E-commerce platform commodity sorting method based on generative adversarial network and cloud platform

Abstract

The invention relates to an e-commerce platform commodity sorting method based on a generative adversarial network and a cloud platform, and belongs to the technical field of e-commerce commodity sorting. Features are extracted by receiving user behavior data, a candidate sorting list is output by a generator, and sorting information is output by a discriminator; the method comprises the following steps: for cold-start commodities, acquiring attribute characteristics of the cold-start commodities to judge an approximate competitive product quantity, crawling associated information to calculate a priority value when the attribute characteristics are lower than a preset value, generating virtual user characteristics in combination with a conversion algorithm, inputting a model, and then outputting updated sorting information, thereby improving the sorting accuracy of the cold-start commodities, enhancing the generalization ability of the model, and improving the sorting efficiency of the cold-start commodities. And the user experience and the platform target are balanced, and remarkable technical innovation and practical value are achieved.

Description

Technical Field

[0001] This invention relates to the field of e-commerce product sorting technology, and in particular to an e-commerce platform product sorting method and cloud platform based on generative adversarial networks. Background Technology

[0002] In e-commerce platforms, product ranking is a core technology affecting user experience and platform operational efficiency. Traditional product ranking methods are mostly based on models such as collaborative filtering and logistic regression, relying primarily on historical user interaction data (such as clicks and purchase behavior) and basic product attributes (such as price and sales volume) for ranking. However, for cold-start products (i.e., newly listed products or products lacking user interaction data), the lack of effective user feedback makes it difficult to accurately assess their market potential and user preferences, leading to biased ranking results. This can result in insufficient exposure for high-quality new products or inaccurate recommendations that negatively impact the user browsing experience. In existing technologies, some solutions attempt to improve the ranking of cold-start products by generating virtual user interaction data through generative adversarial networks (GANs). However, these solutions generally suffer from the following problems: First, the feature mining of cold-start products is not comprehensive enough, relying only on single-dimensional product attributes or simple user behavior simulations, without fully integrating external correlation information such as supply chain data and social media popularity. Second, the collaborative training mechanism between the generator and discriminator is not optimized enough, resulting in differences between the generated virtual data and real user behavior patterns, leading to insufficient generalization ability of the ranking model. Third, the correlation analysis between cold-start products and existing products is relatively crude, lacking a systematic method for priority value calculation and virtual feature generation, making it difficult to ensure ranking diversity while also considering platform business objectives (such as profit and conversion rate). Summary of the Invention

[0003] The main objective of this invention is to provide a product ranking method and cloud platform for e-commerce platforms based on generative adversarial networks, which can improve the accuracy of cold start product ranking, enhance the generalization ability of the model, and balance user experience and platform goals, thus having significant technological innovation and practical value.

[0004] To achieve the above objectives, the present invention provides a product ranking method for e-commerce platforms based on generative adversarial networks, comprising the following steps: Receive user search requests and / or browsing behavior data, extract first user features and first product attribute features, the first user features include time decay features of user historical clicks, add-to-cart, and purchase behavior, and the first product attribute features include product category, price, brand, and review information; Based on a pre-trained adversarial network model, the model includes a generator and a discriminator. The generator is a multi-layer deception neural network that takes a first user feature and a first product attribute feature as input and outputs a candidate product ranking list. The discriminator is a multi-layer recognition neural network that takes a real product ranking list and a candidate product ranking list generated by the generator as input and outputs and displays the ranking information. For cold start products that need to be inserted into the sorting information, the second product attribute features corresponding to the cold start products are obtained, and the number of approximate competitors is judged by the second product attribute features. If the number of approximate competitors is lower than a preset value, the priority value is calculated by crawling the association information of the cold start products using a preset association algorithm. Using a preset conversion algorithm, a virtual second user feature is generated by combining the priority value. The second user feature and the second product attribute feature are then input into the adversarial network model, and finally the sorting information after the cold start product is inserted is re-output and displayed.

[0005] Furthermore, the generator is a multi-layer deception neural network, and the step of taking the first user feature and the first product attribute feature as input and outputting a candidate product ranking list includes: The first user features and the first product attribute features input into the multi-layer deception neural network are normalized to form a set of feature vectors with uniform dimensions. The feature vector set is initially mapped using the underlying network. Different weight matrices are set for each layer of the underlying network. The time decay features are weighted in a time series to obtain the user feature vector. The product category, price, brand and evaluation information are categorized and converted to obtain the product attribute feature vector. The middle layer network cross-concatenates the user feature vector and the product attribute feature vector, and generates user-product interaction features by element-wise multiplication and dimension concatenation. After processing by the activation function, the features are input into the high-level network. The high-level network calculates the ranking score for each product through a fully connected layer, and sorts the candidate products in ascending or descending order based on the ranking scores to generate a candidate product ranking list.

[0006] Furthermore, the discriminator is a multi-layer recognition neural network. The steps of taking the real product ranking list and the candidate product ranking list generated by the generator as input, and outputting and displaying the ranking information, include: Sequence features were extracted from both the real product sorting list and the candidate product sorting list. The extracted features included the position index of each product in the list, the user's historical click conversion rate, the product price gradient, and the brand's market share. The extracted sequence features are standardized by a low-level recognition network. The feature vectors of the real product sorting list and the candidate product sorting list are then input into two parallel sub-networks to obtain the real feature representation and the candidate feature representation. The mid-layer contrastive network performs element-wise difference calculation between the true feature representation and the candidate feature representation to generate a feature difference vector, and then combines the sentiment value of the product evaluation information to perform weighted correction on the difference vector. The high-level evaluation network calculates the ranking quality score through a fully connected layer, outputs ranking information including the ranking quality score and the position of the differentiated products, and displays it on the cloud platform.

