A push optimization method and system based on e-commerce big data feedback

Abstract

This invention discloses a push optimization method and system based on e-commerce big data feedback, relating to the field of product push optimization technology. This push optimization method based on e-commerce big data feedback acquires multi-source e-commerce data for several products; based on the multi-source e-commerce data, it extracts feedback feature sets for each product, including user feedback feature sets and attribute feedback feature sets; based on the feedback feature sets, it analyzes the user preference evaluation value and product competition evaluation value for each product, and calculates the push adaptation evaluation value for each product. This invention optimizes the push for each product through the push adaptation evaluation value, thereby accurately capturing users' real needs and matching the core advantages of the product, reducing invalid pushes, improving product exposure and conversion efficiency, balancing platform operation needs and actual user experience, and ultimately improving push accuracy.

Description

Technical Field

[0001] This invention relates to the field of product recommendation optimization technology, specifically a recommendation optimization method and system based on e-commerce big data feedback. Background Technology

[0002] With the rapid development of internet technology and e-commerce platforms, online shopping has become one of the mainstream consumption methods. In order to improve user experience, enhance user stickiness, and promote platform transaction growth, major e-commerce platforms have generally adopted product push technology. As the core bridge connecting users and products on e-commerce platforms, its adaptability directly affects the platform's operational efficiency, product exposure quality, and user experience. It has become an important tool for e-commerce platforms to enhance their core competitiveness. The core of product push relies on the analysis and mining of massive e-commerce big data to build a push adaptability evaluation system and achieve dynamic optimization and precise implementation of push strategies.

[0003] The limitations of existing technologies include at least the following issues: existing technologies do not integrate two core multi-source e-commerce data types, user feedback data and product attribute data, and do not perform standardized feature extraction processing on the two types of data respectively. They do not construct complete user feedback feature sets and attribute feedback feature sets, and in particular, they do not extract and merge differentiated features from user behavior feedback and textual semantic feedback. They also do not decompose product attribute data into corresponding dimensions for feature synthesis. As a result, the analysis of user preferences and product competitiveness lacks comprehensive and accurate feature support, and the evaluation results are one-sided, making it difficult to truly reflect users' real needs and the actual competitive advantages of products. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a push optimization method and system based on e-commerce big data feedback, which solves the problem that existing technologies do not perform in-depth mining and integration of multi-source data, easily leading to insufficient push accuracy.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a push optimization method based on e-commerce big data feedback, comprising the following steps: acquiring multi-source e-commerce data for several types of goods, wherein the multi-source e-commerce data includes user feedback data and product attribute data; based on the multi-source e-commerce data, extracting feedback feature sets for each type of goods, including user feedback feature sets and attribute feedback feature sets; based on the feedback feature sets, analyzing the user preference evaluation value and product competition evaluation value for each type of goods, and calculating the push adaptation evaluation value for each type of goods; and based on the push adaptation evaluation value, performing push optimization processing on each type of goods.

[0006] Furthermore, the user feedback data includes user behavior data and user feedback text information. The specific steps for extracting the behavior feedback feature set for each product are as follows: Based on the user behavior data, determine the behavior interaction evaluation value for each product; perform semantic parsing processing on the user feedback text information to generate a text semantic value evaluation value for each product; based on the behavior interaction evaluation value and the text semantic value evaluation value, construct the user feedback feature set for each product.

[0007] Furthermore, the specific steps for generating the behavioral interaction evaluation value for each product are as follows: standardizing the user behavior data; generating a behavior weight set based on the standardized user behavior data; and comprehensively processing the standardized user behavior data and the behavior weight set to generate the behavioral interaction evaluation value for each product.

[0008] Further, the specific steps for generating the text semantic value assessment value for each product are as follows: input the user feedback text information into a pre-trained adaptive language understanding model to extract the structured semantic element feature set for each product; perform statistical processing on the structured semantic element feature set to generate the semantic index set for each product; and perform aggregation processing on the semantic index set to generate the text semantic value assessment value for each product.

[0009] Further, the specific steps for extracting the attribute feedback feature set for each product are as follows: preprocessing the product attribute data; performing feature synthesis processing based on the preprocessing results to generate the attribute feedback feature set for each product, including supply guarantee value, market vitality value, and associated sales potential value.

[0010] Further, the specific steps for analyzing the user preference evaluation value of each product are as follows: read the behavioral interaction evaluation value and the text semantic value evaluation value, and perform normalization processing; perform comprehensive processing on the normalized behavioral interaction evaluation value and the text semantic value evaluation value to obtain the user preference evaluation value of each product.

[0011] Further, the specific steps for analyzing the product competition evaluation value of each product are as follows: read the supply guarantee value, the market vitality value, and the associated sales potential value, and perform normalization processing; merge the normalized supply guarantee value, the market vitality value, and the associated sales potential value to obtain the product competition evaluation value of each product.

