A method and system for whole-link management and control of commodity and price data of a procurement e-commerce platform

Abstract

This invention discloses a method and system for end-to-end management and control of product and price data on an e-commerce platform. It relates to the field of data processing technology. The method includes: acquiring product and price data; sequentially performing field standardization verification through a first neural network model and terminology standardization through a second neural network model to obtain basic verification data; sequentially performing category mapping through a third neural network model based on a product category knowledge graph and terminology standardization through a fourth neural network model to obtain standardized category data; generating product data with material codes through a material matching module; sequentially extracting price fluctuation features through a fifth neural network model and performing anomaly detection through a sixth neural network model based on the isolated forest algorithm to obtain price anomaly scores; generating executable rule combinations through a visual rule configuration module; and inputting the above data into the rule-driven model to obtain end-to-end management and control results. This achieves automated data verification and intelligent mapping.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for full-link management and control of product and price data on a procurement e-commerce platform. Background Technology

[0002] Existing e-commerce procurement platforms commonly encounter problems when connecting to multi-source supplier data, such as non-standard data formats, inconsistent category systems, and difficulty in real-time identification of abnormal price fluctuations. Traditional methods rely on manual verification and fixed rules, which are insufficient to handle massive amounts of heterogeneous data. Furthermore, they lack end-to-end collaborative management from data access, category mapping, material matching to price monitoring, resulting in poor data quality, delayed response to anomalies, and complex rule configurations, thus affecting the accuracy and efficiency of procurement decisions. Summary of the Invention

[0003] This application provides a method and system for end-to-end management and control of product and price data on e-commerce platforms. This addresses at least the problems of traditional methods relying on manual verification and fixed rules, which struggle to handle massive amounts of heterogeneous data, and lacking end-to-end collaborative management from data access, category mapping, material matching to price monitoring. These issues result in poor data quality, delayed anomaly response, and complex rule configuration, ultimately affecting the accuracy and efficiency of procurement decisions.

[0004] This application provides a method and system for end-to-end control of product and price data on e-commerce platforms, including:

[0005] Obtain product and price data from e-commerce platforms;

[0006] The commodity and price data are processed sequentially by a first neural network model trained to perform non-empty verification, uniqueness verification, and format verification on entities related to field norms, and a second neural network model that performs terminology standardization on data related to field norms, to obtain basic verification data for completing entity recognition, relation extraction, and terminology standardization.

[0007] The basic verification data is processed sequentially by a trained third neural network model for entity recognition and relationship extraction related to category mapping, and a fourth neural network model for terminology standardization of category mapping-related data. The third neural network model performs neighborhood aggregation encoding and vector similarity matching on supplier categories and internal procurement categories based on a product category knowledge graph to obtain category mapping results. The fourth neural network model performs terminology standardization on the category mapping results to obtain standardized category data.

[0008] Based on the standardized category data, the supplier's product code is mapped to the internal material code through the material matching module to generate product data with material code;

[0009] The price data in the commodity and price data is processed sequentially by a fifth neural network model trained to extract price fluctuation-related features and a sixth neural network model trained to detect anomalies in price fluctuation-related features. The fifth neural network model calculates the first-order and second-order difference sequences of prices based on a multi-time-scale sliding window, and the sixth neural network model performs multi-dimensional outlier detection on price change rate, price acceleration, and trading volume changes based on the isolated forest algorithm to obtain a price anomaly score.

[0010] The visual rule configuration module receives conditional and action components that users combine by dragging and dropping, and generates a directed acyclic graph based on the rule dependency parsing algorithm to obtain executable rule combinations.

[0011] The product data with material codes, the price anomaly score, and the executable rules are combined and input into the rule-driven model to determine the end-to-end control results of the product and price data. The control results include data quality reports, anomaly interception instructions, or price warning information.

[0012] Furthermore, the construction of the product category knowledge graph includes:

[0013] Using supplier product category nodes and internal procurement category nodes as entities, and historically successfully mapped category pairs as edges, the multi-hop neighborhood information of each category node is aggregated and encoded based on the graph isomorphic network in the graph neural network to generate a 200-dimensional category semantic feature vector. For newly added supplier categories, the cosine similarity between its semantic feature vector and the vectors of all internal category nodes in the knowledge graph is calculated, and the top three with the highest similarity are selected as candidate mapping results.

