E-commerce abnormal group adaptive identification and processing method and system

Abstract

The invention relates to an e-commerce abnormal group adaptive identification and processing method and system, and the method comprises the steps: carrying out the cross-modal association of multi-modal features extracted from the interaction data of a multi-modal e-commerce platform, and outputting an abnormal group probability score; according to an association relationship between a historical library matching number and a real-time access amount, an influence relationship of abnormal access acceleration and a proportional relationship between an abnormal shop number and a total number of active shops, constructing a rampant degree quantification module, and calculating a rampant degree index; adaptively adjusting a processing rule updating period according to the rampant degree index; according to the processing rule updating period, the processing rule is updated and executed by adopting a large language model; and after executing according to the updated processing rule, returning real-time data to the rampant degree quantification module, and performing quantitative evaluation on the processing effect as a historical processing effect score. Compared with the prior art, the method has the advantages of being efficient, accurate, reliable and capable of achieving self-adaptive closed-loop adjustment.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce security technology, and in particular to a method and system for adaptively identifying and processing abnormal e-commerce groups. Background Art

[0002] With the rapid development of internet technology, the e-commerce industry has experienced explosive growth. E-commerce platforms provide convenient transaction channels for merchants and consumers, greatly promoting the circulation of goods and economic development. At the same time, the presence of abnormal e-commerce groups on e-commerce platforms poses a serious threat to the healthy development of the e-commerce industry.

[0003] Abnormal e-commerce groups often engage in activities such as selling counterfeit goods, fake orders, and false advertising through organized, collaborative work. This structure is characterized by a clear division of labor, with procurement, warehousing, operations, and fake orders forming a complete industry chain. For example, procurement personnel are responsible for acquiring counterfeit and shoddy goods, warehouse personnel are responsible for storing the goods, operations personnel are responsible for setting up and promoting stores on e-commerce platforms, and fake order operators use fraudulent transactions to boost store sales and reputation. This collaborative and organized division of labor makes the non-compliant activities of these groups more efficient and covert.

[0004] Furthermore, e-commerce anomalous groups rely heavily on technical means to circumvent regulation. They employ crawler technology to gather competitor information, employ AI voice-changing technology to create fake customer service communications, and even employ automated scripts to engage in large-scale fake order fraud. The use of these technical means complicates the identification and crackdown on these anomalous groups. The trans-regional nature of these groups also contributes to their difficulty in tackling. The flow of goods often spans national boundaries and even international borders, creating significant challenges for regulators during investigations and evidence collection. Differences in laws, regulations, and regulatory standards across regions also create challenges for cross-regional enforcement. The drive for high returns is fundamentally driving the existence of these anomalous groups. Selling counterfeit goods or fake orders at low cost can generate substantial profits, with some groups involved in fraudulent activities reaching tens of millions of yuan. These high profits encourage some criminals to take these risks, severely undermining fair competition in the e-commerce market and harming the legitimate rights and interests of consumers.

[0005] Abnormal group behavior on current e-commerce platforms is characterized by industrialization, intelligence, and cross-regionalization. Traditional processing methods have three major technical bottlenecks: 1) Single-dimensional recognition technology is difficult to meet the needs of multi-modal abnormal group feature fusion detection, and cannot effectively identify new methods such as AI voice-changing communication and cross-platform collaborative crime; 2) The static rule update mechanism lacks the ability to dynamically respond to abnormal situations, resulting in significant rule lag; 3) The lack of a closed-loop feedback loop results in a disconnect between rule optimization and execution effectiveness.

[0006] Therefore, there is an urgent need for a more efficient, accurate, reliable and adaptive closed-loop adaptive identification and processing solution for abnormal e-commerce groups to ensure the security and healthy development of e-commerce platforms. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an efficient, accurate, reliable and adaptive closed-loop adjustable e-commerce abnormal group adaptive identification method and system.

[0008] The purpose of the present invention can be achieved by the following technical solutions: According to a first aspect of the present invention, a method for adaptively identifying and processing abnormal e-commerce groups is provided, comprising: Extract features from interactive data on multimodal e-commerce platforms, perform cross-modal correlation on multimodal features through a dynamic attention weight matrix, and output probability scores for abnormal groups. ; Based on the correlation between the number of historical database matches and real-time visits, the impact of abnormal visit growth rate, and the ratio of the number of abnormal stores to the total number of active stores, a rampant degree quantification module is constructed to calculate the rampant degree index. ; According to the rampant index , adaptively adjust the processing rule update cycle ; Score based on the current abnormal group probability , rampant degree index , expert experience rule base and historical processing effect scores, according to the processing rule update cycle , using a large language model to update processing rules; After executing the updated processing rules, the real-time data will be returned to the rampant degree quantification module, and the processing effect will be quantitatively evaluated as the historical processing effect score.

[0009] Preferably, the feature extraction of the multimodal e-commerce platform interaction data is performed, and the multimodal features are cross-modally associated through the dynamic attention weight matrix to obtain the abnormal group probability score , specifically: The Transformer-CNN hybrid architecture is used to extract features from multimodal e-commerce platform interaction data to obtain multimodal features, including text feature vectors. , image feature vector and time series data feature vector ; The multimodal features are used as query feature vectors and key feature vectors Input source, calculate the dynamic attention weight matrix , Softmax is a normalized exponential function, As feature dimension, cross-modal association is performed by weighted summation to obtain fusion features ; The fusion features Input to the fully connected layer and get the probability score of abnormal group through activation function .

[0010] Preferably, the rampant degree quantification module is constructed based on the correlation between the number of historical library matches and the real-time access volume, the influence of abnormal access growth rate, and the proportional relationship between the number of abnormal stores and the total number of active stores to calculate the rampant degree index. , the expression is:

[0011] Where: It is the correlation item between the number of historical database matches and the real-time access volume. represents the number of records in the historical database that match the current suspected abnormal group behavior pattern. Real-time platform visits; is the adjustable weight coefficient; is the impact item of abnormal access growth rate, It is the abnormal access growth rate, is an adjustable weight coefficient; is the ratio of the number of abnormal stores to the total number of active stores, is the number of stores currently detected with abnormal behavior, is the total number of active stores on the e-commerce platform, is an adjustable weight coefficient.