[0012] Furthermore, the specific formula for calculating the push adaptation evaluation value for each product is as follows: ;in, , , The order is number 1 The evaluation scores for product push adaptation, user preference, and product competition are as follows: , , The coefficients are, in order, the preference adjustment coefficient, the competition adjustment coefficient, and the difference adjustment coefficient stored in the database.

[0013] Furthermore, the specific steps for optimizing the push notifications for each product are as follows: sort the push adaptation evaluation values ​​in descending order to generate a product push optimization table; and perform product push operations based on the product push optimization table.

[0014] A push optimization system based on e-commerce big data feedback includes: a data acquisition unit for acquiring multi-source e-commerce data for several types of products, including user feedback data and product attribute data; a feedback extraction unit for extracting feedback feature sets for each product based on the multi-source e-commerce data, including user feedback feature sets and attribute feedback feature sets; a comprehensive evaluation unit for analyzing user preference evaluation values ​​and product competition evaluation values ​​for each product based on the feedback feature sets, and calculating a push adaptation evaluation value for each product; and a push optimization unit for performing push optimization processing on each product based on the push adaptation evaluation value.

[0015] The present invention has the following beneficial effects:

[0016] (1) This push optimization method based on e-commerce big data feedback integrates user feedback data and product attribute data, and performs standardized processing and targeted feature extraction: Based on user behavior data, standardized processing is completed and a set of behavior weights is generated to obtain a behavior interaction evaluation value that reflects the user's interaction intention; the user feedback text is parsed by a pre-trained adaptive language understanding model, semantic elements are extracted, and a text semantic value evaluation value is generated by aggregation. The two are integrated to construct a complete user feedback feature set. At the same time, after standardizing the product attribute data, an attribute feedback feature set is formed. On this basis, the user preference evaluation value and product competition evaluation value are obtained through fusion processing. The push adaptation evaluation value is calculated by combining the preset adjustment coefficient. Finally, the push optimization table is generated by sorting in descending order and the push is executed. This method can accurately capture the user's real needs and match the core advantages of the product, reduce invalid pushes, improve the product exposure and conversion efficiency, and enable users to obtain their favorite products faster. It takes into account the platform operation needs and the user's actual experience, thereby improving the accuracy of the push.

[0017] (2) The push optimization system based on e-commerce big data feedback obtains user feedback data and product attribute data through the data collection unit to ensure that the data is comprehensive and accurate. The feedback extraction unit strictly corresponds to the feature extraction steps in the method and automatically completes the extraction of user feedback feature sets and attribute feedback feature sets without manual intervention in feature processing, which greatly reduces labor costs and ensures the standardization and consistency of feature extraction. The comprehensive evaluation unit efficiently completes the analysis and calculation of user preference evaluation value, product competition evaluation value and push adaptation evaluation value to ensure that the evaluation results are accurate and reliable and avoid human calculation errors. The push optimization unit automatically sorts the push adaptation evaluation values ​​in descending order, generates a push optimization table and executes the push, realizing the automation of push operation, reducing invalid pushes and improving push efficiency.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of a push optimization method based on e-commerce big data feedback according to the present invention.

[0020] Figure 2 This is a block diagram of a push optimization system based on e-commerce big data feedback according to the present invention. Detailed Implementation

[0021] Please see Figure 1 This invention provides a technical solution: a push optimization method based on e-commerce big data feedback, comprising the following steps: acquiring multi-source e-commerce data for several types of goods, including user feedback data and product attribute data; extracting feedback feature sets for each type of goods based on the multi-source e-commerce data, including user feedback feature sets and attribute feedback feature sets; analyzing user preference evaluation values ​​and product competition evaluation values ​​for each type of goods based on the feedback feature sets, and calculating push adaptation evaluation values ​​for each type of goods; and performing push optimization processing for each type of goods based on the push adaptation evaluation values.

[0022] The specific formula for calculating the push adaptation evaluation value for each product is as follows: ;in, , , The order is number 1 The evaluation scores for product push adaptation, user preference, and product competition are as follows: , , The coefficients are, in order, the preference adjustment coefficient, the competition adjustment coefficient, and the difference adjustment coefficient stored in the database.

[0023] It should be noted that, , , The steps to obtain it are as follows:

[0024] A training sample set is constructed by collecting full historical push data, user behavior data, and product competition data from the e-commerce platform's backend database within a preset statistical period (e.g., the past 3 months). Push data includes push records for each product and user feedback data such as clicks, purchases, and blocking after the push. User behavior data includes user preference evaluation values ​​obtained by weighted fusion of quantitative user behavior evaluation values ​​and quantitative user feedback text evaluation values. Product competition data includes product competition evaluation values ​​obtained by weighted fusion of quantitative product attribute evaluation values ​​and competitive landscape data of similar products. Performance metrics data include push click-through rate, conversion rate, and user retention rate, used to measure the actual effect of the push adaptation evaluation values. Using the product's unique identifier as the key, the above data are linked and integrated to form a structured training sample containing user preference evaluation values, product competition evaluation values, and push performance metrics. Abnormal samples (such as invalid records with no user feedback after the push) are removed to ensure the validity and representativeness of the sample set.

