Model drift adaptive sample screening method and device, equipment and medium
By constructing a cross-learning mechanism between sample quality estimation parameters and test models, hidden label flipping drift samples are screened out, which solves the performance degradation problem of deep learning models under concept drift and improves the model's adaptability to new data distributions and data quality.
Patent Information
- Application Number
- CN202611109165.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-11-06
AI Technical Summary
In the long-term deployment of deep learning models, the phenomenon of concept drift leads to a decline in prediction performance. Traditional methods cannot distinguish between hidden drift samples and noise, which disrupts the data distribution pattern and weakens the model's ability to adapt to new data distributions.
By constructing sample quality estimation parameters and utilizing the cross-learning mechanism of the test model, hidden label flipping drift samples are screened out and error propagation noise is removed, while high-value samples are retained to achieve model updates.
It significantly improves data quality and the long-term deployment performance of downstream detectors, and enhances the model's adaptability to new data distributions.
Smart Images
Figure CN123049109A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the technical field of sample screening, and more specifically, to a method, apparatus, device, and medium for model drift adaptation sample screening. Background Technology
[0002] In deep learning-based network traffic detection or data classification tasks, models often face concept drift during long-term deployment. This means that the statistical distribution of test data changes over time, leading to a continuous decline in the predictive performance of the original model. To adapt the model to new data distributions with minimal manual annotation costs, existing active learning frameworks typically select a small, representative set of test samples for manual annotation and use label propagation algorithms to infer and diffuse labels to a massive unlabeled set of test samples based on feature similarity. However, in dynamic network environments, a hidden drift phenomenon frequently occurs where the test sample set is highly similar in features to historical training data, but the underlying true labels have been flipped. Because the true feature-label mapping relationship of such hidden label-flipped drift samples has changed compared to historical data, deep learning models also incur very high loss during prediction. Traditional denoising methods cannot distinguish between genuine error propagation noise and high-value hidden drift samples, incorrectly filtering out high-value drift samples as noise, disrupting the overall distribution of test traffic, and severely weakening the model's ability to incrementally adapt to new data distributions. Summary of the Invention
[0003] One objective of this disclosure is to provide a technical solution for model drift adaptation sample screening.
[0004] According to a first aspect of this disclosure, a method for screening model drift adaptation samples is provided, the method comprising: The test sample set is output to each test model of the same architecture in the Nth update, and the Nth parameter set output by each test model in the Nth update is obtained. The Nth parameter subset in the Nth parameter set whose quality estimation parameters are lower than the set quality value is used as the Nth update condition; where N is greater than or equal to 0 and is an integer. Based on the Nth update condition, the test models that are not related to the output relationship in each test model are updated to obtain the test models updated for the N+1th time. The test sample set is output to each test model in the (N+1)th update to obtain the (N+1)th parameter set output by each test model in the (N+1)th update. Test samples that meet the set conditions from the first parameter group to the (N+1)th parameter group are used as model offset samples.
[0005] Optionally, the quality estimation parameters of any parameter set include loss parameters and confidence parameters; the task parameter set includes the Nth parameter set and the (N+1)th parameter set; after outputting the test sample set to each test model of the same architecture for the Nth update, to obtain the Nth parameter set output by each test model for the Nth update, the method further includes: Extract the sample quality estimation parameters corresponding to the loss parameters and confidence parameters in the Nth parameter group; Identify the Nth subset of parameters whose estimated parameters for the sample instructions are lower than a set quality value, and use the loss parameters of the Nth subset of parameters as the Nth update condition.
[0006] Optionally, the loss parameter is obtained by linearly superimposing the cross-entropy loss value and the ambiguity loss value generated by the test model corresponding to each test model in the Nth update on the test sample set.
[0007] Optionally, the confidence parameter is the maximum probability value selected based on the predicted probabilities generated by the corresponding test models in the Nth update when processing the test sample set.
[0008] Optionally, the sample quality estimation parameter is obtained by linearly superimposing the normalized loss parameter under the first balance factor and the normalized confidence parameter under the second balance factor; wherein the sum of the first balance factor and the second balance factor is always a constant.
[0009] Optionally, the set elements contained in the Nth parameter subset are filtered to a set number by a preset filtering ratio.
[0010] Optionally, the setting conditions include test samples that appear more than or equal to a set threshold number of times in the first parameter group to the (N+1)th parameter group; the set threshold is positively correlated with the number of updates of each test model.