[0012] Preferably, a machine learning model is constructed based on the actual operation of the e-commerce platform and the activity characteristics of abnormal groups, using the rampant degree index to The relationship between the severity of the actual abnormal group behavior incidents is the objective function, and the rampant degree index Adjustable weight coefficients in , and Perform dynamic optimization and adjustment.

[0013] Preferably, according to the rampant degree index , using the logistic model to adaptively adjust the processing rule update cycle , the expression is:

[0014] Where: Indicates the processing rule update cycle; Indicates the maximum value of the processing rule update cycle; Indicates the minimum value of the processing rule update cycle; is the parameter that controls the steepness of the curve; is the threshold parameter; Indicates the degree of rampant.

[0015] Preferably, a large language model is used to update the processing rules , the update expression is:

[0016] Where: For large language models; Represents vector concatenation; It is the expert experience rule base; Is the output of the recognition phase, including the probability score of the abnormal group and the non-compliant information detection results of relevant stores; Score the historical processing effectiveness evaluation.

[0017] Preferably, the updating process of the processing rules includes: For dynamic penalty sub-rules:

[0018]

[0019] Where: Indicates the amount of fines imposed on non-compliant stores; is the basic penalty coefficient; Quantify the degree of prevalence of abnormal group phenomena; Indicates the total sales of the store in the past set time period; For the degree of rampant The relevant penalty adjustment factor is used to control the extent to which the amount of the fine changes with the degree of rampant behavior. is the penalty adjustment coefficient before the update, is the updated penalty adjustment factor, is the adjustment factor; And / or, for the tiered store closure rules:

[0020] Where: The probability of store closure; The control coefficient for the store closure probability curve change is adjusted according to the degree of caution of the e-commerce platform in taking store closure measures; is the store closure threshold; when the parameters are generated by the random number generator satisfy , then execute the store closure operation, according to the store closure probability Set tiered store closure durations; And / or, for the smart current limiting sub-rule:

[0021] Where: The query rate per second after limiting the store's flow; This is the store's original base query rate per second; is the current limiting adjustment coefficient related to the degree of rampantness, which is used to control the amplitude of the current limiting degree changing with the degree of rampantness; And / or, for product removal rules:

[0022] Where: Indicates the percentage of products removed from the store; This is the delisting ratio adjustment coefficient related to the abnormal group probability score, which is used to control the degree to which the product delisting ratio changes with the abnormal group probability score.

[0023] Preferably, in the process of updating the processing rules using the large language model, the processing rule parameters are optimized by the gradient descent method according to the historical processing effect score, and the processing rule parameters include the penalty adjustment coefficient in the dynamic penalty sub-rule , the control coefficient of the store closure probability curve change in the graded store closure sub-rules , the delisting ratio adjustment coefficient in the product delisting rules , Current limiting adjustment coefficient in intelligent current limiting sub-rule .

[0024] Preferably, the historical processing effect score adopts the evaluation index Characterization, specifically:

[0025] Where: To govern the rampant degree before and after The change in ,in The rampant degree index is calculated based on the historical database, real-time visits and real-time store information before processing. The degree of rampantness calculated accordingly after processing; is the processing cycle; The number of related complaints received by the e-commerce platform within a certain period of time after processing; is the number of complaints in the same period before the handling; and is the weight coefficient, which is set according to the importance the e-commerce platform attaches to the changes in rampant levels and the number of complaints. .

[0026] According to a second aspect of the present invention, a system for adaptively identifying and processing abnormal e-commerce groups is provided, applying the above-mentioned method, the system comprising: Multimodal feature fusion recognition module is used to extract features from interactive data of multimodal e-commerce platforms, perform cross-modal correlation on multimodal features through dynamic attention weight matrix, and output abnormal group probability score ; The rampant degree quantification module is used to calculate the rampant degree index based on the correlation between the number of historical database matches and the real-time access volume, the impact of abnormal access growth rate, and the ratio of the number of abnormal stores to the total number of active stores. ; Update cycle adaptive adjustment module, used to adjust the level of rampant indicators , adaptively adjust the processing rule update cycle ; Processing rule update module, used to score according to the probability of the current abnormal group , rampant degree index , expert experience rule base and historical processing effect scores, according to the processing rule update cycle , using a large language model to update processing rules; The processing execution module is used to execute the updated processing rules and feed real-time data back to the rampant degree quantification module; An effect quantification evaluation module is used to quantitatively evaluate the effect of executing the updated processing rules and input the evaluation results into the processing rule update module; The modules of the system communicate asynchronously via message queues.

[0027] Compared with the prior art, the present invention has the following beneficial effects: (1) Based on the correlation between the number of historical database matches and the real-time visit volume, the impact of abnormal visit growth rate, and the ratio of the number of abnormal stores to the total number of active stores, the prevalence of abnormal groups is accurately quantified. The processing rules are updated using the quantified prevalence of abnormal groups, realizing adaptive identification and processing of abnormal groups on e-commerce platforms, which is more accurate and reliable.

[0028] (2) Based on the quantified abnormal group rampant degree index, the rule update cycle is adaptively adjusted, which can achieve a rapid response to the rule update under high rampant degree and meet the requirements of the rule update cycle under different rampant degrees. In addition, the logistic model is used to adaptively adjust the rule update cycle, and its S-shaped curve characteristics are used to avoid the periodic mutation problem of the linear model at the threshold point, further improving the rule update response speed. The setting of the inner curve steepness control parameter can also better balance the response speed and stability, ensuring stable and effective operation under different e-commerce abnormal group behavior situations.