[0007] Further, for cold-start products that need to be inserted into the sorting information, the step of obtaining the second product attribute features corresponding to the cold-start products and using the second product attribute features to determine the approximate number of competitors includes: By using the product category, price, brand, and functional description included in the second product attribute features, a set of candidate competitors that match the product category tags is selected from the historical database of the cloud platform; For each competitor in the candidate competitor set, calculate its overlap with the cold start product in terms of price range, its brand identity correlation, and its semantic similarity to the functional description text. The number of competitors in the candidate competitor set that meet the following criteria is recorded as the approximate competitor count: price overlap ≥ the first threshold, brand relevance ≥ the second threshold, and semantic similarity ≥ the third threshold.

[0008] Furthermore, if the number of similar competing products is lower than a preset value, the step of crawling the association information of cold-start products using a preset association algorithm and calculating the priority value includes: The system crawls supplier cooperation ratings and inventory turnover cycle data for cold start products from the cloud platform supply chain database, and crawls recent mentions and KOL recommendation records of cold start product keywords from pre-bound social media platforms. In addition, it crawls sales growth rate data of complementary product categories over the past three months from the cloud platform sales database to obtain the associated information. The numericalization of the associated information involves mapping the supplier cooperation rating and inventory turnover cycle data to a stability coefficient of 0-1, converting mentions and recommendation records on social media platforms into a market popularity index, and converting the sales growth rate of complementary categories into an association promotion coefficient. The priority value of cold-start products is calculated based on the association algorithm.

[0009] Furthermore, the formula for the association algorithm is: Where P is the priority value, S is the stability coefficient, H is the market popularity index, and C is the correlation promotion coefficient. These are the weight parameters for the corresponding coefficients, and they satisfy... .

[0010] Furthermore, the step of generating a virtual second user feature using a preset conversion algorithm and the priority value includes: Determine the proximity information [n-1, n+1] of the product corresponding to the priority value from the currently displayed sorting information; Based on the time decay characteristics of users' historical clicks, add-to-cart, and purchase behaviors corresponding to the products in the near-domain information, the conversion algorithm is used to generate virtual second user characteristics.

[0011] Furthermore, the conversion algorithm is as follows:

[0012] Where V is the virtual second user feature vector, and N is the near-field information set [n-1, n+1]. For the historical click count, add-to-cart count, and purchase count of product i in the near domain, The time interval between the corresponding action and the current time. The coefficient representing the time decay characteristic. This represents the maximum number of historical actions recorded on the platform. These are the feature dimension basis vectors corresponding to the behavior.

[0013] Further, the step of inputting the second user feature and the second product attribute feature into the adversarial network model, and finally re-outputting and displaying the ranking information after the cold start product update, includes: The second user feature and the second product attribute feature are input into the adversarial network model; Following the same generator and discriminator operating modes as the first user feature and the first product attribute feature, the cold start product is inserted into the corresponding sorting position to obtain the updated sorting information.

[0014] This invention proposes a cloud platform for product ranking on an e-commerce platform based on generative adversarial networks, comprising: The receiving unit is used to receive user search requests and / or browsing behavior data, and extract first user features and first product attribute features. The first user features include time decay features of user historical clicks, add-to-cart, and purchase behaviors. The first product attribute features include product category, price, brand, and evaluation information. The model unit is used to build upon a pre-trained adversarial network model, which includes a generator and a discriminator. The generator is a multi-layer deception neural network that takes a first user feature and a first product attribute feature as input and outputs a candidate product ranking list. The discriminator is a multi-layer recognition neural network that takes a real product ranking list and a candidate product ranking list generated by the generator as input and outputs and displays the ranking information. The request unit is used to obtain the second product attribute features corresponding to the cold start product for the cold start product that is requested to be inserted into the sorting information, and to judge the approximate number of competitors through the second product attribute features. If the approximate number of competitors is lower than a preset value, the priority value is calculated by crawling the association information of the cold start product using a preset association algorithm. The update unit is used to generate a virtual second user feature by combining the priority value with a preset transformation algorithm, and input the second user feature and the second product attribute feature into the adversarial network model, and finally output and display the sorting information after the cold start product update. The product ranking method and cloud platform for e-commerce platforms based on generative adversarial networks provided by this invention have the following beneficial effects: (1) By obtaining the second product attribute characteristics of cold start products and making a judgment on the quantity of similar competitors, and combining supply chain data, social media popularity and sales of complementary categories and other multi-dimensional related information to calculate priority value, the problem of sorting deviation caused by lack of user interaction data for cold start products can be effectively solved, so that new products can get more reasonable exposure opportunities.

