Risk shop station group detection method, device thereof, equipment and medium

Abstract

The invention relates to a risk shop station group detection method and device, equipment and a medium, and the method comprises the steps: marking a shop as a risk shop when the number of commodities with abnormal transactions in the shop meets a preset condition, and collecting a plurality of risk shops to form a risk shop set; the ratio of the total number of commodities of every two risk shops in the risk shop set is calculated, when the ratio of the total number of commodities is located in a preset ratio interval, the two corresponding risk shops are marked as suspected risk shop pairs, and the suspected risk shop pairs are used for forming a suspected risk shop pair set; filtering out a part of suspected risk shop pairs from the suspected risk shop pair set; and calculating the commodity similarity coverage rate of the two risk shops corresponding to the suspected risk shop pair, and determining risk station group shops based on the commodity similarity coverage rates corresponding to all the suspected risk shop pairs in the suspected risk shop pair set. Through a multi-stage screening mechanism, the detection range is gradually reduced, and finally the risk shop station group is accurately identified.

Description

Technical Field

[0001] The present application relates to the field of e-commerce technology, and in particular to a method for detecting risky store clusters and its devices, equipment, and media. Background Art

[0002] E-commerce platforms usually have a large number of online stores, which are actually operated by different business entities. Different operators and their operating methods make the corresponding online stores present different risk levels. The risk level of an online store refers to an abstract description of the operating health of an online store, which is usually reflected in the level of operating credit of the online store. Online stores with low risk levels have better operating credit, while online stores with high risk levels, i.e. risky stores, have relatively poor operating credit. It can be seen that the risk level of online stores on an e-commerce platform basically determines the survival of the platform.

[0003] Existing technologies usually rely on independent analysis of a single store when identifying risky stores. Specifically, existing technologies mainly use preset rules or models to conduct risk assessments and risk level classifications for single stores. However, in order to circumvent e-commerce platform supervision, diversify risks, or engage in other improper behaviors, some operators often set up multiple online stores. These stores appear to be independently operated, with different store names and operating entities, but are actually controlled by the same operating entity. In this case, when a store is at risk, if the store group to which it belongs cannot be identified and controlled in a timely manner, it will cause huge potential risks and losses to the e-commerce platform.

[0004] Therefore, how to design a detection method that can identify risky store clusters has become a technical problem that e-commerce platforms urgently need to solve. Summary of the invention

[0005] The primary purpose of the present application is to solve at least one of the above problems and provide a risky store site group detection method and its device, equipment, and medium.

[0006] In order to meet the various objectives of this application, this application adopts the following technical solutions:

[0007] A risky store group detection method provided for one of the purposes of this application includes the following steps:

[0008] When the number of abnormally traded goods in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set;

[0009] Calculate the ratio of the total number of goods of each pair of risky shops in the risky shop set, and when the ratio of the total number of goods is within a preset ratio interval, mark the corresponding two risky shops as a suspected risky shop pair, and the suspected risky shop pair is used to form a suspected risky shop pair set;

[0010] Based on the quantity of the abnormally traded commodities corresponding to the suspected risky store pairs and a preset similarity determination rule, filtering out some of the suspected risky store pairs from the set of suspected risky store pairs;

[0011] Based on the filtered set of suspected risk store pairs, the product similarity coverage rate of the suspected risk store pairs corresponding to the two risk stores is calculated, and based on the corresponding product similarity coverage rates of all suspected risk store pairs in the set of suspected risk store pairs, the risky site group stores are determined.

[0012] On the other hand, a risky store cluster detection device provided to meet one of the purposes of the present application includes:

[0013] A risky store marking module is used to mark a store as a risky store when the number of abnormally traded goods in the store meets a preset condition, and collect multiple risky stores to form a risky store set;

[0014] A risky store pair formation module is used to calculate the ratio of the total number of goods of each of the risky stores in the risky store set. When the total number of goods ratio is within a preset ratio interval, the corresponding two risky stores are marked as suspected risky store pairs, and the suspected risky store pairs are used to form a suspected risky store pair set;

[0015] A risky store pair filtering module, configured to filter out some of the suspected risky store pairs from the suspected risky store pair set based on the quantity of the abnormally traded goods corresponding to the suspected risky store pairs and a preset similarity determination rule;

[0016] The risky store cluster determination module is used to calculate the product similarity coverage rate of the corresponding two risk stores of the suspected risk store pair based on the filtered set of suspected risk store pairs, and determine the risky store cluster based on the corresponding product similarity coverage rate of all suspected risk store pairs in the set of suspected risk store pairs.

[0017] On the other hand, a computer device provided to meet one of the purposes of the present application includes a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the risky store site cluster detection method described in the present application.

[0018] On the other hand, a computer program product provided to meet another purpose of the present application includes a computer program / instruction, which, when executed by a processor, implements the steps of the method described in any embodiment of the present application.

[0019] The technical solution of this application has many advantages, including but not limited to the following aspects:

[0020] On the one hand, this application comprehensively considers the correlation between multiple risky stores and can more effectively identify risky store clusters controlled by the same business entity. This not only improves the coverage of detection, but also can more accurately capture the behavior of merchants to diversify risks by operating multiple stores.

[0021] On the other hand, the present application gradually narrows the scope of detection through a multi-stage screening mechanism, and finally determines the risky station group stores. Specifically, the present application first screens out risky stores through preset conditions to form a risky store set, and then further screens out suspected risky store pairs to form a suspected risky store pair set based on the ratio of the total number of goods in the risky stores in the risky store set. That is, by quantifying the relationship between the scale of goods between stores, it is possible to preliminarily identify risky store stations that may be controlled by the same business entity. Then, based on the number of abnormally traded goods and the preset similarity judgment rules, the suspected risky store pairs are filtered, and on the basis of the suspected risky store pair set, more strongly associated suspected risky store pairs are screened out, and suspected risky store pairs that do not meet the preset similarity judgment rules are filtered out, reducing the false alarm rate and improving the accuracy of detection. Finally, by calculating the similarity coverage of goods and determining the risky station group stores, the risky store stations controlled by the same business entity can be fully identified.

[0022] This application uses a multi-stage screening mechanism to gradually narrow the scope of detection, and ultimately accurately identifies risky store clusters controlled by the same business entity, effectively reducing the false alarm rate, improving the accuracy and reliability of detection, and providing e-commerce platforms with more efficient risky store cluster detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0024] Figure 1 This is a flowchart of a typical embodiment of the risky store site group detection method of the present application;

[0025] Figure 2 This is a flow chart of marking risky stores according to the proportion of prohibited or restricted sales commodities in an embodiment of the present application;

[0026] Figure 3This is a schematic diagram of a process for marking risky stores according to the fraud risk ratio in an embodiment of the present application;

[0027] Figure 4 This is a flow chart of filtering suspected risky store pairs based on the number of abnormally traded goods in an embodiment of the present application;

[0028] Figure 5 A schematic diagram of a process for determining whether the first similarity judgment condition is met in an embodiment of the present application;

[0029] Figure 6 This is a flow chart of determining risky site group stores based on commodity similarity coverage in an embodiment of the present application;

[0030] Figure 7 A schematic diagram of a process for determining a commodity similarity coverage rate in an embodiment of the present application;

[0031] Figure 8 This is a principle block diagram of the risky store cluster detection device of the present application;

[0032] Fig. 9 A schematic diagram of the structure of a computer device used in this application. DETAILED DESCRIPTION

[0033] A risky store cluster detection method of the present application can be programmed as a computer program product, deployed in a server and implemented. For example, in the exemplary application scenario of the present application, it can be deployed in the server of an e-commerce platform, wherein the e-commerce platform can be an e-commerce platform that opens independent station services. An independent station refers to a new type of official website (website) established on the SaaS technology platform, which has an independent domain name, private content, data, and rights and interests, independent business sovereignty and business entity responsibilities, supported by social cloud computing capabilities, and can independently and freely connect to third-party software tools, publicity and promotion media and channels. Usually, the seller of goods builds and operates an online store on an independent station, so that the online store can display its goods on the independent station, and provide e-commerce services such as searching for goods, adding shopping carts, settlement, logistics delivery, placing orders, and after-sales service on the online store to the buyers of goods (that is, users). The buyers of goods (users) perform various interactive operations with the e-commerce platform on the terminal device.