[0011] According to a second aspect of this disclosure, a model drift adaptation sample screening device is also provided, comprising: The output module is used to output the test sample set to each test model of the same architecture in the Nth update, to obtain the Nth parameter set output by each test model in the Nth update. The Nth parameter set in the Nth parameter set whose quality estimation parameters are lower than a set quality value is used as the Nth update condition; where N is greater than or equal to 0 and is an integer. The first obtaining module is used to update the test models that are not related to the output relationship in each test model based on the Nth update condition, so as to obtain the test models updated for the N+1th time. The second obtaining module is used to output the test sample set to each test model of the N+1th update, and obtain the N+1th parameter set output by each test model of the N+1th update; The filtering module is used to filter test samples from the first parameter group to the N+1 parameter group that meet the set conditions, and use them as model offset samples.
[0012] According to a third aspect of the present disclosure, an electronic device is also provided, including a memory and a processor, the memory being used to store a computer program; the processor being used to execute the computer program to implement the method according to the first aspect of the present disclosure.
[0013] According to a fourth aspect of the present disclosure, a computer-readable storage medium is also provided, on which a computer program is stored, the computer program implementing the method according to the first aspect of the present disclosure when executed by a processor.
[0014] This disclosure embodiment constructs sample quality estimation parameters and utilizes a cross-learning mechanism of the test model to accurately retain hidden label flip-drift samples while filtering out error propagation label noise. This significantly improves the data purity of downstream adaptive updates, greatly enhances the data quality during incremental training, and significantly improves the long-term deployment performance of downstream detectors.
[0015] Other features and advantages of the embodiments of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description
[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the disclosure.
[0017] Figure 1 A schematic diagram of the hardware structure of an electronic device that can be used to implement a model drift adaptation sample screening method according to embodiments of the present disclosure is shown.
[0018] Figure 2 This is a flowchart illustrating a model drift adaptation sample selection method based on some embodiments.
[0019] Figure 3 This is a block schematic diagram of an electronic device according to one embodiment. Detailed Implementation
[0020] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.
[0021] The following description of at least one exemplary embodiment is merely illustrative and is not intended to limit the scope of this disclosure or its application or use.
[0022] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.
[0023] In all the examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.
[0024] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.
[0025] It should be noted that all data acquisition actions in this disclosure were carried out in compliance with the relevant data protection laws and policies of the country where the data is located, and with the authorization of the relevant equipment owner.
[0026] This disclosure provides a novel sample selection scheme for model drift adaptation.
[0027] Figure 1 A schematic diagram of the hardware structure of an electronic device that can be used to implement a model drift adaptation sample screening method according to embodiments of the present disclosure is shown.
[0028] The electronic device 1000 is a device capable of running computer programs. These programs can be local applications installed on the electronic device, or web applications, lightweight applications, or mini-programs, etc., without limitation. The electronic device 1000 can be a mobile phone, tablet computer, PC, etc., without limitation.
[0029] like Figure 1 As shown, the electronic device 1000 may include a processor 1101, a memory 1102, an interface device 1103, a communication device 1104, a display screen 1105, an input device 1106, a microphone 1107, etc. Figure 1 The hardware configuration shown is illustrative only and is not intended to limit this disclosure, its application, or its use.
[0030] The processor 1101 executes computer programs, which can be written using instruction sets of architectures such as x86, Arm, RISC, MIPS, and SSE. The memory 1102 includes, for example, ROM (Read-Only Memory), RAM (Random Access Memory), and non-volatile memory such as a hard disk. The interface device 1103 includes, for example, a USB interface, a network cable interface, and a headphone jack. The communication device 1104 is capable of wired or wireless communication. The communication device 1104 may include at least one short-range communication module, such as any module for short-range wireless communication based on short-range wireless communication protocols such as Hilink, WiFi (IEEE 802.11), Mesh, Bluetooth, ZigBee, Thread, Z-Wave, NFC, UWB, and LiFi. The communication device 1104 may also include a long-range communication module, such as any module for WLAN, GPRS, or 2G / 3G / 4G / 5G long-range communication. The input device 1106 may include, for example, a touchscreen, a keyboard, and various sensors.
[0031] In this embodiment, the memory 1102 of the electronic device 1000 is used to store a computer program that controls the processor 1101 to perform a model drift adaptation sample screening method according to any embodiment of the present disclosure.
[0032] Next, as Figure 1 Using an electronic device 1000 as an example, various embodiments of the computer program update method are explained.
[0033] The following combination Figure 1 The electronic device 1000 illustrates various embodiments of this disclosure.