[0014] Furthermore, the determination of price anomaly scores also includes:

[0015] Based on the first-order and second-order difference sequences of the prices, the quantile thresholds of the rate of change of prices within the 7-day, 30-day, and 90-day sliding windows are calculated respectively. The minimum value of the quantile thresholds of the three windows is weighted and averaged as the dynamic judgment threshold. When the abnormal score output by the isolated forest algorithm is greater than or equal to the dynamic judgment threshold, a price anomaly marker is triggered.

[0016] Furthermore, when a price anomaly flag is triggered, the system calls the price statistics of similar products from a third-party e-commerce platform via API to calculate the deviation of the current product price from the industry average price. If the deviation is less than 15% and the anomaly score is high risk, the risk level is downgraded to medium risk. If the deviation is greater than or equal to 15% and the anomaly score is high risk, a high-risk anomaly attribution report containing supplier information, product information, and deviation value is generated.

[0017] Furthermore, the visual rule configuration module includes:

[0018] The system includes a condition component library, an action component library, and a data privacy label component. The data privacy label component embeds L1 to L4 privacy level labels for each data field and verifies data access permissions in real time when the user drags and drops to combine rule components.

[0019] The automatic rule dependency resolution algorithm includes: resolving the temporal dependencies between rule components, constructing a directed acyclic graph, detecting and rejecting rule combinations with circular dependencies, and generating an executable rule script after verification.

[0020] Furthermore, the rule description input by the user in Chinese natural language is received through the natural language semantic parsing engine, the rule description is parsed into a combination of corresponding conditional components and action components, and displayed on the interface of the visual rule configuration module.

[0021] Furthermore, when training at least one of the first neural network model, the second neural network model, the third neural network model, the fourth neural network model, the fifth neural network model, and the sixth neural network model, the loss function includes a focus loss function or a weighted cross-entropy loss function, and includes a regularization term in the corresponding loss function; the data samples used to train the first neural network model, the second neural network model, the third neural network model, the fourth neural network model, the fifth neural network model, and the sixth neural network model are data-cleaned samples after removing noise, outlier samples, and low-quality samples.

[0022] Furthermore, the training data for the third and fourth neural network models includes category mapping samples constructed based on different suppliers, different category systems, and different historical mapping records, as well as terminology standardization samples constructed based on medical dictionaries or e-commerce category dictionaries; the training data for the isolated forest algorithm includes price change rate, price acceleration, and transaction volume change characteristics in historical price time series, and the labeling threshold for abnormal samples is dynamically determined based on the quantiles of the historical price distribution.

[0023] Furthermore, it also includes pushing the control results to the procurement terminal or supplier terminal in real time via pop-up windows or embedded windows; generating a structured report based on the control results, the structured report listing abnormal data items, triggering rules and original data traceability information; and proposing price negotiation strategies or supplier replacement suggestions to the procurement party based on the price anomaly score.

[0024] Furthermore, a full-chain management and control system for product and price data on an e-commerce platform includes:

[0025] One or more processors;

[0026] One or more non-transitory computer-readable media, the one or more non-transitory computer-readable media collectively storing a computer program, the computer program causing the system to perform the method as described in any of the preceding claims when executed by the one or more processors.

[0027] The application employs the above technical solution and has at least the following beneficial effects:

[0028] This invention employs a neural network model for field standardization and terminology standardization, combined with a product category knowledge graph to achieve intelligent mapping from supplier categories to internal procurement categories and material code matching. It also utilizes multi-timescale difference sequences and isolated forest algorithms for multi-dimensional anomaly detection of price fluctuations. Furthermore, it provides visualized drag-and-drop rule configuration and natural language parsing, achieving fully automated control of product and price data across the entire supply chain. This effectively improves data quality, reduces manual intervention costs, identifies price anomalies in real time and generates attribution reports, significantly enhancing the data governance and risk warning capabilities of the procurement platform.

