A time-evolution-oriented social robot account detection method and system

Abstract

The application relates to the technical field of Internet and discloses a time-evolution-oriented social robot account detection method and system. The method comprises the following steps: obtaining an annotated user account sample, and training a baseline account detection model by using a feature matrix; calculating a significance p value based on the maximum mean difference between the feature spaces of the user account sample and the new user account sample of a replay buffer, and determining whether a significant distribution drift occurs; assigning a weight to a target domain sample based on a domain discriminator and filtering; reading the user account sample from the replay buffer, performing incremental training, and detecting whether the user account is a social robot account based on the obtained account detection model. Through the closed-loop mechanism of drift detection, adversarial refinement, weakly supervised pseudo-labeling and incremental fine-tuning, the application realizes continuous and accurate identification and knowledge accumulation of social robots.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet, in particular to a time-evolution-oriented social robot account detection method and system. BACKGROUND

[0002] With the development of Internet technology and the popularity of mobile Internet, social networks have become an indispensable part of people's daily life. In the vast user group of social networks, the number of social robots (also known as machine accounts) is also growing rapidly, which are automatically controlled by programs rather than real people.

[0003] The current research focus of social robots mainly concentrates on feature extraction-based and graph structure-based detection methods. These methods analyze the behavior patterns, content features, and relationship structures in social networks to identify social robots, but the overall effect is limited due to the time-varying characteristics and incomplete network data.

[0004] 1. Feature extraction-based social robot detection:

[0005] Feature extraction-based social robot detection is a traditional machine learning method that relies on various features extracted from user accounts for classification. These features may include user profile information, posting frequency and time, connection patterns in social networks, and statistical properties of user behavior. Through these features, a classifier can be constructed to distinguish between real users and social robots.

[0006] The advantages of feature extraction-based social robot detection are:

[0007] Explicit feature representation and mature machine learning models: By extracting explicit features, a clear feature space can be constructed, and mature machine learning models can be used for classification. These models have been widely applied and verified in multiple fields.

[0008] Relatively low computational cost and good interpretability: Compared with deep learning methods, feature extraction-based social robot detection methods generally have lower computational cost. At the same time, this method can provide intuitive explanations for why a user is classified as a social robot, which helps to improve the transparency and credibility of the detection system.

[0009] Realizable and easy to integrate and extend: Feature extraction-based social robot detection methods have lower data acquisition thresholds through appropriate feature selection, making them easier to implement in practical engineering applications. At the same time, by regularly updating the feature set and model, feature extraction-based methods can adapt to new robot behaviors and strategies.

[0010] However, this approach has several major limitations:

[0011] Complexity of feature extraction: Extracting useful features from a large amount of user data is a complex and time-consuming process that requires a lot of expertise.

[0012] Temporal relevance of features: Social bot developers continuously improve their strategies to evade detection. This means that the feature set must be constantly updated to adapt to new bot behaviors.

[0013] Limitations in generalization ability: Feature extraction-based social bot detection methods often rely on specific datasets and contexts, limiting their ability to generalize across different social networks and different types of social bots.

[0014] 2. Graph structure-based social bot detection

[0015] Graph structure-based social bot detection methods utilize the graph structure properties of social networks for analysis. This approach typically involves constructing a social network graph where nodes represent users and edges represent social relationships between users. Then, graph analysis techniques, such as graph convolutional neural networks, are used to identify abnormal social patterns that may indicate a user is a social bot.

[0016] The advantages of graph structure-based social bot detection methods are:

[0017] Utilization of global information: Ability to capture global social relationships between users, which is particularly important for identifying advanced social bots that attempt to mimic real user behavior.

[0018] Interpretability: Often provides intuitive explanations for why a user was classified as a social bot, which helps improve the transparency and credibility of the detection system.

[0019] However, graph structure-based social bot detection methods also face some challenges:

[0020] Complexity of graph data: Social network graphs can be very large and complex, making it difficult to collect and analyze graph data.

[0021] Dynamically changing networks: Social networks are dynamically changing, with new connections and users constantly emerging, requiring graph structure methods to be able to update and adapt to changes in the network in real-time.