[0015] (2) By utilizing the adversarial training mechanism between the generator and the discriminator in the generative adversarial network, and combining the time decay characteristics of user historical behavior and the deep feature extraction of product attributes, a candidate ranking list that is closer to real user preferences is generated, thereby improving the model's ability to fit complex user behavior and product features. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a product sorting method for an e-commerce platform based on generative adversarial networks in one embodiment of the present invention. Figure 2 This is a structural block diagram of an e-commerce platform product ranking cloud platform based on generative adversarial networks in one embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Reference Figure 1 This is a flowchart illustrating a product ranking method for e-commerce platforms based on generative adversarial networks proposed in this invention. The method includes the following steps: S1, receive user search requests and / or browsing behavior data, extract first user features and first product attribute features, the first user features include time decay features of user historical clicks, add-to-cart, and purchase behavior, the first product attribute features include product category, price, brand, and evaluation information; S2, based on a pre-trained adversarial network model, the model includes a generator and a discriminator, wherein the generator is a multi-layer deception neural network, which takes a first user feature and a first product attribute feature as input and outputs a candidate product ranking list; the discriminator is a multi-layer recognition neural network, which takes a real product ranking list and a candidate product ranking list generated by the generator as input and outputs and displays the ranking information. S3. For cold start products that need to be inserted into the sorting information, obtain the second product attribute features corresponding to the cold start products, and judge the approximate number of competitors through the second product attribute features. If the approximate number of competitors is lower than a preset value, then use a preset association algorithm to crawl the association information of the cold start products and calculate the priority value. S4. Using a preset conversion algorithm, a virtual second user feature is generated by combining the priority value. The second user feature and the second product attribute feature are then input into the adversarial network model. Finally, the sorting information after the cold start product is inserted is re-output and displayed.

[0019] In one embodiment, a data acquisition module deployed on the e-commerce platform server captures user search requests (such as entered keywords and filter combinations) and browsing behavior data (including timestamps and object IDs for actions such as clicking on product detail pages, adding to cart, and completing purchases) in real time, and stores this data in a user behavior database. For the extraction of the first user feature, an exponential decay model is used to weight the behavior time based on the user's historical clicks, add-to-cart actions, and purchases. Specifically, this is achieved through a time decay function. (where t is the number of days since the event occurred.) Configurable decay rate parameters assign dynamic weights to behaviors at different time points. For example, the weight of click behaviors within the last 7 days is significantly higher than that of similar behaviors 30 days ago, thus forming a user behavior feature vector containing time-sensitive characteristics. For the first product attribute feature, the product category and brand are converted into high-dimensional sparse vectors or dense embedding vectors using one-hot encoding or word embedding technology. Price information is mapped to the [0,1] interval through normalization. Evaluation information is analyzed for sentiment tendency using natural language processing technology (such as building a sentiment classification model based on the bidirectional long short-term memory network BILSTM to output a sentiment score from -1 to 1) and extracting keywords. Finally, all kinds of attribute features are spliced ​​into a product attribute vector of a unified dimension. For example, when a user searches for "laptop", the system simultaneously collects the user's click records on electronic products in the past 30 days (including the timestamp of each click), the IDs of 3 laptops added to the cart, and information on a peripheral product purchased 15 days ago. After calculating the weight coefficient of each behavior through time decay, it is structured with the corresponding product's category (computer-laptop), price (0.82 after normalization of 5999 yuan), brand (Lenovo's corresponding embedded vector), and evaluation sentiment score (0.91) and other attributes to form a standardized feature vector set for the input model.

[0020] In one embodiment, the generator is a multi-layer deception neural network, and the steps of taking a first user feature and a first product attribute feature as input and outputting a candidate product ranking list include: The first user features and the first product attribute features input into the multi-layer deception neural network are normalized to form a set of feature vectors with uniform dimensions. The feature vector set is initially mapped using the underlying network. Different weight matrices are set for each layer of the underlying network. The time decay features are weighted in a time series to obtain the user feature vector. The product category, price, brand and evaluation information are categorized and converted to obtain the product attribute feature vector. The middle layer network cross-concatenates the user feature vector and the product attribute feature vector, and generates user-product interaction features by element-wise multiplication and dimension concatenation. After processing by the activation function, the features are input into the high-level network. The high-level network calculates the ranking score for each product through a fully connected layer, and sorts the candidate products in ascending or descending order based on the ranking scores to generate a candidate product ranking list.

[0021] In practical implementation, the data preprocessing module normalizes the input user features (including time-decaying click, add-to-cart, and purchase behavior weights) and product attribute features (category, price, brand, and review information). Min-Max scaling or Z-Score normalization methods are used to eliminate dimensional differences. Unstructured review text is converted into fixed-dimensional vectors using word embedding techniques (such as Word2Vec), which are then concatenated with structured features to form a unified feature vector set. The underlying network contains multiple fully connected layers, each configured with a trainable weight matrix. For the time-decaying features in user features, a time-series weighting matrix is ​​used to strengthen the weights of recent behaviors (e.g., assigning a weight coefficient of 0.9 to clicks within the last 7 days and 0.3 to behaviors from 30 days ago), generating user feature vectors with time-sensitive characteristics. For product attribute features, a category encoding matrix converts product categories (e.g., "clothing-men's clothing-shirts") into sparse one-hot vectors or dense embedding vectors. Price, brand, and other information are linearly transformed and mapped to a specific feature space to form product attribute feature vectors.

