Shopping clue research and judgment system and early warning method based on artificial intelligence

Abstract

The application relates to the technical field of shopping research and judgment, in particular to a shopping clue research and judgment system and a prewarning method based on artificial intelligence, the prewarning method comprising the following steps: constructing a multi-channel shopping data set by acquiring shopping data of an e-commerce platform and generating compliance shopping data; performing semantic matching on the compliance shopping data and a dangerous goods list database, identifying dangerous goods purchase records, pre-processing related orders, and forming standardized order data; combining an association rule mining algorithm, identifying a dangerous goods raw material combination purchase mode, and generating combination feature data; performing time series analysis on a time interval and a purchase frequency of combination purchase, evaluating scattered purchase behavior, and obtaining commodity correlation analysis results; comprehensively considering a commodity risk grade, purchase behavior abnormal features and historical transaction records, calculating a risk score, comparing the risk score with a preset risk threshold, and generating a risk prewarning result; and the application identifies potential dangerous goods combination purchase behavior.

Description

Technical Field

[0001] This invention relates to the field of shopping analysis technology, specifically to an artificial intelligence-based shopping clue analysis system and early warning method. Background Technology

[0002] With the rapid development of e-commerce platforms and online transactions, commodity transactions are characterized by diversified channels, high transaction frequency, and dispersed transaction entities. While e-commerce platforms greatly facilitate people's lives, they also provide new avenues for the illegal acquisition of some regulated commodities, especially potentially hazardous chemicals, explosive raw materials, and controlled implements. Individuals may circumvent platform rules and manual supervision by purchasing hazardous raw materials at different times, through different channels, and across different commodity categories, thereby posing a public safety hazard.

[0003] Existing e-commerce risk control or regulatory technologies mainly focus on identifying violations of single products or rule matching based on keywords, assessing the risk of a single transaction or a single type of product. These technologies typically rely on static rules or simple threshold judgments, making it difficult to identify potential dangerous product combination purchase behaviors from multiple transactions, multiple product combinations, and time dimensions, especially difficult to detect highly concealed risk patterns such as "diversified purchases" and "combined purchases."

[0004] Therefore, it is necessary to design an AI-based shopping clue analysis system and early warning method to address the problems existing in current technologies. Summary of the Invention

[0005] In view of this, the present invention proposes an artificial intelligence-based shopping clue analysis system and early warning method, which aims to solve the problem of difficulty in identifying potential dangerous product combination purchase behaviors from multiple transactions, multiple product combinations and time dimensions.

[0006] This invention proposes an early warning method for a shopping clue analysis system based on artificial intelligence, comprising:

[0007] Acquire shopping data from e-commerce platforms to generate a multi-channel shopping dataset; process the purchase privacy information in the multi-channel shopping dataset using encryption algorithms and de-identification technologies to generate compliant shopping data;

[0008] Based on the compliant shopping data and the dangerous goods list database, semantic matching is performed to identify dangerous goods purchase records, generate dangerous goods identification results, and based on the dangerous goods identification results, several order data are preprocessed to generate standardized order data;

[0009] Based on the standardized order data, the support and confidence scores between products are calculated using an association rule mining algorithm to generate association rule data. This identifies the combined purchasing patterns of hazardous material raw materials and generates combined feature data. Based on the combined feature data, the purchase time intervals and frequencies of hazardous material raw material combinations are analyzed to generate time series analysis results. The dispersed purchasing behavior of hazardous material raw material combinations is evaluated, generating dispersed purchase evaluation data. Based on the dispersed purchase evaluation data and the combined feature data, a comprehensive judgment is made on the combined purchasing behavior of hazardous material raw materials, generating product association analysis results.

[0010] Based on the product association analysis results, calculate the product risk level index and generate product risk data; based on the product association analysis results, analyze abnormal purchase behavior characteristics and generate abnormal behavior data; based on the product risk data and abnormal behavior data, and combined with historical purchase transaction records, calculate dimensional risk scores; based on the dimensional risk scores, and combined with a weighted fusion algorithm, calculate and generate risk score data; based on the risk score data, compare it with a preset risk threshold to generate a risk warning result.

[0011] Furthermore, when generating compliant shopping data by processing purchase privacy information in the multi-channel shopping dataset using encryption algorithms and de-identification techniques, the following steps are included:

[0012] Based on the multi-channel shopping dataset, extract product name, specifications, transaction time, transaction amount, and user identification information;

[0013] The user identification information is encrypted using the AES encryption algorithm and desensitized using hash desensitization technology.

[0014] Based on the results of encryption and de-identification processing, compliant shopping data is generated.

[0015] Furthermore, when generating standardized order data, the following steps are included:

[0016] Based on the compliant shopping data, extract the text information of the product name to generate product name data; based on the product name data, perform word segmentation and keyword extraction to generate keyword data; based on the keyword data, perform precise matching with the built-in dangerous goods list database; when precise matching fails, use the cosine similarity algorithm for fuzzy semantic matching; based on the results of precise matching or fuzzy semantic matching, identify dangerous goods, mark dangerous goods purchase records, and generate dangerous goods identification results;

[0017] Based on the hazardous materials identification results, hazardous materials order records are extracted to generate hazardous materials order data; based on the hazardous materials order data, the format of product names is standardized to generate standardized name data; product specifications are converted into standardized units of measurement to generate standardized specification data; transaction times are converted into a standardized time format to generate standardized time data.

[0018] Standardized order data is generated by combining the standardized name data, standardized specification data, and standardized time data.

[0019] Furthermore, when generating combined feature data, the following are included:

[0020] Based on the standardized order data, a hazardous materials purchase transaction database is constructed; based on the hazardous materials purchase transaction database, frequent itemsets are mined using the Apriori algorithm; based on the frequent itemsets, the support of product combinations is calculated to generate support data; based on the frequent itemsets, the confidence of product combinations is calculated to generate confidence data; based on the support data and confidence data, a general minimum support threshold and a general minimum confidence threshold are set; based on the general minimum support threshold and the general minimum confidence threshold, preliminary effective association rules are filtered to generate preliminary association rule data.

[0021] Based on the preliminary association rule data and combined with professional knowledge in the field of hazardous materials regulation, a professional support threshold and a professional confidence threshold are set.

[0022] Based on the professional support threshold and professional confidence threshold, high-risk hazardous materials combinations are screened; based on the high-risk hazardous materials combinations, a preset combination pattern is matched to identify hazardous materials raw material combinations.