[0025] Based on the constructed training sample set, a push notification adaptation effect prediction model is built. The actual effect of the push notification adaptation evaluation value (such as click-through rate and conversion rate) is used as the objective variable, and user preference evaluation value and product competition evaluation value are used as independent variables. The model solves the problem by minimizing the error between the predicted and actual values. , , The initial value is used, and in specific implementation, the core optimization objective is the push conversion rate. An error function is constructed, and the gradient descent algorithm or least squares method is used to iteratively optimize the error function to minimize the prediction error. , , The initial solution; constrain and verify the initial solution to ensure that the range of values ​​for each coefficient conforms to the business logic (e.g., ...). , The value is between 0 and 1. The value is between 0 and 1; if it exceeds the range, it will be corrected by a regularization term.

[0026] Specifically, user feedback data includes user behavior data and user feedback text information. The specific steps for extracting the behavior feedback feature set for each product are as follows: Based on user behavior data, determine the behavior interaction evaluation value for each product; perform semantic parsing processing on the user feedback text information to generate a text semantic value evaluation value for each product; based on the behavior interaction evaluation value and the text semantic value evaluation value, construct the user feedback feature set for each product.

[0027] User behavior data includes product click frequency, add-to-cart rate, percentage of items saved to favorites, effective dwell time, browsing conversion rate, and repeat purchase trigger frequency. The specific steps for generating behavioral interaction evaluation values ​​for each product are as follows:

[0028] Standardize user behavior data by mapping the values ​​of each parameter in the user behavior data to between 0 and 1. Standardization can be achieved using the min-max normalization method.

[0029] Based on standardized user behavior data, a set of behavior weights is generated. Specifically, for each product, the information entropy value is extracted based on the standardized product click frequency, add-to-cart rate, collection ratio, effective dwell time, browsing conversion ratio, and repeat purchase trigger count. Then, the corresponding information entropy values ​​are transformed using the reciprocal suppression mapping function f(x)=1 / (1+x), such as 1 / (1+information entropy value of product click frequency), and summed to obtain the information entropy sum. The transformed information entropy values ​​are then compared with the information entropy sum to obtain the weight coefficients corresponding to each parameter, thus generating the set of behavior weights.

[0030] The standardized user behavior data and behavior weight set are comprehensively processed to generate a behavior interaction evaluation value for each product. Specifically, the product click frequency, add-to-cart rate, collection ratio, effective dwell time, browsing conversion ratio, and repeat purchase trigger count for each product are weighted with the corresponding weight coefficients of the behavior weight set to generate a behavior interaction evaluation value for each product (used to represent the overall effectiveness and interest intensity of all explicit interactive behaviors between the user group and the product).

[0031] The product click frequency value is the total number of times all users click on the product-related display content within a preset statistical period. It is used to characterize the product's initial attractiveness to users. The higher the value, the stronger the initial intention of users to pay attention to the product. The product-related display content includes, but is not limited to, the product item in the e-commerce platform's product list page, the product details page, the product link in push messages, and the product display item in the search results page. It can collect the raw click behavior data of all users in real time within the preset statistical period (which can be flexibly set according to the e-commerce business needs, such as daily, weekly, or monthly; in this embodiment, it is preferably 7 days) through the e-commerce platform's backend data collection interface. The raw click behavior data includes at least the user identifier, the clicked product identifier, the click time, and the click scenario (list page / details page / push link, etc.). Secondly, the collected raw click behavior data is filtered to remove abnormal click data (including but not limited to erroneous clicks by the same user clicking the same product more than 3 times in 1 second, bot-simulated clicks, and clicks corresponding to invalid IP addresses). Finally, based on the product identifier as the grouping basis, the total number of click behaviors corresponding to the product identifier in the filtered data is counted. This total number is the product click frequency value of the product.

[0032] The add-to-cart rate is the probability that a user adds a product to their shopping cart after clicking on it within a preset statistical period. It represents the strength of a user's purchase intention for that product. The process involves obtaining the total number of clicks on the product after filtering (i.e., the number of valid clicks corresponding to the product click frequency value). Through the e-commerce platform's backend shopping cart data collection interface, raw data on all users adding the product to their cart within the same preset statistical period is collected. This raw data includes at least the user identifier, the identifier of the product added to the cart, and the time the product was added. The raw add-to-cart data is then filtered to remove abnormal add-to-cart data (including but not limited to canceling the add-to-cart within one minute or invalid add-to-cart data where the same user adds the same product more than twice). The number of valid add-to-cart entries for that product after filtering is then counted. The add-to-cart rate is calculated using the formula: Add-to-cart rate = Number of valid add-to-cart entries ÷ Number of valid clicks. If the number of valid clicks is 0, the add-to-cart rate for that product is 0. The result is rounded to four decimal places, completing the acquisition of the add-to-cart rate.