[0022] The middle-layer network receives user and product feature vectors output from the lower-layer network. It uses element-wise product multiplication to capture fine-grained relationships between user preferences and product attributes (e.g., the matching degree between frequently purchased high-priced branded products and the current product price). Simultaneously, it preserves complete feature information through dimensional concatenation, generating a composite feature vector containing interaction relationships. This vector is then fed into the higher-layer network after being activated by ReLU or LeakyReLU to introduce non-linearity. The higher-layer network constructs a ranking score calculation model through fully connected layers. This model weights and sums the user-product interaction features with preset business objective parameters (e.g., click-through rate weight 0.4, conversion rate weight 0.3, and product profit weight 0.3 set by the platform) to obtain a comprehensive ranking score for each product. Finally, the candidate products are ranked in descending order (or ascending order, configured according to business needs) to generate a ranking list. For example, a user's feature vector contains recent high-frequency clicks on "sports and outdoor - running shoes" products (time decay weight 0.85), and the product attribute vector corresponds to a new running shoe (normalized price value 0.75, cosine similarity between the brand embedding vector and the user's historical preferred brands 0.92, and sentiment score 0.88). After bottom-level mapping, mid-level splicing, and high-level calculation, the product obtains a ranking score of 0.89, ranking high in the candidate list, demonstrating the generator's ability to deeply fit the user's real-time preferences and the comprehensive attributes of the product.

[0023] In one embodiment, the discriminator is a multilayer recognition neural network. The steps of taking a real product ranking list and a candidate product ranking list generated by a generator as input, and outputting and displaying the ranking information, include: Sequence features were extracted from both the real product sorting list and the candidate product sorting list. The extracted features included the position index of each product in the list, the user's historical click conversion rate, the product price gradient, and the brand's market share. The extracted sequence features are standardized by a low-level recognition network. The feature vectors of the real product sorting list and the candidate product sorting list are then input into two parallel sub-networks to obtain the real feature representation and the candidate feature representation. The mid-layer contrastive network performs element-wise difference calculation between the true feature representation and the candidate feature representation to generate a feature difference vector, and then combines the sentiment value of the product evaluation information to perform weighted correction on the difference vector. The high-level evaluation network calculates the ranking quality score through a fully connected layer, outputs ranking information including the ranking quality score and the position of the differentiated products, and displays it on the cloud platform.

[0024] In practice, the sequence feature extraction module performs structured processing on the real product ranking list (i.e., the high-quality ranking results verified by the platform's history) and the candidate ranking list output by the generator. It extracts the position index of each product in the list (such as the 1st position, 5th position, etc.), the user's historical click conversion rate (the ratio of the product's historical clicks to the number of exposures), the product price gradient (the ratio of the absolute value of the price difference between adjacent products to the average price of the list), and the brand market share (the proportion of the brand's sales in the same category on the platform in the past 30 days). This forms a multi-dimensional feature matrix that includes the sequence order and product performance.

[0025] The underlying recognition network contains two parallel sub-network structures (such as multi-layer fully connected layers). The extracted sequence features are standardized (e.g., Z-score standardization) and then input separately. The feature matrices of the true list and the candidate list are encoded by sharing weights to generate true feature representation vectors and candidate feature representation vectors with the same dimension (e.g., compressing 100-dimensional sequence features into 32-dimensional dense vectors), ensuring that the two are in the same feature space for comparison. The mid-level contrastive network performs element-wise difference calculation between the true feature representation and the candidate feature representation (e.g., the absolute value of the difference between the i-th dimension of the true vector and the i-th dimension of the candidate vector) to generate a feature difference vector. At the same time, it introduces the sentiment value of product evaluation information (score in the range of -1 to 1 output by the NLP model) to weight and correct the difference vector. Products with high sentiment values ​​(e.g., a high-quality product with a score of 0.9) have their weight reduced in the difference calculation (e.g., multiplied by 0.8), while products with low sentiment values ​​(e.g., a low-quality product with a score of -0.5) have their weight increased (e.g., multiplied by 1.2), thereby highlighting the impact of the ranking bias of low-rated products on the overall quality.

[0026] The high-level evaluation network constructs a ranking quality assessment model through fully connected layers. This model takes the corrected feature difference vector as input and combines cosine similarity (which measures the overall directional consistency between the real and candidate feature vectors) and the product profit distribution balance (which calculates the positional dispersion of high-profit and low-profit products in the list) for weighted calculation, outputting a ranking quality score in the range of 0-1 (the higher the score, the closer the candidate list is to the true high-quality ranking). Finally, the discriminator packages the ranking quality score with the positions of significantly different products (such as product IDs with ranking differences exceeding 5 positions, along with their original and current positions) into ranking information, which is transmitted to the front-end system via a cloud platform interface for visualization, providing feedback for the generator's parameter optimization. For example, a highly rated product ranked 3rd in a candidate list is ranked 8th in the real list, with a sentiment score of 0.95. After weighted adjustment, the contribution of the feature difference at this position to the total score decreases by 30%. On the other hand, a low-rated product ranked 10th is ranked 2nd in the real list, with a sentiment score of -0.6, and the contribution of the difference increases by 50%. Finally, the discriminator outputs a quality score of 0.68 and marks two significant difference positions to guide the generator to adjust its ranking strategy.

[0027] In one embodiment, for a cold-start product to be inserted into the sorting information, the step of obtaining the second product attribute feature corresponding to the cold-start product and determining the approximate competitor volume using the second product attribute feature includes: By using the product category, price, brand, and functional description included in the second product attribute features, a set of candidate competitors that match the product category tags is selected from the historical database of the cloud platform; For each competitor in the candidate competitor set, calculate its overlap with the cold start product in terms of price range, its brand identity correlation, and its semantic similarity to the functional description text. The number of competitors in the candidate competitor set that meet the following criteria is recorded as the approximate competitor count: price overlap ≥ the first threshold, brand relevance ≥ the second threshold, and semantic similarity ≥ the third threshold.