[0023] Analyze the risk level relationship of the commodities based on the combination of the aforementioned hazardous materials raw materials;

[0024] Based on the aforementioned risk level relationship, the degree of risk of the combination is assessed, and combination characteristic data is generated.

[0025] Furthermore, when generating decentralized purchase assessment data, the following are included:

[0026] Based on the combined feature data, the purchase time sequence of each commodity in the combination of hazardous raw materials by the same user is extracted;

[0027] Based on the purchase time series, calculate the average purchase time interval between products within the bundle; calculate the purchase frequency of products within the bundle to generate purchase frequency data; analyze the dispersed purchase behavior pattern based on the average purchase time interval and purchase frequency data; and match it with a preset abnormal dispersed purchase behavior feature library.

[0028] Based on the matching results, assess the risk level of diversified purchasing behavior and generate diversified purchasing assessment data.

[0029] Furthermore, when generating product association analysis results, the following are included:

[0030] Based on the dispersed purchase assessment data, a dispersed purchase risk score is extracted; based on the combined characteristic data, a hazard assessment data for the combination of hazardous raw materials is extracted; based on the dispersed purchase risk score and hazard assessment data, a two-dimensional risk assessment matrix is ​​constructed, risk assessment rules are set, and a comprehensive risk index for combined purchase behavior is calculated.

[0031] Based on the comprehensive risk index, the results are compared with the preset risk index threshold to generate product association analysis results.

[0032] Furthermore, when calculating the dimensional risk score, the following are included:

[0033] Based on the results of the commodity association analysis, the dangerous goods control level is extracted, the commodity risk level index is calculated, and commodity risk data is generated.

[0034] Based on the product association analysis results, the degree of abnormality in purchase quantity, purchase frequency, and purchase channel is analyzed; based on the degree of abnormality, behavioral abnormality characteristic indicators are calculated, and behavioral abnormality data is generated.

[0035] Based on the product risk data and abnormal behavior data, combined with purchase history transaction records, a dimensional risk score is calculated.

[0036] Furthermore, when calculating and generating risk score data based on the aforementioned dimensional risk scores and using a weighted fusion algorithm, the process includes:

[0037] Based on the risk scores of the aforementioned dimensions, determine the weight coefficients of the product risk sub-model, the behavior anomaly sub-model, and the user history sub-model;

[0038] Based on the weighting coefficients, a weighted average algorithm is applied to calculate the comprehensive risk score; and linear normalization is performed to generate standardized risk score data.

[0039] Furthermore, when generating a risk warning result by comparing the risk score data with a preset risk threshold, the process includes:

[0040] The preset risk threshold includes a first preset risk threshold and a second preset risk threshold, wherein the first preset risk threshold is greater than the second preset risk threshold;

[0041] The risk warning results include high-risk warning results, medium-risk warning results, and low-risk warning results;

[0042] The risk score data is compared with the first preset risk threshold and the second preset risk threshold respectively;

[0043] When the risk score data is greater than or equal to the first preset risk threshold, a high-risk warning result is generated;

[0044] When the risk score data is less than the first preset risk threshold and greater than or equal to the second preset risk threshold, a medium-risk warning result is generated;

[0045] When the risk score data is less than the second preset risk threshold, a low-risk warning result is generated.

[0046] Compared with existing technologies, the beneficial effects of this invention are as follows: By combining encryption algorithms and de-identification techniques on user identification information in multi-channel shopping data, key transaction feature information required for risk assessment is retained while ensuring user privacy and data compliance, thus solving the problem of difficulty in in-depth data analysis due to privacy restrictions. By segmenting product names and extracting keywords, combined with precise matching and fuzzy semantic matching mechanisms from a hazardous materials database, it is possible to identify hazardous materials purchase records under different expressions and naming methods, avoiding the omission problem caused by relying solely on keyword rules, and improving the accuracy and robustness of hazardous materials identification. Based on association rule mining algorithms, frequent itemset analysis is performed on standardized order data, and association rules are further filtered using professional knowledge in the field of hazardous materials supervision. This enables the discovery of potentially dangerous raw material combination purchase patterns from a large number of dispersed orders, overcoming the difficulty in detecting cross-product and cross-order combination risks. By introducing time series analysis, the purchase time interval and frequency of hazardous material raw material combinations are comprehensively analyzed and matched with an abnormal dispersed purchase behavior feature database, identifying deliberately dispersed purchasing and regulatory evasion behaviors, thus improving the ability to detect concealed high-risk behaviors. This system integrates the risk level of the product itself, the hazard level of the product combination, and abnormal purchasing behavior characteristics to construct a multi-dimensional risk scoring system. A weighted fusion algorithm is then used to generate a comprehensive risk score, making risk assessment results more comprehensive and objective, and reducing false positives and false negatives caused by single-dimensional judgments. By comparing the risk score data with multi-level preset risk thresholds, different levels of early warning results (high risk, medium risk, and low risk) can be output, providing a tiered response basis for regulatory authorities or platform risk control systems.

[0047] On the other hand, this application also provides an artificial intelligence-based shopping clue analysis system, for use in applying the aforementioned artificial intelligence-based shopping clue analysis system's early warning method, including:

[0048] The data acquisition unit is configured to acquire shopping data from e-commerce platforms and generate a multi-channel shopping dataset; based on the multi-channel shopping dataset, it processes purchase privacy information through encryption algorithms and de-identification technologies to generate compliant shopping data.

[0049] The preprocessing unit is configured to perform semantic matching between the compliant shopping data and the dangerous goods list database, identify dangerous goods purchase records, generate dangerous goods identification results, and preprocess several order data based on the dangerous goods identification results to generate standardized order data.

[0050] The analysis unit is configured to calculate the support and confidence of products based on the standardized order data and an association rule mining algorithm, generate association rule data, identify the combined purchase pattern of hazardous raw materials, and generate combination feature data; based on the combination feature data, analyze the purchase time interval and purchase frequency of hazardous raw material combinations, generate time series analysis results, evaluate the dispersed purchase behavior of hazardous raw material combinations, and generate dispersed purchase evaluation data; based on the dispersed purchase evaluation data and the combination feature data, comprehensively judge the combined purchase behavior of hazardous raw materials, and generate product association analysis results.

[0051] The early warning unit is configured to: calculate a product risk level index and generate product risk data based on the product association analysis results; analyze abnormal purchase behavior characteristics and generate abnormal behavior data based on the product association analysis results; calculate a dimensional risk score based on the product risk data and abnormal behavior data, combined with historical purchase transaction records; calculate and generate risk score data based on the dimensional risk score and a weighted fusion algorithm; and compare the risk score data with a preset risk threshold to generate a risk warning result.

