Collocation recommendation method and system based on association rule mining

Abstract

The present invention provides a collocation recommendation method and system based on association rule mining, which relates to the technical field of association rule mining, including obtaining user historical transaction data to generate a user commodity transaction rating matrix and standardizing it, and calculating the cosine similarity between users to predict the commodity rating to obtain an initial candidate set; then normalizing the commodity price and combining the category hierarchy structure to calculate the comprehensive similarity of the commodity for clustering and grouping, and screening frequent item sets that meet the support and confidence thresholds; finally extracting the commodity price change sequence to calculate the dynamic price matching degree, constructing a multi-dimensional evaluation index, and optimizing it using a non-dominated sorting genetic algorithm and a deep Q network to achieve dynamic evaluation and recommendation based on scenarios. The present invention improves the accuracy and timeliness of commodity collocation recommendations and enhances the interpretability of recommendation results.

Description

Technical Field

[0001] The present invention relates to association rule mining technology, and in particular to a collocation recommendation method and system based on association rule mining. Background Art

[0002] With the rapid development of e-commerce, product pairing recommendations have become an important means to improve user shopping experience and platform sales. At present, mainstream e-commerce platforms generally use methods based on association rule mining to recommend product pairings. By analyzing user historical transaction data and mining the associations between products, they can provide users with personalized product pairing suggestions. This method can effectively capture users' shopping habits and the rules for pairing between products.

[0003] Traditional association rule mining methods mainly rely on static historical transaction data and fail to fully consider the dynamic changes in commodity prices, resulting in recommendation results that may not match the real-time market conditions. Existing methods often only focus on the similarity measurement of a single dimension when matching commodities, ignoring the comprehensive impact of multi-dimensional evaluation indicators such as price matching and category diversity. Existing recommendation systems lack the ability to dynamically perceive user preferences in different scenarios and are unable to adjust recommendation strategies in a timely manner according to the real-time market environment and user behavior.

[0004] The present invention aims to solve the above technical problems and proposes a combination recommendation method that can comprehensively consider the dynamic changes of commodity prices, multi-dimensional evaluation indicators and scene adaptability. This method constructs a dynamic price matching degree and a multi-dimensional evaluation system, combines a non-dominated sorting genetic algorithm and a deep Q network, and realizes dynamic optimization and scene adaptation of commodity combination recommendations, thereby improving the accuracy and real-time performance of recommendations. Summary of the invention

[0005] The embodiments of the present invention provide methods and systems that can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides a collocation recommendation method based on association rule mining, including:

[0008] Obtain the user's historical transaction data, extract the user number, product number, product category number, product price, purchase time, and transaction score from the historical transaction data, and generate a user product transaction score matrix; calculate the average and standard deviation of the user score, standardize the user product transaction score matrix, eliminate the score deviation, and obtain a standardized score matrix; based on the standardized score matrix, use cosine similarity to calculate the cosine similarity value between users to generate a user similarity matrix; predict the missing score of each product according to the user similarity matrix to obtain a score prediction matrix; filter the product combinations whose predicted scores are lower than a preset score threshold in the score prediction matrix to obtain an initial product candidate set;

[0009] Perform logarithmic normalization on the commodity prices in the initial commodity candidate set to generate a price feature vector; calculate the price distance between commodities based on the price feature vector and construct a price similarity matrix; obtain the commodity category hierarchical structure, calculate the category path distance between commodities and generate a category similarity matrix; perform weighted fusion of the price similarity matrix and the category similarity matrix to obtain a comprehensive commodity similarity matrix; cluster and group the commodities in the initial commodity candidate set based on the comprehensive commodity similarity matrix to obtain a commodity combination set with similar price ranges and categories; calculate the support and confidence of each commodity combination in the commodity combination set, and select the combination that meets the minimum support threshold and the minimum confidence threshold as a frequent item set;

[0010] For the commodity combinations in the frequent item set, firstly, the recent price change sequence of each commodity is extracted, and the correlation coefficient of the price change trend is calculated to generate a dynamic price matching degree; based on the dynamic price matching degree, the frequent item set is reordered to generate a commodity combination sequence that takes into account the dynamic price changes; by constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect, and combining with a non-dominated sorting genetic algorithm, the Pareto optimal solution set is obtained; the deep Q network is used to optimize parameters, and the dynamic adjustment of the evaluation indicator weights is achieved based on scene recognition; the differentiated sliding window update mechanism is used to complete the commodity combination evaluation and make commodity combination matching recommendations.

[0011] Based on the standardized rating matrix, the cosine similarity values ​​between users are calculated using cosine similarity to generate a user similarity matrix; the missing rating of each product is predicted according to the user similarity matrix to obtain a rating prediction matrix including:

[0012] Obtain user product rating data, calculate the average rating and the standard deviation of the rating of each user in the user product rating data; based on the average rating and the standard deviation of the rating, perform standard processing on the user product rating data to obtain a rating standardization matrix; according to the rating vectors of the commonly rated products between the user pairs in the rating standardization matrix, use the cosine similarity formula to calculate the cosine similarity value between the user pairs, the cosine similarity value is calculated by dividing the inner product of the rating vectors between the user pairs by the product of the modulus of the rating vectors, and generate a user similarity matrix;

[0013] Identify the missing rating position in the user's product rating data, where the missing rating position corresponds to the position where the user has not rated the product; for each missing rating position, extract the rating similarity value between the current user and other users from the user similarity matrix; sort other users in descending order according to the rating similarity value, and select K users with the highest rating similarity as the neighbor user set; extract a subset of users from the neighbor user set who have rated the product corresponding to the current missing rating position, and the rating of each user in the user subset needs to be standardized;

[0014] For the users in the user subset, a scoring weight is calculated based on their cosine similarity values ​​in the user similarity matrix, and the scoring weight is obtained by normalizing the similarity value; the scoring weight is weighted and summed with the corresponding standardized score in the user subset to obtain the predicted score of the current missing score position; the predicted score is de-normalized by the average score and the score standard deviation of the current user to obtain the predicted score value under the original score scale;

[0015] Repeat the above prediction process for all missing rating positions, combine the predicted rating values ​​with the original rating data, and construct a complete rating prediction matrix; calculate the confidence of each predicted rating value in the rating prediction matrix, and the confidence is calculated based on the number of neighboring users participating in the prediction, the discreteness of the neighboring user ratings, and the average similarity of the neighboring users; filter the predicted rating values ​​in the rating prediction matrix according to the confidence, mark the predicted rating values ​​with a confidence lower than a preset threshold as invalid predictions, and obtain the final rating prediction matrix.