[0022] Existing social robot detection techniques generally rely on large-scale manual annotation data, and the model remains static once it is trained. When platform policies, user behavior and robot strategies evolve over the years, the data distribution drifts, resulting in a sharp decline in detection accuracy. If full re-annotation and re-training are used, the cost is high and the cycle is long, and it cannot be updated online continuously. In addition, existing incremental methods are prone to catastrophic forgetting when introducing new annual data, and lack an effective preservation mechanism for historical high-value samples. SUMMARY

[0023] To solve the above technical problems, the present application provides a time-evolving social robot account detection method and system, which can realize real-time monitoring and self-adaptive correction of cross-year data distribution drift without additional manual annotation, so that the detection model can maintain high accuracy continuously. During the incremental update process, the present application can make full use of new annual weak supervision information and avoid catastrophic forgetting of historical knowledge, ensuring knowledge accumulation rather than replacement.

[0024] To solve the above technical problems, the present application adopts the following technical solutions:

[0025] In a first aspect, the present application provides a time-evolving social robot account detection method, comprising:

[0026] S1, obtaining labeled user account samples, including user account data and labels indicating whether the user account is a social robot, selecting a plurality of labeled user account samples as source domain samples, and training a baseline account detection model using the corresponding feature matrix, and saving part of the source domain samples to a replay buffer;

[0027] S2, calculating a significance p-value based on the maximum mean difference between the feature space of the user account samples in the replay buffer and the newly added unlabeled user account samples as target domain samples, if the significance p-value is lower than a preset threshold, it is determined that there is a significant distribution drift, and steps S3 to S4 are triggered;

[0028] S3, assigning weights to the target domain samples based on the domain discriminator and filtering, using the baseline account detection model and the auxiliary model to predict the filtered target domain samples, and writing the target domain samples with consistent prediction results, confidence higher than the threshold and weight higher than the weight threshold, and the corresponding pseudo-labels and weights into the replay buffer;

[0029] S4, reading user account samples from the replay buffer, incrementally training based on the baseline account detection model as the initial model, and detecting whether the user account is a social robot account based on the trained account detection model.

[0030] In one of the embodiments, the user account data includes account metadata, user published content, behavior timing data, and relationship network data.

[0031] In one of the embodiments, the feature matrix is obtained by extracting multi-dimensional features of the labeled user account samples as source domain samples.

[0032] The multi-dimensional features include behavior features, timing features, and text features of the user account samples.

[0033] In one of the embodiments, the method further comprises a cross-period continuous learning loop, which comprises the following steps:

[0034] S51, waiting for the arrival of new unlabeled user account samples in the next period, and performing step S2 to perform drift detection:

[0035] S52, if the significance p-value is lower than the preset threshold, sequentially performing steps S3 and S4; replacing the latest account detection model obtained by step S4 with the initial model of the current round of circulation as the initial model of the next round of circulation, and updating the priority of each user account sample in the replay buffer to reserve a specified number of user account samples according to the priority to update the replay buffer;

[0036] S53, if the significance p-value is not lower than the preset threshold, only performing: updating the priority of each user account sample in the replay buffer to reserve a specified number of user account samples according to the priority to update the replay buffer;

[0037] Repeating steps S51 to S53 realizes the self-evolution of the social robot account detection capability across periods and without manual annotation.

[0038] In one of the embodiments, updating the priority of each user account sample in the replay buffer to reserve a specified number of user account samples according to the priority to update the replay buffer comprises:

[0039] The priority of the i-th user account sample in the replay buffer is:

[0040] ;

[0041] wherein, is the information entropy of the i-th user account sample, is the relative year decay amount of the i-th user account sample, and a is an adjustable coefficient;

[0042] Reserving the first N user account samples in descending order of priority to form the replay buffer required for the next round of circulation.

[0043] ​In one of the embodiments, the maximum mean discrepancy between the feature space of the replay buffer-based user account sample and the newly added unlabeled user account sample as the target domain sample calculates the significance p value, specifically comprising:

[0044] The feature space of the user account sample in the replay buffer is taken as the source domain feature, and the feature space of the newly added unlabeled user account sample is taken as the target domain feature;

[0045] The maximum mean discrepancy is used to calculate the drift statistic D:

[0046] ;

[0047] Wherein f is a test function defined on the sample feature space, belonging to the unit ball of the reproducing kernel Hilbert space; Indicates the supremum on the unit ball of the reproducing kernel Hilbert space; The expected output of function f on the distribution After the source domain feature is mapped by the kernel function; The expected output of function f on the distribution After the target domain feature is mapped by the same kernel function; Indicates the feature vector of the user account sample;

[0048] The significance p value is calculated according to the drift statistic D.