[0052] It is understandable that the aforementioned AI-based shopping clue analysis system and early warning method have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0053] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0054] Figure 1 A flowchart illustrating the early warning method of an artificial intelligence-based shopping clue analysis system provided in this embodiment of the invention;

[0055] Figure 2 This is a functional block diagram of an artificial intelligence-based shopping clue analysis system provided in an embodiment of the present invention. Detailed Implementation

[0056] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] For this, please refer to Figure 1 As shown, this application proposes an early warning method for a shopping clue analysis system based on artificial intelligence, including:

[0058] S100: Acquire shopping data from e-commerce platforms to generate multi-channel shopping datasets; process purchase privacy information in the multi-channel shopping datasets using encryption algorithms and de-identification technologies to generate compliant shopping data;

[0059] S200: Based on the compliant shopping data and the dangerous goods list database, semantic matching is performed to identify dangerous goods purchase records, generate dangerous goods identification results, and based on the dangerous goods identification results, several order data are preprocessed to generate standardized order data;

[0060] S300: Based on standardized order data, and combined with association rule mining algorithms, calculate the support and confidence of products to generate association rule data, identify the combined purchase patterns of hazardous raw materials, and generate combined feature data; based on the combined feature data, analyze the purchase time interval and purchase frequency of hazardous raw material combinations to generate time series analysis results, evaluate the dispersed purchase behavior of hazardous raw material combinations, and generate dispersed purchase evaluation data; based on the dispersed purchase evaluation data and combined feature data, comprehensively judge the combined purchase behavior of hazardous raw materials, and generate product association analysis results.

[0061] S400: Based on the product association analysis results, calculate the product risk level index and generate product risk data; based on the product association analysis results, analyze abnormal purchase behavior characteristics and generate abnormal behavior data; based on the product risk data and abnormal behavior data, combined with historical purchase transaction records, calculate the dimensional risk score; based on the dimensional risk score, combine with a weighted fusion algorithm to calculate and generate risk score data; based on the risk score data, compare it with a preset risk threshold to generate a risk warning result.

[0062] Specifically, step S100 constructs a compliant shopping dataset by acquiring shopping data such as product names, transaction times, and transaction amounts from major e-commerce platforms, forming a multi-channel shopping dataset. Advanced encryption algorithms and data anonymization techniques are then used to process user privacy information, ensuring data usage complies with the Personal Information Protection Law and the Data Security Law, generating compliant shopping data that retains analytical value while meeting regulatory requirements. In step S200, the compliant shopping data undergoes multi-level semantic matching with a hazardous materials database. This involves not only precise matching but also fuzzy semantic matching using natural language processing technology to identify hazardous materials purchase records that use aliases, abbreviations, or descriptive language to evade regulation, generating hazardous materials identification results. Related order data is then standardized to lay the foundation for subsequent analysis. Step S300 uses association rule mining algorithms to analyze the purchase relationships between products, identifying combined purchase patterns of hazardous materials raw materials. Simultaneously, time series analysis is used to assess dispersed purchasing behavior, comprehensively judging the risk level of combined hazardous materials raw material purchases. The S400 process involves a multi-dimensional risk assessment, calculating risk scores from two perspectives: product risk and behavioral anomalies. A weighted fusion algorithm is used to generate the final risk score, which is then compared with a preset threshold to generate a tiered early warning result.

[0063] The working process and principle of this application are as follows: Raw data such as product names, specifications, transaction times, transaction amounts, and user identifiers are obtained from mainstream e-commerce platforms through secure API interfaces or compliant data exchange mechanisms to form a multi-channel shopping dataset. To ensure user privacy and security, AES encryption is used to encrypt user identifier information, and hash desensitization technology is applied for secondary processing to generate compliant shopping data that meets regulatory requirements. A hazardous materials identification process is initiated, semantically matching the product names in the compliant shopping data with a hazardous materials directory database. An exact match is attempted, and products directly containing hazardous material keywords are immediately identified. For cases using aliases, abbreviations, or descriptive language, a cosine similarity algorithm is used for fuzzy semantic matching. By calculating the similarity between the product description and the hazardous material feature vector, potential hazardous material transactions are identified. After identifying hazardous material purchase records, related orders are standardized, unifying product name descriptions, specifications, and time formats to eliminate data heterogeneity. In the product association analysis stage, a hazardous material purchase transaction database is constructed, and the Apriori algorithm is applied to mine frequent itemsets, calculating the support and confidence of product combinations. Support reflects the frequency of a specific product combination in all transactions, while confidence represents the probability that a user who buys product A also buys product B. A general minimum support threshold (typically 0.5%) and a general minimum confidence threshold (typically 60%) are set for initial screening. Then, combining expertise in hazardous materials regulation, a more stringent professional threshold (0.1% support, 80% confidence) is set to screen out high-risk hazardous product combinations. For example, the combination of nitric acid and glycerin may suggest the preparation of nitroglycerin, and the combination of potassium permanganate and concentrated sulfuric acid may be used to manufacture explosives. The purchase time intervals and frequencies of these combinations are also analyzed to identify dispersed purchasing behavior—that is, users purchasing hazardous material raw materials in batches at different times and on different platforms to circumvent regulations. By constructing a two-dimensional risk assessment matrix, the risk level of the combination and the risk of dispersed purchasing behavior are comprehensively assessed to generate product association analysis results. Risk scores are calculated from two dimensions: product risk and behavioral anomaly. The product risk dimension considers the control level of hazardous materials and the degree of combination risk; the behavioral anomaly dimension analyzes the degree of anomaly in purchase quantity, frequency, and channels. By combining users' historical transaction records, a comprehensive risk score is generated through a weighted fusion algorithm, and then compared with preset thresholds to generate tiered warnings. When the risk score exceeds the first preset threshold (e.g., 0.85), a high-risk warning is generated, indicating a possible high-risk activity involving the preparation of hazardous materials; when it falls between the first and second thresholds (e.g., 0.65), a medium-risk warning is generated, indicating suspicious activity requiring attention; and when it falls below the second threshold, it is considered low-risk, indicating a normal transaction.