[0029] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart provided in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0033] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, a method for end-to-end control of product and price data on an e-commerce platform includes:

[0035] Obtain commodity and price data of the procurement e-commerce platform, which includes data from multiple supplier systems, ERP systems, supplier portals or third-party price monitoring platforms; the data formats include structured data such as price lists and commodity master data, and unstructured data such as commodity description texts and contract clauses.

[0036] In this embodiment, the method further comprises performing standardized cleaning on the original commodity and price data; unifying field formats, standardizing dates to YYYY-MM-DD, unifying price units to yuan, and normalizing spaces and special characters for commodity category labels; performing cross-source association on the cleaned data through supplier IDs or commodity SKUs, sorting multiple records associated to the same commodity entity according to update timestamps, and only retaining the latest record.

[0037] Sequentially processing commodity and price data through a trained first neural network model for performing non-null check, uniqueness check and format check on field normativity-related entities, and a second neural network model for performing term standardization on field normativity-related data, so as to obtain basic check data that has completed entity recognition, relation extraction and term standardization.

[0038] In this embodiment, the first neural network model is a sequence labeling model based on a bi-directional long short-term memory network and a conditional random field, and is used for identifying entities such as supplier names, commodity codes, specification parameters, and measurement units and the affiliation relationships therebetween from commodity description texts; in this embodiment, the second neural network model is a transformer-based encoder-decoder model, and is used for mapping the identified entity terms to the unified term standard of the platform, for example, unifying and standardizing "zhi", "ge" and "PCS" into "ge".

[0039] In a specific embodiment, for electronic component commodity data uploaded by a certain supplier, the original description is "resistor, 10 , 0603, 5%, zhi", the first neural network model identifies that commodity name = "resistor", specification parameter = "10 , 0603, 5%", measurement unit = "zhi"; the second neural network model standardizes "zhi" into "ge" according to a pre-constructed measurement unit dictionary, and simultaneously standardizes "10 " into "10000 ohms", the non-null check finds that the "brand" field is missing, then the basic check data marks that this commodity lacks a required field and returns an error prompt "the brand field cannot be empty", so as to block unqualified data from entering subsequent processes from the source.

[0040] In another embodiment, the first neural network model and the second neural network model can adopt the lightweight ALiteBERT model to reduce computational resource consumption. During training, the loss function adopts the focus loss function plus the L2 regularization term. The training data comes from historical product data of different suppliers and different categories, which are formed after manually labeling entity boundaries and standardizing terms. Data samples with inconsistent labeling and excessive noise are deleted.

[0041] The basic verification data is processed sequentially by a third neural network model trained for entity recognition and relationship extraction related to category mapping, and a fourth neural network model for terminology standardization of category mapping data. The third neural network model performs neighborhood aggregation encoding and vector similarity matching on supplier categories and internal procurement categories based on the commodity category knowledge graph to obtain category mapping results. The fourth neural network model performs terminology standardization on the category mapping results to obtain standardized category data.

[0042] When constructing a product category knowledge graph, supplier product category nodes and internal procurement category nodes are used as entities, and historically successfully mapped category pairs are used as edges. Based on the graph isomorphic network in the graph neural network, the multi-hop neighborhood information of each category node is aggregated and encoded to generate a 200-dimensional category semantic feature vector. For newly added supplier categories, the cosine similarity between its semantic feature vector and the vectors of all internal category nodes in the knowledge graph is calculated, and the top three with the highest similarity are used as candidate mapping results.

[0043] In one specific embodiment, taking the category "galvanized hexagonal nuts M8" from an MRO industrial products supplier as an example, the third neural network model first extracts the supplier category name from the basic verification data. It then aggregates information from neighboring nodes such as "fasteners," "nuts," and "bolt accessories" around the node through a graph isomorphic network to generate a feature vector. The feature vector is then calculated with the vectors of nodes such as "fasteners-nuts-hexagonal nuts," "hardware-standard parts-nuts," and "mechanical parts-fasteners" in the internal procurement category tree. The similarity scores are 0.92, 0.75, and 0.58, respectively. The system uses "fasteners-nuts-hexagonal nuts" with a similarity score of 0.92 as the recommended mapping and outputs the mapping result to the fourth neural network model for terminology standardization. "galvanized" is standardized to "hot-dip galvanized," and "M8" is standardized to "nominal diameter 8mm."