[0034] <First Embodiment> Figure 2 This is a flowchart illustrating a model drift adaptation sample selection method based on some embodiments. For example, Figure 1 1000 Chinese electronic devices. For example... Figure 2 As shown, the model drift adaptation sample screening method of this embodiment may include the following steps S210 to S230.
[0035] Step S210: Output the test sample set to each test model of the same architecture in the Nth update to obtain the Nth parameter set output by each test model in the Nth update. The Nth parameter subset in the Nth parameter set whose quality estimation parameters are lower than the set quality value is used as the Nth update condition.
[0036] In this embodiment, each test model block can include multiple test models with the same architecture. Two deep neural networks with identical architectures but independent internal parameters are used as the first test model. Second test model The first test model It can be tested using the second test model. The output parameter set serves as the update condition for model updating, and the second test model. The first test model can be used The output parameter set serves as the update condition for model updates. The test sample set with inferred labels is simultaneously and in parallel input to the first test model. Second test model In the process, through forward propagation, the two test models independently output their respective parameter sets. For example, the first test model outputs parameters up to the 0th update of the test sample set. Second test model From this, the first test model can be obtained. Second test model The 0th parameter group of each output. In the first test model. Second test model After each update, the test sample set is output to each test model of the same architecture that has undergone the first update, resulting in the first parameter set output by each test model after the first update. Here, N can be 1, 2, or 5, etc., preferably 2.
[0037] In some embodiments, the quality estimation parameters of any parameter group include loss parameters and confidence parameters; the task parameter group includes the Nth parameter group and the (N+1)th parameter group; after step S210, the method further includes the following steps S211 and S212: Step S211: Extract the sample quality estimation parameters corresponding to the loss parameters and confidence parameters in the Nth parameter group.
[0038] Step S212: Determine the Nth parameter subset whose sample instruction estimation parameters are lower than the set quality value, and use the loss parameters of the Nth parameter subset as the Nth update condition.
[0039] In this embodiment, the set quality value can be preset and is not limited here.
[0040] In this embodiment, the sample quality estimation parameters are represented by loss parameters and confidence parameters. The loss parameter in the parameter set is used to capture the degree of conflict between a given inferred label and the model's prediction, while the confidence parameter is used to indicate whether the decisions made by the model are stable. The parameters of these two complementary dimensions together form the basis for subsequent sample quality assessment.
[0041] In some examples, the loss parameter is obtained by linearly superimposing the cross-entropy loss value and the ambiguity loss value generated by the corresponding test model in each test model in the Nth update on the test sample set.
[0042] Specifically, for each sample i in any set of parameters, its comprehensive loss Cross-entropy loss It is a weighted combination of the loss Hi of the fuzziness output by the quantization network, i.e. .in, , (C represents the total number of categories, and pi,c represents the probability of being predicted as category c). After calculation, the comprehensive loss is normalized to obtain the normalized loss value. By combining cross-entropy loss with ambiguity loss, we can reflect both the degree of agreement between the sample label and the prediction, and quantify the uncertainty of the model output, making the loss parameter a more comprehensive characterization of the sample health. In other examples, the loss parameter can also be simply the cross-entropy loss value, which, after normalization, can be directly used as the evaluation basis for the loss dimension.
[0043] In some examples, the confidence parameter is the maximum probability value selected from the predicted probabilities generated by the corresponding test models in the Nth update when processing the test sample set.
[0044] Specifically, the maximum value of the predicted probability output by the test model is extracted as the confidence score, i.e. Then, normalize the results to obtain the normalized confidence level. The maximum predicted probability directly reflects the test model's confidence in the class to which a sample belongs, effectively indicating the consistency between the sample features and the distribution learned by the current test model. In other examples, the confidence parameter can also be other metrics that measure decision stability, such as the reciprocal of the entropy of the predicted probability distribution.
[0045] Step S220: Based on the Nth update condition, update the test models that are not related to the output relationship in each test model to obtain the test models updated for the N+1th time.
[0046] In some examples, each test model includes a first test model. Second test model and the third test model First test model If parameter set T1 is output, then the test model that corresponds to the non-output relationship of parameter subset T11 of parameter set T1 is the second test model. and the third test model Therefore, the parameter subset T11 of parameter group T1 can be used for the second test model. and the third test model Update the model.
[0047] Step S230: Output the test sample set to each test model in the N+1th update to obtain the N+1th parameter set output by each test model in the N+1th update.
[0048] Step S240: Select test samples from the first parameter group to the (N+1)th parameter group that meet the set conditions as model shift samples. This yields a high-quality subset of data that reflects the flow distribution, retains important hidden drift samples, and removes noise, which is used for subsequent model drift adaptation.