[0033] The collection percentage is the probability that a user adds a product to their favorites after clicking on it within a preset statistical period. It uses the same number of valid clicks as the product's add-to-cart rate to ensure consistent statistical standards. The data is collected from the e-commerce platform's backend collection data collection interface, gathering raw data on all users' collections of the product within the same preset statistical period. This raw data includes at least the user ID, the collected product ID, and the collection time. The raw collection data is then filtered to remove abnormal collection data (including but not limited to uncollecting within 10 minutes of collection or invalid collections from the same user repeatedly collecting the same product). The number of valid collections for the product after filtering is then counted. The collection percentage is calculated using the formula: Collection Percentage = Number of Valid Collections ÷ Number of Valid Clicks. If the number of valid clicks is 0, the collection percentage is 0. The result is rounded to four decimal places to obtain the collection percentage for the product.

[0034] The effective dwell time value is the average effective dwell time of all users browsing the product details page within a preset statistical period. This is achieved by collecting raw data from all users accessing the product details page within the preset statistical period through the e-commerce platform's backend page access data collection interface. The raw data includes at least the user identifier, the accessed product identifier, the time of entering the details page, and the time of leaving the details page. The raw access data is then filtered to remove invalid access data (including but not limited to accidental clicks with a dwell time of less than 3 seconds after entering the details page, abnormal dwell times caused by page loading failures, and duplicate data from the same user accessing the same page repeatedly within the same time period). For each valid access data entry after filtering, the effective dwell time of a single user browsing the product details page is calculated using the formula: Single Effective Dwell Time = Time Leaving Details Page - Time Entering Details Page. The sum of all valid dwell times is then divided by the number of valid access data entries to obtain the effective dwell time value (i.e., average effective dwell time) for the product.

[0035] The browsing conversion ratio is the probability that a user will complete a purchase after browsing a product within a preset statistical period. It is calculated by accessing the data collection interface through the e-commerce platform's backend page to collect valid browsing data for all users browsing the product within the same preset statistical period, and counting the number of valid browsing visits (consistent with the number of valid access data entries used to calculate the effective dwell time, i.e., valid browsing visits = number of valid access data entries). It also collects the original order data for all users purchasing the product within the same statistical period through the e-commerce platform's backend order data collection interface. The original data must include at least the user identifier, the purchased product identifier, the order time, and the order status. The original order data is then filtered to remove invalid orders (including but not limited to canceled orders, returned orders, refunded orders, and abnormal orders), and the number of valid purchases for the corresponding product is counted. The browsing conversion ratio is calculated using the formula: Browsing Conversion Ratio = Number of Valid Purchases ÷ Number of Valid Browsing Visits. If the number of valid browsing visits is 0, the browsing conversion ratio is 0, thus completing the acquisition of the browsing conversion ratio.

[0036] The repeat purchase trigger count is the total number of times all users repurchase the product within a preset statistical period. This is achieved by collecting valid order data (consistent with the valid order data used in the browsing conversion ratio calculation, i.e., orders that have been completed and not returned / refunded) from all users within the preset statistical period through the e-commerce platform's backend order data collection interface. The valid order data is then correlated, using both user and product identifiers as dual filtering conditions to identify multiple valid order records (i.e., repeat purchase order records) from the same user for the same product. The total number of repeat purchase order records is then calculated. Each time a user completes a repeat purchase of the same product (i.e., each valid order other than the initial purchase) is counted as one repeat purchase trigger. The total number of repeat purchase triggers for all users is the repeat purchase trigger count for that product. If no repeat purchase order records exist, the repeat purchase trigger count for that product is 0.

[0037] User feedback text information refers to unstructured text data related to products, actively generated by users. Its main sources include: user review data: evaluation text submitted by users after purchasing products, typically containing subjective descriptions of product attributes, user experience, advantages, and disadvantages; and user Q&A data: questions raised by potential buyers on the product details page regarding product details, functions, and specifications, as well as responses provided by other users or sellers. The specific steps for generating a textual semantic value assessment value for each product are as follows:

[0038] User feedback text information is input into a pre-trained adaptive language understanding model to extract a structured semantic feature set for each product, specifically:

[0039] First, the collected user feedback text information is preprocessed, including word segmentation, stop word removal, part-of-speech tagging, and abnormal text filtering. Abnormal text filtering removes invalid text without actual semantic meaning (including but not limited to pure symbols, repetitive meaningless statements, and chatty text unrelated to the product), while retaining text containing valid semantic meaning such as product attributes, user experience, functional questions, and answers. The preprocessed valid text is then input into a pre-trained adaptive language understanding model (preferably a BERT or ERNIE model finely tuned for the e-commerce domain). This model combines sequence labeling and information extraction techniques to perform deep semantic parsing on each valid text segment, accurately identifying and extracting attribute-evaluation pairs or aspect-opinion pairs, while also labeling the polarity of each element (only the element relationship is labeled).