[0016] The prices of the commodities in the initial commodity candidate set are logarithmically normalized to generate a price feature vector; the price distance between commodities is calculated based on the price feature vector to construct a price similarity matrix; the commodity category hierarchical structure is obtained, the category path distance between commodities is calculated, and a category similarity matrix is ​​generated; the price similarity matrix is ​​weightedly fused with the category similarity matrix to obtain a comprehensive commodity similarity matrix, including:

[0017] Extracting price data of each commodity in the initial commodity candidate set, performing logarithmic transformation on the price data to obtain an initial logarithmic value; calculating the mean and standard deviation of the initial logarithmic value, normalizing the initial logarithmic value based on the mean and standard deviation, and obtaining a price feature vector corresponding to each commodity; calculating price distribution characteristics within the commodity category based on the price feature vector, the price distribution characteristics including the price variation coefficient within the category and the price range of the category;

[0018] Based on the price feature vector, a commodity price distance matrix is ​​constructed, wherein each element of the price distance matrix is ​​obtained by calculating the square root of the sum of the differences between the components of the two commodity price feature vectors; a Gaussian kernel function is applied to the price distance matrix for nonlinear mapping, wherein the bandwidth parameter of the Gaussian kernel function is optimized and determined by a cross-validation method, and a price similarity matrix is ​​generated;

[0019] Obtain the category label of each product in the initial product candidate set and construct a category hierarchical tree structure; calculate the maximum depth of the category hierarchical tree structure and assign a weight coefficient to each depth level, wherein the weight coefficient decays exponentially with increasing hierarchical depth; extract the category path between product pairs and calculate the depth of the nearest common ancestor node; calculate the weighted category path distance between product pairs based on the maximum depth, the weight coefficient and the depth of the nearest common ancestor node, and generate a category similarity matrix;

[0020] Based on the price distribution characteristics, the price variation coefficient within the category is calculated, and the price variation coefficient is obtained by dividing the standard deviation of the commodity prices within the category by the average value; the degree of overlap of the price ranges between commodity pairs is calculated to generate the price range overlap rate; the feature fusion weight is calculated by combining the price variation coefficient and the price range overlap rate; the price similarity matrix and the category similarity matrix are weightedly combined according to the feature fusion weight to obtain a comprehensive commodity similarity matrix.

[0021] Based on the comprehensive similarity matrix of commodities, the commodities in the initial commodity candidate set are clustered and grouped to obtain a commodity combination set with similar price ranges and categories; the support and confidence of each commodity combination in the commodity combination set are calculated, and the combinations that meet the minimum support threshold and the minimum confidence threshold are selected as frequent item sets, including:

[0022] Clustering and grouping the commodities in the initial commodity candidate set based on the commodity comprehensive similarity matrix; the clustering and grouping process comprises: pairing the commodities in the initial commodity candidate set with each other to calculate commodity similarity values, and dividing the commodities into different initial commodity combinations based on the commodity similarity values;

[0023] Obtaining price range information and category information of the commodities in the initial commodity combination, calculating the price mean and price variance of the commodities in each initial commodity combination based on the price range information, and calculating the category similarity of the commodities in each initial commodity combination based on the category information; splitting and reorganizing the initial commodity combination whose price variance is greater than a first price preset threshold or whose category similarity is less than a second category similarity preset threshold, to obtain a set of commodity combinations that meet the price range and have similar categories;

[0024] Counting the support of each commodity combination in the commodity combination set based on the historical transaction data, where the support is the ratio of the number of co-occurrences of the commodity combination in the historical transaction data to the total number of transactions; calculating the confidence of each commodity combination in the commodity combination set, where the confidence is the co-occurrence probability of the commodity combination in the historical transactions;

[0025] A minimum support threshold and a minimum confidence threshold are set, the support and confidence of each commodity combination in the commodity combination set are calculated, and commodity combinations whose support is greater than the minimum support threshold and whose confidence is greater than the minimum confidence threshold are selected as frequent item sets.

[0026] For the commodity combinations in the frequent item set, extract the recent price change sequence of each commodity, calculate the correlation coefficient of the price change trend, and generate a dynamic price matching degree; reorder the frequent item set based on the dynamic price matching degree, and obtain a commodity combination sequence considering the dynamic price change, including:

[0027] For each commodity combination in the frequent item set, extract the recent price change sequence of each commodity in the commodity combination from the historical transaction data, and the price change sequence includes the daily price of the past thirty days; use the maximum and minimum value normalization method to map the price change sequence into a standardized price sequence, and the value range of the standardized price sequence is zero to one;

[0028] Calculating the price correlation coefficients of the price change trends for the standardized price series of each commodity in the commodity combination, and constructing a price correlation coefficient matrix based on the price correlation coefficients, wherein the price correlation coefficient matrix is ​​used to characterize the synchronization of price changes of each pair of commodities in the commodity combination; calculating the statistical eigenvalues ​​of the price correlation coefficient matrix, including the mean, variance and maximum eigenvalue;

[0029] The dynamic price matching degree of the commodity combination is generated based on the statistical eigenvalues ​​of the price correlation coefficient matrix, and the dynamic price matching degree is obtained by weighted calculation of the mean, variance and maximum eigenvalue; a re-ranking scoring function is constructed, and the re-ranking scoring function integrates the dynamic price matching degree with the support and confidence of the commodity combination; and the commodity combinations in the frequent item set are sorted in descending order according to the re-ranking scoring function to obtain a commodity combination sequence that takes into account the dynamic changes in prices.

