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

Abstract

The application relates to an e-commerce platform commodity sorting method and a cloud platform based on a generative adversarial network, and belongs to the technical field of e-commerce commodity sorting. Features are extracted by receiving user behavior data, a generator is used to output a candidate sorting list, and a discriminator is used to output sorting information. For cold-start commodities, attribute features are acquired to determine the approximate number of competitive products. When the number is lower than a preset value, associated information is crawled to calculate a priority value. Virtual user features are generated by combining a conversion algorithm, and updated sorting information is output after being input into a model. The method improves the sorting accuracy of cold-start commodities, enhances the generalization ability of the model, balances user experience and platform goals, and has significant technical innovation and practical value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of e-commerce commodity sorting, and particularly relates to an e-commerce platform commodity sorting method based on a generative adversarial network and a cloud platform. BACKGROUND

[0002] In an e-commerce platform, commodity sorting is a core technology that affects user experience and platform operation efficiency. Traditional commodity sorting methods are mostly based on collaborative filtering, logistic regression and other models, and mainly rely on user historical interaction data (such as click, purchase behavior) and commodity basic attributes (such as price, sales) for sorting, but for cold-start commodities (i.e. new commodities or commodities lacking user interaction data), it is difficult to accurately evaluate their market potential and user preferences due to the lack of effective user feedback information, resulting in biased sorting results, which may either cause insufficient exposure of high-quality new products or affect user browsing experience due to inaccurate recommendations. In the prior art, some schemes attempt to generate virtual user interaction data through a generative adversarial network (GAN) to improve the sorting effect of cold-start commodities, but generally have the following problems: first, the feature mining of cold-start commodities is not comprehensive, relying only on single-dimensional commodity attributes or simple user behavior simulation, without fully combining supply chain data, social media heat and other external associated information; second, the collaborative training mechanism of the generator and the discriminator is not optimized, and the generated virtual data differs from the real user behavior pattern, resulting in insufficient generalization ability of the sorting model; third, the correlation analysis of cold-start commodities and existing commodities is relatively rough, and there is no systematic priority value calculation and virtual feature generation method, making it difficult to balance platform business goals (such as profit, conversion rate) while ensuring sorting diversity. SUMMARY

[0003] The main purpose of the present application is to provide an e-commerce platform commodity sorting method based on a generative adversarial network and a cloud platform, to realize accurate cold-start commodity sorting, enhance model generalization ability, balance user experience and platform goals, and have significant technical innovation and practical value.

[0004] To achieve the above purpose, the e-commerce platform commodity sorting method based on the generative adversarial network provided by the present application comprises the following steps:

[0005] Receiving a user's search request and / or browsing behavior data, extracting first user features and first commodity attribute features, the first user features including time decay features of user historical click, add-to-cart and purchase behavior, and the first commodity attribute features including commodity category, price, brand and evaluation information;

[0006] Based on a pre-trained adversarial network model, the model includes a generator and a discriminator, wherein the generator is a multi-layer fraud-making neural network, inputs first user features and first commodity attribute features, and outputs a candidate commodity ranking list; the discriminator is a multi-layer identification neural network, inputs a real commodity ranking list and a candidate commodity ranking list generated by the generator, and outputs and displays ranking information;

[0007] For a cold-start commodity requiring insertion of the ranking information, second commodity attribute features corresponding to the cold-start commodity are obtained, approximate competitor quantity judgment is performed through the second commodity attribute features, if the approximate competitor quantity is lower than a preset value, a preset association algorithm is used to crawl and calculate a priority value of the association information of the cold-start commodity;

[0008] A preset conversion algorithm is used to generate a virtual second user feature in combination with the priority value, and the second user feature and the second commodity attribute features are input into the adversarial network model, and finally the updated ranking information after insertion of the cold-start commodity is re-output and displayed.

[0009] Further, the step of the generator being a multi-layer fraud-making neural network, inputting first user features and first commodity attribute features, and outputting a candidate commodity ranking list, includes:

[0010] The first user features and the first commodity attribute features input into the multi-layer fraud-making neural network are normalized to form a dimension-unified feature vector set;

[0011] The feature vector set is preliminarily mapped by using a bottom network, different weight matrices are respectively set for each layer of the bottom network, the time decay features are time-weighted to obtain a user feature vector, and the commodity category, price, brand and evaluation information are category-encoded and converted to obtain a commodity attribute feature vector;

[0012] The user feature vector and the commodity attribute feature vector are cross-spliced by using a middle network, a user-commodity interaction feature is generated by using an element-by-element multiplication and a dimension concatenation, and is input into a high network after being processed by an activation function;

[0013] The high network calculates a ranking score of each commodity through a full connection layer, arranges the candidate commodities in ascending or descending order according to the ranking score, and generates a candidate commodity ranking list.

[0014] Further, the step of the discriminator being a multi-layer identification neural network, inputting a real commodity ranking list and a candidate commodity ranking list generated by the generator, and outputting and displaying ranking information, includes:

[0015] The real commodity ranking list and the candidate commodity ranking list are respectively subjected to sequence feature extraction, and the extracted content includes the position index of each commodity in the list, the historical click conversion rate of the user, the price gradient of the commodity, and the brand market share;

[0016] The extracted sequence features are standardized by a bottom layer recognition network, and the feature vectors of the real commodity ranking list and the candidate commodity ranking list are respectively input into two parallel sub-networks to obtain real feature representation and candidate feature representation;

[0017] The middle layer comparison network performs element-by-element difference calculation on the real feature representation and the candidate feature representation to generate a feature difference vector, and the difference vector is weighted and corrected in combination with the sentiment tendency value of the commodity evaluation information;

[0018] The high-level evaluation network calculates the ranking quality score through a fully connected layer, outputs the ranking information containing the ranking quality score and the difference commodity position, and displays it on the cloud platform.

