity, when no drift is detected, low-intensity periodic fine-tuning can be performed as needed to improve robustness, but this fine-tuning is an engineering optimization in the embodiments and does not constitute a limitation of the necessary steps of the present invention.

[0092] In representative experiments, under scenarios involving long-term online operation and experiencing slow changes and sudden drift, this method can recover and significantly improve model accuracy in a shorter time. The specific recovery time and improvement are related to data and hardware conditions; the above descriptions are merely examples and do not constitute a limitation on the effectiveness of this invention. Compared to traditional processes that require extensive manual annotation and offline retraining, this method significantly shortens the latency for the model to adapt to new environments and improves continuous service capabilities.

[0093] In summary, the method of this invention integrates the above steps into an online learning process: the system continuously receives data streams, performs pseudo-label weighting and storage for each batch or each new sample, monitors concept drift, and updates the prediction model when necessary, thereby achieving continuous adaptive optimization of the model. This method can be implemented by computing devices through a combination of hardware and software, and each step can correspond to an independent functional module working collaboratively.

[0094] Please see Figure 2 As shown, the present invention also provides a system for elastic memory drift detection and hot update method based on pseudo-label weighting, comprising:

[0095] The pseudo-label weighting module is used to acquire new samples in the data stream. For unlabeled samples in the new samples, it uses the current prediction model parameters to predict their category as pseudo-labels. At the same time, it calculates the confidence level of the prediction model and uses it as the original confidence weight of the pseudo-label.

[0096] In this embodiment, the system continuously acquires new samples from the data stream. For each new sample, it first determines whether it already has a human label. If the sample has a real label, it is directly used for subsequent model training; for samples without human labels, the pseudo-label generation process begins. Before this, the original samples may optionally be preprocessed (e.g., normalization, missing value imputation) to meet the model's input requirements.

[0097] The storage module is used to store new samples with assigned weights into the elastic memory module; when the number of stored samples reaches the capacity, the sample with the lowest contribution to the current prediction model or the earliest stored sample is removed.

[0098] The drift detection module is used to determine the drift detection index based on the difference between the historical data distribution and the newly received data distribution in the elastic memory module; when the drift detection index exceeds the adaptively changing drift threshold, it is determined that concept drift has occurred.

[0099] The update module is used to trigger an online prediction model update when concept drift is detected. It merges historical data in the elastic memory module with the current new data into a training set, incrementally trains the prediction model, and applies the minimization of the overall loss function to update the prediction model parameters.

[0100] The replacement module is used to seamlessly replace the currently running prediction model with the updated prediction model after training convergence.

[0101] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0102] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for elastic memory drift detection and hot update based on pseudo-label weighting, characterized in that, Applied to the Internet of Things (IoT), it monitors concept drift, performs predictive model updates, and is deployed at edge nodes on the data acquisition end or access network side, including: Acquire new samples from the data stream at the data acquisition terminal or access network side. For unlabeled samples in the new samples, predict their category using the current prediction model parameters as pseudo-labels. At the same time, calculate the confidence level of the prediction model and use it as the original confidence weight of the pseudo-label. The new samples with assigned weights are stored in the elastic memory module; when the number of stored samples reaches the capacity, the sample with the lowest contribution to the current prediction model or the earliest stored sample is removed, and the storage capacity is dynamically adjusted. Based on the difference between the distribution of historical data and the distribution of newly received data in the elastic memory module, a drift detection index is determined; when the drift detection index exceeds the adaptively changing drift threshold, a concept drift is determined to have occurred. When the Internet of Things detects concept drift, it triggers an online prediction model update; historical data in the elastic memory module is merged with the current new data into a training set, the prediction model is incrementally trained, and the prediction model parameters are updated by minimizing the overall loss function. After training converges, the updated prediction model seamlessly replaces the currently running prediction model within the edge node to continuously predict the category to which the sample belongs; The contribution to the current prediction model is obtained using the following formula: ； In the formula, For sample-based Sensitivity to the loss of the current prediction model For gradient, These are the parameters of the current prediction model. The loss of the current prediction model, This is the output of the current prediction model. For the sample The tag, It is a norm; The adaptively changing drift threshold is dynamically updated based on the average error of the current prediction model output of the edge nodes, and is calculated using the following formula: ； In the formula, For adaptively changing drift threshold, As the baseline drift threshold, This is an operator that takes the minimum value among multiple input values. It is an empirical constant. This represents the average error output by the current prediction model based on the edge nodes.