[0029] (3) Based on the current abnormal group probability score, rampant index, expert experience rule base and historical processing effect score, and in accordance with the processing rule update cycle, the large language model is used to update the processing rules, and the specific update generation process of dynamic fine sub-rules, graded store closure sub-rules, product removal sub-rules and intelligent flow control sub-rules is given, thus realizing the dynamic and precise adjustment of the abnormal group processing rules of the e-commerce platform.

[0030] (4) Through the execution and processing effect evaluation feedback setting, stores are subject to dynamic fines, graded store closures, product removal, and intelligent flow control according to the processing rules, and the rule parameters are optimized through the gradient descent method. 、 、 、 , to achieve dynamic adjustment and adaptive optimization of governance rules.

[0031] (5) A Transformer-CNN hybrid architecture is used to process multimodal data, and cross-modal feature association is achieved through a dynamic attention weight matrix, which further improves the accuracy of multimodal recognition of abnormal group data on e-commerce platforms.

[0032] (6) By building a complete “identification → evaluation → update → governance → feedback” automated processing closed-loop system and iteratively setting processing rules based on a large language model, the probability of repeated abnormal group behaviors after processing can be reduced.

[0033] (7) The lightweight design of the model reduces the consumption of system resources. The system modules realize asynchronous communication through message queues, which enables full-process automation and can greatly improve the system throughput. The system throughput can reach 5,000 requests / second, meeting the large-scale data processing needs of e-commerce platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention; Figure 2 Implementation architecture diagram for the system. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0036] Example like Figure 1 As shown, this embodiment provides a method for adaptively identifying and processing abnormal e-commerce groups, including: Extract features from interactive data on multimodal e-commerce platforms, perform cross-modal correlation on multimodal features through a dynamic attention weight matrix, and obtain probability scores for abnormal groups. ; Based on the correlation between the number of historical database matches and real-time visits, the impact of abnormal visit growth rate, and the ratio of the number of abnormal stores to the total number of active stores, a rampant degree quantification module is constructed to calculate the rampant degree index. ; According to the rampant index , adaptively adjust the processing rule update cycle ; Score based on the current abnormal group probability , rampant degree index , expert experience rule base and historical processing effect scores, according to the processing rule update cycle , using a large language model to update processing rules; After executing the updated processing rules, the real-time data will be returned to the rampant degree quantification module, and the processing effect will be quantitatively evaluated as the historical processing effect score.

[0037] Next, the method of this embodiment is introduced in detail.

[0038] 1. Multimodal feature fusion recognition 1.1 Modal Analysis of E-commerce Platform Interaction Data Text modality: Conversations between customer service and consumers may contain key information such as false promises and misleading language. For example, to circumvent regulations on counterfeit sales, customer service may use obscure language to imply that a product is genuine when it is actually a counterfeit. Multimodal feature dimensions This dimension setting is based on the characteristics of the pre-trained language model using whole-word masking (RoBERTa-wwm) based on the RoBERTa model, which can effectively cover the text semantic information.

[0039] Image Modality: Product images may have been manipulated to mislead consumers, such as retouching a low-quality counterfeit image to appear as a high-quality, authentic product.

[0040] Time series data modality: The time series of transaction records can reveal unusual trading patterns, such as a large number of concentrated wash trades within a short period of time. The TCN network uses a void factor of [1, 2, 4] to help capture multi-scale time series characteristics. This setting is based on the different scale patterns that different criminal behaviors may exhibit in the time series.

[0041] 1.2 Model Architecture Selection This example uses a hybrid Transformer-CNN architecture to extract features from interactive data on multimodal e-commerce platforms. Transformers have a powerful self-attention mechanism for processing long text sequences, effectively capturing semantic associations within them. CNNs excel at extracting spatial features from images and are also well-suited to capturing local patterns in time-series data.

[0042] 1.3 Text Feature Extraction For customer service conversation text, use a pre-trained language model such as RoBERTa-wwm. Input the text into the model, encode it through multiple layers of Transformer blocks, and output the text feature vector. . Assume that the input text sequence length is , the vocabulary size is , the embedding dimension is , then the text input first passes through the embedding layer , get the initial embedding representation. Then, Layer Transformer blocks, each layer of Transformer blocks contains a multi-head self-attention mechanism and a feedforward neural network. In the multi-head self-attention mechanism, the query matrix , key matrix , value matrix Obtained from the embedded representation by linear transformation, such as ,in , is the head dimension. The self-attention score is calculated as , multi-head self-attention output ,in For the number of heads, . Finally, after Layer Transformer block gets the text feature vector .

[0043] 1.4 Image Feature Extraction For product images, we use the ResNet-50 network. The input image is processed through a series of operations such as convolutional layers and pooling layers to extract image features. Assume the input image size is ( is the height, is the width, is the number of channels), the convolution layer passes through the convolution kernel Perform convolution on the input, such as ,in is the input feature map, is the output feature map, is the convolution kernel size, is the bias. After multiple layers of convolution and pooling, the image feature vector is obtained .

[0044] 1.5. Time Series Data Feature Extraction For transaction record time series data, a temporal convolutional network (TCN) is used. TCN processes time series data through causal convolution and dilated convolution, which can effectively capture dependencies in long sequences. Assume that the length of the input time series data is , the feature dimension is , after a series of convolutional layers, such as ,in is the convolution kernel weight, is the convolution kernel size, is the activation function, and the time series data feature vector is obtained .

[0045] 1.6 Cross-modal feature association Through the dynamic attention weight matrix To achieve cross-modal feature association, Softmax is a normalized exponential function. Here, the text feature vector , image feature vector , time series data feature vector As query feature vectors , key feature vector The input source (for example, As , and After splicing ), is the feature dimension. After calculation, the dynamic attention weight matrix is ​​obtained , different modal features are fused by weighted summation, such as ,in for The weight parts corresponding to different modes.