[0064] As a preferred embodiment, the specific implementation of this application is as follows: First, 120 million shopping records for the current month are obtained from the platform, and after AES encryption and hash de-identification processing, compliant shopping data is generated. In the hazardous materials identification stage, it is found that a user purchased nitric acid on May 3rd, glycerol (99%) on May 8th, concentrated sulfuric acid on May 12th, and anhydrous ethanol on May 15th. Through semantic matching, it is confirmed that these goods correspond to nitric acid, glycerol, sulfuric acid, and ethanol in the hazardous chemicals list, respectively. In the association analysis stage, a transaction database is constructed, and the Apriori algorithm is applied to calculate the association of product combinations. Analysis shows that the support for the combination of nitric acid and glycerol is 0.08%, with a confidence level of 85%; the support for the combination of sulfuric acid and ethanol is 0.12%, with a confidence level of 78%. Based on professional threshold screening, both of these combinations are identified as high-risk combinations. Further analysis of the purchase time intervals revealed that the purchase interval for nitric acid and glycerin was 5 days, and for sulfuric acid and ethanol it was 3 days, with an average interval of 4 days. This is significantly lower than the normal consumption interval (usually over 30 days), and the purchase frequency is clearly abnormal (2-3 times per month). This highly matches the pre-set abnormal scattered purchase behavior feature database, resulting in a scattered purchase risk score of 0.92. A two-dimensional risk assessment matrix was constructed, with the nitric acid-glycerin combination assessed as 0.85 (high risk) and the sulfuric acid-ethanol combination as 0.78 (medium-high risk), resulting in a comprehensive risk index of 0.83. In the risk scoring stage, the product risk dimension score was 0.87, and the abnormal behavior dimension score was 0.89. Combining this with the user's historical transaction records (the user had no previous records of purchasing dangerous goods), a weighted fusion algorithm was used to calculate a comprehensive risk score of 0.85. Since this score equals the first pre-set risk threshold of 0.85, a high-risk warning result was generated.

[0065] Through the above approach, this application organically combines technologies such as data compliance processing, semantic matching, association rule mining, and time series analysis to construct a complete system for judging dangerous goods purchasing behavior, thereby improving the accuracy of dangerous goods identification; multi-dimensional risk assessment has identified hidden behaviors such as dispersed purchasing and combined purchasing.

[0066] This application further proposes processing purchase privacy information in multi-channel shopping datasets using encryption algorithms and de-identification techniques to generate compliant shopping data, including:

[0067] Based on the multi-channel shopping dataset, extract product name, specifications, transaction time, transaction amount, and user identification information;

[0068] The user identification information is encrypted using the AES encryption algorithm and desensitized using hash desensitization technology.

[0069] Based on the results of encryption and de-identification processing, compliant shopping data is generated.

[0070] Specifically, key information elements are extracted from multi-channel shopping datasets, including basic data fields such as product name, specifications, transaction time, transaction amount, and user identification information. A dual protection mechanism is implemented for user identification information: firstly, the AES-256 encryption algorithm is used to encrypt the user identification information. This algorithm uses a 256-bit key length, providing high-strength data encryption protection to ensure that even if the data is leaked, it cannot be easily decrypted; secondly, hash desensitization technology is applied, using the SHA-256 hash algorithm to convert the user identification information into a fixed-length hash value and adding a random salt value to enhance security, making the original user identification impossible to reverse engineer.

[0071] Through the above technical solutions, the two processing methods in this application complement each other. Encryption preserves data traceability, while anonymization ensures data privacy and security during the analysis process. The processed data is then reintegrated to generate compliant shopping data that retains the key information needed for analysis while meeting data security regulations.

[0072] This application further proposes methods for generating standardized order data, including:

[0073] Based on compliant shopping data, extract product name text information to generate product name data; based on product name data, perform word segmentation and keyword extraction to generate keyword data; based on keyword data, perform precise matching with the built-in hazardous materials list database; when precise matching fails, use cosine similarity algorithm for fuzzy semantic matching; based on the results of precise matching or fuzzy semantic matching, identify hazardous materials, mark hazardous material purchase records, and generate hazardous material identification results.

[0074] Based on the hazardous materials identification results, hazardous materials order records are extracted to generate hazardous materials order data; based on the hazardous materials order data, the format of commodity names is standardized to generate standardized name data; commodity specifications are converted into standardized units of measurement to generate standardized specification data; and transaction times are converted into a standardized time format to generate standardized time data.

[0075] Standardized order data is generated by combining standardized name data, standardized specification data, and standardized time data.

[0076] Specifically, the process involves extracting product name text information from compliant shopping data to form raw product name data. Natural language processing is then applied to this data, using Chinese word segmentation technology to divide long texts into meaningful word units. A keyword extraction algorithm is then used to identify the most representative product feature words. These extracted keywords are precisely matched against a built-in hazardous materials database: when a product name directly contains the standard name of a hazardous material (such as nitric acid or potassium permanganate), it is immediately identified as a hazardous material. For cases using aliases, abbreviations, or descriptive language (e.g., battery water may refer to sulfuric acid, hydrogen peroxide to hydrogen peroxide), a cosine similarity algorithm is used for fuzzy semantic matching. This converts the product description into a feature vector in a vector space model, and the similarity between this vector and the hazardous material feature vector is calculated. When the similarity exceeds a preset threshold (usually set to 0.75), the product is determined to be a potential hazardous material. After identifying dangerous goods purchase records, the dangerous goods order data is standardized: the product names are standardized (e.g., battery water and battery fluid are standardized as sulfuric acid), the specifications are converted into standard units of measurement (e.g., 500ml and 0.5L are standardized as 500 milliliters), and the transaction time is converted into the ISO 8601 standard format (e.g., 2023-05-12T14:30:00+08:00).

[0077] Through the above technical solutions, this application eliminates data heterogeneity and improves the comprehensiveness and accuracy of hazardous materials identification by implementing these standardization processes.

[0078] This application further proposes methods for generating combined feature data, including:

[0079] Based on standardized order data, a hazardous materials purchase transaction database is constructed. Using this database, the Apriori algorithm is employed to mine frequent itemsets. Based on these frequent itemsets, the support of product combinations is calculated, generating support data. Based on the frequent itemsets, the confidence of product combinations is calculated, generating confidence data. Based on the support and confidence data, a general minimum support threshold and a general minimum confidence threshold are set. Based on these thresholds, preliminary effective association rules are selected, generating preliminary association rule data.

[0080] Based on preliminary association rule data and combined with professional knowledge in the field of hazardous materials regulation, professional support threshold and professional confidence threshold are set.

[0081] Based on the professional support threshold and the professional confidence threshold, high-risk hazardous materials combinations are screened; based on the high-risk hazardous materials combinations, preset combination patterns are matched to identify hazardous materials raw material combinations.