[0034] In some embodiments of the present application, commodities with a risk of being banned or restricted for sale or with a risk of fraud are regarded as commodities with abnormal transactions, wherein commodities with a risk of being banned or restricted for sale (hereinafter referred to as banned or restricted commodities) are commodities that are clearly restricted or prohibited from sale by e-commerce platforms, such as contraband, illegal and illegal commodities, dangerous goods, etc. In the field of cross-border e-commerce, banned or restricted commodities may include seafood, agricultural products, alcoholic beverages, etc.; commodities with a risk of fraud (hereinafter referred to as fraud commodities) refer to commodities that are illegally operated by merchants, such as false shipments, inferior goods as good ones, and other illegal business operations. It should be noted that when merchants sell fraud commodities on the shelves, they will attract consumers' attention at a low price far below the market standard price, so that consumers ignore the doubts about the authenticity of the commodities. Among them, fraud commodities are usually counterfeit and shoddy products with extremely low costs. Merchants usually sell them at low prices to quickly obtain profits, circumvent platform supervision or induce consumers to place orders.

[0035] In a typical embodiment of the present application, the number of goods with abnormal transactions in a store is obtained, and whether the store is a risk store is determined based on the number of goods, and multiple risk stores are collected to form a risk store set. The number of goods with abnormal transactions in the risk store in the risk store set accounts for a high proportion of the total number of goods, which meets the characteristics of the risk store corresponding to the risk store station group. And if one of the multiple stores controlled by the same business entity is a risk store, there is a high probability that the other stores controlled by the business entity are also risk stores. The ratio of the total number of goods of the two risk stores in the risk store set is calculated. When the ratio of the total number of goods is within the preset ratio interval, the corresponding two risk stores are marked as suspected risk store pairs. This step takes into account that the number of goods on the shelves of the risk stores in the same risk store station group usually has a certain similarity. When the same business entity controls multiple stores, it often adopts a unified operation strategy and product listing mode, which causes the total number of goods in these stores to fluctuate within a certain range. By calculating the ratio of the total number of goods and setting a reasonable preset ratio interval, these suspected risk store pairs with similar characteristics can be effectively identified.

[0036] Based on the number of goods in the suspected risky store pairs corresponding to abnormal transaction goods and the preset similarity judgment rules, the similarity and correlation between the suspected risky store pairs are further determined to confirm whether the suspected risky store pairs belong to the same risky store station group. In some embodiments, similarity judgment rules are preset based on the first product category, the risk ratio of prohibited and restricted sales goods, the second product category, and the ratio of fraudulent product categories. For specific steps, please refer to the subsequent specific implementation methods, which will not be repeated here. Through this step, some suspected risky store pairs that do not meet the preset similarity judgment rules are filtered out from the set of suspected risky store pairs. Among them, the similarity judgment rules are preset by technical personnel in this field as needed.

[0037] Finally, the risky station group stores are determined by calculating the product similarity coverage of the suspected risky store to the corresponding two risky stores. In one embodiment, each store in the suspected risky store pair set is marked as a node in the risky station group store graph. When the product similarity coverage is greater than the preset coverage threshold, an edge connection is established between the corresponding two stores to form a risky station group store graph for determining the risky station group to which the store belongs. The coverage threshold is set by a technician in this field based on business data or experience.

[0038] A risky store site cluster detection method of the present application can be programmed as a computer program product and deployed in a server for execution. For example, in the exemplary application scenario of the present application, it can be deployed and implemented in the server of an e-commerce customer service platform, thereby accessing an interface opened after the computer program product is running and performing human-computer interaction with the process of the computer program product through a graphical user interface to execute the method.

[0039] See also Figure 1 The risky store cluster detection method of the present application, in its typical embodiment, comprises the following steps:

[0040] Step S5100: When the number of abnormally traded commodities in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set;

[0041] The number of goods with abnormal transactions in the store is obtained. In one embodiment, the goods with prohibited or restricted sales risk are regarded as abnormal transactions. The preset condition is set as the proportion of prohibited or restricted sales goods is greater than the preset ratio threshold, wherein the number of goods with abnormal transactions with prohibited or restricted sales risk in the corresponding store divided by the total number of goods in the store is the proportion of prohibited or restricted sales goods, and the ratio threshold is set by technical personnel in this field as needed. Specifically, all valid goods in the store are tested for prohibited or restricted sales risks, and prohibited or restricted sales risk labels are added to the corresponding valid goods. For details of this step, please refer to the subsequent specific implementation methods, which will not be repeated here. After adding prohibited or restricted sales risk labels to valid goods, the number of prohibited or restricted sales risk labels of the store can be directly counted to obtain the number of goods with abnormal transactions. When the number of goods with abnormal transactions in the store meets the preset conditions, that is, when the proportion of prohibited or restricted sales goods is greater than the preset ratio threshold, the corresponding store will be marked as a risky store.

[0042] In another embodiment, commodities with fraud risk are regarded as abnormal transaction commodities, and the preset condition is set as the proportion of fraud commodities is greater than the preset second ratio threshold. When the number of commodities with abnormal transaction in the store meets the corresponding preset condition, that is, the proportion of fraud commodities is greater than the preset second ratio threshold, the corresponding store is marked as a risky store. In another embodiment, commodities with prohibited or restricted sales risk and commodities with fraud risk are both regarded as abnormal transaction commodities. When the corresponding two preset conditions are met, that is, the proportion of prohibited or restricted sales commodities is greater than the preset ratio threshold and the proportion of fraud commodities is greater than the preset second ratio threshold, the corresponding store is marked as a risky store.

[0043] After marking the stores that need to be tested as risky stores, multiple risky stores are collected to form a risky store set. The construction of the risky store set can be based on the detection and marking of all stores on the platform, which can be used for comprehensive detection when the risky store cluster is determined later. This step uses a screening mechanism based on abnormal transaction products to screen out potential risky stores and form a risky store set, providing a data basis for subsequent risky store cluster detection.

[0044] Step S5200, calculating the ratio of the total number of commodities between two risky shops in the risky shop set, and when the ratio of the total number of commodities is within a preset ratio interval, marking the corresponding two risky shops as a suspected risky shop pair, and the suspected risky shop pair is used to form a suspected risky shop pair set;

[0045] For each pair of risky stores in the risky store set, calculate the ratio of the total number of goods in each pair of risky stores. Compare the calculated ratio of the total number of goods with the preset ratio interval. The preset ratio interval is pre-set according to platform rules and historical data, and is used to determine whether two risky stores have similarities in terms of product scale. When the ratio of the total number of goods is within the preset ratio interval, the corresponding two risky stores are marked as suspected risky store pairs. In this way, suspected risky store pairs with similarities in product scale can be screened out to form a suspected risky store pair set.

[0046] Risky stores in a risky store cluster are operated by the same entity or an associated entity, and the corresponding product quantities are often highly consistent. By screening store pairs whose total product ratios are within a preset range, risky stores that may belong to the same cluster can be effectively identified, thereby providing more accurate data support for subsequent risky cluster detection.

[0047] Step S5300: based on the quantity of the abnormally traded commodities corresponding to the suspected risky store pairs and a preset similarity determination rule, filter out some of the suspected risky store pairs from the set of suspected risky store pairs;

[0048] In one embodiment, based on each pair of shops (suspected risk shop pairs) in the set of suspected risk shop pairs, and based on the number of goods of the corresponding abnormal transaction goods (including the number of prohibited and restricted sales goods and the number of fraudulent goods) of the suspected risk shop pairs, the corresponding prohibited and restricted sales goods risk ratio and the fraudulent goods risk ratio are calculated. Then, based on the preset similarity judgment conditions, each pair of shops is screened in two stages. The first stage is based on the similarity judgment of prohibited and restricted sales goods, and the second stage is based on the similarity judgment of fraudulent goods. Among them, this application does not limit the order of the two stages, that is, the order of the two stages does not affect the embodiment of the creative spirit of this application.

[0049] In the first stage of similarity judgment, the difference in the risk ratio of banned and restricted goods of the two stores is first calculated, that is, the absolute value of the difference in the risk ratio of banned and restricted goods of the two stores. The difference is compared with the preset difference threshold. If the difference is less than the preset difference threshold, it is preliminarily determined that the two stores are similar in the risk ratio of banned and restricted goods. The banned and restricted goods categories of the two stores are further analyzed. Specifically, multiple first product categories with a large number of banned and restricted goods in the two stores are obtained respectively, and the overlap rate of the first product categories of the two stores is determined. The overlap rate is the number of the same categories in the first product categories of the two stores divided by the total number of first product categories of any of the two stores. If the overlap rate is greater than the preset overlap rate threshold, it is determined that the two stores are highly similar in the banned and restricted goods categories, which meets the preset similarity judgment rules. If the difference in the risk ratio of banned and restricted goods of the two stores and the category overlap rate both meet the preset conditions, the second stage of similarity judgment is entered; otherwise, the suspected risk store pair is filtered out from the suspected risk store pair set.