2. The elastic memory drift detection and hot update method based on pseudo-label weighting according to claim 1, characterized in that, The difference between the distribution of historical data and the distribution of newly received data is calculated using Kullback–Leibler divergence, Jensen–Shannon divergence, or a combination of both, to determine the drift detection index.

3. The elastic memory drift detection and hot update method based on pseudo-label weighting according to claim 2, characterized in that, The drift detection metric is determined using the Kullback–Leibler divergence, applying the following formula: ； In the formula, Drift detection metrics determined for Kullback–Leibler divergence For historical data distribution, For the distribution of newly received data, For sample index.

4. The elastic memory drift detection and hot update method based on pseudo-label weighting according to claim 2, characterized in that, The drift detection index is determined using the Jensen–Shannon divergence, applying the following formula: ； ； In the formula, Drift detection index determined for Jensen–Shannon divergence. For historical data distribution, For the distribution of newly received data, for and The differences between them for and The average distribution.

5. The elastic memory drift detection and hot update method based on pseudo-label weighting according to claim 1, characterized in that, The expression for minimizing the overall loss function is: ； In the formula, To minimize the overall loss function, This represents the number of labeled samples. For prediction model parameters, The number of pseudo-labeled samples. For the first A labeled sample The output of the time-ahead prediction model For the first The label of each sample, The tradeoff coefficient for semi-supervised loss, For pseudo-label samples Corresponding to the confidence weights after discounting, For the first A pseudo-label sample, For the first Pseudo-labels for each sample This represents the loss of the current prediction model.

6. The elastic memory drift detection and hot update method based on pseudo-label weighting according to claim 5, characterized in that, Without altering the online service, a parameter importance penalty term is introduced into minimizing the overall loss function to obtain the optimized overall loss function, which is expressed by the following formula: = ； ； In the formula, The optimized function minimizes the overall loss function. The initial coefficient for the penalty term, The attenuation constant of the penalty coefficient, It is a natural constant. For time, For prediction model parameters Importance coefficient, For the first The updated prediction model parameters Predictive model parameters after the previous training stage The value of .

7. The elastic memory drift detection and hot update method based on pseudo-label weighting according to claim 5, characterized in that, The discounted confidence weights are expressed by the following formula: ； ； ； In the formula, The confidence weight after discounting. The original confidence weights, For activation function, Predicting the category to which a sample belongs in the predictive model The probability, These are two predicted probability distributions obtained for the same unlabeled sample under different perturbations. The total number of categories, To predict the entropy value of the probability distribution, The degree of difference between the two predicted distributions. , They are respectively , The weight.

8. A system for elastic memory drift detection and hot update method based on pseudo-label weighting according to any one of claims 1-7, characterized in that, include: The pseudo-label weighting module is used to acquire new samples in the data stream. For unlabeled samples in the new samples, it uses the current prediction model parameters to predict their category as pseudo-labels. At the same time, it calculates the confidence level of the prediction model and uses it as the original confidence weight of the pseudo-label. The storage module is used to store newly weighted samples into the elastic memory module; When the number of stored samples reaches the capacity, remove the sample that contributes the least to the current prediction model or was stored earliest. The drift detection module is used to determine the drift detection index based on the difference between the historical data distribution and the newly received data distribution in the elastic memory module; when the drift detection index exceeds the adaptively changing drift threshold, it is determined that concept drift has occurred. The update module is used to trigger an online prediction model update when concept drift is detected; it merges historical data in the elastic memory module with the current new data into a training set, incrementally trains the prediction model, and applies the minimization of the overall loss function to update the prediction model parameters. The replacement module is used to seamlessly replace the currently running prediction model with the updated prediction model after training convergence.

Images