[0030] The method further comprises:

[0031] Collect the competitive product data of each product in the product combination, convert the product title into a word vector to obtain a semantic feature vector, calculate the cosine similarity of the semantic feature vector to obtain the product similarity, and screen the competitive products based on the preset similarity threshold to form a competitive product price monitoring pool;

[0032] Performing kernel density estimation on the competitive product price data in the competitive product price monitoring pool to obtain a competitive product price distribution curve, calculating the mean and quantile of the competitive product price distribution curve to obtain market price distribution characteristics, and calculating the standard deviation of the competitive product price to obtain market price dispersion;

[0033] Calculate the price deviation rate based on the actual price of the commodity and the market price distribution characteristics, perform weighted combination of the price deviation rate and the market price dispersion to obtain a single product price competitiveness index, and use the historical sales volume of the commodity to weight the single product price competitiveness index to obtain a commodity combination price competitiveness index;

[0034] Calculate the price change sequence of the product and the competing product, use a sliding time window to calculate the sliding coefficient of the price change sequence to obtain the price change trend correlation, and calculate the price linkage index based on the price change trend correlation;

[0035] Inputting the commodity combination price competitiveness index and the price linkage index into a pre-trained multi-layer perceptron network, the multi-layer perceptron network outputs a feature weight coefficient, and using the feature weight coefficient to perform weighted fusion on the commodity combination price competitiveness index and the price linkage index to obtain a comprehensive score for the commodity combination;

[0036] The commodity combination sequence is reordered based on the comprehensive score of the commodity combination to generate an optimized commodity combination recommendation sequence, and the commodity combination recommendation sequence is used for commodity combination recommendation.

[0037] By constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect, and combining with non-dominated sorting genetic algorithm to obtain the Pareto optimal solution set; using deep Q network for parameter optimization, and realizing dynamic adjustment of evaluation indicator weights based on scene recognition; using differentiated sliding window update mechanism to complete product combination evaluation and make product combination matching recommendations, including:

[0038] Construct multi-dimensional evaluation indicators, obtain the user's historical purchase price and the current product combination price, calculate the Mahalanobis distance between the two price distributions through the covariance matrix to obtain the price adaptation index; extract the category tree path of each product in the product combination to construct the category distribution vector, and calculate the category distribution uncertainty based on information entropy to obtain the diversity index; collect the historical promotion records of the product within 90 days to establish the conversion rate curve, and calculate the promotion effect index in combination with the current promotion discount depth;

[0039] Normalizing the price adaptation index, the diversity index and the promotion effect index to obtain normalized index values, constructing an initial population for commodity combination evaluation based on the normalized index values, and iteratively optimizing the initial population using a non-dominated sorting genetic algorithm, wherein in each iteration, the crowding distance of individuals in the population is calculated and a new population is generated through tournament selection, crossover and mutation operations until convergence to obtain a Pareto optimal solution set of the commodity combination;

[0040] Based on the Pareto optimal solution set, the state space and action space of the deep Q network are constructed, wherein the state space includes the normalized index value, the historical commodity conversion data and the market environment characteristics, and the action space includes the adjustment scheme of the evaluation parameters. The reward function that comprehensively considers the click rate, conversion rate and rate of return is used for parameter optimization training, and the ε-greedy strategy is used to select the parameter adjustment action during the training process;

[0041] Input the parameter optimization results obtained from the deep Q network training into the scene recognition model, map the current scene to a predefined scene category based on time attributes, user group characteristics and promotion activity attributes, assign differentiated weight initial values ​​to the normalized index values ​​according to the scene category, and dynamically adjust the weight coefficients of each index through an online learning method based on real-time conversion effect feedback data;

[0042] Differentiated sliding time windows are set according to the dynamically adjusted indicator weight coefficients, and the price adaptation index, diversity index and promotion effect index are incrementally updated using corresponding time window sizes. When the fluctuation of any indicator exceeds a preset threshold, the latest indicator value is re-input into the deep Q network for parameter optimization.

[0043] According to a second aspect of the embodiments of the present invention,

[0044] Provides a collocation recommendation system based on association rule mining, including:

[0045] The first unit is used to obtain the historical transaction data of the user, extract the user number, product number, product category number, product price, purchase time, and transaction score from the historical transaction data, and generate a user product transaction score matrix; calculate the average value and standard deviation of the user score, standardize the user product transaction score matrix, eliminate the score deviation, and obtain a standardized score matrix; based on the standardized score matrix, use cosine similarity to calculate the cosine similarity value between users to generate a user similarity matrix; predict the missing score of each product according to the user similarity matrix to obtain a score prediction matrix; filter the product combinations whose predicted scores are lower than a preset score threshold in the score prediction matrix to obtain an initial product candidate set;

[0046] The second unit is used to perform logarithmic normalization processing on the commodity prices in the initial commodity candidate set to generate a price feature vector; calculate the price distance between commodities based on the price feature vector to construct a price similarity matrix; obtain the commodity category hierarchical structure, calculate the category path distance between commodities, and generate a category similarity matrix; perform weighted fusion of the price similarity matrix and the category similarity matrix to obtain a comprehensive commodity similarity matrix; cluster and group the commodities in the initial commodity candidate set based on the comprehensive commodity similarity matrix to obtain a commodity combination set with similar price ranges and categories; calculate the support and confidence of each commodity combination in the commodity combination set, and select the combination that meets the minimum support threshold and the minimum confidence threshold as a frequent item set;

[0047] The third unit is used to generate a dynamic price matching degree for the commodity combinations in the frequent item sets by first extracting the recent price change sequence of each commodity and calculating the correlation coefficient of the price change trend; reordering the frequent item sets based on the dynamic price matching degree to generate a commodity combination sequence that takes into account the dynamic price changes; obtaining the Pareto optimal solution set by constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect and combining the non-dominated sorting genetic algorithm; optimizing parameters using a deep Q network and dynamically adjusting the evaluation indicator weights based on scene recognition; and using a differentiated sliding window update mechanism to complete commodity combination evaluation and make commodity combination matching recommendations.

[0048] According to a third aspect of the embodiments of the present invention,

[0049] An electronic device is provided, comprising:

[0050] processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0053] A fourth aspect of the embodiments of the present invention is:

[0054] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0055] The beneficial effects of this application are as follows:

[0056] Improve recommendation accuracy: By combining multi-dimensional information such as user ratings, price similarity, category similarity, etc. to recommend product combinations, we can more accurately capture the relationship between user needs and products, thereby improving the accuracy of recommendations, avoiding recommending irrelevant products, and improving user experience.