[0050] In the second stage of similarity judgment, further screening is performed based on the similarity of fraudulent goods. Obtain multiple second product categories with a large number of fraudulent goods in the two stores, and calculate the risk ratio of fraudulent goods. By analyzing the second product categories of the two stores and their fraudulent product risk ratios, determine whether they meet the preset similarity judgment rules. For example, you can set a difference threshold corresponding to the difference in category overlap rate and fraudulent product risk ratio. If the two stores meet the preset conditions in both the fraudulent product category and the risk ratio, they are considered to meet the preset similarity judgment rules; otherwise, the store pair is filtered out from the set of suspected risky store pairs. Please refer to the subsequent specific implementation methods for the specific steps, which will not be repeated here.

[0051] Through the above-mentioned double screening mechanism, suspected risky store pairs that do not meet the preset similarity judgment conditions can be effectively eliminated, and suspected risky store pairs that are highly similar in both banned and restricted goods and fraudulent goods can be retained. This refined screening process not only improves the accuracy of risky store cluster detection, but also provides a more reliable data basis for subsequent risky cluster determination. At the same time, the preset difference threshold, overlap rate threshold and other parameters can be adjusted according to actual business needs to adapt to the risk detection needs in different scenarios, further improving the flexibility and applicability of this application.

[0052] Step S5400: based on the filtered set of suspected risk store pairs, calculate the product similarity coverage rate of the suspected risk store pairs corresponding to the two risk stores, and determine the risky site group stores based on the corresponding product similarity coverage rates of all suspected risk store pairs in the set of suspected risk store pairs.

[0053] In one embodiment, for each pair of stores in the filtered set of suspected risk store pairs (suspected risk store pairs), the corresponding product titles are converted into vector representations to obtain product title vectors. The conversion of product title vectors can be achieved through natural language processing technology, for example, using a word embedding model (such as Word2Vec, BERT, etc.) to map the words in the product title into vectors, and obtain the vector representation of each product title through weighted averaging or pooling operations to obtain product title vectors. Then, based on the product title vectors, the cosine similarity between the product title vectors of the two stores is calculated to obtain a product similarity matrix. The cosine similarity is calculated as follows: the dot product of the two product title vectors divided by the corresponding module length product, with a value range of [-1,1]. The closer the value is to 1, the more similar the semantics of the two product titles are.

[0054] Based on the obtained product similarity matrix, the product similarity coverage of the two stores is determined. Specifically, for each store's product title vector, the maximum cosine similarity between the store's product title vector and another store's product title vector is determined in the product similarity matrix, and the maximum cosine similarity is compared with the preset similarity threshold. The number of products whose maximum cosine similarity is greater than the similarity threshold is counted, and the number of products is divided by the total number of products in any of the two stores to obtain the product similarity coverage of the two stores. The product similarity coverage reflects the degree of semantic similarity between the two stores in the product titles. The higher the coverage, the more likely it is that the products of the two stores belong to the same category or have the same operating characteristics.

[0055] In another embodiment, a risky station group store graph is constructed based on the commodity similarity coverage of all suspected risky store pairs. Each store is regarded as a node in the graph, and when the commodity similarity coverage of two stores is greater than a preset coverage threshold, an edge connection is established between the corresponding two stores. In this way, pairs of stores with high similarity are connected to form a risky station group store graph. Based on the risky station group store graph, a densely connected subgraph can be identified through a graph analysis algorithm (such as a community detection algorithm) to determine a potential set of risky station group stores. Through this graph construction method based on commodity similarity coverage, risky store clusters operated by the same entity or related entities can be effectively identified, providing strong support for risk management of e-commerce platforms.

[0056] In another embodiment, since a large number of stores need to be detected simultaneously, including a huge number of commodities, directly calculating the commodity vector representation in real time and performing commodity similarity coverage analysis between two stores will result in low computational efficiency. In this embodiment, all commodity titles are first preprocessed, and the corresponding commodity title vectors are obtained for offline storage. At the same time, an index is established for the commodity based on the store ID and commodity serial number of the commodity for quick retrieval. When performing risky store cluster detection, the commodity data is divided into historical commodity data and newly listed commodity data. For historical commodity data, the offline stored commodity title vector is directly obtained through the commodity index; for newly listed commodity data, the corresponding commodity title vector is obtained, and the commodity title vector corresponding to the newly listed commodity data is stored offline, and the corresponding index is updated at the same time. In this way, the time cost of real-time calculation is significantly reduced, providing efficient technical support for large-scale risky store cluster detection.

[0057] It can be known from the typical embodiments of the present application that the technical solution of the present application has many advantages, including but not limited to the following aspects:

[0058] On the one hand, this application comprehensively considers the correlation between multiple risky stores and can more effectively identify risky store clusters controlled by the same business entity. This not only improves the coverage of detection, but also can more accurately capture the behavior of merchants to diversify risks by operating multiple stores.

[0059] On the other hand, this application gradually narrows the scope of detection through a multi-stage screening mechanism, and finally determines the risky station group stores. Specifically, this application first screens out risky stores through preset conditions to form a risky store set, and then further screens out suspected risky store pairs to form a suspected risky store pair set based on the total number of goods in the risky store set. That is, by quantifying the relationship between the scale of goods between stores, it is possible to preliminarily identify the risky store station group that may be controlled by the same business entity. Then, based on the number of abnormal transaction goods and the preset similarity judgment rules, the suspected risky store pairs are filtered, and the suspected risky store pairs with stronger associations are screened out on the basis of the suspected risky store pair set, and the suspected risky store pairs that do not meet the preset similarity judgment rules are filtered out, reducing the false alarm rate and improving the accuracy of the detection. Finally, by calculating the similarity coverage of goods and determining the risky station group stores, the risky store station group controlled by the same business entity can be fully identified. This application gradually narrows the scope of detection through a multi-stage screening mechanism, and finally accurately identifies the risky store station group controlled by the same business entity, effectively reducing the false alarm rate, improving the accuracy and reliability of detection, and providing more efficient risky store station group detection for e-commerce platforms.

[0060] In addition, by storing the product title vector offline, the need for real-time computing can be greatly reduced, especially when processing large-scale product data, avoiding repeated computing and waste of resources. In addition, the introduction of offline storage and indexing mechanisms also enhances the scalability and flexibility of this application. As the number of products on the platform continues to increase, the accuracy and timeliness of risky store cluster detection can be ensured by regularly updating the offline stored product title vectors and indexes. This enables the technical solution of this application to adapt to the ever-changing data scale and business needs of e-commerce platforms, and provide continuous and reliable risk management support for the platform.

[0061] For further examples, see Figure 2 When the number of abnormally traded commodities in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set, including the following steps:

[0062] Step S5110: Obtain the product titles of all valid products in the store, where the product titles include the product main title and the product subtitle;

[0063] You can access the product information of the target store through the data interface or database of the e-commerce platform, and extract the product titles of all valid products. Valid products refer to products that are currently on sale, that is, products that have not been removed from the shelves or deleted. The product title consists of the main title and the subtitle. The main title is the core description information of the product, which is used to summarize the main features of the product, such as brand, model, function, etc.; the subtitle is a supplement to the main title, including promotional information, specification parameters or other detailed descriptions. When extracting product titles, you need to ensure that both the main title and the subtitle are obtained to ensure the completeness and accuracy of the product information. For some products, the subtitle may be empty, in which case only the main title is used as the product title.

[0064] In one embodiment, the title and subtitle of the product are spliced ​​together to form a new text. Taking cross-border e-commerce as an example, when the product title is in English, all texts are lowercase to eliminate the impact of uppercase and lowercase differences on text analysis. Then, the text can be segmented by BPE (Byte Pair Encoding), and the segmentation results can be mapped into a numerical sequence through a predefined vocabulary. Each integer in the numerical sequence corresponds to a word in the original text, thereby converting the text into a numerical form that can be processed by the subsequent model. Finally, the processed text is input into a preset text pre-training model to extract text features and represent them with vectors. The text pre-training model adopts a contrastive learning framework to make similar texts closer in the vector space. The vector obtained by the text pre-training model is L2 normalized so that the sum of the squares of the values ​​of all positions is 1, thereby ensuring the consistency and comparability of the vector in subsequent calculations.