[0044] In another embodiment, if the highest similarity is below 0.60, the system will refuse automatic mapping and switch to manual mapping. At the same time, the manually confirmed mapping pair will be updated to the knowledge graph as new knowledge feedback, so as to realize the continuous learning of the knowledge graph.

[0045] Based on standardized category data, the material matching module maps supplier product codes to internal material codes, generating product data with material codes. The material matching module can use a combination of rule matching and vector retrieval. First, it locks the material candidate set according to the standardized category, and then performs precise or approximate matching with the material attribute library according to the product specifications such as size, material, rated power, etc. When the matching degree exceeds the preset threshold, the material code is automatically bound; otherwise, the material administrator is prompted to create or expand the material code.

[0046] In one embodiment, the material matching module also maintains a material mapping cache table. For the mapping relationship between the supplier product code and the internal material code that has been matched, the same supplier product code can be directly reused the next time it appears, thereby improving processing efficiency.

[0047] The price data in the commodity and price data are processed sequentially by a fifth neural network model trained to extract features related to price fluctuations and a sixth neural network model trained to detect anomalies related to price fluctuations. The fifth neural network model calculates the first-order and second-order difference sequences of prices based on a multi-timescale sliding window, while the sixth neural network model performs multi-dimensional outlier detection on price change rate, price acceleration, and trading volume changes based on the isolated forest algorithm to obtain price anomaly scores.

[0048] Specifically, for each product SKU, the fifth neural network model constructs a time series by obtaining its daily price data from the past 90 days. Calculate the first-order difference sequences within the 7-day, 30-day, and 90-day sliding windows, respectively. and second-order difference sequences Then calculate the rate of price change within each window. and price acceleration At the same time, the number of daily orders or inquiries is counted as a characteristic of changes in transaction volume. .

[0049] The sixth neural network model uses the Isolation Forest algorithm for anomaly detection. The input feature vector is [current price change rate, current price acceleration, percentage change in trading volume within the current window]. The Isolation Forest consists of 100 isolated trees, and each tree randomly selects a feature and a splitting value to partition the data. Anomaly score is then assigned. ,in Let x be the average path length across all trees. The normalization constant for the average path length when the number of samples is n is given. The closer the anomaly score is to 1, the more likely it is to be an anomaly.

[0050] Determining price anomaly scores also includes calculating the quantile thresholds of price change rates within 7-day, 30-day, and 90-day sliding windows based on the first-order and second-order difference sequences of prices, respectively. For example, the 95th percentile can be used as the anomaly threshold for that window. , , Then, the minimum value of the three window quantile thresholds is weighted and averaged to obtain the dynamic judgment threshold. When the anomaly score s(x) output by the Isolation Forest algorithm is greater than or equal to When an anomaly occurs, a price anomaly flag is triggered, and the anomaly score and trigger window are recorded simultaneously.

[0051] In one embodiment, for seasonal goods such as Christmas supplies and air conditioners, the fifth neural network model can add an annual sliding window to compare the current price with the price of the same period last year and calculate the year-on-year change rate as an additional feature input to the sixth neural network model.

[0052] When a price anomaly flag is triggered, the method also includes calling third-party e-commerce platforms, such as similar B2B platforms and industry price index publishing platforms, via API to retrieve statistical data on similar products and calculate the deviation between the current product price and the industry average price. If the absolute value of the deviation is less than 15% and the abnormal score is high-risk and greater than 0.8, the risk level will be downgraded to medium risk, and the cause will be recorded as overall industry price fluctuations. If the deviation is greater than or equal to 15% and the abnormal score is high-risk, a high-risk abnormal attribution report containing supplier information, product information, and deviation value will be generated. The report will include supplier name, product SKU, current price, industry average price, deviation, abnormal score, and suggested verification actions, such as suggesting that the purchaser renegotiate or initiate a supplier price comparison process.