[0082] Analyze the risk level relationships of commodities based on the combination of hazardous raw materials;

[0083] Based on the risk level relationship, assess the risk level of the portfolio and generate portfolio characteristic data.

[0084] Specifically, a hazardous materials purchase transaction database is constructed based on standardized order data. Each user's hazardous materials purchase record is considered a transaction, containing all hazardous materials purchased by the user. The Apriori algorithm is applied to mine frequent itemsets. This algorithm iteratively expands from single itemets to combinations of multiple itemsets, retaining only itemsets with support exceeding a threshold in each iteration. Support represents the frequency of a specific item combination appearing in all transactions, calculated as: Support = (Number of transactions containing this item combination) / (Total number of transactions). Confidence represents the probability that a user who purchases item A also purchases item B, calculated as: Confidence = Support (A∪B) / Support (A). A general minimum support threshold (typically 0.5%) and a general minimum confidence threshold (typically 60%) are set for initial screening to generate preliminary association rule data. Then, combining expertise in hazardous materials regulation, stricter professional thresholds are set: a professional support threshold (typically set to 0.1%) and a professional confidence threshold (set to 80%), to filter out truly high-risk hazardous material combinations. For example, the combination of nitric acid and glycerin may suggest the preparation of nitroglycerin, while the combination of potassium permanganate and concentrated sulfuric acid may be used to manufacture explosives. High-risk combinations are screened and matched against a pre-defined hazardous material combination pattern library to identify hazardous raw material combinations. For identified combinations, the risk level relationship between the commodities is analyzed; for example, some chemicals have low risk when used alone, but the risk increases significantly when used in combination. Based on the hazard level, legal and regulatory control level, and potential hazards of the combination, the overall hazard level of the combination is assessed, generating combination characteristic data.

[0085] Through the above technical solution, this application has achieved systematic identification of dangerous goods raw material combination purchasing patterns, and improved the detection rate of dangerous goods combinations.

[0086] This application further proposes methods for generating decentralized purchase assessment data, including:

[0087] Based on the combined feature data, extract the purchase time series of each commodity in the combination of hazardous raw materials by the same user;

[0088] Based on the purchase time series, calculate the average purchase time interval between products within the bundle; calculate the purchase frequency of products within the bundle to generate purchase frequency data; analyze the scattered purchase behavior pattern based on the average purchase time interval and purchase frequency data; and match it with a pre-set abnormal scattered purchase behavior feature library.

[0089] Based on the matching results, assess the risk level of diversified purchasing behavior and generate diversified purchasing assessment data.

[0090] Specifically, based on the combined feature data, the complete purchase time series of each item within the hazardous materials raw material combination by the same user is extracted, including the timestamp of each purchase, item type, and quantity. The average purchase time interval between items within the combination is calculated using the formula: Average Time Interval = (Last Purchase Time - First Purchase Time) / (Number of Purchases - 1). The purchase frequency of items within the combination is calculated, generating purchase frequency data, including the number of purchases per unit time and the average quantity per purchase. These indicators are analyzed to identify dispersed purchasing behavior patterns: normal consumers typically purchase the required quantity at once, while users attempting to circumvent regulations will disperse the purchases of the same combination across different times and platforms. The analysis results are matched against a pre-set abnormal dispersed purchasing behavior feature library, which stores known dispersed purchasing behavior patterns (such as purchasing in batches at intervals of 3-7 days, and each purchase quantity being close to but not exceeding the regulatory threshold). The risk level of dispersed purchasing behavior is assessed based on the degree of matching. For example, when the purchase interval is 4-6 days, each purchase quantity is 80%-95% of the regulatory threshold, and purchases are dispersed across more than 3 platforms, it is determined to be high-risk dispersed purchasing behavior. A risk score is assigned to each diversified purchasing pattern, ranging from 0 to 1, with higher values ​​indicating higher risk. This process ultimately generates diversified purchasing assessment data.

[0091] Through the above technical solution, this application identifies the decentralized purchasing behavior of hazardous raw materials and improves the detection rate of hazardous transactions that are highly concealed.

[0092] This application further proposes that when generating product association analysis results, the following should be included:

[0093] Based on the decentralized purchase assessment data, a decentralized purchase risk score is extracted; based on the combination characteristic data, a hazard assessment data for the combination of hazardous raw materials is extracted; based on the decentralized purchase risk score and hazard assessment data, a two-dimensional risk assessment matrix is ​​constructed, risk assessment rules are set, and a comprehensive risk index for combination purchase behavior is calculated.

[0094] Based on the comprehensive risk index, the results of commodity correlation analysis are generated by comparing it with the preset risk index threshold.

[0095] Specifically, a diversified purchase risk score is extracted from the diversified purchase assessment data. This score reflects the degree of abnormality in the user's diversified purchase behavior, ranging from 0 to 1, with higher values ​​indicating higher risk. Risk assessment data for the combination of hazardous raw materials is extracted from the combination feature data. This data reflects the inherent risk of the combination itself, also ranging from 0 to 1. A two-dimensional risk assessment matrix is ​​constructed, with the horizontal axis representing the diversified purchase risk score and the vertical axis representing the combination risk level, dividing the risk space into four regions: low-risk zone (both low), medium-risk zone (one high), and high-risk zone (both high). Risk assessment rules are set: when both the diversified purchase risk score and the combination risk level are higher than 0.7, it is considered high-risk; when one indicator is higher than 0.7 and the other is higher than 0.5, it is considered medium-risk; otherwise, it is considered low-risk. A comprehensive risk index for the combination purchase behavior is calculated based on these two indicators, using the formula: Comprehensive Risk Index = 0.6 × Combination Risk Level + 0.4 × Diversified Purchase Risk Score. This weighting reflects that the inherent risk of the combination is more critical than the diversified purchase behavior itself. The calculated comprehensive risk index is compared with a preset risk index threshold (usually set to 0.65). When the comprehensive risk index is greater than or equal to the threshold, a high-risk commodity association analysis result is generated; otherwise, a low-risk result is generated.

[0096] Through the above technical solution, this application realizes a comprehensive assessment of the purchasing behavior of dangerous raw material combinations, organically combining the inherent danger of the combination with the abnormality of the purchasing behavior, thereby improving the accuracy of the assessment.

[0097] This application further proposes the following for calculating the dimensional risk score:

[0098] Based on the results of the commodity association analysis, the hazardous materials control level is extracted, the commodity risk level index is calculated, and commodity risk data is generated.