[0019] Further, for a cold start commodity required to be inserted into the ranking information, the second commodity attribute feature corresponding to the cold start commodity is obtained, and the step of judging the approximate competitor quantity through the second commodity attribute feature includes:

[0020] Through the commodity category, price, brand and function description included in the second commodity attribute feature, a candidate competitor set consistent with the commodity category label is screened from the cloud platform historical database;

[0021] For each competitor in the candidate competitor set, the overlap degree of its price interval value with the cold start commodity, the association degree of the brand identifier, and the semantic similarity of the function description text are calculated;

[0022] The number of competitors in the candidate competitor set that satisfy the price overlap degree ≥ the first threshold value, the brand association degree ≥ the second threshold value, and the semantic similarity ≥ the third threshold value is counted, and is recorded as the approximate competitor quantity.

[0023] Further, if the approximate competitor quantity is lower than a preset value, the step of calculating a priority value by crawling the association information of the cold start commodity using a preset association algorithm includes:

[0024] The supplier cooperation score and the inventory turnover period data of the cold start commodity are crawled from the cloud platform supply chain database, the recent mention quantity of the cold start commodity keyword and the KOL recommendation record data are crawled from the pre-bound social media platform, and the near three-month sales growth rate data of complementary product category commodities are crawled from the cloud platform sales database to obtain the association information;

[0025] The correlation information is valued, the supplier cooperation score and the inventory turnover period data are mapped to a stability coefficient of 0-1, the mention amount and the recommendation record in the social media platform are converted into a market heat index, and the complementary category sales growth rate is converted into a correlation promotion coefficient;

[0026] The priority value of the cold start commodity is calculated according to the correlation algorithm.

[0027] Further, the correlation algorithm formula is: Wherein, P is the priority value, S is the stability coefficient, H is the market heat index, C is the correlation promotion coefficient, are weight parameters of corresponding coefficients, and satisfy .

[0028] Further, the step of generating a virtual second user feature by using a preset conversion algorithm in combination with the priority value comprises:

[0029] The near-field information [n-1, n+1] of the priority value corresponding commodity is determined from the currently displayed sorting information;

[0030] Based on the time decay characteristics of the user historical click, add-to-cart, and purchase behaviors of the near-field information corresponding commodity, the conversion algorithm is used to generate a virtual second user feature.

[0031] Further, the conversion algorithm is:

[0032]

[0033] Wherein, V is the virtual second user feature vector, N is the near-field information set [n-1, n+1], is the historical click count, add-to-cart count, and purchase count of the near-field commodity i, is the time interval from the corresponding behavior to the current time, is the coefficient of the time decay characteristic, is the maximum value of the platform historical behavior count, is the feature dimension base vector of the corresponding behavior.

[0034] Further, the step of finally re-outputting the updated sorting information of the cold start commodity display insertion by inputting the second user feature and the second commodity attribute feature into the adversarial network model comprises:

[0035] The second user feature and the second commodity attribute feature are input into the adversarial network model;

[0036] The cold start commodity is inserted into a corresponding ranking position according to the same working mode of the generator and the discriminator as the first user feature and the first commodity attribute feature, to obtain updated ranking information.

[0037] The application provides an e-commerce platform commodity ranking cloud platform based on a generative adversarial network, which comprises:

[0038] A receiving unit is configured to receive a search request and / or browsing behavior data of a user, and extract first user features and first commodity attribute features, wherein the first user features comprise time decay features of historical click, add-to-cart and purchase behaviors of the user, and the first commodity attribute features comprise commodity categories, prices, brands and evaluation information;

[0039] A model unit is configured to use a pre-trained adversarial network model, wherein the model comprises a generator and a discriminator, the generator is a multi-layer fraud detection neural network, the first user features and the first commodity attribute features are input into the generator, and a candidate commodity ranking list is output; the discriminator is a multi-layer identification neural network, a real commodity ranking list and the candidate commodity ranking list generated by the generator are input into the discriminator, and ranking information is output and displayed;

[0040] A request unit is configured to obtain second commodity attribute features corresponding to a cold start commodity which needs to be inserted into the ranking information, perform approximate competitor quantity judgment through the second commodity attribute features, and if the approximate competitor quantity is lower than a preset value, calculate a priority value by using a preset association algorithm to crawl association information of the cold start commodity;

[0041] An updating unit is configured to use a preset conversion algorithm to generate virtual second user features in combination with the priority value, input the second user features and the second commodity attribute features into the adversarial network model, and finally output and display updated ranking information of the cold start commodity after insertion

[0042] The e-commerce platform commodity ranking method and cloud platform based on the generative adversarial network have the following beneficial effects:

[0043] (1) The second commodity attribute features of the cold start commodity are obtained, approximate competitor quantity judgment is performed, the priority value is calculated in combination with multi-dimensional association information such as supply chain data, social media heat and complementary product category sales, the ranking deviation problem of the cold start commodity caused by the lack of user interaction data is effectively solved, and the new product obtains more reasonable exposure opportunities.

[0044] (2) The adversarial training mechanism of the generator and the discriminator in the generative adversarial network is used, the time decay features of the historical behaviors of the user and the deep feature extraction of the commodity attributes are combined, a candidate ranking list closer to the real user preferences is generated, and the fitting ability of the model to complex user behaviors and commodity features is improved. Attached Figure Description

[0045] 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.