[0065] Step S5120: input the product title into a pre-trained sales restriction risk detection model to determine whether the valid product corresponding to the product title has a sales restriction risk. The valid product with a sales restriction risk is characterized as a product that is restricted or prohibited from sale by the e-commerce platform where the store is located;

[0066] The obtained product title (including the main title and subtitle of the product) is input into the pre-trained prohibited and restricted sales risk detection model. In one embodiment, the prohibited and restricted sales risk detection model is constructed based on natural language processing technology and machine learning algorithms, and can perform semantic analysis and prohibited and restricted sales risk classification on the product title. The training data of the prohibited and restricted sales risk detection model includes a large number of labeled product title samples, where the positive samples are product titles that are restricted or prohibited from sale by the platform, and the negative samples are compliant product titles. Through training, the prohibited and restricted sales risk detection model can learn the key features and semantic patterns of prohibited and restricted sales product titles, such as vocabulary and descriptions involving contraband, illegal and illegal products, dangerous goods, etc. The product title obtained in the previous step is input into the prohibited and restricted sales risk detection model, and the prohibited and restricted sales risk detection model outputs a risk probability value, indicating the possibility that the product title has a prohibited and restricted sales risk. According to the preset risk threshold, the risk probability value is compared with the risk threshold. If the risk probability value is greater than the risk threshold, it is determined that the valid product corresponding to the product title has a prohibited and restricted sales risk, and a prohibited and restricted sales risk label is added to the corresponding valid product. Among them, the risk threshold is set by technicians in this field as needed. Through this step, the prohibited or restricted sales commodities that may exist in the store can be identified efficiently and accurately.

[0067] Step S5130: Count the number of commodities with abnormal transactions that are at risk of being banned or restricted in sale in the store, and obtain the proportion of banned or restricted commodities in the store based on the number of commodities with abnormal transactions and the total number of valid commodities in the store;

[0068] In one embodiment, all valid commodities in the store are traversed to check whether the valid commodities are marked as having a risk of prohibited or restricted sales. By querying the prohibited or restricted sales risk label, the number of commodities with abnormal transactions in the store that have a risk of prohibited or restricted sales is counted. Based on the number of prohibited or restricted sales commodities obtained by statistics and the total number of valid commodities, the proportion of prohibited or restricted sales commodities is calculated. (The number of abnormal transactions with a risk of prohibited or restricted sales / the total number of valid commodities in the store) × 100% is the proportion of prohibited or restricted sales commodities. The proportion of prohibited or restricted sales commodities reflects the relative proportion of prohibited or restricted sales commodities in the store, and is an important indicator for assessing the risk level of the store.

[0069] Step S5140: When the proportion of the prohibited or restricted sales commodities is greater than a preset ratio threshold, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set.

[0070] The calculated proportion of prohibited or restricted sales commodities is compared with the preset proportion threshold. The preset proportion threshold is set by technicians in this field according to the rules, historical data and / or risk management requirements of the e-commerce platform, and is used to determine whether a store has a higher risk of prohibited or restricted sales. For example, assuming that the preset proportion threshold is 3%, if the proportion of prohibited or restricted sales commodities in a store is 5%, it exceeds the preset proportion threshold, indicating that the store has a higher risk of prohibited or restricted sales, and the store is marked as a risky store. In one embodiment, a risky store label is added to the store whose proportion of prohibited or restricted sales commodities is greater than the preset proportion threshold. Traverse all stores on the platform, filter out all stores whose proportion of prohibited or restricted sales commodities is greater than the preset threshold, and form a set of risky stores.

[0071] In this embodiment, by taking restricted or prohibited goods as abnormal transaction goods, the risk level of the corresponding store is reflected according to the proportion of restricted or prohibited goods, and the risk level is quantified to determine whether the corresponding store is a risky store, and finally multiple risky stores are collected to form a risky store set. This risk assessment mechanism based on the proportion of restricted or prohibited goods can accurately identify potential high-risk stores.

[0072] For further examples, see Figure 3 When the number of abnormally traded commodities in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set, including the following steps:

[0073] Step S5150: Obtain the commodity prices of all valid commodities in the store and a preset standard price library;

[0074] In one embodiment, the product information of the store is accessed through the data interface or database of the e-commerce platform, and the product prices of all valid products are extracted. The product price includes one or more price information such as the selling price, promotional price, discount price, etc. of the product. And obtain a preset standard price library, which contains the standard price range of various products on the platform, and is set based on historical sales data, market conditions and platform rules. The construction of the standard price library can be achieved through data mining and statistical analysis technology to ensure that it can reflect the reasonable price range of similar products on the market. Through this step, it is ensured that the identification process of risky stores has a reliable price basis.

[0075] Step S5160: using a fraud risk detection model to determine whether the commodity price of each valid commodity is falsely low according to a preset standard price library; if the commodity price is falsely low, confirming the valid commodity corresponding to the commodity price as having a fraud risk, and marking the valid commodity as an abnormally traded commodity;

[0076] In one embodiment, based on the commodity title corresponding to the commodity price, the standard price range of the corresponding commodity is retrieved from the standard price library, and the commodity price and its corresponding standard price range are input into a pre-trained fraud risk detection model. The fraud risk detection model can determine the degree of deviation between the commodity price and the standard price range, and obtain the degree of deviation by comparing with the lowest price in the standard price range, and determine whether the commodity price of each valid commodity is falsely low in combination with the preset risk threshold, so as to determine whether the commodity has a fraud risk. For example, if the commodity price of a valid commodity is lower than the risk threshold (e.g., 30%) of the standard price of the corresponding commodity in the standard price library, the commodity is marked as having a fraud risk, and the valid commodity is marked as an abnormal transaction commodity. Merchants usually put commodities on the shelves at extremely low prices to attract consumers, but the physical objects received by consumers after purchase are often different from the descriptions on the e-commerce platform. It should be noted that this embodiment is usually applied to standard products, that is, commodities with clear standards and rules, and the standard price range is related to the corresponding commodity brand and specifications. For example, due to the different sizes, the corresponding prices of a 64-inch TV and a 32-inch TV will be significantly different.

[0077] Through this embodiment, the fraud risk model will be tested, and a fraud risk label will be added to the corresponding valid product with fraud risk. At the same time, the valid product will be marked as an abnormal transaction product. The fraud risk label is used to count the number of abnormal transaction products with fraud risk in the store.

[0078] Step S5170: Count the number of products with abnormal transactions that have fraud risks in the store, determine whether the corresponding store is a risky store based on the number of products with abnormal transactions and the preset conditions, and collect multiple risky stores to form a risky store set.

[0079] In one embodiment, all valid goods in the store are traversed to check whether all valid goods are marked as having fraud risks. By querying the fraud risk label, the number of goods with abnormal transactions that have fraud risks in the store is counted. Then the total number of valid goods in the store is obtained, and the proportion of fraudulent goods is calculated based on the number of fraud risk goods and the total number of valid goods obtained by statistics. The proportion of fraudulent goods reflects the relative proportion of fraudulent goods in the store. The calculated proportion of fraudulent goods is compared with a preset proportion threshold. The preset proportion threshold is set according to the rules of the e-commerce platform and historical data, and is used to determine whether the store has a high fraud risk. If the proportion of fraudulent goods is greater than the preset proportion threshold, the store is marked as a risky store, and all stores whose proportion of fraudulent goods is greater than the preset threshold are screened out to form a risky store set.

[0080] In this embodiment, by combining the standard price library and the fraud risk detection model, it is possible to efficiently identify commodities with falsely low prices as commodities with fraud risks, treat commodities with fraud risks as commodities with abnormal transactions, reflect the risk level of the corresponding store based on the proportion of fraudulent commodities, and quantify the risk level to determine whether the corresponding store is a risky store, and finally collect multiple risky stores to form a risky store set. This risk assessment mechanism based on the proportion of fraudulent commodities can accurately identify potential high-risk stores.

[0081] For further examples, see Figure 4 , based on the quantity of the abnormally traded commodities corresponding to the suspected risky store pairs and a preset similarity determination rule, filtering out some of the suspected risky store pairs from the suspected risky store pair set, comprising the following steps:

[0082] Step S5310: Based on the suspected risky store and the corresponding two risky stores, obtain the prohibited or restricted sale commodity categories of the two risky stores, including a plurality of first commodity categories with a large number of commodities, and the prohibited or restricted sale commodity risk ratios;

[0083] In one embodiment, for two risky stores in a suspected risky store pair, all valid goods in the risky stores are traversed respectively to check whether the goods are marked as having a risk of prohibited or restricted sales. By querying the prohibited or restricted sales risk label, the category distribution of prohibited or restricted sales goods in each store is counted, that is, the number of prohibited or restricted sales goods under each category. Specifically, the product title, description text or other relevant information is analyzed by natural language processing technology to identify the category to which the product belongs, and the identified category is matched with the preset prohibited or restricted sales category list. The prohibited or restricted sales category list is constructed based on platform rules, laws and regulations, and historical data, and contains commodity categories that are explicitly prohibited or restricted from sale, such as "aquatic products", "agricultural products", "wine and beverages", "electronic products", "cosmetics", etc. in the cross-border e-commerce field. If the commodity category matches a category in the prohibited or restricted sales category list, the corresponding prohibited or restricted sales category label is added to the commodity. For example, if a commodity belongs to the "aquatic products" category, a "prohibited or restricted sales-aquatic products" label is added to the corresponding commodity. It should be noted that the list of prohibited and restricted sales categories can be dynamically updated according to changes in platform rules, laws and regulations, and actual business needs. When the list is updated, rescan the products in the store to check whether the corresponding product categories match the new list, and update the corresponding prohibited and restricted sales category tags.