[0057] Enhanced personalization and diversity of recommendations: Taking into account the dynamic price trends of commodities and combining multi-dimensional evaluation indicators such as price adaptability, category diversity and promotion effect, more personalized commodity combination recommendations are achieved, which can meet users' diverse shopping needs and provide more abundant choices.

[0058] Optimize recommendation efficiency: The deep Q network is used for parameter optimization and scene recognition, and a differentiated sliding window update mechanism is adopted to dynamically adjust the evaluation index weights and efficiently complete product combination evaluation and recommendation, thereby improving recommendation efficiency and reducing the consumption of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of a collocation recommendation method based on association rule mining according to an embodiment of the present invention;

[0060] Figure 2 The figure is a schematic diagram of the structure of a collocation recommendation system based on association rule mining according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a collocation recommendation method based on association rule mining according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S101. Obtain the historical transaction data of the user, extract the user number, product number, product category number, product price, purchase time, and transaction score from the historical transaction data, and generate a user product transaction score matrix; calculate the average and standard deviation of the user scores, standardize the user product transaction score matrix, eliminate the score deviation, and obtain a standardized score matrix; based on the standardized score matrix, use cosine similarity to calculate the cosine similarity values ​​between users to generate a user similarity matrix; predict the missing score of each product according to the user similarity matrix to obtain a score prediction matrix; filter the product combinations whose predicted scores are lower than a preset score threshold in the score prediction matrix to obtain an initial product candidate set;

[0065] S102. Perform logarithmic normalization on the commodity prices in the initial commodity candidate set to generate a price feature vector; calculate the price distance between commodities based on the price feature vector to construct a price similarity matrix; obtain the commodity category hierarchical structure, calculate the category path distance between commodities, and generate a category similarity matrix; perform weighted fusion of the price similarity matrix and the category similarity matrix to obtain a comprehensive commodity similarity matrix; cluster and group the commodities in the initial commodity candidate set based on the comprehensive commodity similarity matrix to obtain a commodity combination set with similar price ranges and categories; calculate the support and confidence of each commodity combination in the commodity combination set, and select the combination that meets the minimum support threshold and the minimum confidence threshold as a frequent item set;

[0066] S103. For the commodity combinations in the frequent item set, firstly extract the recent price change sequence of each commodity, calculate the correlation coefficient of the price change trend and generate a dynamic price matching degree; reorder the frequent item set based on the dynamic price matching degree to generate a commodity combination sequence that takes into account the dynamic price changes; construct multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect, and obtain the Pareto optimal solution set in combination with a non-dominated sorting genetic algorithm; use a deep Q network to optimize parameters, and realize dynamic adjustment of the evaluation indicator weights based on scene recognition; use a differentiated sliding window update mechanism to complete the commodity combination evaluation and make commodity combination matching recommendations.

[0067] In an optional implementation, based on the standardized rating matrix, cosine similarity is used to calculate the cosine similarity values ​​between users to generate a user similarity matrix; the missing rating of each product is predicted according to the user similarity matrix to obtain a rating prediction matrix including:

[0068] Obtain user product rating data, calculate the average rating and the standard deviation of the rating of each user in the user product rating data; based on the average rating and the standard deviation of the rating, perform standard processing on the user product rating data to obtain a rating standardization matrix; according to the rating vectors of the commonly rated products between the user pairs in the rating standardization matrix, use the cosine similarity formula to calculate the cosine similarity value between the user pairs, the cosine similarity value is calculated by dividing the inner product of the rating vectors between the user pairs by the product of the modulus of the rating vectors, and generate a user similarity matrix;

[0069] Identify the missing rating position in the user's product rating data, where the missing rating position corresponds to the position where the user has not rated the product; for each missing rating position, extract the rating similarity value between the current user and other users from the user similarity matrix; sort other users in descending order according to the rating similarity value, and select K users with the highest rating similarity as the neighbor user set; extract a subset of users from the neighbor user set who have rated the product corresponding to the current missing rating position, and the rating of each user in the user subset needs to be standardized;

[0070] For the users in the user subset, a scoring weight is calculated based on their cosine similarity values ​​in the user similarity matrix, and the scoring weight is obtained by normalizing the similarity value; the scoring weight is weighted and summed with the corresponding standardized score in the user subset to obtain the predicted score of the current missing score position; the predicted score is de-normalized by the average score and the score standard deviation of the current user to obtain the predicted score value under the original score scale;

[0071] Repeat the above prediction process for all missing rating positions, combine the predicted rating values ​​with the original rating data, and construct a complete rating prediction matrix; calculate the confidence of each predicted rating value in the rating prediction matrix, and the confidence is calculated based on the number of neighboring users participating in the prediction, the discreteness of the neighboring user ratings, and the average similarity of the neighboring users; filter the predicted rating values ​​in the rating prediction matrix according to the confidence, mark the predicted rating values ​​with a confidence lower than a preset threshold as invalid predictions, and obtain the final rating prediction matrix.

[0072] A user collaborative filtering recommendation method and system is used to predict user ratings of products and improve the accuracy of the recommendation system.

[0073] First, obtain the user product rating data. Next, calculate the average rating and standard deviation of each user. For example, the average rating of user 1 is (5+3) / 2=4, and the standard deviation is 1. Perform the same calculation for all user ratings. Then, based on the calculated average and standard deviation, standardize the user product rating data. The standardization method is to subtract the average rating of each user from each rating of each user, and then divide it by the standard deviation of the user.

[0074] Next, calculate the similarity between users. For every two users, find the items they have rated together, and then use the standardized rating vectors of these two users for these items to calculate the cosine similarity. Cosine similarity is obtained by dividing the inner product of two vectors by the product of the two vector moduli. For example, user 1 and user 2 have rated movie A and movie B together. The standardized rating vector of user 1 is [1,-1], and the standardized rating vector of user 2 is [0, (missing values ​​cannot be calculated)]. Due to the existence of missing values, only the similarity part for movie A can be calculated here. Calculate the similarity between all users to obtain the user similarity matrix.