[0046] 1.7 Abnormal Group Probability Score Output The fused feature vector Input to the fully connected layer, and through the activation function (such as Sigmoid function) to obtain the probability score of the abnormal group . Let the weight of the fully connected layer be , the bias is ,but , It is an activation function. The score indicates the probability that the subject corresponding to the current data belongs to an abnormal group and is used for subsequent analysis and processing.

[0047] 2. Madness Quantification Module In order to achieve adaptive adjustment of the update cycle of the rules for handling abnormal groups in e-commerce, it is necessary to accurately quantify the rampant degree of abnormal group phenomena. This embodiment comprehensively considers multiple key factors related to abnormal group crimes in e-commerce platforms and integrates these factors into a quantitative indicator through a specific mathematical formula. , which reflects the rampant degree of abnormal group phenomena on the platform and provides a basis for subsequent periodic adaptive adjustment based on the logistic model.

[0048] 2.1 Quantitative indicators of rampant severity This embodiment selects key information such as historical database, real-time visit volume, and real-time store information, and constructs a rampant degree quantification module based on the correlation between the number of historical database matches and real-time visit volume, the impact of abnormal visit growth rate, and the ratio of the number of abnormal stores to the total number of active stores to calculate the rampant degree index. , the expression is:

[0049] in: The correlation between the number of historical database matches and the real-time access volume ( ): Indicates the number of records in the historical database that match the current suspected abnormal group behavior pattern. Abnormal group behavior on e-commerce platforms has a certain pattern repetitiveness. The more matching numbers in the historical database, the more frequent similar behaviors are currently occurring, and the more rampant abnormal group activities may be. The value range is ,and is an integer; Represents the platform's real-time visits, reflecting the overall activity of the platform. Dividing the number of historical database matches by the real-time visits can, to a certain extent, measure the relative proportion of suspected abnormal group behavior under the current platform activity level; is an adjustable weight coefficient, which is used to adjust the importance of this part in the quantification of the degree of rampantness. The value range is For example, when the platform's historical data is relatively reliable and the abnormal group behavior patterns are recorded in detail, the to highlight the impact of the number of historical library matches on the degree of rampantness.

[0050] The impact of abnormal access growth rate ( ): It is the abnormal access growth rate, which reflects the growth rate of abnormal access behaviors on the e-commerce platform. Abnormal groups often create a large number of abnormal accesses in a short period of time. By monitoring the abnormal access growth rate, we can capture the dynamic changes of abnormal group activities and The logarithm is taken because the growth rate may change greatly in a short period of time. Taking the logarithm can make the data more stable and facilitate the reasonable reflection of its impact in quantitative indicators. The value range is ; It is the adjustable weight coefficient of this part. If the platform has been frequently attacked by abnormal access recently, it can be appropriately increased. value, making the abnormal access growth rate play a more important role in the quantification of rampant degree, and the value range is .

[0051] The ratio of the number of abnormal stores to the total number of active stores ( ): Refers to the number of stores currently detected with abnormal behaviors. These abnormal stores are important carriers for abnormal groups to implement non-compliant behaviors. The more abnormal stores there are, the wider the coverage of abnormal group behaviors at the store level may be. The value range is ,and is an integer; Represents the total number of active stores on the platform. By comparing the number of abnormal stores with the total number of active stores, we can get the proportion of abnormal stores in the active stores on the platform. This is the adjustable weight coefficient for this part. When the platform is more sensitive to abnormal group activities at the store level, it can be appropriately increased. To emphasize the impact of the proportion of abnormal stores on the degree of rampantness. The value range is The initial value of the adjustable weight coefficient can be set to , , .

[0052] 2.2 Weight coefficient adjustment strategy Adjustable weight coefficient 、 and It needs to be dynamically adjusted according to the actual operation of the e-commerce platform and the characteristics of abnormal group activities. For example, in a certain period of time, the platform finds that the abnormal group mainly conducts order-brushing by creating a large number of abnormal visits. At this time, it can be appropriately increased. The value of makes the abnormal access growth rate play a greater role in quantifying the degree of rampantness.

[0053] The weight coefficient can be adjusted in the following two ways: Adjustment based on expert experience: The security experts of the e-commerce platform manually adjust the weight coefficient based on the long-term observation and analysis of criminal activities on the platform and the characteristics of abnormal group activities in different periods. For example, if the experts find that the recent counterfeit sales group has used new methods to distribute goods to a large number of stores, resulting in a significant increase in the number of abnormal stores, the weight coefficient can be appropriately increased. The value of .

[0054] Automatic adjustment based on data analysis: By collecting historical platform data, building machine learning models (such as linear regression models or more complex neural network models), and quantifying indicators at a high level The relationship between the severity of the actual abnormal group crime incidents is used as the objective function to optimize the adjustable weight coefficient. At the same time, the data is updated every hour. The sliding window is 6 hours to reflect the changes in abnormal access growth rate in real time. The model automatically adjusts by continuously learning the patterns in historical data. 、 、 The value of makes the degree of rampant quantitative indicators It can more accurately reflect the actual rampant extent of abnormal group phenomena.

[0055] Through the above comprehensive and detailed quantitative indicator construction and weight coefficient adjustment strategy, we can accurately quantify the rampant extent of abnormal group phenomena on e-commerce platforms, providing a solid foundation for subsequent periodic adaptive adjustments and processing rule updates.

[0056] 3. Processing rule periodic adaptive update Traditional static rule update mechanisms, such as fixed-period update methods, are unable to adapt to the dynamic changes in the behavior patterns of abnormal groups on e-commerce platforms, resulting in rule lags and difficulties in effectively handling them. Linear adjustment strategies, on the other hand, experience a significant drop in response speed when the occurrence rate of behavior exceeds a threshold. Therefore, a method is needed that can dynamically adjust the rule update cycle based on the prevalence of abnormal groups on e-commerce platforms, enabling faster responses as the prevalence increases.