[0040] For example, from the user Q&A "Does this refrigerator have a frost-free function?", we can analyze the following: {Aspect: frost-free function, Evaluation: very useful, Polarity: positive}. From the user comment "The body is relatively thin, saving space", we can analyze the following: {Aspect: The body is thick, Evaluation: relatively thin, Polarity: positive}.

[0041] The semantic elements parsed from all valid text related to the product are summarized, and duplicate and redundant semantic elements are removed. A structured semantic element feature set for each product is constructed in list form. This feature set clearly presents all the core focus of user discussions around the product and related evaluations.

[0042] Statistical processing is performed on the structured semantic element feature set to generate a semantic index set for each product, specifically as follows:

[0043] Based on the structured semantic element feature set, quantitative statistics and information measurement methods are used to calculate the coverage breadth value, information concentration value, and element uniqueness value, which together constitute the semantic index set for each product. The specific statistical calculation methods for each index are as follows:

[0044] Coverage breadth value is the total number of different product attributes or discussion aspects involved in the structured semantic element feature set. During the statistical process, synonymous attributes need to be merged (for example, battery life and battery life capability are merged into the same attribute). The higher the number, the more comprehensive the focus of user discussions around the product, and the higher the coverage breadth index value.

[0045] Information concentration is calculated by counting the total number of words in all valid user feedback texts and then counting the total number of valid attribute-evaluation pairs or aspect-opinion pairs in the structured semantic element feature set. It is obtained by the formula Information Concentration = Total number of valid semantic elements ÷ Total number of valid text words × 100. That is, the number of valid semantic elements contained in every 100 words of valid text. The higher the value, the higher the effective density of text information.

[0046] The uniqueness value of an element is achieved by combining the information entropy algorithm with frequency statistics. First, the frequency of occurrence of each semantic element in the structured semantic element feature set is counted. Then, the structured semantic element feature set of other competing products in the same category is obtained, and the frequency of occurrence of each semantic element in the competing product text is counted. The uniqueness coefficient of each semantic element is calculated using the information entropy formula. The average value of the uniqueness coefficients of all semantic elements is taken as the uniqueness value of the element for the product. If a certain semantic element appears frequently in the product text but has a very low frequency in the competing product text, the higher the uniqueness coefficient of the element, the higher the overall uniqueness index value of the element.

[0047] The semantic indicator set is aggregated to generate a textual semantic value assessment value for each product. Specifically, the coverage breadth value, information density value, and element uniqueness value of the semantic indicator set for each product are standardized and mapped to a value between 0 and 1. Then, a weighted average is applied (the weight coefficients of each parameter in this weighted average are consistent with the generation logic of the behavioral weight set) to obtain the textual semantic value assessment value for each product. The higher the value, the richer, more unique, and more value-densed the textual information generated by the user around the product is, and the higher the consideration weight should be given to the product in subsequent push decisions.

[0048] This implementation plan specifies detailed acquisition methods and abnormal data removal rules for all parameters of user behavior data. After standardization, a behavior weight set is generated using information entropy to make the influence weight of each parameter more reasonable. This ensures that the behavior interaction evaluation value can truly reflect the user's interest in the product and purchase intention. For user feedback text, invalid content is first pre-processed to remove it, and then semantic elements are extracted using a model finely tuned in the e-commerce field. The text value is quantified by coverage breadth, information concentration, and element uniqueness. The weight logic during aggregation is consistent with the behavior data to ensure unified processing standards. This approach can comprehensively capture the user's interaction intention and the key points of text feedback while avoiding interference from invalid data. This makes the constructed user feedback feature set more realistic and provides accurate and reliable data support for subsequent push adaptation evaluation, effectively helping to optimize push implementation.

[0049] Specifically, the product attribute data includes product value, cumulative sales volume, inventory quantity, number of updates / items, number of associated products (complementary / substitute products), delivery timeliness, and after-sales guarantee level. The specific steps for extracting the attribute feedback feature set for each product are as follows:

[0050] Preprocess the product attribute data, that is: standardize each parameter in the product attribute data (to remove the unit and map its value between 0 and 1).

[0051] Based on the preprocessing results, feature synthesis is performed to generate attribute feedback feature sets for each product, including supply guarantee value, market vitality value, and associated sales potential value, specifically:

[0052] The inventory quantity, delivery timeliness, and after-sales guarantee level of each product after preprocessing are weighted to obtain its corresponding supply guarantee value. It should be noted that in this weighting process, the delivery timeliness value needs to be transformed using the reciprocal suppression mapping function f(x)=1 / (1+x).

[0053] The cumulative sales value and update iteration number of each product after preprocessing are weighted to obtain its corresponding market vitality value;

[0054] The preprocessed commodity value and related commodity quantity values ​​are weighted to obtain the corresponding associated sales potential value. It should be noted that in this weighting process, the commodity value needs to be transformed using the reciprocal suppression mapping function f(x)=1 / (1+x).

[0055] It should be noted that in the above weighting process, the weight coefficients corresponding to each parameter are consistent with the logic for obtaining the behavior weight set.