[0075] Next, predict the missing ratings. For example, to predict user 1's rating for movie C, first extract the similarity values ​​between user 1 and other users from the user similarity matrix. Then, sort the other users in descending order according to the similarity value, and select the K users with the highest similarity (assuming K=2) as the set of neighboring users. Assume that user 2 and user 3 are the neighboring users of user 1. Since user 2 did not rate movie C, the final user subset is only user 3. Then, calculate the rating weights of the neighboring users. The weight is obtained by dividing the similarity value of each neighboring user by the sum of the similarity values ​​of all neighboring users. In this example, the weight of user 3 is 1.

[0076] Next, the standardized scores of neighboring users and their corresponding weights are weighted and summed to obtain the predicted standardized score. In this example, user 1's predicted standardized score for movie C is 1*5=5. Then, the predicted standardized score is de-standardized to obtain the predicted score under the original rating scale. The de-standardization method is to multiply the predicted standardized score by the standard deviation of the current user, and then add the average score of the current user. In this example, user 1's predicted score for movie C is 5*1+4=9. Since the rating range is 1-5, truncation is required, and the final predicted score is 5.

[0077] Repeat the above prediction process for all missing rating positions to obtain a complete rating prediction matrix. Finally, calculate the confidence of each predicted rating. The confidence can be calculated based on the number of neighboring users participating in the prediction, the discreteness of the neighboring user ratings, and the mean similarity of the neighboring users. For example, if the number of users participating in the prediction is large, the discreteness of the ratings is small, and the mean similarity is high, the confidence is high. Filter the predicted ratings according to the confidence, mark the predicted ratings with a confidence lower than the preset threshold as invalid predictions, and finally obtain the rating prediction matrix.

[0078] This application can:

[0079] Improve prediction accuracy: Through standardization and weighted average based on similarity, users' ratings of products can be predicted more accurately. Personalized recommendations: By predicting users' ratings of unrated products, users can be better provided with personalized product recommendations. Improve user experience: More accurate recommendations can help users discover products they are interested in and improve user experience.

[0080] In an optional implementation, logarithm normalization is performed on the commodity prices in the initial commodity candidate set to generate a price feature vector; the price distance between commodities is calculated based on the price feature vector to construct a price similarity matrix; the commodity category hierarchy structure is obtained, the category path distance between commodities is calculated, and a category similarity matrix is ​​generated; the price similarity matrix is ​​weightedly fused with the category similarity matrix to obtain a comprehensive commodity similarity matrix, which includes:

[0081] Extracting price data of each commodity in the initial commodity candidate set, performing logarithmic transformation on the price data to obtain an initial logarithmic value; calculating the mean and standard deviation of the initial logarithmic value, normalizing the initial logarithmic value based on the mean and standard deviation, and obtaining a price feature vector corresponding to each commodity; calculating price distribution characteristics within the commodity category based on the price feature vector, the price distribution characteristics including the price variation coefficient within the category and the price range of the category;

[0082] Based on the price feature vector, a commodity price distance matrix is ​​constructed, wherein each element of the price distance matrix is ​​obtained by calculating the square root of the sum of the differences between the components of the two commodity price feature vectors; a Gaussian kernel function is applied to the price distance matrix for nonlinear mapping, wherein the bandwidth parameter of the Gaussian kernel function is optimized and determined by a cross-validation method, and a price similarity matrix is ​​generated;

[0083] Obtain the category label of each product in the initial product candidate set and construct a category hierarchical tree structure; calculate the maximum depth of the category hierarchical tree structure and assign a weight coefficient to each depth level, wherein the weight coefficient decays exponentially with increasing hierarchical depth; extract the category path between product pairs and calculate the depth of the nearest common ancestor node; calculate the weighted category path distance between product pairs based on the maximum depth, the weight coefficient and the depth of the nearest common ancestor node, and generate a category similarity matrix;

[0084] Based on the price distribution characteristics, the price variation coefficient within the category is calculated, and the price variation coefficient is obtained by dividing the standard deviation of the commodity prices within the category by the average value; the degree of overlap of the price ranges between commodity pairs is calculated to generate the price range overlap rate; the feature fusion weight is calculated by combining the price variation coefficient and the price range overlap rate; the price similarity matrix and the category similarity matrix are weightedly combined according to the feature fusion weight to obtain a comprehensive commodity similarity matrix.

[0085] A method for calculating product similarity is used in scenarios such as product recommendation on e-commerce platforms, and can more accurately measure the similarity between products.

[0086] First, collect product data. Extract the initial product candidate set from the database, which contains the price and category information of each product. For example, collect 1,000 products and record the price and category of each product.

[0087] Then, generate the price feature vector. Perform a logarithmic transformation on the price of each commodity. For example, the price of commodity A is 100 yuan, and the logarithmic transformation is 2 (base 10). After that, calculate the mean and standard deviation of the logarithmic transformed prices of all commodities. Assume that the mean is 2 and the standard deviation is 0.5. Subtract the mean from the logarithmic transformed price of each commodity and divide it by the standard deviation to get the price feature value of the commodity. For example, the price feature value of commodity A is (2-2) / 0.5=0. The price feature values ​​of all commodities are combined into a vector, which is the price feature vector of the commodity. In this example, since only the price dimension is considered, the price feature vector of each commodity has only one element. If multiple price-related features are considered, such as historical prices, discount prices, etc., the price feature vector of each commodity will contain multiple elements.

[0088] Next, calculate the price similarity. Based on the price feature vector obtained above, calculate the price distance between each pair of products. For example, if the price feature value of product A is 0 and the price feature value of product B is 1, then their price distance is the absolute value of the difference between the two, that is, |0-1|=1. The price distances between all pairs of products are combined into a matrix, called the price distance matrix. Apply the Gaussian kernel function to the price distance matrix for nonlinear mapping. The bandwidth parameter of the Gaussian kernel function is optimized and determined by the cross-validation method. For example, it was found through experiments that when the bandwidth parameter is 0.5, the recommendation effect of the model is the best. Substitute each element in the price distance matrix into the Gaussian kernel function for calculation to obtain the price similarity matrix.