[0084] Sort the categories according to the number of banned or restricted goods, and select multiple first product categories with a large number of banned or restricted goods. The number of first product categories can be set according to actual needs, for example, the top three categories with the largest number of banned or restricted goods are selected. The first product category reflects the main risk distribution of the store on banned or restricted goods, and is the basis for judging the similarity of stores. And obtain the risk ratio of banned or restricted goods. The risk ratio of banned or restricted goods reflects the concentration of risks of the store on banned or restricted goods, and is a key indicator for assessing the risk level of the store. For example, risk store A has a total of 1,000 valid products, of which 100 products are marked as banned or restricted goods, that is, the risk ratio of banned or restricted goods in the risk store is 10%, and among these 100 banned or restricted goods, by statistically analyzing the distribution of banned or restricted goods categories, multiple first product categories with a large number of banned or restricted goods are selected. For example, assuming that among the 100 banned or restricted goods, there are 40 goods in the "aquatic products" category, 30 goods in the "agricultural products" category, 20 goods in the "alcoholic beverages" category, 5 goods in the "electronic products" category, and 5 goods in the "cosmetics" category, then sort by the number of banned or restricted goods, and select the top three first commodity categories as "aquatic products", "agricultural products" and "alcoholic beverages". This step can obtain the data on the banned or restricted goods categories and risk ratios of the two risky stores in the suspected risky store pair, providing a comprehensive and accurate basis for subsequent similarity judgment.

[0085] Step S5320: Based on the first commodity category and the risk ratio of the prohibited or restricted commodities, determine whether the suspected risky store pair meets a preset first similarity judgment condition; if not, filter the corresponding suspected risky store pair from the suspected risky store pair set;

[0086] In one embodiment, assuming that risky store A has a total of 1,000 valid commodities, of which 100 commodities are marked as prohibited or restricted commodities, the corresponding prohibited or restricted commodity risk ratio is 10%; risky store B has a total of 800 valid commodities, of which 80 commodities are marked as prohibited or restricted commodities, and the prohibited or restricted commodity risk ratio is 10%. Calculate the difference in the prohibited or restricted commodity risk ratios of the two stores, i.e., |10%-10%|=0%, and compare the difference with a preset first difference threshold (e.g., 5%). If the difference is less than the preset first difference threshold, it is preliminarily determined that the two stores have similarities in the prohibited or restricted commodity risk ratios.

[0087] When the difference is less than the preset first difference threshold, the prohibited or restricted sale categories of the two stores are further analyzed. Obtain multiple first commodity categories with a large number of prohibited or restricted sale commodities in the two stores. For example, the prohibited or restricted sale commodities of risk store A are mainly distributed in the three categories of "aquatic products" (40), "agricultural products" (30) and "alcohol and beverages" (20); the prohibited or restricted sale commodities of risk store B are mainly distributed in the three categories of "aquatic products" (35), "agricultural products" (25) and "alcohol and beverages" (15). Calculate the overlap rate of the two first commodity categories, and the overlap rate is calculated by dividing the number of the same categories in the first commodity categories of the two stores by the total number of the first commodity categories of one of the stores. In this example, the first commodity categories of the two stores are "aquatic products", "agricultural products" and "alcohol and beverages", the number of the same categories is 3, and the total number of the first commodity categories of one of the stores is 3, so the overlap rate is 3 / 3×100%=100%. The overlap rate is compared with a preset first overlap rate threshold (for example, 80%). If the overlap rate is greater than the preset first overlap rate threshold, it is determined that the two stores are highly similar in the categories of prohibited or restricted goods. That is, when the difference in the risk ratios of prohibited or restricted goods of the two stores is less than the preset first difference threshold and the overlap rate of the corresponding first product category is greater than the preset first overlap rate threshold, it is determined to meet the first similarity judgment condition.

[0088] If the difference in the risk ratio of banned or restricted goods and the category overlap rate of the two stores meet the preset conditions, the suspected risk store pair will be retained; otherwise, the suspected risk store pair will be filtered out from the suspected risk store pair set. Through the above steps, it is possible to effectively screen out suspected risk store pairs that are highly similar in the risk ratio of banned or restricted goods and category distribution.

[0089] Step S5330: When the suspected risky store meets the preset first similarity judgment condition, based on the suspected risky store for the corresponding two risky stores, the fraudulent product categories of the two risky stores are obtained, including a plurality of second product categories with a large number of products and a fraudulent product risk ratio, wherein the fraudulent product risk ratio is the ratio of the number of products with abnormal transactions having a fraudulent risk in the corresponding store to the total number of valid products in the store;

[0090] For the two risky stores in the suspected risky store pair that meet the first similarity judgment condition, the fraud risk labels of all valid products in the risky stores are extracted respectively. The fraud risk labels are generated by the pre-trained fraud risk detection model. According to the fraud risk labels, the number of fraudulent products in each store is counted and classified and summarized according to the product category. For example, the fraudulent products of risky store A are mainly distributed in the three categories of "luxury goods" (30), "electronic products" (20) and "health products" (10); the fraudulent products of risky store B are mainly distributed in the three categories of "luxury goods" (25), "electronic products" (15) and "virtual goods" (10). Multiple second product categories with a large number of fraudulent products are screened out, for example, the top three categories are selected as the second product categories. At the same time, the fraudulent product risk ratio of each store is calculated, that is, the ratio of the number of abnormal transaction products with fraudulent risks in the corresponding store to the total number of valid products in the store.

[0091] Step S5340: Based on the second commodity category and the risk ratio of the fraudulent commodity, determine whether the suspected risk store pair meets the preset second similarity judgment condition; if not, filter the corresponding suspected risk store pair from the suspected risk store pair set.

[0092] In one embodiment, for two risky stores in a suspected risky store pair, the difference in the risk ratio of fraudulent goods in the two risky stores is calculated, that is, the absolute value of the difference in the risk ratio of fraudulent goods in the two stores, and the difference is compared with the preset second difference threshold. For example, the risk ratio of fraudulent goods in risky store A is 5%, the risk ratio of fraudulent goods in risky store B is 6%, and the difference is 1%. If the preset difference threshold is 2%, the difference is less than the preset second difference threshold, and it is preliminarily determined that the two stores have similarities in the risk ratio of fraudulent goods. Then, the distribution of fraudulent goods categories in the two stores is further analyzed, and multiple second commodity categories with a large number of fraudulent goods are obtained, and the overlap rate of the second commodity categories is calculated. For example, the second commodity categories of risky store A are "luxury goods", "electronic products" and "health products", and the second commodity categories of risky store B are "luxury goods", "electronic products" and "virtual goods". The number of the same categories is 2, and the total number of second commodity categories in one store is 3, then the overlap rate is 2 / 3≈66.7%. Compare the overlap rate with the preset second overlap rate threshold (for example, 60%). If the overlap rate is greater than the preset second overlap rate threshold, it is determined that the two stores are highly similar in the fraudulent product category and meet the second similarity judgment condition. If the difference in the risk ratio of fraudulent products and the category overlap rate of the two stores meet the preset conditions, the suspected risk store pair is retained; otherwise, the suspected risk store pair is filtered out from the suspected risk store pair set. For example, if the risk ratio of fraudulent products of risk store A is 5%, and the risk ratio of fraudulent products of risk store B is 8%, the difference is 3%, which is greater than the preset difference threshold, or the category overlap rate is 50%, which is less than the preset overlap rate threshold, then the suspected risk store pair is filtered out from the suspected risk store pair set. Through the above steps, suspected risk store pairs with high similarity in the risk ratio of fraudulent products and category distribution can be effectively screened out.