[0028] In practice, for cold-start products (i.e., newly listed products or products with missing interaction data) that require the insertion of sorting information, the specific implementation process for judging the number of approximate competitors is as follows: First, the second product attribute characteristics of the cold-start product are obtained through the product information collection module, including structured product category tags (such as "Home Appliances - Kitchen Appliances - Coffee Machines"), price range values ​​(such as 200-300 yuan), brand identifiers (such as "XX Brand"), and unstructured functional description text (such as "15Bar high-pressure extraction, stainless steel body, supports mobile APP control"). Based on the product category tags, all products belonging to the "Home Appliances - Kitchen Appliances" category are retrieved from the historical database of the cloud platform to form a candidate competitor set (for example, including 50 coffee machines and related kitchen appliances that have already been listed).

[0029] For each product in the candidate competitor set, three matching metrics are calculated: First, the overlap of price ranges, calculated using the formula... Calculate (where For cold start product price range, The three factors are: 1) the price range of competing products, reflecting the degree of overlap between the two price ranges; 2) brand association, quantified by brand hierarchy (0.8 for brands under the same parent company, 0.6 for competing brands in the same positioning, and 0 for unrelated brands) or brand user group similarity (based on the overlap ratio of historical buyers); and 3) semantic similarity of functional description text, which uses a pre-trained language model (such as BERT) to vectorize the functional descriptions of cold-start products and competing products, and calculates the semantic matching degree (value range 0-1) through cosine similarity.

[0030] After calculating the three indicators, the number of competitors in the candidate competitor set that meet the following criteria is recorded as the number of approximate competitors: price overlap ≥ the first threshold (e.g., 0.6), brand relevance ≥ the second threshold (e.g., 0.5), and semantic similarity ≥ the third threshold (e.g., 0.7). For example, a cold start product is a "new Italian coffee machine from XX brand," priced between 250-350 yuan, with features including "intelligent temperature control and manual milk frothing." Among the 30 coffee machines in the candidate competitor set, 8 have a price overlap ≥ 0.6, brand relevance ≥ 0.5, and semantic similarity ≥ 0.7, resulting in 8 approximate competitors. This step, through multi-dimensional feature matching, accurately identifies existing products highly similar to the cold start product in terms of category, price, function, and brand positioning. In one embodiment, if the number of similar competing products is lower than a preset value, the step of crawling the association information of cold-start products using a preset association algorithm and calculating the priority value includes: The system crawls supplier cooperation ratings and inventory turnover cycle data for cold start products from the cloud platform supply chain database, and crawls recent mentions and KOL recommendation records of cold start product keywords from pre-bound social media platforms. In addition, it crawls sales growth rate data of complementary product categories over the past three months from the cloud platform sales database to obtain the associated information. The numericalization of the associated information involves mapping the supplier cooperation rating and inventory turnover cycle data to a stability coefficient of 0-1, converting mentions and recommendation records on social media platforms into a market popularity index, and converting the sales growth rate of complementary categories into an association promotion coefficient. The priority value of cold-start products is calculated based on the association algorithm.

[0031] In practice, the data interface module crawls relevant information from internal and external data sources on the cloud platform, including: obtaining supplier cooperation ratings (e.g., a maximum score of 100 points, reflecting on-time delivery rate, quality pass rate, etc.) and inventory turnover cycle (the average number of days from goods entering the warehouse to being sold) from the supply chain database; obtaining recent mentions of product keywords (e.g., product name, model) and KOL recommendation records (the number of times key opinion leaders with ≥100,000 followers recommend the product) from pre-bound social media platforms (e.g., Weibo, Douyin) via API interfaces; and extracting the sales growth rate of complementary category products (e.g., products that are functionally complementary or scenario-related to the cold start product, such as "phone cases" and "smartphones") in the past three months (compared to the sales growth rate in the same period last year) from the sales database.

[0032] After acquiring the relevant information, the data preprocessing module performs numerical transformation: For supplier cooperation ratings and inventory turnover cycles, a linear mapping method is used to convert them into a stability coefficient of 0-1—the higher the supplier rating (e.g., 90 points corresponds to a coefficient of 0.9) and the shorter the inventory turnover cycle (e.g., 30 days corresponds to a coefficient of 0.8), the higher the stability coefficient, reflecting the reliability of the supply chain; social media mentions and recommendation records are converted into a market popularity index through normalization (e.g., Max-Min scaling), with 1000 mentions corresponding to an index of 0.7 and 5 KOL recommendations corresponding to an index of 0.9, reflecting market exposure and user attention; the sales growth rate of complementary categories is converted into a correlation promotion coefficient through Z-Score standardization, with a growth rate 20% higher than the industry average corresponding to a coefficient of 0.6, and a growth rate lower than the average decreasing proportionally, reflecting the market driving effect of related products.