[0046] 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;

[0047] 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

[0048] 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.

[0049] 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:

[0050] 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;

[0051] 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.

[0052] 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.

[0053] 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.

[0054] In one embodiment, the data acquisition module deployed in the e-commerce platform server captures the user's search request (such as the input keyword, the combination of filtering conditions) and browsing behavior data (including the timestamp of clicking the product detail page, adding to the shopping cart, completing the purchase, etc., and the operation object ID, etc.) on the platform in real time, and stores them in the user behavior database. For the extraction of the first user feature, for the user's historical click, add-to-cart, and purchase behavior, an exponential decay model is used to calculate the weight of the behavior time, specifically through a time decay function (where t is the number of days from the behavior occurrence time to the current time, configurable decay rate parameter), to give dynamic weights to behaviors at different time points, for example, the weight of the click behavior in the recent 7 days is significantly higher than that of the same behavior 30 days ago, thereby 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, the price information is mapped to the [0, 1] interval through normalization processing, and the evaluation information is analyzed for sentiment using natural language processing technology (such as building a sentiment classification model based on a bidirectional long short-term memory network BILSTM, outputting a sentiment score of -1 to 1) and keyword extraction, and finally the various attribute features are concatenated into a unified dimension product attribute vector. For example, when a user searches for "laptop", the system synchronously collects the user's click records (including the timestamp of each click) on electronic products in the past 30 days, 3 laptop IDs added to the cart, and the information of a peripheral product purchased 15 days ago, calculates the weight coefficients of each behavior after time decay, and performs structured processing on the attribute features such as the category (computer-laptop) and price (5999 yuan normalized to 0.82) of the corresponding product, brand (Lenovo corresponding embedding vector), and evaluation sentiment score (0.91), to form a standardized feature vector set for input model.

[0055] In one embodiment, the generator is a multi-layer fraud generation neural network, which inputs the first user feature and the first product attribute feature, and outputs a candidate product ranking list, including:

[0056] The first user feature and the first product attribute feature input into the multi-layer fraud generation neural network are normalized to form a feature vector set with uniform dimensions;

[0057] The feature vector set is preliminarily mapped using a bottom network, different weight matrices are set for each layer of the bottom network, the time decay feature is weighted in time sequence to obtain a user feature vector, and the product category, price, brand, and evaluation information are converted to obtain a product attribute feature vector;

[0058] The middle layer network cross-splices the user feature vector and the commodity attribute feature vector, adopts element-by-element multiplication and dimension concatenation to generate a user-commodity interaction feature, and inputs the user-commodity interaction feature into the high layer network after being processed by an activation function;

[0059] The high layer network calculates a ranking score of each commodity through a fully connected layer, ranks the candidate commodities in ascending or descending order according to the ranking score, and generates a candidate commodity ranking list.

[0060] In a specific implementation, the data preprocessing module is used to normalize the input user features (time-decay click, add-to-cart, and purchase behavior weights) and commodity attribute features (category, price, brand, and evaluation information), eliminate dimension differences by using a Min-Max scaling or Z-Score standardization method, convert unstructured evaluation text into a fixed-dimension vector through a word embedding technology (such as Word2Vec), splice the structured features to form a dimension-uniform feature vector set. The bottom layer network includes multiple fully connected layers, each layer is configured with a trainable weight matrix, the time-decay feature in the user feature is strengthened by using a time-weighting matrix to emphasize recent behavior weights (for example, a 0.9 weight coefficient is assigned to a click behavior in the last 7 days, and a 0.3 weight coefficient is assigned to a behavior 30 days ago), to generate a user feature vector containing time-sensitive characteristics; for the commodity attribute features, a category encoding matrix is used to convert a commodity category (such as “clothing-menswear-shirt”) into a sparse one-hot vector or a dense embedding vector, and price, brand, and other information are mapped to a specific feature space after being linearly transformed to form a commodity attribute feature vector.

[0061] The middle layer network receives the user and commodity feature vectors output by the bottom layer, uses element-wise product to capture the fine-grained association between user preferences and commodity attributes (such as the matching degree between the high-frequency purchased high-price brand commodity of the user and the current commodity price), and generates a composite feature vector containing interaction relationships through dimension concatenation (Concatenation) to retain complete feature information. After the introduction of nonlinearity by the ReLU or LeakyReLU activation function, the high layer network is input. The high layer network builds a ranking score calculation model through a fully connected layer. This model performs weighted summation on the user-commodity interaction features and the preset business target parameters (such as the click-through rate weight 0.4, conversion rate weight 0.3, and commodity profit weight 0.3 set by the platform), obtains the comprehensive ranking score of each commodity, and finally arranges the candidate commodity ranking list in descending (or ascending, depending on business requirements) order according to the score. For example, a user feature vector contains high-frequency clicks on "sports outdoor - running shoes" type commodities in the near future (time decay weight 0.85), and the commodity attribute vector corresponds to a new running shoe (price normalization value 0.75, brand embedding vector cosine similarity with user historical preference brand 0.92, and evaluation sentiment score 0.88). After bottom layer mapping, middle layer splicing, and high layer calculation, the commodity obtains a ranking score of 0.89 and ranks in the front position in the candidate list, reflecting the generator's deep fitting ability for user real-time preferences and commodity comprehensive attributes.