[0056] The product pricing value refers to the regular selling price (i.e., the marked price or crossed-out price) displayed on the e-commerce platform's sales page, excluding short-term promotional discounts. This price is retrieved from the price field or marked price field of the platform's backend product information management database. To ensure consistency, the product's promotional status field must be considered when retrieving this value. If the product is currently participating in a limited-time discount promotion, its regular selling price will still be used, not the promotional price.

[0057] The cumulative sales value is the total sales quantity of all successfully completed orders for this product within a preset statistical period. This is calculated by querying the platform's order transaction database. Using the product's unique identifier (SKU_ID) as the key, records with order statuses of "successful transaction" or "received" are filtered, and the sales quantity field is summed. The result is the cumulative sales value. Orders canceled due to refunds must be excluded from the calculation.

[0058] The inventory quantity value refers to the actual physical inventory quantity of this product available for sale in the warehouse at the current moment. It is obtained by calling or querying the real-time interface or database table of the inventory management system. This value is usually stored in the available inventory field in the inventory center. It is important to distinguish between available inventory and locked inventory (orders placed but not yet shipped).

[0059] The update iteration count is the number of records where key attributes (including but not limited to title, main image, details page, and specifications) of a product are substantially modified and successfully submitted within a preset statistical period. This is achieved by setting a counter for each product in the product information editing log table. Each time product information is approved and an update operation is completed, the counter for that product is incremented by 1. The update iteration count can be obtained directly from this counter.

[0060] The associated product quantity value is the sum of the number of complementary products and substitute products automatically matched and marked by the e-commerce platform's product association engine within a preset statistical period. Complementary products refer to those used in conjunction with the product in complementary scenarios, while substitute products refer to products of the same category and function that are interchangeable. When retrieving this value, the platform's product association database is queried using the product's unique identifier (SKU_ID) as the key. Valid records with association types of complementary and substitute products are filtered, and duplicate associations and invalid products not yet listed are removed. The number of product entries meeting the criteria is then counted, and the result is the associated product quantity value.

[0061] The shipping timeliness value is the standard promised time by the merchant for the product, from the time the user completes payment to the time the product is shipped. It represents the product's shipping response speed, and the unit is uniformly in hours. It is obtained by: reading the corresponding shipping timeliness field for the product from the platform's merchant delivery management database or the product shipping promise table; if the product does not have a separately set timeliness, then reading the default shipping promise timeliness of its respective store; converting descriptions such as "ship within 24 hours" and "ship within 48 hours" into numerical values ​​(e.g., 24, 48); assigning values ​​based on the platform's default standard timeliness for items that cannot be explicitly promised; the final value obtained is the shipping timeliness value.

[0062] The after-sales guarantee level is a standardized service level categorized by e-commerce platforms based on the official after-sales benefits enabled for the product. It is typically divided into three levels: basic guarantee, intermediate guarantee, and advanced guarantee. This level is used to quantify the quality of after-sales service. To obtain the guarantee level, the platform's after-sales guarantee service activation table is queried using the product ID or store ID as the key. This identifies whether the product has activated services such as 7-day no-reason return, shipping insurance, genuine product guarantee, extended warranty, and door-to-door return pickup. The corresponding guarantee level is matched according to the platform's preset level determination rules, and the level is converted into a quantitative value (e.g., basic = 1, intermediate = 2, advanced = 3). This quantitative value is then directly read as the after-sales guarantee level value.

[0063] The specific steps for analyzing the competitive evaluation value of each product are as follows: First, read the supply guarantee value, market vitality value, and associated sales potential value, and normalize them. Second, fuse the normalized supply guarantee value, market vitality value, and associated sales potential value to obtain the competitive evaluation value of each product. Specifically, this involves weighted fusion of the supply guarantee value, market vitality value, and associated sales potential value to obtain the competitive evaluation value of each product. Furthermore, during the weighting process, the weight coefficients corresponding to each parameter are consistent with the logic for obtaining the behavioral weight set.

[0064] In this implementation plan, the parameters are first standardized to eliminate differences in units and values. Then, through reasonable feature synthesis, supply guarantee, market vitality, and associated sales potential are extracted. This simplifies the data dimensions while fully preserving the key characteristics of the product. Dedicated mapping functions are used to adjust delivery timeliness and product pricing to adapt the data to the actual business scenario. The weight calculation logic and behavioral data are kept consistent throughout the process to ensure the consistency of the overall solution. Finally, the product competitiveness evaluation value is obtained through normalization and weighted fusion, which can accurately reflect the actual level of the product in the market and avoid the bias caused by single data. This makes the push adaptation judgment more in line with the actual operation of e-commerce.

[0065] Specifically, the steps for analyzing the user preference evaluation value of each product are as follows: Read the behavioral interaction evaluation value and the text semantic value evaluation value, and perform normalization processing; perform comprehensive processing on the normalized behavioral interaction evaluation value and the text semantic value evaluation value to obtain the user preference evaluation value of each product. Specifically, the normalized behavioral interaction evaluation value and the text semantic value evaluation value are used as two independent feature dimensions and input into the attention fusion module. The core of this module is a lightweight attention network layer, whose function is to simulate the attention allocation mechanism in the human cognitive process, automatically identify and emphasize the feature signals that are more critical to the current user preference judgment.