[0099] Based on the product association analysis results, analyze the degree of anomalies in purchase quantity, purchase frequency, and purchase channels; based on the degree of anomalies, calculate behavioral anomaly characteristic indicators and generate behavioral anomaly data;

[0100] Based on product risk data and abnormal behavior data, combined with purchase history transaction records, a dimensional risk score is calculated.

[0101] Specifically, based on the product association analysis results, the hazard control level is extracted. This level is determined according to the National Hazardous Chemicals List and is divided into four levels: extremely hazardous, high-risk, medium-risk, and low-risk, quantified into values ​​ranging from 0.2 to 1.0. The product risk level index is calculated using the formula: Product Risk Level Index = Combined Hazard Degree × Control Level Coefficient, where the control level coefficient is set according to the hazard level (extremely hazardous 1.0, high-risk 0.8, medium-risk 0.6, low-risk 0.4). The generated product risk data reflects the inherent hazard level of the product. Simultaneously, the abnormality of purchase quantity, purchase frequency, and purchase channel is analyzed. Abnormal purchase quantity refers to a single purchase quantity approaching but not exceeding the regulatory threshold; abnormal purchase frequency refers to purchase intervals significantly shorter than the normal consumption cycle; abnormal purchase channel refers to purchases dispersed across multiple platforms to circumvent the supervision of a single platform. An abnormality degree index is calculated for each type of abnormality, ranging from 0 to 1, and then weighted and summed to generate a behavioral abnormality characteristic index. The generated behavioral abnormality data reflects the suspiciousness of user behavior. Based on users' historical transaction records, a dimensional risk score is calculated: for newly emerging dangerous goods purchases, historical records have a lower weight; for users with a history of dangerous goods purchases, historical records have a higher weight. The formula for calculating the dimensional risk score is: Dimensional Risk Score = α × Commodity Risk Level Index + β × Abnormal Behavioral Characteristics Index + γ × Historical Risk Factor, where α, β, and γ are dynamic weighting coefficients, summing to 1, and adjusted according to specific circumstances.

[0102] Through the above technical solutions, this application achieves a multi-dimensional quantitative assessment of the risks of purchasing dangerous goods, making the risk assessment more comprehensive and accurate; the dual-dimensional assessment of commodity risk and behavioral abnormalities distinguishes between high-risk and low-risk transactions.

[0103] This application further proposes a method for generating risk score data based on dimensional risk scores and a weighted fusion algorithm, including:

[0104] Based on the dimensional risk scores, determine the weight coefficients of the product risk sub-model, the behavior anomaly sub-model, and the user history sub-model;

[0105] Based on the weighting coefficients, a weighted average algorithm is applied to calculate the comprehensive risk score; and linear normalization is then performed to generate standardized risk score data.

[0106] Specifically, the weight coefficients of the three sub-models are determined. For example, the product risk sub-model reflects the inherent danger of the dangerous goods themselves, and its weight coefficient is typically set to 0.5; the abnormal behavior sub-model reflects the suspiciousness of purchasing behavior, and its weight coefficient is typically set to 0.3; the user history sub-model reflects the cumulative risk of a user's historical behavior, and its weight coefficient is typically set to 0.2. These weights are dynamically adjusted according to the current analysis scenario. For example, the weight of the product risk sub-model is increased when identifying new combinations of dangerous goods, and the weight of the user history sub-model is increased when detecting habitual offenders. A weighted average algorithm is applied to calculate the comprehensive risk score, with the formula: Comprehensive Risk Score = w1 × Product Risk Score + w2 × Abnormal Behavior Score + w3 × Historical Risk Score, where w1, w2, and w3 are the weight coefficients of the three sub-models, respectively. To ensure the comparability of scores across different scenarios, the comprehensive risk score is linearly normalized, mapping the original score to a standard range of 0-1, with the formula: Standardized Risk Score = (Original Score - Minimum Possible Score) / (Maximum Possible Score - Minimum Possible Score).

[0107] Through the above technical solution, the normalized risk score data of this application eliminates the score differences in different scenarios and periods, making the risk assessment results consistent and comparable, and providing a unified standard for generating graded early warning results.

[0108] This application further proposes that when generating a risk warning result by comparing risk score data with a preset risk threshold, the following should be included:

[0109] The preset risk thresholds include a first preset risk threshold and a second preset risk threshold, wherein the first preset risk threshold is greater than the second preset risk threshold;

[0110] Risk warning results include high-risk warning results, medium-risk warning results, and low-risk warning results;

[0111] The risk score data is compared with the first preset risk threshold and the second preset risk threshold respectively;

[0112] When the risk score data is greater than or equal to the first preset risk threshold, a high-risk warning result is generated;

[0113] When the risk score data is less than the first preset risk threshold and greater than or equal to the second preset risk threshold, a medium-risk warning result is generated.

[0114] When the risk score data is less than the second preset risk threshold, a low-risk warning result is generated.

[0115] Specifically, two preset risk thresholds are set: a first preset risk threshold (usually set to 0.85) and a second preset risk threshold (usually set to 0.65), with the first preset risk threshold being greater than the second. Standardized risk score data is precisely compared to these two thresholds, and a three-level warning is generated based on the comparison results: when the risk score data is greater than or equal to the first preset risk threshold, it indicates an extremely high risk, generating a high-risk warning result, indicating a possible high-risk activity involving the preparation of hazardous materials, requiring immediate action; when the risk score data is less than the first preset risk threshold but greater than or equal to the second preset risk threshold, it indicates a moderate risk, generating a medium-risk warning result, indicating suspicious activity requiring close monitoring and analysis; when the risk score data is less than the second preset risk threshold, it indicates a low risk, generating a low-risk warning result, indicating normal transactions or controllable risk, requiring only routine monitoring.

[0116] Through the above technical solution, this application's tiered early warning mechanism enables the precise allocation of regulatory resources, avoiding resource waste or underreporting caused by a one-size-fits-all approach.

[0117] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this invention also proposes an artificial intelligence-based shopping clue analysis system, and an early warning method for applying the artificial intelligence-based shopping clue analysis system, including:

[0118] The data acquisition unit is configured to acquire shopping data from e-commerce platforms and generate multi-channel shopping datasets; based on the multi-channel shopping datasets, it processes purchase privacy information through encryption algorithms and de-identification technologies to generate compliant shopping data.

[0119] The preprocessing unit is configured to perform semantic matching between compliant shopping data and the dangerous goods list database, identify dangerous goods purchase records, generate dangerous goods identification results, and preprocess several order data based on the dangerous goods identification results to generate standardized order data.