[0053] Taking the price monitoring of a certain brand of laptop on a procurement platform as an example, the price of this product has remained stable at 5000±100 yuan over the past 30 days. The fifth neural network model calculates the price change rate within a 7-day window and finds that the price suddenly jumps to 6500 yuan on a certain day, with a change rate of 30%. The price acceleration is positive and relatively large. There are no obvious abnormalities in the transaction volume. The sixth neural network model has an anomaly score s(x)=0.85, and the dynamic threshold is... This triggered a high-risk anomaly flag. Subsequently, the system called a third-party e-commerce API to obtain the market average price of the same model of laptop, which was 5100 yuan. The deviation was d=27.5%>15%. The system generated a high-risk anomaly attribution report and pushed it to the purchasing specialist. The report pointed out that "the price of the laptop from supplier XX is 6500 yuan, which deviates from the market average price of 5100 yuan by 27.5%. It is recommended to check the reasonableness of the quotation or conduct price negotiations."

[0054] In another embodiment, if the absolute value of the deviation is less than 15% but the anomaly score is greater than 0.8, the system will downgrade the risk level to medium risk but still send an alert, indicating that "the price fluctuation is consistent with the industry trend, but the fluctuation range is large, so we suggest paying attention."

[0055] In addition, the method also includes receiving conditional components and action components combined by the user through drag-and-drop via a visual rule configuration module, and generating a directed acyclic graph based on an automatic rule dependency parsing algorithm to obtain executable rule combinations.

[0056] The visual rule configuration module includes a condition component library, an action component library, and a data privacy label component. The condition component library provides preset conditions such as "price greater than", "category equal to", "field is empty", "time range is", and "regular expression matching". Each condition component has a configurable parameter panel. The action component library provides preset actions such as "intercept data", "mark as abnormal", "send email alert", "write to the abnormal table", and "call Webhook". The data privacy label component embeds L1 to L4 privacy level labels for each data field. L1 is for public data such as product name, L2 is for internal data such as supplier name, L3 is for sensitive data such as purchase price, and L4 is for highly sensitive data such as contract terms. When users drag and drop to combine rule components, the system verifies in real time whether the user's permissions are not lower than the privacy level of the accessed field. If the permissions are insufficient, dragging is prohibited and a prompt is displayed.

[0057] The automatic rule dependency resolution algorithm includes resolving the temporal dependencies between rule components. For example, if the triggering condition of rule A depends on the output of rule B, then rule B must be executed before rule A. The algorithm traverses the data flow dependencies between all rule components and constructs a directed acyclic graph (DAG). If a circular dependency is detected, such as A depending on B, B depending on C, and C depending on A, the rule combination is rejected and the user is prompted to modify it. After verification, the system generates an executable rule script based on the topological sorting of the DAG. The script format can be Drools rule language or a custom JSON serialized rule set.

[0058] In one embodiment, a business user configures a rule by dragging and dropping: the condition components are "product category equals 'electronic components'" and "price change rate is greater than 30%", and the action component is "mark as abnormal and send an email to the purchasing team leader". The system automatically parses the order of these two condition components to be parallel and without dependency conflicts, generates a DAG, outputs the rule script, and then deploys it to the rule engine for real-time execution.

[0059] The method also includes receiving rule descriptions input by users in Chinese natural language through a natural language semantic parsing engine, parsing the rule descriptions into combinations of corresponding conditional components and action components, and displaying them on the interface of the visual rule configuration module. The semantic parsing engine adopts a rule understanding model based on dependency parsing and semantic role labeling. For example, if the user inputs "If the price of a product exceeds twice the highest price in the past 30 days of its category, issue an alert", the engine extracts the conditional entities as "product price", "highest price in the past 30 days of its category", and "twice", the operation relationship is "exceeds", and the action is "issue an alert", which is automatically converted into the conditional component "price > (highest historical price in the category × 2)" and the action component "send an alert message".

[0060] Finally, the product data with material codes, price anomaly scores, and executable rules are combined and input into the rule-driven model to determine the end-to-end control results of product and price data. The control results include data quality reports, anomaly interception instructions, or price warning information.