[0062] In one embodiment, the discriminator is a multi-layer identification neural network that inputs the real commodity ranking list and the candidate commodity ranking list generated by the generator, and outputs and displays the ranking information. The steps include:

[0063] The real commodity ranking list and the candidate commodity ranking list are respectively subjected to sequence feature extraction, and the extracted content includes the position index of each commodity in the list, the user historical click-through rate, the commodity price gradient, and the brand market share;

[0064] The extracted sequence features are standardized by the bottom layer identification network, and the feature vectors of the real commodity ranking list and the candidate commodity ranking list are respectively input into two parallel sub-networks to obtain real feature representations and candidate feature representations;

[0065] The middle layer comparison network performs element-wise difference calculation on the real feature representations and the candidate feature representations to generate a feature difference vector, and combines the sentiment tendency value of the commodity evaluation information to weight and correct the difference vector;

[0066] The high layer evaluation network calculates the ranking quality score through a fully connected layer, outputs the ranking information containing the ranking quality score and the difference commodity position, and displays it on the cloud platform.

[0067] In a specific implementation, the sequence feature extraction module is used to structure the real commodity ranking list (i.e., a high-quality ranking result verified by the platform history) and the candidate ranking list output by the generator, extract the position index (such as the ordinal feature of the first place, the fifth place, etc.) of each commodity in the list, the historical click conversion rate of the user (the ratio of the historical click times to the exposure times of the commodity), the price gradient of the commodity (the ratio of the absolute value of the price difference of adjacent commodities to the average price of the list), and the brand market share (the proportion of the platform sales of the brand in the same category in the last 30 days), and form a multi-dimensional feature matrix containing sequence order and commodity performance.

[0068] The bottom layer recognition network includes two parallel sub-network structures (such as multi-layer fully connected layers), which are respectively input after standardization (such as Z-Score standardization) of the extracted sequence features. The feature matrix of the real list and the candidate list is encoded through shared weights to generate real feature representation vectors and candidate feature representation vectors with consistent dimensions (for example, compressing 100-dimensional sequence features into 32-dimensional dense vectors), ensuring that both are in the same feature space for comparison. The middle layer comparison network performs element-by-element difference calculation on the real feature representation and the candidate feature representation (such as the absolute value of the difference between the i-th dimension of the real vector and the i-th dimension of the candidate vector), generates a feature difference vector, and introduces the sentiment tendency value of the commodity evaluation information (the score in the interval of -1 to 1 output by the NLP model) to weight and correct the difference vector — commodities with high sentiment tendency values (such as high-quality evaluation commodities with a score of 0.9) have reduced weights in the difference calculation (such as multiplied by 0.8), and commodities with low sentiment tendency values (such as poor evaluation commodities with a score of -0.5) have increased weights (such as multiplied by 1.2), thereby highlighting the impact of low-evaluation commodity ranking deviation on the overall quality.

[0069] The high-level evaluation network builds a ranking quality evaluation model through a fully connected layer, which takes the corrected feature difference vector as input, combines the cosine similarity (measuring the overall direction consistency of the true and candidate feature vectors) and the balance degree of commodity profit distribution (calculating the position dispersion degree of high-profit and low-profit commodities in the list), and performs weighted calculation to output the ranking quality score in the interval of 0-1 (the higher the score, the closer the candidate list to the real high-quality ranking). Finally, the discriminator packs the ranking quality score and the significant commodity position (such as the commodity ID and the original position, the current position whose ranking difference exceeds 5 positions) into ranking information, which is transmitted to the front-end system through the cloud platform interface for visualization, providing feedback basis for parameter optimization of the generator. For example, a high-evaluation commodity ranked 3rd in a certain candidate list is ranked 8th in the real list, and its sentiment tendency value is 0.95. After weighted correction, the contribution of the feature difference of this position to the total score decreases by 30%, while the low-evaluation commodity ranked 10th is ranked 2nd in the real list, and its sentiment tendency value is -0.6. The difference contribution increases by 50%, and the final discriminator outputs a quality score of 0.68 and marks two significant difference positions, guiding the generator to adjust the ranking strategy.

[0070] In one embodiment, for a cold-start commodity requiring insertion of the ranking information, the second commodity attribute feature corresponding to the cold-start commodity is obtained, and the step of performing approximate competitor quantity judgment through the second commodity attribute feature includes:

[0071] A candidate competitor set consistent with the commodity category label is screened from the cloud platform historical database through the commodity category, price, brand, and function description included in the second commodity attribute feature.

[0072] For each competitor in the candidate competitor set, the overlap degree of its price interval value with the cold-start commodity, the association degree of its brand identifier, and the semantic similarity of its function description text are calculated.

[0073] The number of competitors in the candidate competitor set that satisfy the conditions of price overlap degree ≥ first threshold value, brand association degree ≥ second threshold value, and semantic similarity ≥ third threshold value is counted, and is recorded as the approximate competitor quantity.

[0074] In a specific implementation, for a cold-start commodity that needs to insert sorting information (i.e., a new commodity or an interactive data missing commodity), the specific implementation process of the approximate competitor quantity judgment is as follows: first, the second commodity attribute characteristics of the cold-start commodity are obtained through the commodity information collection module, including the structured commodity category label (such as "household appliances - kitchen appliances - coffee machine"), the price interval value (such as 200-300 yuan), the brand identification (such as "XX brand"), and the unstructured function description text (such as "15Bar high pressure extraction, stainless steel body, support for mobile phone APP control"). Based on the commodity category label, all commodities belonging to the "household appliances - kitchen appliances" category are retrieved from the cloud platform historical database to form a candidate competitor set (for example, containing 50 coffee machines and related kitchen appliances that have been listed).