[0049] Specifically, each test model is used as the first test model. Second test model Both tests used the example of no model update, with the first test model... Output the 0th parameter set for all test samples in the current batch. Then, based on the loss parameter and confidence parameter in the 0th parameter set, calculate the corresponding sample quality estimation parameter for each sample, and filter out samples whose sample quality estimation parameter is lower than the set quality value to form the 0th parameter subset. First test model The selected subset of parameters 0 Transmitted to the second test model Second test model Using the 0th parameter subset The loss parameters are backpropagated and the internal parameters are updated. Second test model. After the update is complete, input all test sample sets into the updated second test model. Calculate the sample quality estimation parameter for each sample, and select the first parameter subset whose sample quality estimation parameter is lower than the set quality value. The second test model after one update With the first parameter subset The loss parameters are backpropagated and the internal parameters are updated. The second test model is then updated once. It is the second test model Through the 0th parameter subset The loss parameters are obtained by performing backpropagation and updating the internal parameters. With two updates, two sets of parameter 0, two sets of parameter 1, and two sets of parameter 2 are finally obtained. Test samples that appear more than or equal to a set threshold in the two sets of parameter 0, two sets of parameter 1, and two sets of parameter 2 are selected as model offset samples.
[0050] In some examples, the sample quality estimation parameters are obtained by linearly superimposing the normalized loss parameter under the first balancing factor and the normalized confidence parameter under the second balancing factor; wherein the sum of the first balancing factor and the second balancing factor is always a constant.
[0051] Specifically, the sample quality estimation parameter, namely the LCS (Loss-Confidence Selection) index, is calculated using the following formula: .in, The pre-defined balancing factors are α and (1-α), with their sum always equal to 1. Increasing α prioritizes the feature-label mapping relationship (loss term); decreasing α shifts the evaluation focus to feature distribution consistency (confidence term). A lower LCS score indicates that the sample is less likely to be pure label noise. The core advantage of this metric is that for hidden label flipped drift samples, although the label flip leads to a higher loss term, the strong feature space consistency with historical data distribution results in extremely high confidence in the outputs of both test models. The high confidence term strongly compensates for the high loss term, keeping the LCS score low and thus preventing misclassification as noise. In these examples, by introducing an adjustable balancing factor α to dynamically coordinate the contribution weights of the two dimensions, the evaluation strategy can be flexibly adjusted according to the actual noise rate and drift intensity, significantly improving the accuracy of sample identification. In other examples, the sample quality estimation parameters can also be implemented using other evaluation functions that weight and fuse loss and confidence.
[0052] In some examples, the elements contained in the Nth parameter subset are filtered to a set number using a preset filtering ratio. Specifically, the size of the filtered Nth parameter subset is determined by the preset filtering ratio β, satisfying... ,in The filtering ratio β controls the proportion of samples retained in each round relative to the test sample set. By adjusting the filtering ratio β, a balance can be struck between the intensity of noise filtering and the integrity of drift sample retention. In these examples, by setting a preset filtering ratio, the problems of over-filtering leading to drift sample loss or over-retention leading to noise residue can be effectively prevented.
[0053] In some examples, the first test model With the second test model The above cross-learning process is continuously executed. After multiple iterations (e.g., 3 iterations), the iteration stops when both networks meet the set update conditions (e.g., the number of iterations reaches a preset threshold, or the subsets selected by the two networks tend to stabilize). In these examples, multiple iterations allow the two networks to gradually develop stable discriminative abilities. Real noise samples are continuously filtered out because they exhibit high LCS scores on both networks, while hidden drift samples are stably retained due to high confidence compensation. The final output is a high-purity test dataset that retains drift samples.
[0054] As can be seen from steps S210 to S240 above, this embodiment inputs the test sample set into two test models with independent parameters, jointly evaluates the sample quality from two complementary dimensions: loss parameter and confidence parameter, uses the LCS index to achieve accurate identification and retention of hidden drift samples, and continuously optimizes the screening results through a dual-network cross-learning mechanism, finally outputting a high-purity test dataset containing hidden drift samples, providing a high-quality data foundation for subsequent incremental adaptive training of the model.
[0055] <Second Embodiment> This embodiment provides an electronic device. Figure 3 A schematic diagram of the hardware structure of the electronic device is shown.
[0056] like Figure 3 As shown, the electronic device 500 includes a processor 510 and a memory 520. The memory can be used to store computer programs, and the processor can be used to retrieve the computer programs from the memory to execute any method embodiment of this disclosure. The processor can be one or more, and these processors can execute instructions individually or jointly. Similarly, the memory can be one or more, and these memories can store the aforementioned computer programs individually or jointly.