[0061] The rule-driven model employs a generative rule system based on forward reasoning. The rule base stores multiple IF-THEN rules, such as "IF - Non-empty error exists in basic verification data - THEN - Intercept the data and return an error message", "IF - Price anomaly score is high risk - AND - Market deviation > 15% - THEN - Generate a high-risk warning and push it to the purchaser", and "IF - Material matching failed - THEN - Record pending tasks and notify the material administrator". The rule-driven model receives all input feature vectors and scores, traverses the rule base to match trigger conditions, executes the conclusion part of the matching rules, and finally outputs the comprehensive control results.

[0062] In one embodiment, the control results are pushed to the procurement terminal or supplier terminal in real time via pop-up or embedded window. The procurement terminal displays "The price of a supplier's product is abnormal, it is recommended to check" and the supplier terminal displays "Your product price exceeds the platform's warning threshold, please adjust".

[0063] The method also includes generating structured reports based on the control results. The structured reports list abnormal data items, triggering rules, and original data traceability information. For example, the report title is "Commodity Data Quality Report for March 2025". The table lists the commodity code, the anomaly type "price fluctuation anomaly", the triggering rule "isolated forest anomaly score 0.85 higher than dynamic threshold 0.75", and the original data "price on March 15, 2025: 6,500 yuan, average price of the previous 30 days: 5,100 yuan". At the same time, based on the price anomaly score, the method proposes price negotiation strategies or supplier replacement suggestions to the purchaser, such as "it is recommended to negotiate with supplier XX to the range of 5,200-5,400 yuan" or "it is recommended to introduce alternative suppliers".

[0064] The following describes in detail the implementation process of the method and system according to the embodiments of this disclosure, using a real-world e-commerce platform operation scenario as an example:

[0065] A certain enterprise's procurement e-commerce platform needs to connect with the product and price data of more than 300 suppliers and process about 500,000 product information entries every day. The platform has deployed a full-link management and control system according to an embodiment of this disclosure.

[0066] First, the system automatically extracts incremental product data from the supplier's FTP, API interface, and data platform every day at midnight. After data cleaning and deduplication, the first and second neural network models perform field standardization checks on approximately 80,000 newly added or changed product data. They found that 1,200 products were missing the "brand" field, 350 products had non-standard units of measurement such as mixing "Pcs" and "pieces", and 80 products had incorrect "specification parameters" format. These unqualified data were directly intercepted and error reports were returned to the corresponding supplier's system interface. After receiving the report, the supplier corrected the data and resubmitted it.

[0067] Product data that passes basic verification enters the category mapping stage. The third neural network model calls the product category knowledge graph for matching. The knowledge graph already contains more than 100,000 mapping records. For the category "anti-impact goggles" provided by the new supplier, the graph mapping similarity score is 0.88, and it is automatically mapped to the internal category "safety protection-eye-goggles". After the fourth neural network model completes the terminology standardization, it outputs the standardized category. For a few new categories with a similarity of less than 0.6, the system creates a manual mapping task, which is completed by the category administrator in the background interface. The mapping results are fed back to the knowledge graph for incremental training.

[0068] The material matching module binds supplier product codes to internal material codes based on standardized categories and specifications. In cases where matching fails, such as when a new material has not yet been registered in the material library, the system automatically generates a material extension request form and pushes it to the material management department.

[0069] The price monitoring module processes commodity price data in real time. For an industrial switch of model "ABC-123", the fifth neural network model calculates its price change rate as 35% within a 7-day window, 28% within a 30-day window, and 20% within a 90-day window. The isolated forest model gives an anomaly score of 0.92, and the dynamic threshold... This triggered a high-risk anomaly. Subsequently, the system called the price interface of a third-party B2B platform via API to obtain the industry average price of the same model of switch, which was 1,850 yuan. The supplier's current quote was 2,450 yuan, with a deviation of 32.4%. The system generated a high-risk anomaly attribution report and pushed it to the procurement specialist's workbench.