[0049] In one of the embodiments, the domain discriminator is used to assign weights to the target domain samples and filter, specifically comprising:

[0050] The domain class of each labeled user account sample as the source domain sample is marked as 0, and the domain class of each unlabeled user account sample as the target domain sample is marked as 1, to construct a domain discrimination training set;

[0051] A multilayer perceptron is used as a domain discriminator, and the domain discrimination training set is trained to output the predicted probability of each target domain sample belonging to the domain class 1 ;

[0052] The target domain sample is assigned a weight , and the target domain sample with w lower than the preset threshold is filtered.

[0053] In one of the embodiments, the baseline account detection model and the auxiliary model are used to predict the filtered target domain sample, and the target domain sample with consistent prediction results, confidence higher than the threshold, and weight higher than the weight threshold, and the corresponding pseudo label and weight are written into the replay buffer, specifically comprising:

[0054] predicting the target domain samples by using the baseline account detection model to obtain a first prediction result and a first confidence;

[0055] predicting the target domain samples by using the auxiliary model to obtain a second prediction result and a second confidence; the prediction result is whether the target domain sample is a social robot;

[0056] retaining the target domain samples whose first prediction result and second prediction result are consistent, whose first confidence and second confidence are both higher than respective thresholds, and whose weight is higher than a preset weight threshold, as candidate target domain samples, and taking the corresponding first prediction result and second prediction result as pseudo labels of the candidate target domain samples;

[0057] performing k-fold cross-validation on the candidate target domain samples, wherein the macro-average F1 value of each fold is calculated by weighting the weight of the target domain sample, and if the macro-average F1 value is not lower than 0.85, the batch of candidate target domain samples, the corresponding pseudo labels and the weight are written into a replay buffer; if the macro-average F1 value is lower than 0.85, the batch of candidate target domain samples is discarded.

[0058] In one of the embodiments, the user account samples are read from the replay buffer, and the baseline account detection model is used as an initial model for incremental training, specifically including:

[0059] reading the historical user account samples and the newly added target domain samples from the replay buffer to construct an incremental training set; using the baseline account detection model as an initial model, the incremental training set and the gradient boosting decision tree algorithm are used for further training, and the objective function is:

[0060] ;

[0061] Here, is a binary logarithmic loss, , is a regularization coefficient, is the number of leaf nodes of the i-th tree, is the weight vector of the leaf node of the i-th tree; represents the true label of the i-th user account sample, represents the prediction value of the i-th user account sample. In a second aspect, the present application provides a computer system comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method of any one of the embodiments of the first aspect when executing the computer program.

[0062] In a second aspect, the present application provides a computer system comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method of any one of the embodiments of the first aspect when executing the computer program.

[0063] ​The application realizes continuous and accurate identification and knowledge accumulation of social robots in cross-year and zero-labelling scenarios through a closed-loop mechanism of "drift detection-anti-refining-weakly supervised pseudo-label-incremental fine-tuning", a memory replay strategy of information entropy and time decay cooperation. Specifically, the following significant effects are achieved:

[0064] First, cross-year zero-labelling continuous evolution, through the closed-loop mechanism of "drift detection-anti-refining-weakly supervised pseudo-label-incremental fine-tuning", the system can automatically complete model upgrading using public data under the condition of no new artificial labelling, significantly reducing the labelling cost.

[0065] Second, memory replay prevents forgetting, introduces a priority replay buffer based on information entropy and time decay factor, which not only retains high-value historical samples, but also eliminates outdated samples through exponential decay, effectively inhibiting catastrophic forgetting and ensuring knowledge accumulation rather than replacement.