[0057] The logistic model's unique S-shaped curve gives it a significant advantage over other models in addressing the relationship between rule update cycles and crime severity. Linear models experience sudden changes in the cycle at thresholds, which can lead to unstable system responses and prevent smooth adaptation to the required rule update cycles at varying crime severity levels. The logistic model, however, avoids this sudden change. Its S-shaped curve ensures that the rule update cycle is relatively long at low crime severity levels and gradually shortens as crime severity increases, without experiencing sudden jumps. This ensures the system operates stably and effectively across diverse crime scenarios.

[0058] In this embodiment, the following logistic model is used to implement adaptive adjustment of the rule update cycle:

[0059] The variables are defined as follows: : Indicates the rule update cycle, in days. This is calculated through the model and is adjusted in real time based on the prevalence of abnormal groups on the e-commerce platform.

[0060] : Represents the maximum value of the rule update cycle, expressed in days, when the prevalence of abnormal groups on the e-commerce platform is minimal. In practical e-commerce governance scenarios, setting a longer upper limit on the update cycle can prevent unnecessary consumption of system resources from frequent rule updates and provide sufficient time for governance measures to be effective.

[0061] : This represents the minimum rule update interval, expressed in days, when abnormal e-commerce behavior is rampant. This minimum ensures a rapid response and timely rule updates to address changes in the face of severe abnormal behavior.

[0062] : It represents the quantitative value of the rampant degree of abnormal group phenomena in e-commerce, through the formula This value comprehensively reflects the rampant non-compliant behavior of abnormal groups on e-commerce platforms and is the key basis for periodic adaptive adjustments. 、 、 The value range is and , is the number of historical database matches, For real-time visits, For abnormal access growth rate, is the number of abnormal stores, is the total number of active stores.

[0063] : is a parameter that controls the steepness of the curve. It determines the steepness of the curve. Changes in rules update cycle The rate of change. When the value is large, the curve is steeper, which means that a small change in the rampant level will cause a large change in the update period; when When the value is small, the curve is relatively flat, and the update cycle responds slowly to the change in rampant intensity. When , a better balance can be achieved between response speed and system stability.

[0064] : is the threshold parameter. It represents a critical value of rampant degree. Value Logistic models enable rule update cycles when nearby changes occur. There have been more obvious changes to adapt to different levels of crime.

[0065] In order to verify the effectiveness of the adaptive processing rule update cycle based on the logistic model, this embodiment conducted a series of experiments. In the experiments, the activities of abnormal groups with different levels of rampant activity were simulated, and the corresponding rule update cycle changes were recorded.

[0066] For example, setting sky, sky, , .when from Increase to You can see the update cycle from The nonlinearity decreases to The specific data are shown in Table 1 below.

[0067] Table 1

[0068] This result shows that as the prevalence of abnormal groups on e-commerce platforms increases, the model can effectively shorten the rule update cycle and quickly respond to changes in the behavior of abnormal groups, thereby providing strong support for timely adjustment of processing strategies.

[0069] Through the above detailed design and verification, the adaptive setting of the processing rule update cycle based on the logistic model can accurately, stably and efficiently adjust the rule update cycle according to the rampant degree of abnormal groups on the e-commerce platform, effectively making up for the shortcomings of traditional methods in dealing with dynamic abnormal behavior trends, and providing an innovative solution for the security governance of e-commerce platforms.

[0070] 4. Update and iterate processing rules based on large language models 4.1. Rule generation function construction In the scenario of abnormal group governance on e-commerce platforms, in order to achieve accurate and effective rule updates, this embodiment constructs a rule generation function .

[0071] in: This expert experience rule base contains over 200 basic rules. It is a set of rules based on the knowledge and practical experience accumulated by e-commerce experts over a long period of time. It includes various rules for handling common illegal and criminal behaviors of abnormal e-commerce groups. Score the probability of abnormal groups output by the multimodal feature fusion recognition module As well as the specific violation information of related stores and other detection results, these results reflect the actual illegal and criminal situation of the current e-commerce abnormal groups. It is the governance effect evaluation index given by the governance execution and effect feedback module, which is used to measure the effectiveness of previous governance rules. Representation vector concatenation: By concatenating these three pieces of information into a single vector and inputting it into the Large Language Model (LLM), the model can generate new governance rules based on rich contextual information. .

[0072] 4.2 Processing Rule Updates 1) Dynamic penalty sub-rules:

[0073] Variable definition: Indicates the amount of fines imposed on illegal stores; The base penalty coefficient is a fixed value pre-set based on the general regulations of the e-commerce platform and the nature of different types of illegal and criminal behaviors. For example, for minor counterfeit sales, the base is set at 1,000 yuan, and for serious fake order behavior, the base is set at 5,000 yuan. For the degree of rampant The relevant penalty adjustment coefficient can be set to 0.5 initially, and then adjusted dynamically based on the governance effect to control the extent to which the fine amount changes with the degree of rampant behavior. For example, when hour, Automatically increased to 1.2, resulting in a 20% increase in fines; Quantify the degree of prevalence of abnormal group phenomena; Indicates the total sales of the store in the past 7 days, which can be obtained from the transaction record database of the e-commerce platform.

[0074] Principle: This formula shows that the amount of the fine is not only related to the store’s recent sales, but also increases dynamically as the rampant level of abnormal e-commerce groups increases. The larger the amount, that is, the more rampant the abnormal group phenomenon is, the greater the increase in the fine amount based on the product of the basic fine and sales, thereby increasing the intensity of economic penalties for rampant illegal and criminal acts. For example, when When increasing from 0.5 to 0.8, assuming , , The initial value is 0.5. hour, ;when When adjusting according to the governance effect is 0.8, , the amount of fines increased significantly.