[0057] This disclosure also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the model drift adaptation sample selection method in any embodiment of this disclosure. Optionally, the computer-readable storage medium may be a non-transitory storage medium, but is not limited thereto; it may also be a temporary storage medium.
[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0059] This disclosure may be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having a computer-readable program loaded thereon for causing a processor to implement any of the methods in the foregoing embodiments of this disclosure.
[0060] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media may include, for example, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), compact disc-read-only memory (CD-ROM), digital versatile disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any combination thereof. The computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.
[0061] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include one or more of copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to computer-readable storage media in the respective computing / processing device.
[0062] The computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source or object programs written in any combination of one or more programming languages, including object-oriented programming languages (such as Smalltalk, C++, etc.) and conventional procedural programming languages (such as the "C" language or similar programming languages). The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network (e.g., a local area network or a wide area network), or it may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays, or programmable logic arrays, can execute computer-readable program instructions to implement various aspects of the embodiments of this disclosure by utilizing state information from the computer-readable program instructions.
[0063] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.
[0064] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.
[0065] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.
[0066] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It should be noted that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are all equivalent.
[0067] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of this disclosure is defined by the appended claims.
Claims
1. A method for selecting samples to adapt to model drift, characterized in that, The method includes: The test sample set is output to each test model of the same architecture in the Nth update, and the Nth parameter set output by each test model in the Nth update is obtained. The Nth parameter subset in the Nth parameter set whose quality estimation parameters are lower than the set quality value is used as the Nth update condition; where N is greater than or equal to 0 and is an integer. Based on the Nth update condition, the test models that are not related to the output relationship in each test model are updated to obtain the test models updated for the N+1th time. The test sample set is output to each test model in the (N+1)th update to obtain the (N+1)th parameter set output by each test model in the (N+1)th update. Test samples that meet the set conditions from the first parameter group to the (N+1)th parameter group are used as model offset samples.
2. The method according to claim 1, characterized in that, The quality estimation parameters for any parameter set include loss parameters and confidence parameters; the task parameter set includes the Nth parameter set and the (N+1)th parameter set; after outputting the test sample set to each test model updated for the Nth time with the same architecture to obtain the Nth parameter set output by each test model updated for the Nth time, the method further includes: Extract the sample quality estimation parameters corresponding to the loss parameters and confidence parameters in the Nth parameter group; Identify the Nth subset of parameters whose estimated parameters for the sample instructions are lower than a set quality value, and use the loss parameters of the Nth subset of parameters as the Nth update condition.
3. The method according to claim 2, characterized in that, The loss parameter is obtained by linearly superimposing the cross-entropy loss value and the ambiguity loss value generated by the corresponding test model in each test model in the Nth update on the test sample set.
4. The method according to claim 1, characterized in that, The confidence parameter is the maximum probability value selected based on the predicted probabilities generated by the test models in the Nth update when processing the test sample set.
5. The method according to claim 1, characterized in that, The sample quality estimation parameters are obtained by linearly superimposing the normalized loss parameter under the first balance factor and the normalized confidence parameter under the second balance factor; wherein the sum of the first balance factor and the second balance factor is always a constant.
6. The method according to claim 1, characterized in that, The elements contained in the Nth parameter subset are filtered to a set quantity by a preset filtering ratio.
7. The method according to claim 1, characterized in that, The set conditions include test samples that appear more than or equal to a set threshold number of times in the first parameter group to the (N+1)th parameter group; the set threshold is positively correlated with the number of updates of each test model.
8. A model drift adaptation sample screening device, characterized in that, include: The output module is used to output the test sample set to each test model of the same architecture in the Nth update, to obtain the Nth parameter set output by each test model in the Nth update. The Nth parameter set in the Nth parameter set whose quality estimation parameters are lower than a set quality value is used as the Nth update condition; where N is greater than or equal to 0 and is an integer. The first obtaining module is used to update the test models that are not related to the output relationship in each test model based on the Nth update condition, so as to obtain the test models updated for the N+1th time. The second obtaining module is used to output the test sample set to each test model of the N+1th update, and obtain the N+1th parameter set output by each test model of the N+1th update; The filtering module is used to filter test samples from the first parameter group to the N+1 parameter group that meet the set conditions, and use them as model offset samples.
9. An electronic device, characterized in that, It includes a memory and a processor, the memory being used to store a computer program; the processor being used to execute the computer program to implement the method according to any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the method according to any one of claims 1-7.