[0093] In this embodiment, through multi-dimensional similarity judgment based on banned or restricted goods and fraudulent goods, store pairs with highly similar risk characteristics are further screened out from the set of suspected risky store pairs, significantly improving the accuracy and reliability of risky store cluster detection. This multi-level similarity judgment mechanism not only improves the accuracy of screening, but also avoids the misjudgment that may be caused by single-dimensional judgment. By filtering out suspected risky store pairs that do not meet the similarity conditions, stores that may belong to the same risky station group can be accurately identified, providing more reliable data support for subsequent risky station group detection. This embodiment significantly improves the comprehensiveness and robustness of risky store cluster detection through multi-dimensional and multi-level similarity judgments.

[0094] For further examples, see Figure 5, based on the first commodity category and the risk ratio of the prohibited or restricted commodities, determining whether the suspected risky store pair meets a preset first similarity judgment condition includes the following steps:

[0095] Step S5321: Calculate the difference in risk ratio of banned or restricted goods sold by the suspected risky store to the corresponding two risky stores;

[0096] For the two risky stores in the suspected risky store pair, calculate the difference in the risk ratio of banned or restricted goods of the two stores, that is, the absolute value of the difference in the risk ratio of banned or restricted goods of the two stores. This difference reflects the degree of difference in the risk ratio of banned or restricted goods of the two stores, and is a key indicator for judging whether the two stores are similar. By calculating the difference, a quantitative basis can be provided for subsequent similarity judgments, thereby screening out store pairs with consistent risk characteristics of banned or restricted goods.

[0097] Step S5322: when the difference is less than a preset difference threshold, obtain the number of identical first commodity categories of the two risk stores, divide the number by the total number of first commodity categories corresponding to any of the two risk stores, and obtain the overlap rate of the first commodity categories of the two risk stores;

[0098] Determine the first commodity categories corresponding to the two risky stores respectively, and the first commodity categories are obtained by counting multiple categories with a large number of prohibited or restricted commodities in each store. When the difference is less than the preset difference threshold (i.e., the first difference threshold), compare the first commodity categories of the two risky stores to find the same number of categories. Then, divide the number of the same categories by the total number of first commodity categories of any of the two risky stores. Obtain the overlap rate of the first commodity categories of the two risky stores, which reflects the degree of similarity between the two stores in the prohibited or restricted commodity categories. The higher the overlap rate, the more similar the distribution of the two stores in the prohibited or restricted commodity categories, thereby providing a quantitative basis for subsequent similarity judgments.

[0099] Step S5323: When the overlap rate is greater than a preset overlap rate threshold, it is determined that the pair of suspected risky stores meets a preset first similarity judgment condition.

[0100] The preset overlap rate threshold is pre-set based on actual business needs and historical data, and is used to measure whether the similarity between two stores in the categories of banned or restricted goods meets acceptable standards. For example, assuming that the preset overlap rate threshold is 60%, if the calculated overlap rate of the first product category of the two risky stores is 66.7%, then the overlap rate is greater than the preset overlap rate threshold, indicating that the two stores are highly similar in the distribution of banned or restricted goods categories. At this point, it is determined that the suspected risk store pair meets the preset first similarity judgment condition, that is, the two stores meet the similarity requirements in the risk ratio and category distribution of banned or restricted goods, and thus the store pair is retained for subsequent risk site group detection and analysis.

[0101] In this embodiment, by calculating the difference in the risk ratio of banned or restricted goods and the overlap rate of the first commodity category of the two stores in the suspected risk store pair, and performing similarity judgment based on the preset difference threshold and overlap rate threshold, the store pairs with similar risk characteristics are retained, and the store pairs that do not meet the conditions are filtered out, which significantly improves the accuracy and reliability of the screening of suspected risk store pairs. This similarity judgment mechanism based on quantitative indicators not only improves the accuracy and objectivity of the screening, but also avoids the deviation that may be caused by subjective judgment.

[0102] For further examples, see Figure 6 , based on the filtered set of suspected risky store pairs, calculating the commodity similarity coverage rate of the suspected risky store pairs of the corresponding two risky stores, and determining the risky station group stores based on the corresponding commodity similarity coverage rates of all suspected risky store pairs in the set of suspected risky store pairs, including the following steps:

[0103] Step S5410: based on the suspected risky store pairs in the filtered set of suspected risky store pairs, convert all product titles in the two risky stores corresponding to the suspected risky store pairs into vector representations to obtain product title vectors;

[0104] For the suspected risk shop, all the product titles in the corresponding two risk shops are extracted. In one embodiment, the extracted product titles are subjected to text preprocessing, including text cleaning, word segmentation, and removal of stop words. After the preprocessing is completed, the processed product titles are converted into vector representations using a pre-trained natural language processing model (such as Word2Vec, BERT, etc.). For each product title, each word in the title is mapped to a high-dimensional vector through word embedding technology. Then, these word vectors are aggregated into a product title vector of a fixed dimension through weighted averaging or pooling operations (such as mean pooling or maximum pooling). Finally, the generated product title vector is subjected to L2 normalization processing so that the modulus length of all vectors is 1, ensuring consistency in the subsequent calculation of cosine similarity. Through the above steps, all product titles of each store are converted into vector representations, providing a high-quality vector data foundation for subsequent product similarity calculations and risk station group detection.

[0105] Step S5420, calculating the cosine similarity between the product title vectors of the two risky stores to obtain a product similarity matrix of the risky stores;

[0106] For each pair of product title vectors in the two risky stores corresponding to the suspected risky store pair, calculate the cosine similarity between the two risky stores. The formula for calculating cosine similarity is the dot product of the two vectors divided by the corresponding module length product. The value range is [-1,1]. The closer the value is to 1, the more similar the semantics of the two product titles are. Specifically, for the i-th product title vector of store A and the j-th product title vector of store B, calculate the cosine similarity and store the result in the i-th row and j-th column of the similarity matrix. By traversing all pairs of product title vectors of the two stores, a product similarity matrix is ​​finally generated, in which each element represents the semantic similarity of the corresponding product titles in the two stores.

[0107] Step S5430: Based on the product similarity matrix, determine the product similarity coverage rate of the suspected risky store to the corresponding two risky stores;

[0108] In one embodiment, for each row in the commodity similarity matrix (i.e., each commodity title vector of store A), the maximum cosine similarity value between all commodity title vectors of store A and store B is found, and the maximum value is compared with a preset similarity threshold. If the maximum cosine similarity value is greater than the similarity threshold, it is considered that the commodity of store A is highly similar to a certain commodity in store B in semantics, and is counted in the number of similar commodities. Then, the number of all similar commodities that meet the conditions in store A is counted, and the number of similar commodities is divided by the total number of commodities in store A to obtain the commodity similarity coverage rate of store A relative to store B. Similarly, for each column in the commodity similarity matrix (i.e., each commodity title vector of store B), the maximum cosine similarity value between all commodity title vectors of store B and store A is found, and the same comparison and statistics are performed to obtain the commodity similarity coverage rate of store B relative to store A. Finally, the commodity similarity coverage rates of the two stores are comprehensively calculated (such as taking the average or weighted average) to obtain the commodity similarity coverage rate of the suspected risk store to the corresponding two risk stores. The product similarity coverage rate reflects the overall similarity of the product titles of the two stores in terms of semantics, and provides an important quantitative basis for subsequent risk site cluster detection.

[0109] Step S5440: mark each store in the set of suspected risky store pairs as a node in the risky site group store graph; when the product similarity coverage is greater than a preset coverage threshold, establish an edge connection between the corresponding two stores to form a risky site group store graph for determining the risky site group to which the store belongs.

[0110] All the stores in the set of suspected risk store pairs are taken as nodes to construct an initial undirected graph, where each node represents a risky store. Traverse all suspected risky store pairs, obtain the corresponding product similarity coverage for each pair of stores, and compare the product similarity coverage with the preset coverage threshold. If the product similarity coverage is greater than the preset coverage threshold, an undirected edge is established between the corresponding store nodes, indicating that the two stores have a high degree of similarity in product semantics and may belong to the same risky station group. By traversing all store pairs and completing the edge connection, a risky station group store graph is finally constructed. The nodes in the graph represent risky stores, and the edges represent the semantic similarity of products between stores. Through graph analysis algorithms (such as community detection algorithms), the densely connected subgraphs in the graph can be further identified to determine the potential risky station group store set.

[0111] In this embodiment, by converting the product titles of risky stores into vector representations, and calculating the product similarity coverage between stores based on the product title vectors to construct a risky site group store map, a screening mechanism based on product similarity coverage can effectively identify suspected risky store pairs that are highly similar in product operation characteristics, thereby more accurately locating risky store sites that may be operated by the same entity or related entities.