[0075] 2) Rules for tiered store closures:

[0076] Variable definition: The probability of a store being closed; is the store closure probability adjustment coefficient related to the degree of rampant activity. It is an adjustable parameter with a value range of [1, 2], which is used to control the sensitivity of the store closure probability to changes in the degree of rampant activity. Quantify the degree of prevalence of abnormal group phenomena; The critical rampant degree value of the change in the probability of store closure is a threshold set according to the actual operation situation and risk tolerance of the e-commerce platform, for example, it can be set to 0.6.

[0077] Principle: This formula is based on the logistic function. Less than When the store is closed Relatively low; with Gradually increase and exceed , the probability of store closure It will rise rapidly and approach 1. In this way, a graded closure strategy is implemented for different stores according to the rampant degree of abnormal group phenomena. For stores in areas with high rampant degree, the probability of closure is higher, achieving precise strikes. , For example, when hour, ;when hour, , the probability of closing the store has increased significantly. Value, the probability of closing the store Follow Changing data.

[0078] Table 2

[0079] 3) Product removal rules:

[0080] Variable definition: Indicates the proportion of store products that need to be removed from shelves; The coefficient for adjusting the product removal rate related to the probability score of the abnormal group is an adjustable parameter with a value range of [0.5, 1]. It is used to control the degree to which the product removal rate changes with the probability score of the abnormal group. This is the abnormal group probability score output by the multimodal feature fusion recognition module, reflecting the possibility that the store belongs to the e-commerce abnormal group.

[0081] Principle: This formula determines the product removal ratio based on the probability that the identified store belongs to the e-commerce anomaly group. The higher the value, the more likely the store is to be part of a group of illegal and criminal abnormalities, and the proportion of products removed from shelves is higher. The larger the value, the higher the value, but the maximum value cannot exceed 1 (i.e., all products are removed from shelves). In this way, for stores suspected of being abnormal e-commerce groups, a corresponding proportion of products will be removed from shelves according to their degree of suspicion, limiting the scope of their illegal and criminal activities. The parameter setting is based on the fact that different product categories have different risk levels. For high-risk categories, such as beauty products, since they have a greater impact on consumer health, It can be set to 0.8 to ensure a higher removal rate; for general risk categories, such as daily necessities, Can be set to 0.5. For example, when , hour, , that is, 64.8% of the products were removed from the shelves. In actual testing, it was found that for most product categories, when When it is 0.8, it can effectively combat abnormal counterfeiting behavior by e-commerce groups while preventing excessive removal of products from causing excessive impact on normal operating stores.

[0082] 4) Intelligent current limiting sub-rules:

[0083] Variable definition: The query rate per second after the store is throttled (Queries Per Second); The original basic query rate per second of the store, that is, the access rate allowed under normal circumstances; is the current limiting adjustment coefficient related to the degree of rampantness. It is an adjustable parameter with a value range of [0.2, 0.5], which is used to control the amplitude of the current limiting degree changing with the degree of rampantness. A quantitative indicator of the prevalence of abnormal group phenomena.

[0084] Principle: This formula shows that as the abnormal group phenomenon becomes more rampant The increase, The value of will decrease, so that the query rate per second after current limiting Reduce. That is, for shops in areas with high levels of rampant crime, stricter flow control measures will be implemented to reduce their traffic and reduce the scope and influence of their illegal and criminal activities. For example, when , When, if , , the current limiting range is 21.4%. The value is 0.2, When it increases from 0.4 to 0.8, They are , , , , , the current limiting range gradually increases from 7.7% to 14.8%; when The value is 0.5, When it increases from 0.4 to 0.8, They are , , , , The current limit gradually increases from 18.1% to 33%, quantifying the different Value, as Variable current limit amplitude.

[0085] 4.3. Innovation and effectiveness of rule updates Taking into account the organizational, technology-dependent, and cross-regional nature of e-commerce anomaly groups, this rule iteration module based on a large language model is both innovative and effective. Its innovation lies in leveraging the powerful language understanding and generation capabilities of the large language model to integrate multi-source information to generate rules, breaking through the limitations of traditional static rules. Its effectiveness lies in dynamically adjusting key governance measures such as fines, store closure probability, product removal rate, and flow control level, closely integrating the prevalence of anomaly groups with actual detection and governance results. This allows for more precise targeting of illegal and criminal activities by e-commerce anomaly groups, improving governance efficiency, and mitigating the impact these groups pose on e-commerce platforms.

[0086] 5. Governance implementation and treatment effect feedback 5.1 Governance Execution According to the governance rules generated by the rule iteration module, specific handling measures are implemented for relevant stores.

[0087] 1) Dynamic fines: according to the rules Execute fines. Among them, The basic fine amount can be pre-set according to different violation types, platform policies, etc. For example, for minor counterfeit sales, Set to 1,000 yuan, for serious fraudulent behavior, Set to 5,000 yuan, etc. For the degree of rampant The relevant penalty adjustment coefficient is initially set to 0.5, and will be adjusted adaptively based on the governance effect. The adjustment mechanism is as follows: Set initial .

[0088] Defining the Adjustment Factor , Dynamic calculation is performed based on the changing trend of the rampant degree of abnormal e-commerce groups on the platform over a period of time (such as a week) and the feedback on the governance effect of penalty measures.

[0089] , when the governance effect is good, that is, the degree of rampant When the number of complaints decreases significantly, Take a smaller value (such as 0.1); when the governance effect is not good, When there is no significant decrease or the number of complaints increases, Take a larger value (such as 0.3). The sales figures for the store over the past seven days can be obtained from the e-commerce platform's transaction record database. This way, the fine amount is linked to the prevalence of the abnormal group and the store's recent business volume, increasing the deterrent effect of the fine.