[0033] Finally, according to the preset association algorithm formula: Where P is the priority value, S is the stability coefficient, H is the market popularity index, and C is the correlation promotion coefficient. These are the weight parameters for the corresponding coefficients, and they satisfy... This can be configured by platform operators according to business objectives, such as setting it when prioritizing supply chain stability. Marketing Promotion When setting up related sales The stability coefficient S, market popularity index H, and correlation promotion coefficient C are weighted and summed to obtain the priority value P (range 0-1) for cold-start products. For example, a new wireless earphone has a supplier cooperation score of 85 (stability coefficient 0.85), an inventory turnover cycle of 45 days (coefficient 0.75), 1500 mentions on social media (popularity index 0.8), 3 KOL recommendations (index 0.7), and a sales growth rate of 15% for the complementary product "earphone charging case" (promotion coefficient 0.65). These factors are weighted... , , Calculate the priority value P = 0.3 0.8 + 0.4 0.75 + 0.3 0.65 = 0.735. The higher this value, the stronger the overall potential of the product. It provides a quantitative basis for the subsequent generation of virtual user characteristics and sorting insertion, effectively solving the problem of insufficient sorting basis for cold-start products due to the lack of direct competitor data.

[0034] In one embodiment, the step of generating a virtual second user feature using a preset conversion algorithm and the priority value includes: Determine the proximity information [n-1, n+1] of the product corresponding to the priority value from the currently displayed sorting information; Based on the time decay characteristics of users' historical clicks, add-to-cart, and purchase behaviors corresponding to the products in the near-domain information, the conversion algorithm is used to generate virtual second user characteristics.

[0035] The above-mentioned conversion algorithm is as follows:

[0036] Where V is the virtual second user feature vector, and N is the near-field information set [n-1, n+1]. For the historical click count, add-to-cart count, and purchase count of product i in the near domain, The time interval between the corresponding action and the current time. The coefficient representing the time decay characteristic. This represents the maximum number of historical actions recorded on the platform. These are the feature dimension basis vectors corresponding to the behavior.

[0037] Specifically, the target insertion position n of the cold-start product in the current ranking information is determined based on its priority value P (e.g., priority value 0.735 corresponds to insertion at position 8). The nearest neighbor information [n-1, n+1] (i.e., products at positions 7 and 9) of that position is extracted from the displayed ranking information to form a nearest neighbor product set N = {product 7, product 9}. Nearby products usually have potential associations with the cold-start product in terms of category, price, or function (e.g., if the cold-start product is wireless headphones, near neighbor products may be Bluetooth speakers or headphone cases at the same price). Their user history behavior data can reflect the preference patterns of the target user group.

[0038] For each product in the near-domain product set, extract the time decay features of its user's historical clicks, add-to-cart, and purchase behaviors: the time decay weight of click behavior is determined by... ( The number of days since the click. =0.1 is the attenuation coefficient), and the attenuation weight of the add-on purchase behavior. Decay weight of purchasing behavior These figures reflect the intensity of user attention to nearby products in the recent period. For example, product 7 in the nearby domain was clicked 5 days ago (…). Added to cart 10 days ago ( Purchased 15 days ago ), Nearby product 9 was clicked 3 days ago. Added to cart 8 days ago No purchase record , Subsequently, the conversion algorithm generates virtual user features based on the time decay characteristics of near-domain products: a weighted average is calculated for clicks, add-to-cart actions, and purchase actions of near-domain products, and the virtual number of clicks, add-to-cart actions, and purchase actions are combined with time decay weights to generate virtual interaction records with timestamps. These virtual records are then assigned exponential decay weights according to their time proximity. Finally, these virtual records are converted into vectors with dimensions consistent with real user features (such as 128-dimensional user behavior embedding vectors) as virtual second user features.

[0039] In one embodiment, the step of inputting the second user feature and the second product attribute feature into an adversarial network model, and finally re-outputting and displaying the ranking information after the cold-start product update, includes: The second user feature and the second product attribute feature are input into the adversarial network model; Following the same generator and discriminator operating modes as the first user feature and the first product attribute feature, the cold start product is inserted into the corresponding sorting position to obtain the updated sorting information.

[0040] In the specific implementation process, the virtually generated second user features (such as time decay vectors containing virtual clicks, add-to-cart, and purchase behaviors of nearby products) are dimensionally aligned and concatenated with the second product attribute features of the cold-start product (such as encoded vectors of category, price, brand, and functional description) to form a comprehensive input feature vector for the cold-start product (e.g., a 128-dimensional user-product joint feature vector). This feature vector, along with the first user features and first product attribute features of other products acquired in real time (such as mature product features with existing user interaction data), are input into the generator and processed in the same working mode as the first feature: the bottom layer network performs temporal weighting on the time decay features using a trainable weight matrix and performs category encoding transformation on the product attributes; the middle layer network generates user-product interaction features through element-wise multiplication and dimensional concatenation; and the top layer network calculates the ranking score for each product (including the cold-start product) (e.g., a comprehensive score combining click-through rate, conversion rate, and profit).

[0041] The generator generates a candidate ranking list containing cold-start products based on the ranking score (e.g., the original list is [Product A, Product B, Product C], and after inserting the cold-start product D, it becomes [Product A, Product D, Product B, Product C]). This candidate list is then input into the discriminator along with a real product ranking list (e.g., high-quality ranking results verified historically by the platform). Following the same working mode as the first feature processing, the discriminator extracts sequence features (e.g., product location index, click-through rate, price gradient) from the candidate and real lists. A low-level recognition network generates feature representations, a mid-level contrast network calculates the difference vector and corrects it using sentiment values, and a high-level evaluation network outputs a ranking quality score (e.g., 0.82, indicating a high degree of matching between the candidate list and the real high-quality rankings).