[0120] The analysis unit is configured to calculate the support and confidence of products based on standardized order data and an association rule mining algorithm, generate association rule data, identify the combined purchase patterns of hazardous raw materials, and generate combination feature data; based on the combination feature data, analyze the purchase time interval and purchase frequency of hazardous raw material combinations, generate time series analysis results, evaluate the dispersed purchase behavior of hazardous raw material combinations, and generate dispersed purchase evaluation data; based on the dispersed purchase evaluation data and the combination feature data, comprehensively judge the combined purchase behavior of hazardous raw materials, and generate product association analysis results.

[0121] The early warning unit is configured to: calculate product risk level indicators and generate product risk data based on product association analysis results; analyze abnormal purchasing behavior characteristics and generate abnormal behavior data based on product association analysis results; calculate dimensional risk scores based on product risk data and abnormal behavior data, combined with historical purchase transaction records; calculate risk score data based on dimensional risk scores and a weighted fusion algorithm; and compare the risk score data with a preset risk threshold to generate a risk warning result.

[0122] In summary, by combining encryption algorithms and anonymization techniques on user identification information in multi-channel shopping data, key transaction feature information required for risk assessment is retained while ensuring user privacy and data compliance, thus solving the problem of difficulty in in-depth data analysis due to privacy restrictions. By segmenting product names and extracting keywords, combined with precise matching and fuzzy semantic matching mechanisms from a hazardous materials database, it is possible to identify hazardous materials purchase records under different expressions and naming methods, avoiding the omissions caused by relying solely on keyword rules and improving the accuracy and robustness of hazardous materials identification. Based on association rule mining algorithms, frequent itemset analysis is performed on standardized order data, and association rules are further filtered using professional knowledge in the field of hazardous materials supervision. This enables the discovery of potentially hazardous raw material combination purchase patterns from a large number of dispersed orders, overcoming the difficulty in detecting risks across product and order combinations. By introducing time series analysis, the purchase time intervals and frequencies of hazardous material raw material combinations are comprehensively analyzed and matched with an abnormal dispersed purchase behavior feature database, identifying deliberately dispersed purchasing behavior to evade supervision and improving the ability to detect concealed high-risk behaviors. This system integrates the risk level of the product itself, the hazard level of the product combination, and abnormal purchasing behavior characteristics to construct a multi-dimensional risk scoring system. A weighted fusion algorithm is then used to generate a comprehensive risk score, making risk assessment results more comprehensive and objective, and reducing false positives and false negatives caused by single-dimensional judgments. By comparing the risk score data with multi-level preset risk thresholds, different levels of early warning results (high risk, medium risk, and low risk) can be output, providing a tiered response basis for regulatory authorities or platform risk control systems.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An early warning method for a shopping clue analysis system based on artificial intelligence, characterized in that, include: Acquire shopping data from e-commerce platforms to generate a multi-channel shopping dataset; process the purchase privacy information in the multi-channel shopping dataset using encryption algorithms and de-identification technologies to generate compliant shopping data; Based on the compliant shopping data and the dangerous goods list database, semantic matching is performed to identify dangerous goods purchase records, generate dangerous goods identification results, and based on the dangerous goods identification results, several order data are preprocessed to generate standardized order data; Based on the standardized order data, the support and confidence of products are calculated using an association rule mining algorithm to generate association rule data, identify the combined purchase pattern of hazardous raw materials, and generate combined feature data. Based on the combined characteristic data, the purchase time interval and purchase frequency of the combination of hazardous raw materials are analyzed to generate time series analysis results, evaluate the decentralized purchase behavior of the combination of hazardous raw materials, and generate decentralized purchase assessment data; based on the decentralized purchase assessment data and the combined characteristic data, the purchase behavior of the combination of hazardous raw materials is comprehensively judged to generate commodity association analysis results. Based on the product association analysis results, calculate the product risk level index and generate product risk data; Based on the product association analysis results, analyze abnormal purchasing behavior characteristics and generate abnormal behavior data; Based on the aforementioned product risk data and abnormal behavior data, combined with purchase history transaction records, a dimensional risk score is calculated. Risk score data is generated by calculating the risk scores based on the aforementioned dimensions and using a weighted fusion algorithm. Based on the risk score data, a risk warning result is generated by comparing it with a preset risk threshold. When generating combined feature data, the following are included: Based on the standardized order data, a hazardous materials purchase transaction database is constructed; based on the hazardous materials purchase transaction database, frequent itemsets are mined using the Apriori algorithm; based on the frequent itemsets, the support of product combinations is calculated to generate support data; based on the frequent itemsets, the confidence of product combinations is calculated to generate confidence data; based on the support data and confidence data, a general minimum support threshold and a general minimum confidence threshold are set; based on the general minimum support threshold and the general minimum confidence threshold, preliminary effective association rules are filtered to generate preliminary association rule data. Based on the preliminary association rule data and combined with professional knowledge in the field of hazardous materials regulation, a professional support threshold and a professional confidence threshold are set. Based on the professional support threshold and professional confidence threshold, high-risk hazardous materials combinations are screened; based on the high-risk hazardous materials combinations, a preset combination pattern is matched to identify hazardous materials raw material combinations. Analyze the risk level relationship of the commodities based on the combination of the aforementioned hazardous materials raw materials; Based on the aforementioned risk level relationship, assess the degree of risk of the combination and generate combination characteristic data; When generating decentralized purchase assessment data, the following are included: Based on the combined feature data, the purchase time sequence of each commodity in the combination of hazardous raw materials by the same user is extracted; Based on the purchase time series, calculate the average purchase time interval between products within the bundle; calculate the purchase frequency of products within the bundle to generate purchase frequency data; analyze the dispersed purchase behavior pattern based on the average purchase time interval and purchase frequency data; and match it with a preset abnormal dispersed purchase behavior feature library. Based on the matching results, assess the risk level of diversified purchasing behavior and generate diversified purchasing assessment data.

2. The early warning method of the shopping clue analysis system based on artificial intelligence according to claim 1, characterized in that, When processing purchase privacy information in the multi-channel shopping dataset using encryption algorithms and de-identification techniques to generate compliant shopping data, the following steps are included: Based on the multi-channel shopping dataset, extract product name, specifications, transaction time, transaction amount, and user identification information; The user identification information is encrypted using the AES encryption algorithm and desensitized using hash desensitization technology. Based on the results of encryption and de-identification processing, compliant shopping data is generated.