[0075] For each commodity in the candidate competitor set, three matching indicators are calculated respectively: first, the price interval overlap degree is calculated by the formula (wherein is the price interval of the cold-start commodity, is the price interval of the competitor), reflecting the degree of coincidence of the two price bands; second, the brand correlation degree is quantified by the brand hierarchy relationship (such as the same parent company brand value 0.8, the same positioning competitor brand value 0.6, and the irrelevant brand value 0) or the brand user group similarity (based on the overlap proportion of historical purchase users); third, the semantic similarity of the function description text is calculated by using a pre-trained language model (such as BERT) to vectorize the function description of the cold-start commodity and the competitor, and the semantic matching degree is calculated by cosine similarity (the value range is 0-1).

[0076] After the calculation of the three indicators is completed, the number of competitors that satisfy the price overlap degree ≥ the first threshold value (such as 0.6), the brand correlation degree ≥ the second threshold value (such as 0.5), and the semantic similarity ≥ the third threshold value (such as 0.7) in the candidate competitor set is counted, and is recorded as the approximate competitor quantity. For example, a cold-start commodity is "XX brand new Italian coffee machine", the price interval is 250-350 yuan, and the function description contains "intelligent temperature control, manual milk frothing". Among the 30 coffee machines in the candidate competitor set, 8 have a price overlap degree ≥ 0.6, a brand correlation degree ≥ 0.5, and a semantic similarity ≥ 0.7, so the approximate competitor quantity is 8. This step accurately identifies existing commodities that are highly similar to the cold-start commodity in terms of category, price, function, and brand positioning,

[0077] In one embodiment, if the approximate competitor quantity is lower than the preset value, the step of calculating the priority value after crawling the associated information of the cold-start commodity using a preset association algorithm includes:

[0078] crawl the supplier cooperation score, inventory turnover period data of the cold-start commodity from the cloud platform supply chain database, and the recent mention amount of the cold-start commodity keyword, KOL recommendation record data from the pre-bound social media platform, and the complementary category commodity sales growth rate data in the past three months from the cloud platform sales database to obtain the association information;

[0079] The association information is numerized, the supplier cooperation score and the inventory turnover period data are mapped to a stability coefficient of 0-1, the mention amount and the recommendation record in the social media platform are converted into a market heat index, and the complementary category sales growth rate is converted into an association promotion coefficient;

[0080] According to the correlation algorithm, the priority value of the cold-start commodity is calculated.

[0081] In specific implementation, the data interface module crawls the association information from the internal and external data sources of the cloud platform, including: obtaining the supplier cooperation score (such as 100 full marks, reflecting the supplier on-time delivery rate, quality pass rate, etc.) and the inventory turnover period (the average number of days from warehousing to sale of the commodity) from the supply chain database, obtaining the recent mention amount (the number of posts containing the keyword in the past 30 days) and the KOL recommendation record (the number of recommendations of the commodity by key opinion leaders with a fan base of more than 100,000) of the commodity keyword (such as the commodity name, model) from the pre-bound social media platform (such as Weibo, Douyin) through the API interface, and extracting the sales growth rate of complementary category commodities (such as commodities that are functionally complementary or scene-related to the cold-start commodity, for example, "phone case" and "smartphone") in the past three months (comparing the sales growth rate with that of the same period last year).

[0082] After obtaining the association information, the data preprocessing module is used for numerical conversion: for the supplier cooperation score and the inventory turnover period, a linear mapping method is used to convert them into a stability coefficient of 0-1 — the higher the supplier score (such as 90 corresponding to a coefficient of 0.9) and the shorter the inventory turnover period (such as 30 days corresponding to a coefficient of 0.8), the higher the stability coefficient, reflecting the reliability of the supply chain; the mention amount and the recommendation record in the social media are converted into a market heat index through normalization processing (such as Max-Min scaling) — 1000 mentions correspond to an index of 0.7, and 5 KOL recommendations correspond to an index of 0.9, reflecting market exposure and user attention; the complementary category sales growth rate is converted into an association promotion coefficient through Z-Score standardization — a growth rate higher than the industry average by 20% corresponds to a coefficient of 0.6, and a growth rate lower than the average is attenuated in proportion, reflecting the market driving effect of the associated commodities.

[0083] Finally, according to the preset association algorithm formula: Wherein, P is the priority value, S is the stability coefficient, H is the market heat index, C is the associated promotion coefficient, respectively, the weight parameters of the corresponding coefficients, and satisfy The platform operator can configure according to the business target, such as setting when focusing on supply chain stability, when promoting the market, when associated sales, The stability coefficient S, the market heat index H, and the associated promotion coefficient C are weighted and summed to obtain the priority value P (value range 0-1) of the cold start commodity. For example, the supplier cooperation score of a new wireless earphone is 85 points (stability coefficient 0.85), the inventory turnover period is 45 days (coefficient 0.75), the social media mention volume is 1500 times (heat index 0.8), the KOL recommendation is 3 times (index 0.7), and the complementary product "earphone charging box" sales growth rate is 15% (promotion coefficient 0.65). According to the weight , , The priority value P=0.3 0.8+0.4 0.75+0.3 0.65=0.735, which indicates that the higher the value, the stronger the comprehensive potential of the commodity, providing a quantitative basis for subsequent generation of virtual user features and sorting insertion, effectively solving the problem of insufficient sorting basis caused by the lack of direct competitor data for cold start commodities.

[0084] In one embodiment, characterized in that the step of generating a virtual second user feature by combining the priority value with a preset conversion algorithm includes:

[0085] Determining the near-field information [n-1, n+1] of the priority value corresponding commodity from the currently displayed sorting information;

[0086] Based on the time decay characteristics of the user historical click, add-to-cart, and purchase behavior of the near-field information corresponding commodity, a virtual second user feature is generated using the conversion algorithm.