[0066] The attention network layer performs nonlinear transformations on the two input features to generate their respective feature representation vectors and calculates the importance score for each feature. This process is achieved by evaluating the correlation between each feature representation and a learnable context vector. The higher the correlation, the more important the feature is in the current preference evaluation task. Then, the Softmax function is used to normalize these importance scores and transform them into two dynamic attention weight coefficients. The sum of these two weight coefficients is one, and the magnitude of each coefficient is completely determined by the content of the input feature itself. This adaptively reflects the relative contribution of behavioral interaction data and textual semantic data to this specific evaluation.

[0067] Using these two dynamically generated attention weight coefficients, the original behavioral interaction evaluation value and text semantic value evaluation value are scaled respectively. Finally, the two scaled values ​​are added together to obtain a preliminary comprehensive score. This comprehensive score is the user preference intensity representation after the attention mechanism is fused. To make it more suitable for subsequent ranking and recommendation tasks, this comprehensive score can be input into a Sigmoid activation function for smoothing and range constraints, and finally output a user preference evaluation value between zero and one.

[0068] In this implementation plan, the evaluation values ​​for behavioral interaction and text semantics are first normalized to eliminate numerical bias and lay the foundation for subsequent fusion. Then, a lightweight attention fusion module simulates human attention allocation logic, automatically identifying the importance of the two dimensions and generating dynamic weights. This eliminates the need for manual setting and adaptively matches the user feedback characteristics of different products, avoiding biases caused by fixed weights. Finally, Softmax normalization of the weights and Sigmoid constraint on the numerical range stabilize the evaluation values ​​between 0 and 1, perfectly suited for subsequent push ranking tasks. This approach balances the intuitiveness of user interaction behavior with the deeper needs of text feedback, more realistically reflecting user preferences and providing accurate support for subsequent push adaptation judgments.

[0069] Specifically, the steps for optimizing push notifications for each product are as follows: The push adaptation evaluation values ​​are sorted in descending order to generate a product push optimization table, which is as follows:

[0070] All products to be recommended are sorted in descending order of their push compatibility evaluation scores, forming an initial global ranking list. The product ranked first in this list has the highest push compatibility evaluation score. An improved roulette wheel algorithm is then used to intelligently process this ranking, introducing controllable diversity while respecting the overall ranking.

[0071] Each ranking position in the initial list is assigned a base weight, which follows the principle of exponential decay. That is, the higher the ranking, the higher the weight. However, the rate at which the weight decreases as the ranking decreases is controlled by a configurable decay adjustment factor. Specifically, the first-ranked product has the highest weight, the second-ranked product's weight is the first-ranked product's weight multiplied by a decay coefficient less than one, the third-ranked product's weight is the second-ranked product's weight multiplied by the same decay coefficient, and so on. The value of this decay adjustment factor is stored in the system database, and administrators can adjust it to control the conservatism (biased towards high-scoring products) or the exploratory nature (giving more opportunities to lower-ranked products) of the final recommendation list.

[0072] The base weights of all products calculated based on their rankings are summed to obtain the total weight. Then, the base weight of each product is divided by this total weight to calculate a standardized selection probability for each product. Thus, each product obtains a probability value that is positively correlated with its ranking and the sum of the probabilities of all products is one.

[0073] Generate a new sequence (i.e., a product recommendation optimization table) containing a specified number of products, which is accomplished through multiple probability samplings:

[0074] During each sampling, all items are treated as a roulette wheel, and the area occupied by each item is proportional to its calculated selection probability.

[0075] Generate a random number between zero and one. The area of ​​the roulette wheel where the random number lands will select the corresponding product.

[0076] To ensure that the same product does not appear repeatedly in the final list, once a product is selected, it will be temporarily masked in subsequent sampling rounds of this list construction (that is, its probability will be temporarily set to zero), while the selection probability of the remaining products will be immediately recalculated and normalized to ensure that the sum of the probabilities of subsequent sampling is still one.

[0077] Repeat the above sampling process until a predetermined number of items are selected. These items, in the order they are selected, constitute the initial optimized sequence.

[0078] The optimization sequence generated by the above process is the final product recommendation optimization table;

[0079] Based on the product push optimization table, the product push operation is carried out in the following manner: pushes are made sequentially according to the sorting order of the product push optimization table.

[0080] In this implementation plan, the push compatibility evaluation values ​​are sorted in descending order to ensure that products with the highest compatibility scores are prioritized for push notifications, maintaining the core principle of accurate pushes. An improved roulette wheel algorithm is then introduced, allocating base weights according to an exponential decay principle. Administrators can adjust the decay factor to flexibly control whether pushes favor high-scoring products or give more opportunities to lower-scoring products, adapting to different operational needs. Sampling is based on probability selection, avoiding duplicate product appearances, and probabilities are recalculated to ensure rationality. The final optimized push table prioritizes pushes of highly compatible products while introducing controllable diversity to avoid pushes being too homogeneous and improve conversion efficiency.