[0089] At the same time, the category similarity calculation is performed. According to the category tags of the products, a category hierarchical tree structure is constructed. For example, "clothing" is used as a first-level category, "men's clothing" and "women's clothing" are used as second-level categories, and "shirts" and "T-shirts" are used as third-level categories under "men's clothing". The maximum depth of the category hierarchical tree structure is calculated. For example, in this example, the maximum depth is 3. A weight coefficient is assigned to each depth level, and the weight coefficient decays exponentially as the depth of the level increases. For example, the weight coefficient of the first-level category is 1, the weight coefficient of the second-level category is 0.5, and the weight coefficient of the third-level category is 0.25. The category path between the product pairs is extracted, and the depth of the nearest common ancestor node is calculated. For example, product A belongs to "men's clothing-shirts" and product B belongs to "men's clothing-T-shirts". Their nearest common ancestor node is "men's clothing" and the depth is 2. Based on the maximum depth, weight coefficient, and depth of the nearest common ancestor node, the weighted category path distance between the product pairs is calculated. For example, the weighted category path distance between product A and product B is (1-0.5)=0.5. The weighted category path distances between all pairs of products are combined into a matrix, which is called the category similarity matrix.

[0090] Finally, similarity fusion is performed. The price variation coefficient within the category is calculated based on the price feature vector. The price variation coefficient is obtained by dividing the standard deviation of the product prices within the category by the average value. For example, the standard deviation of all product prices under the "Men's Wear" category is 100, and the average value is 500, so the price variation coefficient of this category is 0.2. Calculate the degree of overlap of the price ranges between pairs of products and generate the price range overlap rate. For example, the price range of product A is [100,200], and the price range of product B is [150,250], then their price range overlap rate is 50 / 150=0.33. Calculate the feature fusion weight by combining the price variation coefficient and the price range overlap rate. The price similarity matrix and the category similarity matrix are weighted and combined according to the feature fusion weight to obtain the comprehensive product similarity matrix.

[0091] This application can:

[0092] Improve recommendation accuracy: By combining price and category information, the similarity of products can be more comprehensively described, thereby improving the accuracy of product recommendations. Improve user experience: More accurate product recommendations can help users find their favorite products faster and improve their shopping experience. Enhance platform competitiveness: Accurate product recommendations can improve the platform's conversion rate and user stickiness, and enhance the platform's competitiveness.

[0093] In an optional implementation, based on the comprehensive similarity matrix of commodities, the commodities in the initial commodity candidate set are clustered and grouped to obtain a commodity combination set with similar price ranges and categories; the support and confidence of each commodity combination in the commodity combination set are calculated, and the combination that meets the minimum support threshold and the minimum confidence threshold is selected as a frequent item set, including:

[0094] Clustering and grouping the commodities in the initial commodity candidate set based on the commodity comprehensive similarity matrix; the clustering and grouping process comprises: pairing the commodities in the initial commodity candidate set with each other to calculate commodity similarity values, and dividing the commodities into different initial commodity combinations based on the commodity similarity values;

[0095] Obtaining price range information and category information of the commodities in the initial commodity combination, calculating the price mean and price variance of the commodities in each initial commodity combination based on the price range information, and calculating the category similarity of the commodities in each initial commodity combination based on the category information; splitting and reorganizing the initial commodity combination whose price variance is greater than a first price preset threshold or whose category similarity is less than a second category similarity preset threshold, to obtain a set of commodity combinations that meet the price range and have similar categories;

[0096] Counting the support of each commodity combination in the commodity combination set based on the historical transaction data, where the support is the ratio of the number of co-occurrences of the commodity combination in the historical transaction data to the total number of transactions; calculating the confidence of each commodity combination in the commodity combination set, where the confidence is the co-occurrence probability of the commodity combination in the historical transactions;

[0097] A minimum support threshold and a minimum confidence threshold are set, the support and confidence of each commodity combination in the commodity combination set are calculated, and commodity combinations whose support is greater than the minimum support threshold and whose confidence is greater than the minimum confidence threshold are selected as frequent item sets.

[0098] For ease of understanding and explanation, some example products and values ​​are used in the following description. These values ​​are only examples and may be adjusted according to specific circumstances in actual applications.

[0099] First, the preparation work includes collecting product data and historical transaction data. Product data contains information such as product ID, product name, price, category, etc. Historical transaction data contains user transaction records, such as which products the user purchased. Suppose we collected 10,000 historical transaction data and selected 100 products as the initial product candidate set.

[0100] Next, construct a comprehensive similarity matrix for products. For any two products in the initial product candidate set, calculate their similarity. Product similarity can be calculated based on a variety of factors, such as product name similarity, number of co-purchases, belonging to the same category, etc. For example, the name similarity of product A and product B is 0.8, the number of co-purchases is 500, and they belong to the same category, then their comprehensive similarity can be calculated based on the preset weights. Calculate the similarity between all products to form a 100x100 matrix, which is the comprehensive similarity matrix for products.

[0101] Then, cluster the products according to the comprehensive similarity matrix. First, pair the products in the initial product candidate set and calculate the product similarity value. For example, the similarity between product A and product B is 0.9, the similarity between product A and product C is 0.2, and the similarity between product B and product C is 0.3. Then, set a similarity threshold, such as 0.5. Group the products with similarity greater than the threshold into the same initial product combination. In this example, the similarity between product A and product B is greater than 0.5, so they are grouped into the same initial product combination.

[0102] Get the price range information and category information of the initial product combination. For example, the price of product A is 10 yuan, and the price of product B is 12 yuan, and they belong to the "food" category. Calculate the price mean and price variance of the products in each initial product combination. In this example, the price mean of A and B is 11 yuan, and the price variance is 1. Calculate the category similarity of the products in each initial product combination. Since A and B belong to the same category, their category similarity is 1.

[0103] Set the first price preset threshold and the second category similarity preset threshold. For example, the first price preset threshold is 5, and the second category similarity preset threshold is 0.8. Split and reorganize the initial product combinations whose price variance is greater than the first price preset threshold or whose category similarity is less than the second category similarity preset threshold. In this example, the price variance of A and B is less than 5, and the category similarity is greater than 0.8, so they do not need to be split and reorganized. Assume that the price of product C is 100 yuan and it also belongs to the "food" category. If C, A, and B are in the same initial combination, the price variance will be large and exceed the threshold, and C needs to be separated from the combination. Finally, a set of product combinations that meet the price range and are similar in category is obtained.