[0112] For further examples, see Figure 7 , based on the commodity similarity matrix, determining the commodity similarity coverage rate of the suspected risky store to the corresponding two risky stores, including:

[0113] Step S5431, determining the maximum cosine similarity value of each row in the product similarity matrix, and comparing the maximum cosine similarity value with a preset similarity threshold;

[0114] In one embodiment, for each row in the product similarity matrix (i.e., each product title vector of store A), all matrix elements in the row are traversed to determine the maximum cosine similarity value between all product title vectors of store A and store B. The maximum cosine similarity value represents the highest degree of semantic similarity between the current product of store A and a product in store B. The maximum cosine similarity value found is then compared with a preset similarity threshold. The preset similarity threshold is pre-set based on actual business needs and historical data, and is used to measure whether the two product titles have a sufficiently high semantic similarity. If the maximum cosine similarity value is greater than the preset similarity threshold, it is considered that the product of store A is highly similar to a product in store B in terms of semantics, and is included in the subsequent statistics of the number of similar products; otherwise, it is considered that the product does not have a sufficiently high semantic similarity with the product in store B, and is not counted.

[0115] Step S5432, count the number of times the maximum cosine similarity value is greater than the similarity threshold, divide the number by the total number of goods in any of the two risky stores, and obtain the similarity coverage rate of goods of the suspected risky store to the corresponding two risky stores.

[0116] In one embodiment, through the comparison in the previous step, the number of stores A with the maximum cosine similarity value relative to store B greater than the similarity threshold is counted, and the number is divided by the total number of goods in store A to obtain the corresponding product similarity coverage. Similarly, for each column in the product similarity matrix, the number of times the corresponding maximum cosine similarity value is greater than the preset similarity threshold is counted, that is, the number of goods in store B with highly similar product titles in semantics to those in store A, and the number is divided by the total number of goods in store B to obtain the product similarity coverage of store B relative to store A. Finally, the product similarity coverage of the two stores is comprehensively calculated (such as taking the average or weighted average) to obtain the product similarity coverage of the suspected risk store to the corresponding two risk stores.

[0117] In this embodiment, by calculating the maximum cosine similarity of each row in the product similarity matrix and comparing it with the preset similarity threshold, it is possible to accurately identify pairs of products with highly similar product titles in two stores, and count the number of products with the maximum cosine similarity greater than the similarity threshold. The product similarity coverage is calculated based on the statistical results, and the overall similarity of the product semantics of the two stores can be quantified. Since risky station cluster stores are often operated by the same entity or related entities, the corresponding product selection strategies, operating models and target customer groups are highly consistent, that is, the products on the shelves show significant convergence in categories, titles, descriptions and functions. This embodiment quantifies the similarity of product titles between two stores through product similarity coverage, thereby determining the correlation between the two stores and more accurately locating potential risky station clusters.

[0118] See also Figure 8, a risky store cluster detection device provided to meet one of the purposes of the present application is a functional embodiment of the risky store cluster detection method of the present application. On the other hand, the device is a risky store cluster detection device provided to meet one of the purposes of the present application, including a risky store marking module 5100, a risky store pair composition module 5200, a risky store pair filtering module 5300 and a risky store cluster determination module 5400, wherein the risky store marking module 5100 is used to mark the store as a risky store when the number of abnormally traded goods in the store meets the preset conditions, and collect multiple risky stores to form a risky store set; the risky store pair composition module 5200 is used to calculate the ratio of the total number of goods in each of the risky stores in the risky store set, When the ratio of the total number of commodities is within a preset ratio interval, the corresponding two risky stores are marked as a suspected risky store pair, and the suspected risky store pair is used to constitute a suspected risky store pair set; a risky store pair filtering module 5300 is used to filter out some of the suspected risky store pairs from the suspected risky store pair set based on the quantity of the commodities of the abnormally traded commodities corresponding to the suspected risky store pairs and a preset similarity determination rule; a risky store station group determination module 5400 is used to calculate the commodity similarity coverage rate of the two risky stores corresponding to the suspected risky store pair based on the filtered suspected risky store pair set, and determine the risky station group stores based on the corresponding commodity similarity coverage rates of all suspected risk store pairs in the suspected risky store pair set.

[0119] In a further embodiment, the risky store marking module 5100 includes: a product title acquisition submodule, which is used to obtain the product titles of all valid products in the store, and the product titles include the product main title and the product subtitle; a prohibited or restricted sale risk determination submodule, which is used to input the product title into a pre-trained prohibited or restricted sale risk detection model to determine whether the valid product corresponding to the product title has a prohibited or restricted sale risk, and the valid product with a prohibited or restricted sale risk is characterized as a product that is restricted or prohibited from sale by the e-commerce platform where the store is located; a prohibited or restricted sale product ratio calculation submodule, which is used to count the number of abnormal transaction products with prohibited or restricted sale risks in the store, and obtain the prohibited or restricted sale product ratio of the store based on the number of abnormal transaction products and the total number of valid products in the store; a risky store marking submodule, which is used to mark the store as a risky store when the proportion of prohibited or restricted sales products is greater than a preset ratio threshold, and collect multiple risky stores to form a risky store set.

[0120] In a further embodiment, the risky store marking module 5100 includes: a product price acquisition submodule, which is used to obtain the product prices of all valid products in the store and a preset standard price library; a fraud risk determination submodule, which is used to use a fraud risk detection model to determine whether the product price of each of the valid products is falsely low according to the preset standard price library. When the product price is falsely low, the valid product corresponding to the product price is confirmed to have a fraud risk, and the valid product is marked as an abnormal transaction product; a risky store determination submodule, which is used to count the number of abnormal transaction products with fraud risks in the store, and determine whether the corresponding store is a risky store based on the number of abnormal transaction products and the preset conditions, and collect multiple risky stores to form a risky store set.

[0121] In a further embodiment, the risky store pair filtering module 5300 includes: a first commodity category acquisition submodule, which is used to obtain, based on the two risky stores corresponding to the suspected risky store pair, the prohibited and restricted sale commodity categories of the two risky stores, including multiple first commodity categories with a large number of commodities and the prohibited and restricted sale commodity risk ratios; a first risky store pair filtering submodule, which is used to determine whether the suspected risky store pair meets the preset first similarity judgment condition based on the first commodity category and the prohibited and restricted sale commodity risk ratios, and if not, filter the corresponding suspected risky store pair from the suspected risky store pair set; a second commodity category acquisition submodule, which is used to filter the suspected risky store pair when the suspected risky store pair meets the preset first similarity judgment condition. When a store meets the preset first similarity judgment condition, based on the corresponding two risk stores of the suspected risk store pair, the fraudulent product categories of the two risk stores are obtained, including multiple second product categories with a large number of products and the fraudulent product risk ratio, wherein the fraudulent product risk ratio is the ratio of the number of products with abnormal transactions with fraud risks in the corresponding store to the total number of valid products in the store; the second risk store pair filtering submodule is used to determine whether the suspected risk store pair meets the preset second similarity judgment condition based on the second product category and the fraudulent product risk ratio, and if not, filter the corresponding suspected risk store pair from the suspected risk store pair set.

[0122] In a further embodiment, the first risk store pair filtering submodule includes: a proportion difference calculation submodule, used to calculate the difference in risk proportion of banned or restricted goods of the suspected risk store to the corresponding two risk stores; an overlap rate calculation submodule, used to obtain the number of the same first commodity categories of the two risk stores when the difference is less than a preset difference threshold, and divide the number by the total number of first commodity categories corresponding to any of the two risk stores to obtain the overlap rate of the first commodity categories of the two risk stores; a condition compliance judgment submodule, used to judge whether the suspected risk store pair meets the preset first similarity judgment condition when the overlap rate is greater than the preset overlap rate threshold.

[0123] In a further embodiment, the risky store cluster determination module 5400 includes: a product title vector conversion submodule, which is used to convert all product titles in the two risk stores corresponding to the suspected risk store pair into vector representations based on the suspected risk store pairs in the filtered set of suspected risk store pairs, so as to obtain product title vectors; a product similarity matrix determination submodule, which is used to calculate the cosine similarity between the product title vectors of the two risk stores, so as to obtain the product similarity matrix of the risk stores; a product similarity coverage rate determination submodule, which is used to determine the product similarity coverage rate of the two risk stores corresponding to the suspected risk store pair based on the product similarity matrix; a risk cluster store graph determination submodule, which is used to mark each store in the set of suspected risk store pairs as a node in the risk cluster store graph, and when the product similarity coverage rate is greater than a preset coverage rate threshold, an edge connection is established between the corresponding two stores to form a risk cluster store graph for determining the risk cluster to which the store belongs.