[0090] 2) Tiered store closures: Based on the formula Determine the probability of store closure. To control the parameters of the store closure probability curve, the value will be adjusted according to the platform's cautiousness in taking store closure measures, for example, a value of 1.5; The store closure threshold is set to 0.6, for example, when the rampant When this threshold is exceeded, the probability of store closure increases significantly. , through a random number generator (e.g. in Generate uniformly distributed random numbers in the interval ,like , then execute the store closure operation) to decide whether to close the store. At the same time, the store closure duration is graded according to the store closure probability, for example, when When the store is closed for 1 week; when When the store is closed for 1 month; when The store closure period is set to more than 3 months to achieve precise management. The following is an example of the store closure probability change curve data: when , When Value from Start by increasing each time , calculate the corresponding value.

[0091] hour, .

[0092] hour, .

[0093] hour, .

[0094] hour, .

[0095] hour, .

[0096] 3) Product removal: Based on Determine the ratio of products to be removed from shelves. To control the parameters of the delisting ratio, the setting basis is as follows: For high-risk cosmetics categories, since cosmetics products come into direct contact with the human body, the quality and safety risks are high. Once fakes are sold, they may cause great harm to consumers. Set it to 0.8 to more strictly control the sales of beauty products in stores suspected of abnormal groups.

[0097] For daily necessities with general risks, the potential harm to consumers is relatively small. Set it to 0.5 to minimize the impact on normal operations while ensuring a certain level of control. It is the probability score of abnormal groups output by the multimodal feature fusion recognition module. , select a corresponding proportion of products from the store's product list according to certain rules (such as random selection or order of product sales from high to low) and remove them from the shelves.

[0098] 4) Intelligent current limiting: According to the formula Limit the flow of customers to the store. This is the store's original upper limit on requests per second, which can be set based on factors such as the store's size and reputation level. For example, a large store with a good reputation 1000, small new store is 100; For the degree of rampant The related flow limiting adjustment coefficient can be adjusted according to the platform's flow control strategy, such as a value of 0.3. The quantitative flow limiting range is as follows: when , hour, , the current limiting range is approximately ,Right now .

[0099] when , hour, , the current limiting range is approximately ,Right now .

[0100] when , hour, , the current limiting range is approximately ,Right now By setting , limit the store’s access traffic and reduce the impact of its illegal behavior.

[0101] 5.2 Effect Feedback Constructing evaluation indicators To evaluate the governance effect and provide feedback to the rule update part to optimize the rules.

[0102] Variable definition: To govern the rampant degree before and after The change in ,in The degree of rampantness calculated based on historical database, real-time visits, real-time store information, etc. before management, The degree of rampantness calculated accordingly after governance; The governance cycle can be set to, for example, one week or one month, and adjusted according to actual conditions; The number of complaints related to the store or similar abnormal groups received by the platform within a certain period of time (e.g., one month) after the governance; The number of complaints in the same period before governance; and is the weight coefficient, which is set according to the importance the platform attaches to the changes in rampant levels and the number of complaints, for example , .

[0103] Optimize rule parameters: Optimize rule parameters by gradient descent method . With dynamic fines For example, calculate about Gradient , in each iteration, according to renew The value of is the learning rate, for example, it is 0.01. By continuously iteratively updating the parameters, the evaluation index It gradually increases, that is, the governance effect is continuously optimized, and dynamic adjustment and adaptive optimization of governance rules are realized to better deal with violations by abnormal e-commerce groups.

[0104] like Figure 2 As shown, this embodiment provides an e-commerce abnormal group adaptive identification and processing system, applying the above method, the system includes: Multimodal feature fusion recognition module is used to extract features from interactive data of multimodal e-commerce platforms, perform cross-modal association of multimodal features through dynamic attention weight matrix, and output abnormal group probability scores. ; The rampant degree quantification module is used to calculate the rampant degree index based on the correlation between the number of historical library matches and the real-time access volume, the impact of abnormal access growth rate, and the ratio of the number of abnormal stores to the total number of active stores. ; Update cycle adaptive adjustment module, used to adjust the level of rampant indicators , adaptively adjust the processing rule update cycle ; Processing rule update module, used to score according to the probability of the current abnormal group , rampant degree index , expert experience rule base and historical processing effect scores, according to the processing rule update cycle , using a large language model to update processing rules; The processing execution module is used to execute the updated processing rules and feed real-time data back to the rampant degree quantification module; The effect quantification evaluation module is used to quantitatively evaluate the effect after executing the updated processing rules and input the evaluation results into the processing rule updating module.

[0105] This embodiment establishes a closed-loop structure of "identification → assessment → update → governance → feedback" to achieve efficient and intelligent governance of abnormal e-commerce groups. System modules communicate asynchronously via message queues, effectively decoupling each module and improving the system's concurrent processing capabilities and stability. In actual testing, in a simulated large-scale data processing scenario for an e-commerce platform, the system achieved a throughput of 5,000 requests per second, meeting the large-scale data processing needs of e-commerce platforms during daily operations and peak traffic periods.

[0106] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for adaptively identifying and processing abnormal e-commerce groups, characterized by: include: Extract features from interactive data on multimodal e-commerce platforms, perform cross-modal correlation on multimodal features through a dynamic attention weight matrix, and obtain probability scores for abnormal groups. ; Based on the correlation between the number of historical database matches and real-time visits, the impact of abnormal visit growth rate, and the ratio of the number of abnormal stores to the total number of active stores, a rampant degree quantification module is constructed to calculate the rampant degree index. ; According to the rampant index , adaptively adjust the processing rule update cycle ; Score based on the current abnormal group probability , rampant degree index , expert experience rule base and historical processing effect scores, according to the processing rule update cycle , using a large language model to update processing rules; After executing the updated processing rules, the real-time data will be returned to the rampant degree quantification module, and the processing effect will be quantitatively evaluated as the historical processing effect score.

2. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 1 is characterized in that: The method extracts features from interactive data of multimodal e-commerce platforms, and uses dynamic attention weight matrix to perform cross-modal association on multimodal features to obtain abnormal group probability scores. , specifically: The Transformer-CNN hybrid architecture is used to extract features from multimodal e-commerce platform interaction data to obtain multimodal features, including text feature vectors. , image feature vector and time series data feature vector ; The multimodal features are used as query feature vectors and key feature vectors Input source, calculate the dynamic attention weight matrix , Softmax is a normalized exponential function, As feature dimension, cross-modal association is performed by weighted summation to obtain fusion features ; The fusion features Input to the fully connected layer and get the probability score of abnormal group through activation function .

3. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 1 is characterized in that: According to the correlation between the number of historical database matches and the real-time access volume, the impact of abnormal access growth rate, and the ratio of the number of abnormal stores to the total number of active stores, a rampant degree quantification module is constructed to calculate the rampant degree index. , the expression is: ， Where: It is the correlation item between the number of historical database matches and the real-time access volume. represents the number of records in the historical database that match the current suspected abnormal group behavior pattern. Real-time platform visits; is the adjustable weight coefficient; is the impact item of abnormal access growth rate, Is the abnormal access growth rate, is an adjustable weight coefficient; is the ratio of the number of abnormal stores to the total number of active stores, is the number of stores currently detected with abnormal behavior, is the total number of active stores on the e-commerce platform, is an adjustable weight coefficient.

4. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 3 is characterized in that: According to the actual operation of the e-commerce platform and the activity characteristics of abnormal groups, a machine learning model is built to identify the rampant degree indicators. The relationship between the severity of the actual abnormal group behavior incidents is the objective function, and the rampant degree index Adjustable weight coefficients in , and Perform dynamic optimization and adjustment.

5. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 1 is characterized in that: According to the rampant index , using the logistic model to adaptively adjust the processing rule update cycle , the expression is: ， Where: Indicates the processing rule update cycle; Indicates the maximum value of the processing rule update cycle; Indicates the minimum value of the processing rule update cycle; is the parameter that controls the steepness of the curve; is the threshold parameter; Indicates the degree of rampant.

6. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 1 is characterized in that: Use a large language model to update processing rules , the update expression is: ， Where: For large language models; Represents vector concatenation; It is the expert experience rule base; Is the output of the recognition phase, including the probability score of the abnormal group and the non-compliant information detection results of the relevant stores; Score the historical processing performance evaluation.

7. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 6 is characterized in that: The updating process of the processing rules includes: For dynamic penalty sub-rules: ， ， Where: Indicates the amount of fines imposed on non-compliant stores; is the basic penalty coefficient; Quantify the degree of prevalence of abnormal group phenomena; Indicates the total sales of the store in the past set time period; For the degree of rampant The relevant penalty adjustment factor is used to control the extent to which the amount of the fine changes with the degree of rampant behavior. is the penalty adjustment coefficient before the update, is the updated penalty adjustment factor, is the adjustment factor; And / or, for the tiered store closure rules: ， Where: The probability of store closure; The control coefficient for the store closure probability curve change is adjusted according to the degree of caution of the e-commerce platform in taking store closure measures; is the store closure threshold; when the parameters are generated by the random number generator satisfy , then execute the store closure operation, according to the store closure probability Set tiered store closure durations; And / or, for the smart current limiting sub-rule: ， Where: The query rate per second after limiting the store's flow; This is the store's original base query rate per second; is the current limiting adjustment coefficient related to the degree of rampantness, which is used to control the amplitude of the current limiting degree changing with the degree of rampantness; And / or, for product removal rules: ， Where: Indicates the percentage of products removed from shelves in a store; This is the delisting ratio adjustment coefficient related to the abnormal group probability score, which is used to control the degree to which the product delisting ratio changes with the abnormal group probability score.

8. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 7 is characterized in that: In the process of updating the processing rules using the large language model, the processing rule parameters are optimized by the gradient descent method based on the historical processing effect score. The processing rule parameters include the penalty adjustment coefficient in the dynamic penalty sub-rule. , the control coefficient of the store closure probability curve change in the graded store closure sub-rules , the delisting ratio adjustment coefficient in the product delisting rules , Current limiting adjustment coefficient in intelligent current limiting sub-rule .

9. The method for adaptively identifying and processing abnormal e-commerce groups according to claim 1, characterized in that: The historical processing effect score adopts the evaluation index Characterization, specifically: ， Where: To govern the rampant degree before and after The change in ,in The rampant degree index is calculated based on the historical database, real-time visits and real-time store information before processing. The degree of rampantness calculated accordingly after processing; is the processing cycle; The number of related complaints received by the e-commerce platform within a certain period of time after processing; is the number of complaints in the same period before the handling; and is the weight coefficient, which is set according to the importance the e-commerce platform attaches to the changes in rampant levels and the number of complaints. .

10. An e-commerce abnormal group adaptive identification and processing system, characterized by: Applying the method of claim 1, the system comprises: Multimodal feature fusion recognition module is used to extract features from interactive data of multimodal e-commerce platforms, perform cross-modal correlation on multimodal features through dynamic attention weight matrix, and output abnormal group probability score ; The rampant degree quantification module is used to calculate the rampant degree index based on the correlation between the number of historical database matches and the real-time access volume, the impact of abnormal access growth rate, and the ratio of the number of abnormal stores to the total number of active stores. ; Update cycle adaptive adjustment module, used to adjust the level of rampant indicators , adaptively adjust the processing rule update cycle ; Processing rule update module, used to score according to the probability of the current abnormal group , rampant degree index , expert experience rule base and historical processing effect scores, according to the processing rule update cycle , using a large language model to update processing rules; The processing execution module is used to execute the updated processing rules and feed real-time data back to the rampant degree quantification module; An effect quantification evaluation module is used to quantitatively evaluate the effect of executing the updated processing rules and input the evaluation results into the processing rule update module; The modules of the system communicate asynchronously via message queues.

Images