3. The early warning method of the shopping clue analysis system based on artificial intelligence according to claim 2, characterized in that, When generating standardized order data, the following are included: Based on the compliant shopping data, extract the text information of the product name to generate product name data; based on the product name data, perform word segmentation and keyword extraction to generate keyword data; based on the keyword data, perform precise matching with the built-in dangerous goods list database; when precise matching fails, use the cosine similarity algorithm for fuzzy semantic matching; based on the results of precise matching or fuzzy semantic matching, identify dangerous goods, mark dangerous goods purchase records, and generate dangerous goods identification results; Based on the hazardous materials identification results, hazardous materials order records are extracted to generate hazardous materials order data; based on the hazardous materials order data, the format of product names is standardized to generate standardized name data; product specifications are converted into standardized units of measurement to generate standardized specification data; transaction times are converted into a standardized time format to generate standardized time data. Standardized order data is generated by combining the standardized name data, standardized specification data, and standardized time data.

4. The early warning method of the shopping clue analysis system based on artificial intelligence according to claim 3, characterized in that, When generating product association analysis results, the following are included: Based on the dispersed purchase assessment data, a dispersed purchase risk score is extracted; based on the combined characteristic data, a hazard assessment data for the combination of hazardous raw materials is extracted; based on the dispersed purchase risk score and hazard assessment data, a two-dimensional risk assessment matrix is ​​constructed, risk assessment rules are set, and a comprehensive risk index for combined purchase behavior is calculated. Based on the comprehensive risk index, the results are compared with the preset risk index threshold to generate product association analysis results.

5. The early warning method of the shopping clue analysis system based on artificial intelligence according to claim 4, characterized in that, When calculating the dimensional risk score, the following are included: Based on the results of the commodity association analysis, the dangerous goods control level is extracted, the commodity risk level index is calculated, and commodity risk data is generated. Based on the product association analysis results, the degree of abnormality in purchase quantity, purchase frequency, and purchase channel is analyzed; based on the degree of abnormality, behavioral abnormality characteristic indicators are calculated, and behavioral abnormality data is generated. Based on the product risk data and abnormal behavior data, combined with purchase history transaction records, a dimensional risk score is calculated.

6. The early warning method of the shopping clue analysis system based on artificial intelligence according to claim 5, characterized in that, When generating risk score data based on the aforementioned dimensional risk scores and a weighted fusion algorithm, the following are included: Based on the risk scores of the aforementioned dimensions, determine the weight coefficients of the product risk sub-model, the behavior anomaly sub-model, and the user history sub-model; Based on the weighting coefficients, a weighted average algorithm is applied to calculate the comprehensive risk score; and linear normalization is performed to generate standardized risk score data.

7. The early warning method of the shopping clue analysis system based on artificial intelligence according to claim 6, characterized in that, When generating a risk warning result by comparing the risk score data with a preset risk threshold, the process includes: The preset risk threshold includes a first preset risk threshold and a second preset risk threshold, wherein the first preset risk threshold is greater than the second preset risk threshold; The risk warning results include high-risk warning results, medium-risk warning results, and low-risk warning results; The risk score data is compared with the first preset risk threshold and the second preset risk threshold respectively; When the risk score data is greater than or equal to the first preset risk threshold, a high-risk warning result is generated; When the risk score data is less than the first preset risk threshold and greater than or equal to the second preset risk threshold, a medium-risk warning result is generated; When the risk score data is less than the second preset risk threshold, a low-risk warning result is generated.

8. An artificial intelligence-based shopping clue analysis system, used for applying the early warning method of the artificial intelligence-based shopping clue analysis system as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is configured to acquire shopping data from e-commerce platforms and generate multi-channel shopping datasets. Based on the aforementioned multi-channel shopping dataset, purchase privacy information is processed using encryption algorithms and de-identification techniques to generate compliant shopping data; The preprocessing unit is configured to perform semantic matching between the compliant shopping data and the dangerous goods list database, identify dangerous goods purchase records, generate dangerous goods identification results, and preprocess several order data based on the dangerous goods identification results to generate standardized order data. The analysis unit is configured to calculate the support and confidence between products based on the standardized order data and an association rule mining algorithm, generate association rule data, identify the combination purchase pattern of dangerous raw materials, and generate combination feature data. Based on the combined characteristic data, the purchase time interval and purchase frequency of the combination of hazardous raw materials are analyzed to generate time series analysis results, evaluate the decentralized purchase behavior of the combination of hazardous raw materials, and generate decentralized purchase assessment data; based on the decentralized purchase assessment data and the combined characteristic data, the purchase behavior of the combination of hazardous raw materials is comprehensively judged to generate commodity association analysis results. When generating combined feature data, the following are included: Based on the standardized order data, a hazardous materials purchase transaction database is constructed; based on the hazardous materials purchase transaction database, frequent itemsets are mined using the Apriori algorithm; based on the frequent itemsets, the support of product combinations is calculated to generate support data; based on the frequent itemsets, the confidence of product combinations is calculated to generate confidence data; based on the support data and confidence data, a general minimum support threshold and a general minimum confidence threshold are set; based on the general minimum support threshold and the general minimum confidence threshold, preliminary effective association rules are filtered to generate preliminary association rule data. Based on the preliminary association rule data and combined with professional knowledge in the field of hazardous materials regulation, a professional support threshold and a professional confidence threshold are set. Based on the professional support threshold and professional confidence threshold, high-risk hazardous materials combinations are screened; based on the high-risk hazardous materials combinations, a preset combination pattern is matched to identify hazardous materials raw material combinations. Analyze the risk level relationship of the commodities based on the combination of the aforementioned hazardous materials raw materials; Based on the aforementioned risk level relationship, assess the degree of risk of the combination and generate combination characteristic data; When generating decentralized purchase assessment data, the following are included: Based on the combined feature data, the purchase time sequence of each commodity in the combination of hazardous raw materials by the same user is extracted; Based on the purchase time series, calculate the average purchase time interval between products within the bundle; calculate the purchase frequency of products within the bundle to generate purchase frequency data; analyze the dispersed purchase behavior pattern based on the average purchase time interval and purchase frequency data; and match it with a preset abnormal dispersed purchase behavior feature library. Based on the matching results, assess the risk level of diversified purchasing behavior and generate diversified purchasing assessment data; The early warning unit is configured to: calculate a product risk level index and generate product risk data based on the product association analysis results; analyze abnormal purchase behavior characteristics and generate abnormal behavior data based on the product association analysis results; calculate a dimensional risk score based on the product risk data and abnormal behavior data, combined with historical purchase transaction records; calculate and generate risk score data based on the dimensional risk score and a weighted fusion algorithm; and compare the risk score data with a preset risk threshold to generate a risk warning result.

Images