[0042] If the ranking quality score is higher than a preset threshold (e.g., 0.7), the cold-start product's current position in the candidate list is retained. If the score is insufficient, the generator parameters are adjusted through backpropagation (e.g., optimizing the ranking score calculation weights of the high-level network), and the candidate list is regenerated until the quality score meets the threshold. Finally, the candidate list verified by the discriminator is determined as the updated ranking information and transmitted to the e-commerce front-end system through the cloud platform interface. The real-time ranking results containing the cold-start product are displayed on the user's search or browsing page (e.g., when a user searches for "wireless headphones," the new cold-start headphones D appears in the second position, labeled "New Product Recommendation" below it).

[0043] For example, the second user feature of a cold-start wireless earphone D includes the time decay vectors of virtual clicks (22.7 times) and add-to-cart (9.1 times) of nearby products (Bluetooth speakers and earphone cases at the same price). The second product attribute features are the category "Consumer Electronics - Earphones - Wireless" (embedded vector), price 399 yuan (normalized value 0.7), brand "XX" (embedded vector), and functional description semantic vector (cosine similarity 0.85). After processing by the generator, its ranking score is calculated to be 0.89 (higher than 0.82 for product B and 0.75 for product C in the original list), and the candidate list is updated to [Product A (0.92), Product D (0.89), Product B (0.82), Product C (0.75)]. The discriminator verifies that the quality score of the list is 0.85 (higher than the threshold 0.7), and finally, the cold-start product D is stably inserted in the second position, and users can see its display position on the search results page, realizing the reasonable ranking and effective exposure of cold-start products based on virtual features.

[0044] Reference Appendix Figure 2 This invention presents a structural block diagram of a product ranking cloud platform for e-commerce platforms based on generative adversarial networks. The cloud platform includes: The receiving unit is used to receive user search requests and / or browsing behavior data, and extract first user features and first product attribute features. The first user features include time decay features of user historical clicks, add-to-cart, and purchase behaviors. The first product attribute features include product category, price, brand, and evaluation information. The model unit is used to build upon a pre-trained adversarial network model, which includes a generator and a discriminator. The generator is a multi-layer deception neural network that takes a first user feature and a first product attribute feature as input and outputs a candidate product ranking list. The discriminator is a multi-layer recognition neural network that takes a real product ranking list and a candidate product ranking list generated by the generator as input and outputs and displays the ranking information. The request unit is used to obtain the second product attribute features corresponding to the cold start product for the cold start product that is requested to be inserted into the sorting information, and to judge the approximate number of competitors through the second product attribute features. If the approximate number of competitors is lower than a preset value, the priority value is calculated by crawling the association information of the cold start product using a preset association algorithm. The update unit is used to generate a virtual second user feature by combining the priority value with a preset transformation algorithm, and input the second user feature and the second product attribute feature into the adversarial network model, and finally re-output and display the sorting information after the cold start product is updated.

[0045] In summary, this method addresses the challenge of cold-start product ranking. It extracts features from user behavior data, generates a candidate ranking list using a generator, and outputs ranking information using a discriminator. For cold-start products, it acquires attribute features to determine the number of similar competitors; if the number is below a preset value, it crawls related information to calculate priority values. Combined with a conversion algorithm, it generates virtual user features, which are then input into the model to output updated ranking information. This method improves the accuracy of cold-start product ranking, enhances the model's generalization ability, and balances user experience with platform goals, demonstrating significant technological innovation and practical value.

[0046] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0047] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0048] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A product ranking method for an e-commerce platform based on generative adversarial networks, characterized in that, Includes the following steps: Receive user search requests and / or browsing behavior data, extract first user features and first product attribute features, the first user features include time decay features of user historical clicks, add-to-cart, and purchase behavior, and the first product attribute features include product category, price, brand, and review information; Based on a pre-trained adversarial network model, the model includes a generator and a discriminator. The generator is a multi-layer deception neural network that takes a first user feature and a first product attribute feature as input and outputs a candidate product ranking list. The discriminator is a multi-layer recognition neural network that takes a real product ranking list and a candidate product ranking list generated by the generator as input and outputs and displays the ranking information. For cold start products that need to be inserted into the sorting information, the second product attribute features corresponding to the cold start products are obtained, and the number of approximate competitors is judged by the second product attribute features. If the number of approximate competitors is lower than a preset value, the priority value is calculated by crawling the association information of the cold start products using a preset association algorithm. Using a preset conversion algorithm, a virtual second user feature is generated by combining the priority value. The second user feature and the second product attribute feature are then input into the adversarial network model, and finally the sorting information after the cold start product is inserted is re-output and displayed.

2. The product ranking method for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that, The generator is a multi-layer deception neural network. The steps of taking a first user feature and a first product attribute feature as input and outputting a candidate product ranking list include: The first user features and the first product attribute features input into the multi-layer deception neural network are normalized to form a set of feature vectors with uniform dimensions. The feature vector set is initially mapped using the underlying network. Different weight matrices are set for each layer of the underlying network. The time decay features are weighted in a time series to obtain the user feature vector. The product category, price, brand and evaluation information are categorized and converted to obtain the product attribute feature vector. The middle layer network cross-concatenates the user feature vector and the product attribute feature vector, and generates user-product interaction features by element-wise multiplication and dimension concatenation. After processing by the activation function, the features are input into the high-level network. The high-level network calculates the ranking score for each product through a fully connected layer, and sorts the candidate products in ascending or descending order based on the ranking scores to generate a candidate product ranking list.