[0087] The above-mentioned conversion algorithm is:

[0088]

[0089] Wherein, V is the virtual second user feature vector, N is the near-field information set [n-1, n+1], is the historical click count, add-to-cart count, and purchase count of the near-field commodity i, is the time interval corresponding to the behavior from the current, is the coefficient of the time decay characteristic, The maximum value of the platform historical behavior times, The characteristic dimension basis vector of the corresponding behavior.

[0090] In detail, the target insertion position n of the cold-start commodity in the current ranking information is determined according to the priority value P of the cold-start commodity (for example, the priority value 0.735 corresponds to insertion into the 8th position), and the near-field information [n-1, n+1] (that is, the commodities at the 7th and 9th positions) of the position is extracted from the displayed ranking information to form a near-field commodity set N={commodity 7, commodity 9}. The near-field commodities usually have potential correlations with the cold-start commodity in terms of categories, prices, or functions (for example, the cold-start commodity is a wireless earphone, and the near-field commodities can be a Bluetooth sound box or an earphone protective sleeve at the same price), and the user historical behavior data thereof can reflect the preference patterns of the target user group.

[0091] For each commodity in the near-field commodity set, the time decay features of the user historical click, add-to-cart, and purchase behaviors thereof are extracted: the time decay weight of the click behavior is calculated by ( The number of days from the click time to the current time, =0.1 is the decay coefficient), the decay weight of the add-to-cart behavior is , and the decay weight of the purchase behavior is , which respectively reflect the recent attention intensity of the user to the near-field commodities. For example, the near-field commodity 7 was clicked 5 days ago , added to the cart 10 days ago , and purchased 15 days ago , the near-field commodity 9 was clicked 3 days ago , added to the cart 8 days ago , and has no purchase record 、

[0092] Subsequently, the conversion algorithm generates virtual user features based on the time decay features of the near-field commodities: the weighted average values of the click, add-to-cart, and purchase behaviors of the near-field commodities are calculated, respectively, the virtual number of clicks, the virtual number of add-to-carts, and the virtual number of purchases, the virtual number is combined with the time decay weight to generate virtual interaction records with time stamps, and the exponential decay weight is given according to the time proximity, and finally these virtual records are converted into vectors consistent with the real user feature dimensions (such as 128-dimensional user behavior embedding vectors) as virtual second user features.

[0093] In one embodiment, the second user features and the second commodity attribute features are input into an adversarial network model, and finally the step of re-outputting the ranking information of the cold-start commodity after the display insertion update includes:

[0094] inputting the second user feature and the second commodity attribute feature into an adversarial network model;

[0095] According to the same working mode of the generator and the discriminator as the first user feature and the first commodity attribute feature, the cold start commodity is inserted into the corresponding ranking position to obtain updated ranking information.

[0096] In the process of specific implementation, the virtually generated second user feature (such as a time decay vector containing near-field commodity virtual clicks, adds, and purchase behaviors) is dimensionally aligned and spliced with the second commodity attribute feature (such as a category, price, brand, and function description coding vector) of the cold start commodity to form a comprehensive input feature vector (for example, a 128-dimensional user-commodity joint feature vector) of the cold start commodity. The feature vector and the first user feature and the first commodity attribute feature of other commodities (such as mature commodity features of existing user interaction data) obtained in real time are input into the generator, which is processed in the same working mode as the first feature: the bottom layer network performs time sequence weighting on the time decay feature and category coding conversion on the commodity attribute through a trainable weight matrix, the middle layer network generates user-commodity interaction features through element-by-element multiplication and dimension concatenation, and the high layer network calculates the ranking score (such as a comprehensive score combining click rate, conversion rate, and profit) of each commodity (including the cold start commodity).

[0097] The generator generates a candidate ranking list containing the cold start commodity based on the ranking score (for example, the original list is [commodity A, commodity B, commodity C], and the candidate ranking list is generated as [commodity A, commodity D, commodity B, commodity C] after inserting the cold start commodity D), and then inputs the candidate list and the real commodity ranking list (such as a high-quality ranking result verified by the platform history) into the discriminator. The discriminator extracts the sequence features (such as commodity position index, click conversion rate, and price gradient) of the candidate list and the real list according to the same working mode as processing the first feature, generates feature representations through the bottom recognition network, calculates difference vectors through the middle comparison network, and combines evaluation sentiment values for correction, and the high-level evaluation network outputs the ranking quality score (such as 0.82, indicating that the matching degree of the candidate list and the real high-quality ranking is high).

[0098] If the sorting quality score is higher than the preset threshold (such as 0.7), the cold start commodity is retained in the current position in the candidate list; if the score is insufficient, the generator parameters are adjusted through back propagation (such as optimizing the sorting score calculation weight of the high-level network), and the candidate list is regenerated until the quality score meets the standard. Finally, the candidate list verified by the discriminator is determined as the updated sorting information, which is transmitted to the e-commerce front-end system through the cloud platform interface, and the real-time sorting result containing the cold start commodity is displayed in the user search or browsing page (for example, when the user searches for “wireless earphones”, the new cold start earphone D appears in the second position, and a “new product recommendation” label is marked below it).