[0070] After reviewing the report, the procurement specialist initiated the price negotiation process. At the same time, the system automatically generated negotiation suggestions based on historical price data and industry average prices: "Target price range 1850-2050 yuan". After receiving the warning notification, the supplier proactively lowered the price to 2100 yuan. The system re-evaluated the price anomaly score at 0.65, which was below the dynamic threshold, and lifted the anomaly warning.

[0071] In terms of rule configuration, the platform's rule administrators dragged and dropped the following rules through the visual rule configuration module: "IF - Product category is 'Industrial Electrical' - AND - Price anomaly rating is high risk - AND - Supplier rating is lower than B - THEN - Automatically intercept the product data for secondary review", and "IF - Price deviation > 20% - AND - Transaction volume change is negative - THEN - Send price warning to the purchaser and copy to the audit department". The rule dependency parsing algorithm detected that these two rules had no dependency conflicts, generated a DAG, and deployed it to the rule engine.

[0072] For natural language rule configuration, the platform also provides a semantic input box. Purchasing personnel can input "Notify me if the monthly sales volume of a certain SKU exceeds 1,000 units and the price fluctuates by more than 10%", and the system will automatically parse and generate the corresponding condition components and action components, which greatly reduces the technical threshold for rule configuration.

[0073] The system according to the embodiments of this disclosure also includes a user feedback module. When experts correct the price anomaly judgment result, the correction record will be used to incrementally train the sixth neural network model. For example, if a price anomaly is manually judged as a false alarm, the system will change the label of this sample from abnormal to normal and add it to the dataset of the next round of training, thereby continuously improving the accuracy of anomaly detection.

[0074] In an alternative embodiment, the first to sixth neural network models can be all or partly lightweight models such as ALiteBERT or TinyBERT, deployed on edge gateway devices, to achieve localized fast verification and anomaly interception for supplier data access scenarios with limited network conditions.

[0075] In another alternative embodiment, price anomaly detection is not only based on historical price data, but can also integrate external features such as supplier credit scores, order fulfillment rates, and market sentiment indices, and use a multimodal deep learning model to make a comprehensive anomaly judgment.

[0076] In another alternative embodiment, the condition components of the visual rule configuration module support nested combination. For example, a user can use "category belongs to A or B" as a composite condition component and "price is greater than the threshold" as another condition. The two are combined by "AND" logic, and the resulting complex rule can still be correctly decomposed into DAG by the parsing algorithm.

[0077] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for end-to-end control of product and price data on an e-commerce platform, characterized in that, include: Obtain product and price data from e-commerce platforms; The commodity and price data are processed sequentially by a first neural network model trained to perform non-empty verification, uniqueness verification, and format verification on entities related to field norms, and a second neural network model that performs terminology standardization on data related to field norms, to obtain basic verification data for completing entity recognition, relation extraction, and terminology standardization. The basic verification data is processed sequentially by a trained third neural network model for entity recognition and relation extraction related to category mapping, and a fourth neural network model for terminology standardization of category mapping related data. The third neural network model performs neighborhood aggregation encoding and vector similarity matching on supplier categories and internal procurement categories based on the commodity category knowledge graph to obtain category mapping results. The fourth neural network model performs terminology standardization on the category mapping results to obtain standardized category data. Based on the standardized category data, the supplier's product code is mapped to the internal material code through the material matching module to generate product data with material code; The price data in the commodity and price data is processed sequentially by a fifth neural network model trained to extract price fluctuation-related features and a sixth neural network model trained to detect anomalies in price fluctuation-related features. The fifth neural network model calculates the first-order and second-order difference sequences of prices based on a multi-time-scale sliding window, and the sixth neural network model performs multi-dimensional outlier detection on price change rate, price acceleration, and trading volume changes based on the isolated forest algorithm to obtain a price anomaly score. The visual rule configuration module receives conditional and action components that users combine by dragging and dropping, and generates a directed acyclic graph based on the rule dependency parsing algorithm to obtain executable rule combinations. The product data with material codes, the price anomaly score, and the executable rules are combined and input into the rule-driven model to determine the end-to-end control results of the product and price data. The control results include data quality reports, anomaly interception instructions, or price warning information.