[0066] Third, universality and scalability, the method does not depend on specific platforms or languages and can run with only user-level public metadata; feature extraction and model training are decoupled, and new data sources or more advanced base learners can be flexibly connected in the future, with good horizontal expansion ability. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The method flowchart in the embodiment of the application is shown in the figure;

[0068] Figure 2 The overall framework diagram in the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0069] A preferred embodiment of the application will be described in detail below with reference to the accompanying drawings.

[0070] As shown in the figure, a time-evolution-oriented social robot account detection method in the application includes the following steps: Figure 1

[0071] S1, obtaining labelled user account samples, including user account data and labels indicating whether the user account is a social robot, selecting a plurality of labelled user account samples as source domain samples, training a baseline account detection model using the corresponding feature matrix, and saving part of the source domain samples to a replay buffer;

[0072] S2, calculating a significance p-value based on the maximum mean difference between the feature space of the user account samples in the replay buffer and the newly added unlabelled user account samples as target domain samples, if the significance p-value is lower than a preset threshold, it is determined that there is a significant distribution drift, and steps S3 to S4 are triggered;

[0073] ​S3, based on the domain discriminator, weights and filters the target domain samples, uses the baseline account detection model and the auxiliary model to predict the filtered target domain samples, and writes the target domain samples with consistent prediction results, confidence higher than the threshold and weight higher than the weight threshold, and the corresponding pseudo labels and weights into the replay buffer;

[0074] S4, reads the user account samples from the replay buffer, and uses the baseline account detection model as the initial model to perform incremental training, and uses the trained account detection model to detect whether the user account is a social robot account.

[0075] As shown in Figure 2 The technical solutions of the present application are further introduced in the following several parts.

[0076] 1. Initial training phase.

[0077] First, the training data needs to be prepared and processed.

[0078] User account data and labels indicating whether the user account is a social robot are obtained from multiple public scientific research data sets, and the fields of different data sets are uniformly mapped to generate a standardized data set with consistent structure.

[0079] User account data includes:

[0080] Account metadata, including account registration time, nickname, avatar, introduction, email, mobile phone number, etc.

[0081] User published content, including user published content.

[0082] Behavioral timing data, including timestamp sequences of events such as posting, forwarding, commenting, liking, and private messaging.

[0083] Relationship network data, including user follow and follow relationships.

[0084] Based on the standardized data set, multi-dimensional features are extracted, including behavior features, timing features and text features, to form a feature matrix.

[0085] Among them, the behavior feature is used to represent the interaction mode and social relationship structure of the user on the social platform; the timing feature represents the activity regularity, periodicity and burstiness of the user on different time scales; the text feature represents the semantics, style and generation traces of the user published content.

[0086] A plurality of user account samples with the latest timestamp are used as source domain samples, and a baseline account detection model is trained using the corresponding feature matrix.

[0087] The baseline account detection model can adopt a machine learning-based classification model (such as logistic regression, support vector machine, random forest, gradient boosting decision tree), a deep learning-based neural network model (such as multilayer perceptron, recurrent neural network, convolutional neural network, Transformer), or a graph neural network-based model (such as graph convolutional network, graph attention network) to adapt to the account detection task in different feature spaces.

[0088] A part of the user account samples in the source domain samples are saved as a replay buffer for subsequent cross-year incremental learning.

[0089] 2. Cross-year distribution drift detection phase.

[0090] Based on the trained baseline account detection model and the replay buffer, the following operations are performed:

[0091] The newly added unlabeled user account samples of the year are used as target domain samples, and the unlabeled user account samples only contain user account data without labels indicating whether the user account is a social robot. The feature space of the user account samples in the replay buffer is used as the source domain feature, and the feature space of the to-be-tested newly added unlabeled user account samples of the year is used as the target domain feature.

[0092] The maximum mean difference is used to calculate the drift statistic D, and the formula is:

[0093] ;

[0094] Where f is a test function defined on the sample feature space, and belongs to the unit ball of the reproducing kernel Hilbert space (RKHS); represents the supremum on the unit ball of the reproducing kernel Hilbert space; is the expected output of the function f on the distribution after the source domain feature is mapped by the kernel function; is the expected output of the function f on the distribution after the target domain feature is mapped by the same kernel function.

[0095] The significance p-value is calculated according to the drift statistic D. If the p-value is lower than a preset threshold, it is determined that there is a significant distribution drift, and the subsequent process is triggered.