[0104] Count the support and confidence of each product combination in the product combination set.

[0105] Set the minimum support threshold and the minimum confidence threshold. For example, the minimum support threshold is 0.05 and the minimum confidence threshold is 0.6. Filter the product combinations whose support is greater than the minimum support threshold and whose confidence is greater than the minimum confidence threshold as frequent item sets. In this example, the support of {A, B} is 0.08 and the confidence is 0.8, both of which meet the threshold requirements, so {A, B} is identified as a frequent item set.

[0106] This application can:

[0107] Improve recommendation efficiency: By clustering, grouping and filtering, the number of candidate product combinations is reduced, thereby improving recommendation efficiency. Improve recommendation accuracy: By considering the price range and category similarity of products, users' potential needs can be identified more accurately, thereby improving recommendation accuracy. Discover potential associations: By mining historical transaction data, potential associations between products can be discovered, thereby recommending product combinations that users may be interested in.

[0108] In an optional implementation, for the commodity combinations in the frequent item set, extract the recent price change sequence of each commodity, calculate the correlation coefficient of the price change trend, and generate a dynamic price matching degree; reorder the frequent item sets based on the dynamic price matching degree to obtain a commodity combination sequence considering the dynamic price change, including:

[0109] For each commodity combination in the frequent item set, extract the recent price change sequence of each commodity in the commodity combination from the historical transaction data, and the price change sequence includes the daily price of the past thirty days; use the maximum and minimum value normalization method to map the price change sequence into a standardized price sequence, and the value range of the standardized price sequence is zero to one;

[0110] Calculating the price correlation coefficients of the price change trends for the standardized price series of each commodity in the commodity combination, and constructing a price correlation coefficient matrix based on the price correlation coefficients, wherein the price correlation coefficient matrix is ​​used to characterize the synchronization of price changes of each pair of commodities in the commodity combination; calculating the statistical eigenvalues ​​of the price correlation coefficient matrix, including the mean, variance and maximum eigenvalue;

[0111] The dynamic price matching degree of the commodity combination is generated based on the statistical eigenvalues ​​of the price correlation coefficient matrix, and the dynamic price matching degree is obtained by weighted calculation of the mean, variance and maximum eigenvalue; a re-ranking scoring function is constructed, and the re-ranking scoring function integrates the dynamic price matching degree with the support and confidence of the commodity combination; and the commodity combinations in the frequent item set are sorted in descending order according to the re-ranking scoring function to obtain a commodity combination sequence that takes into account the dynamic changes in prices.

[0112] Get frequent item sets containing product combinations and their support and confidence. For example, frequent item sets are mined from historical transaction data using the Apriori algorithm or the FP-Growth algorithm. An example frequent item set is as follows: {(milk, bread), support: 0.3, confidence: 0.6}, {(milk, eggs, butter), support: 0.2, confidence: 0.5}, {(diapers, beer), support: 0.25, confidence: 0.7}.

[0113] For each product combination in the frequent itemset, extract the recent price change sequence of each product in the combination. For example, extract the daily prices of the past 30 days to form the price sequence of each product. Assuming the product combination is (milk, bread), extract the daily prices of milk and bread in the past 30 days.

[0114] Use the maximum and minimum normalization method to map the price change sequence of each commodity to the range of 0 to 1. For each price sequence, find the maximum and minimum values, and then convert each price value in the sequence according to the formula (current price-minimum value) / (maximum-minimum value).

[0115] For each commodity combination, calculate the price correlation coefficient of the standardized price series of each two commodities. The price correlation coefficient is used to measure the similarity of the price change trends of two commodities. For example, calculate the correlation coefficient of the standardized price series of milk and bread.

[0116] Based on the calculated price correlation coefficients, a price correlation coefficient matrix is ​​constructed. The rows and columns of the matrix represent the commodities in the commodity combination, and each element in the matrix represents the price correlation coefficient of the corresponding two commodities. For example, for the commodity combination (milk, bread, eggs), a 3x3 matrix is ​​constructed, in which the element (1,2) represents the price correlation coefficient of milk and bread, the element (2,3) represents the price correlation coefficient of bread and eggs, and so on.

[0117] Calculate the statistical eigenvalues ​​of the price correlation coefficient matrix, including the mean, variance, and maximum eigenvalue. The mean is the average of all price correlation coefficients, the variance measures the dispersion of the price correlation coefficients, and the maximum eigenvalue represents the main characteristics of the matrix.

[0118] The dynamic price matching degree of the product combination is calculated based on the statistical eigenvalues ​​of the price correlation coefficient matrix. The dynamic price matching degree is obtained by weighted summing the mean, variance and maximum eigenvalue. For example, different weights can be assigned to the mean, variance and maximum eigenvalue, and then they are weighted summed to obtain the final dynamic price matching degree.

[0119] Construct a re-ranking scoring function to integrate the dynamic price matching, product combination support and confidence. The re-ranking scoring function can be designed according to specific needs. For example, the three can be weighted summed.

[0120] According to the reordering score function, the product combinations in the frequent item set are sorted in descending order to obtain a product combination sequence that takes into account the dynamic changes in prices. For example, according to the calculated reordering score, {(milk, bread), (milk, eggs, butter), (diapers, beer)} are sorted to obtain a new product combination sequence.

[0121] This application can:

[0122] Improve the accuracy of product combination recommendations: By considering the dynamic changes in product prices, you can more accurately identify related product combinations, thereby improving the accuracy of recommendations. Promote sales growth: By recommending product combinations with high price matching, you can guide consumers to buy more products, thereby increasing sales. Enhance user experience: By providing product combination recommendations that better meet user needs, you can improve user satisfaction and shopping experience.