2. The method for end-to-end control of product and price data on e-commerce platforms according to claim 1, characterized in that, The construction of the product category knowledge graph includes: Using supplier product category nodes and internal procurement category nodes as entities, and historically successfully mapped category pairs as edges, the multi-hop neighborhood information of each category node is aggregated and encoded based on the graph isomorphic network in the graph neural network to generate a 200-dimensional category semantic feature vector. For newly added supplier categories, the cosine similarity between its semantic feature vector and the vectors of all internal category nodes in the knowledge graph is calculated, and the top three with the highest similarity are selected as candidate mapping results.

3. The method for end-to-end control of product and price data on e-commerce platforms according to claim 1, characterized in that, The determination of the price anomaly score also includes: Based on the first-order and second-order difference sequences of the prices, the quantile thresholds of the rate of change of prices within the 7-day, 30-day, and 90-day sliding windows are calculated respectively. The minimum value of the quantile thresholds of the three windows is weighted and averaged as the dynamic judgment threshold. When the abnormal score output by the isolated forest algorithm is greater than or equal to the dynamic judgment threshold, a price anomaly marker is triggered.

4. The method for full-link control of product and price data on e-commerce platforms according to claim 3, characterized in that, Also includes: When a price anomaly flag is triggered, the system calls the price statistics of similar products from a third-party e-commerce platform via API to calculate the deviation of the current product price from the industry average price. If the deviation is less than 15% and the anomaly score is high risk, the risk level will be downgraded to medium risk; if the deviation is greater than or equal to 15% and the anomaly score is high risk, a high-risk anomaly attribution report containing supplier information, product information, and deviation value will be generated.

5. The method for end-to-end control of product and price data on e-commerce platforms according to claim 1, characterized in that, The visualization rule configuration module includes: The system includes a condition component library, an action component library, and a data privacy label component. The data privacy label component embeds L1 to L4 privacy level labels for each data field and verifies data access permissions in real time when the user drags and drops to combine rule components. The automatic rule dependency resolution algorithm includes: resolving the temporal dependencies between rule components, constructing a directed acyclic graph, detecting and rejecting rule combinations with circular dependencies, and generating an executable rule script after verification.

6. The method for full-link control of product and price data on e-commerce platforms according to claim 5, characterized in that, Also includes: The system receives rule descriptions input by users in Chinese natural language through a natural language semantic parsing engine, parses the rule descriptions into combinations of corresponding conditional components and action components, and displays them on the interface of the visual rule configuration module.

7. The method for full-link control of product and price data on e-commerce platforms according to claim 1, characterized in that: When training at least one of the first, second, third, fourth, fifth, and sixth neural network models, the loss function includes a focus loss function or a weighted cross-entropy loss function, and includes a regularization term in the corresponding loss function; the data samples used to train the first, second, third, fourth, fifth, and sixth neural network models are data-cleaned samples after removing noise, outliers, and low-quality samples.

8. The method for full-link control of product and price data on a procurement e-commerce platform according to claim 1, characterized in that: The training data for the third and fourth neural network models includes category mapping samples constructed based on different suppliers, different category systems, and different historical mapping records, as well as standardized terminology samples constructed based on medical dictionaries or e-commerce category dictionaries; the training data for the isolated forest algorithm includes price change rate, price acceleration, and transaction volume change characteristics in historical price time series, and the labeling threshold for outlier samples is dynamically determined based on the quantiles of the historical price distribution.

9. The method for full-link control of product and price data on a procurement e-commerce platform according to claim 1, characterized in that: It also includes pushing the control results to the procurement terminal or supplier terminal in real time via pop-up windows or embedded windows; generating a structured report based on the control results, the structured report listing abnormal data items, triggering rules and original data traceability information; and proposing price negotiation strategies or supplier replacement suggestions to the procurement party based on the price anomaly score.

10. A system for end-to-end management and control of product and price data on an e-commerce platform, characterized in that, include: One or more processors; One or more non-transitory computer-readable media, the one or more non-transitory computer-readable media collectively storing a computer program, the computer program causing the system to perform the method as described in any one of claims 1 to 9 when executed by the one or more processors.

Images