3. The e-commerce platform product ranking method based on generative adversarial networks according to claim 1, characterized in that, The discriminator is a multi-layer recognition neural network. The steps of taking the real product ranking list and the candidate product ranking list generated by the generator as input, and outputting and displaying the ranking information, include: Sequence features were extracted from both the real product sorting list and the candidate product sorting list. The extracted features included the position index of each product in the list, the user's historical click conversion rate, the product price gradient, and the brand's market share. The extracted sequence features are standardized by a low-level recognition network. The feature vectors of the real product sorting list and the candidate product sorting list are then input into two parallel sub-networks to obtain the real feature representation and the candidate feature representation. The mid-layer contrastive network performs element-wise difference calculation between the true feature representation and the candidate feature representation to generate a feature difference vector, and then combines the sentiment value of the product evaluation information to perform weighted correction on the difference vector. The high-level evaluation network calculates the ranking quality score through a fully connected layer, outputs ranking information including the ranking quality score and the position of the differentiated products, and displays it on the cloud platform.

4. The product ranking method for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that, For cold-start products that need to be inserted into the sorting information, the step of obtaining the second product attribute features corresponding to the cold-start products and determining the approximate competitor volume based on the second product attribute features includes: By using the product category, price, brand, and functional description included in the second product attribute features, a set of candidate competitors that match the product category tags is selected from the historical database of the cloud platform; For each competitor in the candidate competitor set, calculate its overlap with the cold start product in terms of price range, its brand identity correlation, and its semantic similarity to the functional description text. The number of competitors in the candidate competitor set that meet the following criteria is recorded as the approximate competitor count: price overlap ≥ the first threshold, brand relevance ≥ the second threshold, and semantic similarity ≥ the third threshold.

5. The e-commerce platform product ranking method based on generative adversarial networks according to claim 4, characterized in that, If the number of similar competing products is lower than a preset value, the step of crawling the association information of the cold-start products using a preset association algorithm and calculating the priority value includes: The system crawls supplier cooperation ratings and inventory turnover cycle data for cold start products from the cloud platform supply chain database, and crawls recent mentions and KOL recommendation records of cold start product keywords from pre-bound social media platforms. In addition, it crawls sales growth rate data of complementary product categories over the past three months from the cloud platform sales database to obtain the associated information. The numericalization of the associated information involves mapping the supplier cooperation rating and inventory turnover cycle data to a stability coefficient of 0-1, converting mentions and recommendation records on social media platforms into a market popularity index, and converting the sales growth rate of complementary categories into an association promotion coefficient. The priority value of cold-start products is calculated based on the association algorithm.

6. The product ranking method for e-commerce platforms based on generative adversarial networks according to claim 5, characterized in that, The formula for the association algorithm is: Where P is the priority value, S is the stability coefficient, H is the market popularity index, and C is the correlation promotion coefficient. These are the weight parameters for the corresponding coefficients, and they satisfy... .

7. The e-commerce platform product ranking method based on generative adversarial networks according to claim 5, characterized in that, The step of generating a virtual second user feature using a preset conversion algorithm and the priority value includes: Determine the proximity information [n-1, n+1] of the product corresponding to the priority value from the currently displayed sorting information; Based on the time decay characteristics of users' historical clicks, add-to-cart, and purchase behaviors corresponding to the products in the near-domain information, the conversion algorithm is used to generate virtual second user characteristics.

8. The e-commerce platform product ranking method based on generative adversarial networks according to claim 7, characterized in that, The conversion algorithm is as follows: Where V is the virtual second user feature vector, and N is the near-field information set [n-1, n+1]. For the historical click count, add-to-cart count, and purchase count of product i in the near domain, The time interval between the corresponding action and the current time. The coefficient representing the time decay characteristic. This represents the maximum number of historical actions recorded on the platform. These are the feature dimension basis vectors corresponding to the behavior.

9. The product ranking method for e-commerce platforms based on generative adversarial networks according to claim 7, characterized in that, The steps of inputting the second user feature and the second product attribute feature into the adversarial network model, and finally re-outputting and displaying the ranking information after the cold start product update, include: The second user feature and the second product attribute feature are input into the adversarial network model; Following the same generator and discriminator operating modes as the first user feature and the first product attribute feature, the cold start product is inserted into the corresponding sorting position to obtain the updated sorting information.

10. A cloud platform for product ranking on an e-commerce platform based on generative adversarial networks, characterized in that the cloud platform... include: The receiving unit is used to receive user search requests and / or browsing behavior data, and extract first user features and first product attribute features. The first user features include time decay features of user historical clicks, add-to-cart, and purchase behaviors. The first product attribute features include product category, price, brand, and evaluation information. The model unit is used to build upon a pre-trained adversarial network model, which includes a generator and a discriminator. The generator is a multi-layer deception neural network that takes a first user feature and a first product attribute feature as input and outputs a candidate product ranking list. The discriminator is a multi-layer recognition neural network that takes a real product ranking list and a candidate product ranking list generated by the generator as input and outputs and displays the ranking information. The request unit is used to obtain the second product attribute features corresponding to the cold start product for the cold start product that is requested to be inserted into the sorting information, and to judge the approximate number of competitors through the second product attribute features. If the approximate number of competitors is lower than a preset value, the priority value is calculated by crawling the association information of the cold start product using a preset association algorithm. The update unit is used to generate a virtual second user feature by combining the priority value with a preset transformation algorithm, and input the second user feature and the second product attribute feature into the adversarial network model, and finally re-output and display the sorting information after the cold start product is updated.

Images