[0081] Please see Figure 2This invention provides a technical solution: a push optimization system based on e-commerce big data feedback, comprising: a data acquisition unit for acquiring multi-source e-commerce data of several products, including user feedback data and product attribute data; a feedback extraction unit for extracting feedback feature sets for each product based on the multi-source e-commerce data, including user feedback feature sets and attribute feedback feature sets; a comprehensive evaluation unit for analyzing user preference evaluation values ​​and product competition evaluation values ​​for each product based on the feedback feature sets, and calculating push adaptation evaluation values ​​for each product; and a push optimization unit for performing push optimization processing on each product based on the push adaptation evaluation values.

[0082] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A push optimization method based on e-commerce big data feedback, characterized in that, Includes the following steps: Acquire multi-source e-commerce data for several types of products, including user feedback data and product attribute data; Based on the multi-source e-commerce data, feedback feature sets for each product are extracted, including user feedback feature sets and attribute feedback feature sets. Based on the feedback feature set, the user preference evaluation value and product competition evaluation value of each product are analyzed, and the push adaptation evaluation value of each product is calculated. Based on the aforementioned push adaptation evaluation value, push optimization is performed for each product.

2. The push optimization method based on e-commerce big data feedback according to claim 1, characterized in that, The user feedback data includes user behavior data and user feedback text information. The specific steps for extracting the behavior feedback feature set for each product are as follows: Based on the user behavior data, determine the behavioral interaction evaluation value for each product; The user feedback text information is semantically parsed to generate a textual semantic value assessment value for each product; Based on the behavioral interaction evaluation value and the text semantic value evaluation value, a user feedback feature set is constructed for each product.

3. The push optimization method based on e-commerce big data feedback according to claim 2, characterized in that, The specific steps for generating behavioral interaction evaluation values ​​for each product are as follows: The user behavior data is standardized. Based on the standardized user behavior data, a set of behavior weights is generated. The standardized user behavior data and the behavior weight set are combined to generate a behavior interaction evaluation value for each product.

4. The push optimization method based on e-commerce big data feedback according to claim 2, characterized in that, The specific steps for generating the textual semantic value assessment value for each product are as follows: The user feedback text information is input into a pre-trained adaptive language understanding model to extract the structured semantic element feature set for each product; The structured semantic element feature set is statistically processed to generate a semantic index set for each product; The semantic index set is aggregated to generate a textual semantic value assessment value for each product.

5. The push optimization method based on e-commerce big data feedback according to claim 1, characterized in that, The specific steps for extracting the attribute feedback feature set for each product are as follows: The product attribute data is preprocessed; Based on the preprocessing results, feature synthesis processing is performed to generate attribute feedback feature sets for each product, including supply guarantee value, market vitality value, and associated sales potential value.

6. The push optimization method based on e-commerce big data feedback according to claim 2, characterized in that, The specific steps for analyzing user preference evaluation values ​​for each product are as follows: Read the behavioral interaction evaluation value and the text semantic value evaluation value, and perform normalization processing; The normalized behavioral interaction evaluation value and the text semantic value evaluation value are combined to obtain the user preference evaluation value for each product.

7. The push optimization method based on e-commerce big data feedback according to claim 5, characterized in that, The specific steps for analyzing the competitive evaluation value of each product are as follows: Read the supply guarantee value, the market vitality value, and the associated sales potential value, and perform normalization processing; The normalized supply security value, market vitality value, and associated sales potential value are merged to obtain the product competition evaluation value for each product.

8. The push optimization method based on e-commerce big data feedback according to claim 1, characterized in that, The specific formula for calculating the push adaptation evaluation value for each product is as follows: ； in, , , The order is number 1 The evaluation scores for product push adaptation, user preference, and product competition are as follows: , , The coefficients are, in order, the preference adjustment coefficient, the competition adjustment coefficient, and the difference adjustment coefficient stored in the database.

9. The push optimization method based on e-commerce big data feedback according to claim 1, characterized in that, The specific steps for optimizing push notifications for each product are as follows: The push adaptation evaluation values ​​are sorted in descending order to generate a product push optimization table; Based on the product push optimization table, product push operations are performed.

10. A push optimization system based on e-commerce big data feedback, employing the push optimization method based on e-commerce big data feedback as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire multi-source e-commerce data for several types of products, including user feedback data and product attribute data. The feedback extraction unit is used to extract feedback feature sets for each product based on the multi-source e-commerce data, including user feedback feature sets and attribute feedback feature sets. The comprehensive evaluation unit is used to analyze the user preference evaluation value and product competition evaluation value of each product based on the feedback feature set, and to calculate the push adaptation evaluation value of each product. The push optimization unit is used to perform push optimization processing on each product based on the push adaptation evaluation value.

Images