[0099] For example, the second user feature of a certain cold start wireless earphone D includes the time decay vector of the virtual click (22.7 times) and the add-to-cart (9.1 times) behavior of the nearby goods (Bluetooth speakers and earphone protective sleeves of the same price), and the second commodity attribute feature is the “consumer electronics - earphones - wireless” category (embedding vector), the price of 399 yuan (normalized value 0.7), the brand “XX” (embedding vector), and the function description semantic vector (cosine similarity 0.85). After the generator processes, the sorting score is calculated as 0.89 (higher than 0.82 of commodity B and 0.75 of commodity C in the original list), and the candidate list is updated to [commodity A (0.92), commodity D (0.89), commodity B (0.82), commodity 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 commodity D is inserted into the second position, and the user can see its display position in the search result page, realizing the reasonable sorting and effective exposure of the cold start commodity based on virtual features.

[0100] Reference is made to the accompanying drawings Figure 2 A structural block diagram of an e-commerce platform commodity sorting cloud platform based on a generative adversarial network is proposed for the present application, and the cloud platform comprises:

[0101] A receiving unit is configured to receive a user's search request and / or browsing behavior data, extract first user features and first commodity attribute features, the first user features including time decay features of user historical click, add-to-cart, and purchase behavior, and the first commodity attribute features including commodity category, price, brand, and evaluation information;

[0102] A model unit is configured to use a pre-trained adversarial network model, which includes a generator and a discriminator. The generator is a multi-layer deception neural network, which inputs the first user features and the first commodity attribute features and outputs a candidate commodity sorting list. The discriminator is a multi-layer identification neural network, which inputs the real commodity sorting list and the candidate commodity sorting list generated by the generator, and outputs and displays the sorting information.

[0103] The request unit is configured to, for a cold-start commodity requiring insertion of the sorting information, acquire a second commodity attribute feature corresponding to the cold-start commodity, perform approximate competitor quantity judgment through the second commodity attribute feature, and if the approximate competitor quantity is lower than a preset value, calculate a priority value by using a preset association algorithm to crawl association information of the cold-start commodity.

[0104] The updating unit is configured to generate a virtual second user feature by using a preset conversion algorithm in combination with the priority value, and input the second user feature and the second commodity attribute feature into a generative adversarial network model, and finally output updated sorting information of the cold-start commodity after display insertion.

[0105] In summary, the cold-start commodity sorting problem is solved. The features are extracted by receiving user behavior data, and the generator outputs a candidate sorting list and the discriminator outputs sorting information. For the cold-start commodity, the attribute feature is acquired to judge the approximate competitor quantity, the association information is crawled to calculate the priority value when the approximate competitor quantity is lower than a preset value, the virtual user feature is generated by using the conversion algorithm, and the updated sorting information is output after inputting into the model. The method improves the sorting accuracy of the cold-start commodity, enhances the generalization ability of the model, balances the user experience and the platform target, and has significant technical innovation and practical value.