[0096] 3. Adversarial refinement and sample weighting phase.

[0097] After it is determined that there is a significant distribution drift, this phase is immediately started, which specifically includes:

[0098] The domain class of the source domain sample is marked as 0, and the domain class of the target domain sample is marked as 1 to construct a domain discrimination training set. A multilayer perceptron is used as a domain discriminator to train the domain discrimination training set to output a prediction probability P that each target domain sample belongs to the domain class 1.

[0099] Assigning weights to target domain samples , and Target domain samples below the preset threshold are considered high domain difference samples and are filtered. The filtered target domain samples are used for weakly supervised pseudo-label generation and incremental learning.

[0100] 4. Weakly supervised pseudo-label generation and quality verification stage.

[0101] Based on the filtered target domain samples and their weights, the following operations are performed:

[0102] Using the baseline account detection model to predict the target domain samples obtains a first prediction result and a first confidence;

[0103] Using an auxiliary model to predict the same batch of target domain samples obtains a second prediction result and a second confidence;

[0104] Only samples with consistent first and second prediction results and both first and second confidences higher than their respective thresholds are retained as candidate target domain samples. At the same time, the target domain sample weight w is used as a reliability coefficient, and only candidate target domain samples with weight w higher than the preset weight threshold are retained to form the final candidate set.

[0105] The auxiliary model is an independent account detection model different in structure or training method from the baseline account detection model. It can be a machine learning-based classifier (such as random forest, support vector machine), a deep learning-based time series model (such as LSTM), a text model (such as Transformer), or a graph neural network model (GNN). The prediction results of the auxiliary model and the baseline model on the same target domain sample are consistent, which is used to improve the accuracy of writing pseudo-label samples to the replay buffer.

[0106] The candidate target domain samples are subjected to k-fold cross-validation, where the macro-average F1 value of each fold is calculated by weighting the target domain sample weight w. If the macro-average F1 value is not less than 0.85, the batch of candidate target domain samples, their pseudo-labels, and corresponding weights w are written to the replay buffer; otherwise, the batch of candidate target domain samples is discarded and returned to the cross-year distribution drift detection stage, waiting for the next batch of target domain samples.

[0107] 5. Incremental fine-tuning stage.

[0108] After putting the target domain samples with pseudo labels into the replay buffer, the following operations are performed:

[0109] The historical samples (i.e., the user account samples in the replay buffer before the target domain samples with pseudo labels are put in) and the target domain samples with pseudo labels are read from the replay buffer to construct an incremental training set.

[0110] The baseline account detection model that has completed training is used as the initial model, and gradient boosting decision tree algorithm is used for further training, where the objective function is:

[0111] ;

[0112] is the binary logarithmic loss, 、 is the regularization coefficient, is the number of leaf nodes of the i-th tree, is the weight vector of the leaf nodes of the i-th tree. In the t-th iteration, a new tree is fitted by the negative gradient until early stopping is triggered when the loss on the validation set does not decrease for 100 consecutive rounds:

[0113]

[0114] ;

[0115] is the negative gradient value of the i-th user account sample under the prediction result of the t-th iteration, which is equivalent to the residual, and is the fitting target when constructing the i-th tree. After training is completed, the priority of each sample in the replay buffer is updated according to the following formula:

[0116] ;

[0117] ;

[0118] wherein, is the information entropy of the i-th sample, is the relative year decay amount of the i-th sample, and α is an adjustable coefficient.

[0119] The top N samples are retained in descending order of scoreᵢ to form the replay buffer required for the next round of circulation.

[0120] 6. Cross-year continuous learning circulation phase.

[0121] Continuous learning reinforcement can be achieved through the following circulation:

[0122] ​​​​​a) waiting for the next calendar year or the next batch of newly added unlabeled user account samples to arrive, and performing cross-year distribution shift detection;

[0123] b) if the significance p-value in the cross-year distribution shift detection is lower than a preset threshold, sequentially performing the phases of adversarial refinement and sample weighting, weakly supervised pseudo-label generation and quality verification, and incremental fine-tuning;

[0124] c) if the significance p-value in the cross-year distribution shift detection is not lower than the preset threshold, skipping the phases of adversarial refinement and sample weighting, weakly supervised pseudo-label generation and quality verification, and only updating the priority of each sample in the replay buffer to update the replay buffer by retaining a specified number of samples according to the priority;

[0125] d) replacing the original baseline account detection model with the latest account detection model trained in the incremental fine-tuning phase as the initial model for the next round of circulation;

[0126] e) repeating a)-d) to achieve self-evolution of cross-year, continuous and unlabeled social robot account detection capability.