[0124] In a further embodiment, the product similarity coverage determination submodule includes: a cosine similarity comparison submodule, used to determine the maximum cosine similarity of each row in the product similarity matrix, and compare the maximum cosine similarity with a preset similarity threshold; a product similarity coverage calculation submodule, used to count the number of the maximum cosine similarities greater than the similarity threshold, and divide the number by the total number of products in any of the two risk stores to obtain the product similarity coverage of the suspected risk store to the corresponding two risk stores.

[0125] In order to solve the above technical problems, the present application also provides a computer device. Fig. 9As shown, a schematic diagram of the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. Among them, the computer-readable storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The database may store a control information sequence. When the computer-readable instructions are executed by the processor, the processor may implement a risky store cluster detection method. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor may execute the risky store cluster detection method of the present application. The network interface of the computer device is used to connect and communicate with a terminal. Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0126] In this embodiment, the processor is used to execute Figure 8 The memory stores the program code and various data required to execute the above modules or submodules. The network interface is used to transmit data between user terminals or servers. The memory in this embodiment stores the program code and data required to execute all modules / submodules in the risk store cluster detection device of this application, and the server can call the program code and data of the server to execute the functions of all submodules.

[0127] The present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the risky store cluster detection method of any embodiment of the present application.

[0128] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the present application can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0129] It will be understood by those skilled in the art that the various operations, methods, steps, measures, and schemes in the processes discussed in this application may be alternated, changed, combined, or deleted. Furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application may also be alternated, changed, rearranged, decomposed, combined, or deleted. Furthermore, the steps, measures, and schemes in the various operations, methods, and processes in the prior art that are open source and in this application may also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0130] The above description is only a partial implementation method of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for detecting risky store clusters, characterized in that: The steps include: When the number of abnormally traded goods in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set; Calculate the ratio of the total number of goods of each pair of risky shops in the risky shop set, and when the ratio of the total number of goods is within a preset ratio interval, mark the corresponding two risky shops as a suspected risky shop pair, and the suspected risky shop pair is used to form a suspected risky shop pair set; Based on the quantity of the abnormally traded commodities corresponding to the suspected risky store pairs and a preset similarity determination rule, filtering out some of the suspected risky store pairs from the set of suspected risky store pairs; Based on the filtered set of suspected risk store pairs, the product similarity coverage rate of the suspected risk store pairs corresponding to the two risk stores is calculated, and based on the corresponding product similarity coverage rates of all suspected risk store pairs in the set of suspected risk store pairs, the risky site group stores are determined.

2. The risky store site group detection method according to claim 1 is characterized in that: When the number of abnormally traded commodities in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set, including: Get the product titles of all valid products in the store, including the product title and product subtitle; Input the product title into a pre-trained sales restriction risk detection model to determine whether the valid product corresponding to the product title has a sales restriction risk. The valid product with a sales restriction risk is characterized as a product that is restricted or prohibited from sale by the e-commerce platform where the store is located; Counting the number of products with abnormal transactions that are at risk of being banned or restricted in sale in the store, and obtaining a percentage of banned or restricted products in the store based on the number of products with abnormal transactions and the total number of valid products in the store; When the proportion of the banned or restricted sales commodities is greater than a preset ratio threshold, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set.

3. The risky store cluster detection method according to claim 1 is characterized in that: When the number of abnormally traded commodities in a store meets a preset condition, the store is marked as a risky store, and multiple risky stores are collected to form a risky store set, including: Get the product prices of all valid products in the store and the preset standard price library; A fraud risk detection model is used to determine whether the price of each valid commodity is artificially low according to a preset standard price library. If the commodity price is artificially low, the valid commodity corresponding to the commodity price is confirmed to have a fraud risk, and the valid commodity is marked as an abnormal transaction commodity; The number of commodities with abnormal transactions and fraud risks in the store is counted, and based on the number of commodities with abnormal transactions and the preset conditions, whether the corresponding store is a risky store is determined, and multiple risky stores are collected to form a risky store set.

4. The risky store cluster detection method according to claim 1 is characterized in that: Based on the quantity of the abnormally traded commodities corresponding to the suspected risky store pairs and a preset similarity determination rule, filtering out some of the suspected risky store pairs from the set of suspected risky store pairs includes: Based on the suspected risky store and the corresponding two risky stores, obtaining the prohibited or restricted sale commodity categories of the two risky stores, including a plurality of first commodity categories with a large number of commodities and the prohibited or restricted sale commodity risk ratios; Based on the first commodity category and the risk ratio of the prohibited or restricted commodities, determining whether the suspected risky store pair meets a preset first similarity judgment condition; if not, filtering the corresponding suspected risky store pair from the suspected risky store pair set; When the suspected risky store meets the preset first similarity judgment condition, based on the suspected risky store for the corresponding two risky stores, the fraudulent product categories of the two risky stores are obtained, including multiple second product categories with a large number of products and the fraudulent product risk ratio, where the fraudulent product risk ratio is the ratio of the number of products with abnormal transactions and fraudulent risks in the corresponding store to the total number of valid products in the store; Based on the second commodity category and the risk ratio of the fraudulent commodities, determine whether the suspected risky store pair meets a preset second similarity judgment condition; if not, filter the corresponding suspected risky store pair from the suspected risky store pair set.

5. The risky store cluster detection method according to claim 4 is characterized in that: Based on the first commodity category and the risk ratio of the prohibited or restricted commodities, determining whether the suspected risky store pair meets a preset first similarity judgment condition includes: Calculate the difference in risk ratio of banned or restricted goods sold by the suspected risky store to the corresponding two risky stores; When the difference is less than a preset difference threshold, the number of identical first commodity categories of the two risk stores is obtained, and the number is divided by the total number of first commodity categories corresponding to any of the two risk stores to obtain the overlap rate of the first commodity categories of the two risk stores; When the overlap rate is greater than a preset overlap rate threshold, it is determined that the pair of suspected risky stores meets a preset first similarity judgment condition.

6. The risky store cluster detection method according to claim 1 is characterized in that: Based on the filtered set of suspected risky store pairs, calculating the commodity similarity coverage rate of the suspected risky store pairs of corresponding two risky stores, and determining the risky station group stores based on the corresponding commodity similarity coverage rates of all suspected risky store pairs in the set of suspected risky store pairs, including: Based on the suspected risky store pairs in the filtered set of suspected risky store pairs, convert all product titles in two risky stores corresponding to the suspected risky store pairs into vector representations to obtain product title vectors; Calculate the cosine similarity between the product title vectors of the two risky stores to obtain a product similarity matrix of the risky stores; Based on the product similarity matrix, determining the product similarity coverage rate of the suspected risky store to the corresponding two risky stores; Each store in the set of suspected risky store pairs is marked as a node in the risky site group store graph. When the product similarity coverage is greater than a preset coverage threshold, an edge connection is established between the corresponding two stores to form a risky site group store graph for determining the risky site group to which the store belongs.

7. The risky store cluster detection method according to claim 6 is characterized in that: Based on the product similarity matrix, determining the product similarity coverage rate of the suspected risky store to the corresponding two risky stores includes: Determine the maximum value of the cosine similarity of each row in the product similarity matrix, and compare the maximum value of the cosine similarity with a preset similarity threshold; The number of times the maximum cosine similarity value is greater than the similarity threshold is counted, and the number is divided by the total number of commodities in any of the two risky stores to obtain the commodity similarity coverage rate of the suspected risky store to the corresponding two risky stores.

8. A risky store cluster detection device, characterized in that: include: A risky store marking module is used to mark a store as a risky store when the number of abnormally traded goods in the store meets a preset condition, and collect multiple risky stores to form a risky store set; A risky store pair formation module is used to calculate the ratio of the total number of goods of each of the risky stores in the risky store set. When the total number of goods ratio is within a preset ratio interval, the corresponding two risky stores are marked as suspected risky store pairs, and the suspected risky store pairs are used to form a suspected risky store pair set; A risky store pair filtering module, configured to filter out some of the suspected risky store pairs from the suspected risky store pair set based on the quantity of the abnormally traded goods corresponding to the suspected risky store pairs and a preset similarity determination rule; The risky store cluster determination module is used to calculate the product similarity coverage rate of the corresponding two risk stores of the suspected risk store pair based on the filtered set of suspected risk store pairs, and determine the risky store cluster based on the corresponding product similarity coverage rate of all suspected risk store pairs in the set of suspected risk store pairs.

9. A computer device comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: It stores a computer program implemented according to the method described in any one of claims 1 to 7 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

Images