[0106] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, storage, database or other medium provided by the present application and used in the 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. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0107] It is to be understood that the terminology "including", "comprising", or any other variation thereof, is intended to cover a non-exclusive inclusion such that process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0108] The above description is merely the preferred embodiments of the present application, and is not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made according to the content of the present application specification and drawings, or directly or indirectly applied to other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method for ranking e-commerce platform goods based on a generative adversarial network, characterized in that, The method comprises the following steps: Receiving a search request and / or browsing behavior data of a user, extracting first user features and first commodity attribute features, the first user features including time decay features of user historical clicks, adds-to-cart, and purchase behaviors, and the first commodity attribute features including commodity categories, prices, brands, and evaluation information; Based on a pre-trained adversarial network model, the model including a generator and a discriminator, wherein the generator is a multi-layered fraud-making neural network, inputting the first user features and the first commodity attribute features, and outputting a candidate commodity ranking list; the discriminator is a multi-layered identification neural network, inputting a real commodity ranking list and the candidate commodity ranking list generated by the generator, and outputting and displaying ranking information; For a cold-start commodity requiring insertion into the ranking information, obtaining second commodity attribute features corresponding to the cold-start commodity, performing approximate competitor quantity judgment through the second commodity attribute features, and if the approximate competitor quantity is lower than a preset value, calculating a priority value by crawling and calculating the association information of the cold-start commodity using a preset association algorithm; Using a preset conversion algorithm, generating virtual second user features in combination with the priority value, and inputting the second user features and the second commodity attribute features into the adversarial network model, and finally re-outputting and displaying updated ranking information with the cold-start commodity inserted; wherein, Determining the near-field information [n-1, n+1] of the commodity corresponding to the priority value from the currently displayed ranking information, generating virtual second user features using the conversion algorithm based on the time decay features of user historical clicks, adds-to-cart, and purchase behaviors of the commodities corresponding to the near-field information; and The conversion algorithm is: ； wherein V is a virtual second user feature vector, N is a near-field information set [n-1, n+1], is the historical click times, add times, purchase times of the near-field commodity i, is the time interval from the current behavior, is the coefficient of the time decay feature, is the maximum value of the platform historical behavior times, is the feature dimension base vector of the corresponding behavior, and the time decay function wherein t is the number of days from the behavior occurrence time to the current time, configurable decay rate parameter. 2.The method of claim 1, wherein, The step of the generator being a multi-layered fraud-making neural network, inputting the first user features and the first commodity attribute features, and outputting a candidate commodity ranking list, comprises: Performing normalization processing on the first user features and the first commodity attribute features input into the multi-layered fraud-making neural network to form a dimension-unified feature vector set; Using a bottom-layer network to preliminarily map the feature vector set, setting different weight matrices for each layer of the bottom-layer network, performing time-series weighting on the time decay features to obtain a user feature vector, and performing category encoding conversion on the commodity category, price, brand, and evaluation information to obtain a commodity attribute feature vector; The middle-layer network cross-splices the user feature vector and the commodity attribute feature vector, generates user-commodity interaction features in a way of element-by-element multiplication and dimension concatenation, and inputs the user-commodity interaction features into a high-layer network after activation function processing; The high-layer network calculates a ranking score of each commodity through a fully connected layer, arranges the candidate commodities in ascending or descending order according to the ranking score, and generates a candidate commodity ranking list. 3.The method of claim 1, wherein, The step of the discriminator being a multi-layered identification neural network, inputting a real commodity ranking list and a candidate commodity ranking list generated by the generator, and outputting and displaying ranking information, comprises: Performing sequence feature extraction on the real commodity ranking list and the candidate commodity ranking list respectively, and the extracted content including position indexes of commodities in the list, user historical click conversion rates, commodity price gradients, and brand market shares; The sequence features extracted are standardized by a bottom layer recognition network, and the feature vectors of the real commodity ranking list and the candidate commodity ranking list are input into two parallel sub-networks respectively to obtain real feature representation and candidate feature representation; The middle layer comparison network performs element-by-element difference calculation on the real feature representation and the candidate feature representation to generate a feature difference vector, and the difference vector is weighted and corrected in combination with the sentiment tendency value of the commodity evaluation information; The high-level evaluation network calculates the ranking quality score through a fully connected layer, outputs the ranking information including the ranking quality score and the difference commodity position, and displays it on the cloud platform. 4.The method of claim 1, wherein, For a cold-start commodity required to be inserted into the ranking information, the second commodity attribute feature corresponding to the cold-start commodity is obtained, and the approximate competitor quantity judgment step is performed through the second commodity attribute feature, including: Filtering a candidate competitor set consistent with the commodity category label from the cloud platform historical database through the commodity category, price, brand and function description included in the second commodity attribute feature; For each competitor in the candidate competitor set, the overlap degree of the price interval value, the association degree of the brand identifier, and the semantic similarity of the function description text of the cold-start commodity are calculated; The number of competitors in the candidate competitor set that satisfy the price overlap degree ≥ the first threshold value, the brand association degree ≥ the second threshold value, and the semantic similarity ≥ the third threshold value is counted, and is recorded as the approximate competitor quantity. 5.The method of claim 4, wherein, If the approximate competitor quantity is lower than the preset value, the step of calculating the priority value of the cold-start commodity by crawling the associated information of the cold-start commodity using a preset association algorithm, including: Crawling the supplier cooperation score, inventory turnover period data of the cold-start commodity from the cloud platform supply chain database, and crawling the recent mention quantity of the cold-start commodity keyword, KOL recommendation record data from the pre-bound social media platform, and crawling the complementary product category commodity sales growth rate data in the past three months from the cloud platform sales database to obtain the associated information; Numericalizing the associated information, which maps the supplier cooperation score and inventory turnover period data into a stability coefficient of 0-1, converts the mention quantity and recommendation record in the social media platform into a market heat index, and converts the complementary product category sales growth rate into an association promotion coefficient; According to the association algorithm, the priority value of the cold-start commodity is calculated. 6.The method of claim 5, wherein, The correlation algorithm formula is: Wherein, P is the priority value, S is the stability coefficient, H is the market heat index, C is the correlation promotion coefficient, The corresponding coefficient weight parameters satisfy . 7.The method of claim 1, wherein, The step of finally re-outputting the updated ranking information of the cold-start commodity inserted into the ranking information by inputting the second user feature and the second commodity attribute feature into the adversarial network model, including: Inputting the second user feature and the second commodity attribute feature into the adversarial network model; According to the same generator and discriminator working mode as the first user feature and the first commodity attribute feature, the cold-start commodity is inserted into the corresponding ranking position to obtain the updated ranking information.

8. An e-commerce platform product ranking cloud platform based on a generative adversarial network, characterized in that the cloud platform Including: The receiving unit is configured to receive a search request and / or browsing behavior data of a user, extract first user features and first commodity attribute features, the first user features including time decay features of user historical click, add-to-cart, and purchase behaviors, and the first commodity attribute features including commodity categories, prices, brands, and evaluation information; A model unit is configured to generate a candidate product ranking list based on a pre-trained adversarial network model, which includes a generator and a discriminator. The generator is a multi-layered fraud-making neural network, which takes first user features and first product attribute features as inputs and outputs the candidate product ranking list. The discriminator is a multi-layered identification neural network, which takes a real product ranking list and the candidate product ranking list generated by the generator as inputs and outputs and displays ranking information. A request unit is configured to obtain second product attribute features corresponding to a cold-start product that requires to be inserted into the ranking information, perform approximate competitor quantity judgment based on the second product attribute features, and calculate a priority value of the cold-start product by crawling associated information of the cold-start product using a preset association algorithm if the approximate competitor quantity is lower than a preset value. An update unit is configured to generate virtual second user features by combining the priority value using a preset conversion algorithm, input the second user features and the second product attribute features into the adversarial network model, and finally output and display the ranking information of the cold-start product after the update. The priority value corresponding to the near-field information [n-1, n+1] of the product is determined from the ranking information displayed at the moment. The virtual second user features are generated using the conversion algorithm based on the time decay characteristics of the user historical click, add-to-cart, and purchase behaviors of the product corresponding to the near-field information. And The conversion algorithm is ； wherein V is a virtual second user feature vector, N is a near-field information set [n-1, n+1], is the historical click times, add times, purchase times of the near-field commodity i, is the time interval from the current behavior, is the coefficient of the time decay feature, is the maximum value of the platform historical behavior times, is the feature dimension base vector of the corresponding behavior, and the time decay function wherein t is the number of days from the behavior occurrence time to the current time, a configurable decay rate parameter.

Images