[0127] The present application realizes automatic upgrading of robot detection model year by year without the need for new artificial labeling through four-step closed loop of maximum mean difference shift detection triggering, adversarial field refinement empowerment, dual model consistency pseudo-label filtering, and early stopping incremental fine-tuning. The present application uses the product of sample prediction entropy and year decay factor as the priority to dynamically retain and discard historical samples, thereby preventing catastrophic forgetting and avoiding accumulation of outdated noise.

[0128] The terms used herein are merely used to describe specific embodiments, and are not intended to limit the present application. The terms "include", "contain" and the like used herein indicate the existence of the described features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0129] It should be understood that although each step in the flowchart of the drawings is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart of the drawings can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0130] In one of the embodiments, the present application provides a computer system which can be a server. The computer system comprises a processor, a memory and a network interface connected by a system bus. The processor of the computer system is configured to provide computing and control capabilities. The memory of the computer system comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer system is configured to store data used in the above method. The network interface of the computer system is configured to communicate with an external terminal through a network connection. The computer program is configured to be executed by the processor to implement the above method.

[0131] The technical features of the above embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they should be considered as falling within the scope of the present application.

[0132] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and can be realized in other concrete forms without departing from the spirit or essential characteristics of the present application. Accordingly, the embodiments are to be considered in all respects as illustrative and not restrictive, the scope of the present application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. It is also to be understood that the use of "comprising", "including", "containing", "consisting" and their conjugates herein, "comprising", "including", "containing", "consisting", "consisting of" and their conjugates are not intended to imply that the application necessarily excludes other elements or steps.

[0133] Furthermore, it should be understood that although the description is made according to the embodiments, not every embodiment contains only one independent technical solution, and the description is made in this way only for the sake of clarity, and those skilled in the art should consider the description as a whole, and the technical solutions in each embodiment can be combined appropriately to form other embodiments which can be understood by those skilled in the art.

Claims

1. A time-evolving oriented social robot account detection method, characterized in that, Comprise: S1, obtain labeled user account samples, including user account data and labels indicating whether the user account is a social robot, select a plurality of labeled user account samples as source domain samples, and train a baseline account detection model using the corresponding feature matrix, and save part of the source domain samples to a replay buffer; S2, calculate a significance p-value based on the maximum mean difference between the feature space of the user account samples in the replay buffer and the feature space of the newly added unlabeled user account samples as target domain samples: the feature space of the user account samples in the replay buffer is the source domain feature, and the feature space of the newly added unlabeled user account samples is the target domain feature; Calculating drift statistics using maximum mean discrepancy ; f is a test function defined on the sample feature space, belonging to the unit ball of the reproducing kernel Hilbert space; is the supremum in the unit ball of the reproducing kernel Hilbert space; is the expected output of f on the distribution after the source domain features are mapped by the kernel function; is the expected output of f on the distribution after the target domain features are mapped by the same kernel function; is the feature vector of the user account sample; calculate the significance p-value according to D; If the significance p-value is lower than the preset threshold, it is determined that there is a significant distribution drift, and steps S3 to S4 are triggered; S3, assign weights to the target domain samples based on the domain discriminator and filter them, use the baseline account detection model and the auxiliary model to predict the filtered target domain samples, and write the target domain samples with consistent prediction results, confidence higher than the threshold and weight higher than the weight threshold, and the corresponding pseudo-label and weight into the replay buffer; S4, read the user account samples from the replay buffer, and perform incremental training based on the baseline account detection model as the initial model, and detect whether the user account is a social robot account based on the trained account detection model. 2.The time-evolving oriented social robot account detection method of claim 1, wherein, The user account data includes account metadata, user published content, behavior timing data, and relationship network data. 3.The time-evolving oriented social robot account detection method of claim 1, wherein, The feature matrix is obtained by extracting multi-dimensional features of the labeled user account samples as source domain samples; The multi-dimensional features include behavior features, timing features, and text features of the user account samples.

4. The time-evolving oriented social robot account detection method of claim 1, wherein, Further comprising a cross-cycle continuous learning loop process, specifically comprising the following steps: S51, wait for the newly added unlabeled user account samples of the next cycle to arrive, and perform step S2 for drift detection: S52, if the significance p-value is lower than the preset threshold, sequentially execute steps S3 and S4; replace the initial model of the current round of circulation with the latest account detection model obtained by step S4 as the initial model of the next round of circulation, and update the priority of each user account sample in the replay buffer, and reserve a specified number of user account samples according to the priority to update the replay buffer; S53, if the significance p-value is not lower than the preset threshold, only update the priority of each user account sample in the replay buffer, and reserve a specified number of user account samples according to the priority to update the replay buffer; Repeat steps S51 to S53 to realize the self-evolution of the cross-cycle and unlabeled social robot account detection capability.

5. The time-evolving oriented social robot account detection method of claim 4, wherein, The priority of each user account sample in the replay buffer is updated, and a specified number of user account samples are reserved according to the priority to update the replay buffer, specifically including: priority of the i-th user account sample in the playback buffer is: ； wherein, H (i) is the information entropy of the i-th user account sample, is the relative year decay of the i-th user account sample, and a is an adjustable coefficient. Reserve the first N user account samples in descending order of priority to form the replay buffer required for the next round of circulation.

6. The time-evolving oriented social robot account detection method of claim 1, wherein, The weight is assigned to the target domain samples based on the domain discriminator and filtered, specifically including: Mark the domain category of each labeled user account sample as 0 as a source domain sample, and mark the domain category of each unlabeled user account sample as 1 as a target domain sample, to construct a domain discrimination training set; training the domain discriminator using a multi-layer perceptron as the domain discriminator on the domain discrimination training set to output a predicted probability that each target domain sample belongs to domain class 1 ; Assigning weights to target domain samples and filtering target domain samples with w below a preset threshold.

7. The time-evolving oriented social robot account detection method of claim 1, wherein, The method comprises the following steps: The baseline account detection model and the auxiliary model are used to predict the filtered target domain samples, and the target domain samples with consistent prediction results, confidence higher than a threshold, and weight higher than a weight threshold are written into a replay buffer together with corresponding pseudo labels and weights, and the method comprises the following steps: The baseline account detection model is used to predict the target domain samples, and a first prediction result and a first confidence are obtained; The same batch of target domain samples is predicted by using an auxiliary model, and a second prediction result and a second confidence are obtained; the prediction result is whether the target domain sample is a social robot; The target domain samples with consistent first prediction result and second prediction result, first confidence and second confidence higher than respective thresholds, and weight higher than a preset weight threshold are reserved as candidate target domain samples, and the corresponding first prediction result and second prediction result are taken as pseudo labels of the candidate target domain samples; 8. The time-evolving oriented social robot account detection method of claim 1, wherein, K-fold cross-validation is performed on the candidate target domain samples, the macro average F1 value of each fold is calculated by weighting the weight of the target domain sample, if the macro average F1 value is not less than 0.85, the batch of candidate target domain samples, corresponding pseudo labels and weights are written into the replay buffer together; if the macro average F1 value is less than 0.85, the batch of candidate target domain samples is discarded. The user account sample of the history read from the replay buffer and the newly added target domain sample are used to construct an incremental training set; a baseline account detection model is used as an initial model, and the incremental training set and a gradient boosting decision tree algorithm are used for further training, and a target function is : ； Here, is the binary log loss, , is the regularization coefficient, is the number of leaf nodes of the th tree, is the weight vector of the leaf nodes of the th tree; denotes the true label of the i-th user account sample, denotes the predicted value for the i-th user account sample.

9. A computer system comprising a memory and a processor, said memory storing a computer program, characterized in that, The user account samples are read from the replay buffer, and an incremental training is performed on the baseline account detection model as an initial model, and the method comprises the following steps: The processor executes the computer program to realize the steps of the method in any one of claims 1 to 8.

Images