[0123] In an optional implementation, the method further includes:

[0124] Collect the competitive product data of each product in the product combination, convert the product title into a word vector to obtain a semantic feature vector, calculate the cosine similarity of the semantic feature vector to obtain the product similarity, and screen the competitive products based on the preset similarity threshold to form a competitive product price monitoring pool;

[0125] Performing kernel density estimation on the competitive product price data in the competitive product price monitoring pool to obtain a competitive product price distribution curve, calculating the mean and quantile of the competitive product price distribution curve to obtain market price distribution characteristics, and calculating the standard deviation of the competitive product price to obtain market price dispersion;

[0126] Calculate the price deviation rate based on the actual price of the commodity and the market price distribution characteristics, perform weighted combination of the price deviation rate and the market price dispersion to obtain a single product price competitiveness index, and use the historical sales volume of the commodity to weight the single product price competitiveness index to obtain a commodity combination price competitiveness index;

[0127] Calculate the price change sequence of the product and the competing product, use a sliding time window to calculate the sliding coefficient of the price change sequence to obtain the price change trend correlation, and calculate the price linkage index based on the price change trend correlation;

[0128] Inputting the commodity combination price competitiveness index and the price linkage index into a pre-trained multi-layer perceptron network, the multi-layer perceptron network outputs a feature weight coefficient, and using the feature weight coefficient to perform weighted fusion on the commodity combination price competitiveness index and the price linkage index to obtain a comprehensive score for the commodity combination;

[0129] The commodity combination sequence is reordered based on the comprehensive score of the commodity combination to generate an optimized commodity combination recommendation sequence, and the commodity combination recommendation sequence is used for commodity combination recommendation.

[0130] In order to optimize the product combination recommendation effect, a product combination recommendation method based on price competitiveness and linkage is proposed. This method comprehensively evaluates the competitiveness of product combinations by analyzing the relationship between product prices and market competitor prices, as well as the linkage trend between product prices and competitor prices, and makes ranking recommendations accordingly.

[0131] First, collect the competitive product data of each product in the product portfolio. For example, for product A, collect product data on the e-commerce platform that is similar to its function, category, and target user through crawlers and other methods, including product title, price and other information. Then, convert the product title into a word vector to obtain a semantic feature vector. For example, use the Word2Vec model to convert the product title into a numerical vector representation. Next, calculate the cosine similarity of the semantic feature vector to obtain the product similarity. For example, calculate the cosine similarity between product A and other product title vectors, and the larger the value, the higher the similarity. Filter competing products based on a preset similarity threshold (for example, 0.8) to form a competitive product price monitoring pool.

[0132] Next, perform kernel density estimation on the competitive product price data in the competitive product price monitoring pool to obtain the competitive product price distribution curve. For example, use the Gaussian kernel function to perform kernel density estimation on the competitive product price to obtain a smooth probability density curve. Then, calculate the mean and quantile (such as the 25th quantile and the 75th quantile) of the competitive product price distribution curve to obtain the market price distribution characteristics. For example, calculate the average value of the competitive product price, 25% of the prices are lower than a certain value, and 75% of the prices are lower than a certain value. At the same time, calculate the standard deviation of the competitive product price to obtain the market price dispersion. The larger the standard deviation, the greater the price fluctuation.

[0133] Then, the price deviation rate is calculated based on the actual price of the product and the market price distribution characteristics. For example, the ratio of the difference between the actual price of the product and the average price of the competing products to the average price of the competing products is calculated to indicate the degree to which the product price deviates from the market average. The price deviation rate and the market price dispersion are weighted and combined to obtain the single product price competitiveness index. For example, different weights are assigned to the price deviation rate and the market price dispersion, and they are linearly combined to obtain the single product price competitiveness index, and the weight can be adjusted according to the actual situation. Next, the single product price competitiveness index is weighted by the historical sales volume of the product to obtain the product combination price competitiveness index. For example, the historical sales volume of the product is used as the weight, multiplied by the single product price competitiveness index, and then summed or averaged to obtain the product combination price competitiveness index. The higher the historical sales volume, the greater the weight.

[0134] At the same time, the price change sequence of the product and the competitor is calculated. For example, the price change data of the product and the competitor over a period of time is recorded to form a time series. The sliding coefficient of the price change sequence is calculated using a sliding time window to obtain the correlation of the price change trend. For example, using a sliding window with a length of 7 days, the Pearson correlation coefficient of the price change sequence of the product and the competitor is calculated to indicate the similarity of their price change trends. The price linkage index is calculated based on the correlation of the price change trend. For example, the correlation coefficient is converted into a price linkage index. The higher the correlation coefficient, the stronger the price linkage.

[0135] Next, the commodity combination price competitiveness index and price linkage index are input into the pre-trained multi-layer perceptron network. The network has been pre-trained using a large amount of commodity combination data. The multi-layer perceptron network outputs feature weight coefficients. The feature weight coefficients are used to perform a weighted fusion of the commodity combination price competitiveness index and price linkage index to obtain a comprehensive score for the commodity combination.

[0136] Finally, the product combination sequence is reordered based on the comprehensive score of the product combination to generate an optimized product combination recommendation sequence for product combination recommendation. For example, the product combinations are sorted from high to low according to the comprehensive score, and the combinations with higher scores are ranked higher and recommended to users.

[0137] This application can:

[0138] Improve the accuracy of product combination recommendations: By comprehensively considering price competitiveness and price linkage, the attractiveness of product combinations can be more accurately evaluated, thereby improving the accuracy of recommendations and better meeting user needs. Enhance the competitiveness of product combination recommendations: By analyzing market prices, help merchants develop more competitive pricing strategies, improve the competitiveness of product combinations in the market, and promote sales growth. Optimize the diversity of product combination recommendations: This method is not limited to a single indicator, but comprehensively considers multiple factors, which can avoid the recommendation results being too single, provide more diverse product combination options, and improve user experience.

[0139] In an optional implementation, a multi-dimensional evaluation index of price adaptability, category diversity and promotion effect is constructed, and a Pareto optimal solution set is obtained by combining a non-dominated sorting genetic algorithm; a deep Q network is used for parameter optimization, and the evaluation index weight is dynamically adjusted based on scene recognition; a differentiated sliding window update mechanism is used to complete the product combination evaluation and make product combination matching recommendations, including:

[0140] Construct multi-dimensional evaluation indicators, obtain the user's historical purchase price and the current product combination price, calculate the Mahalanobis distance between the two price distributions through the covariance matrix to obtain the price adaptation index; extract the category tree path of each product in the product combination to construct the category distribution vector, and calculate the category distribution uncertainty based on information entropy to obtain the diversity index; collect the historical promotion records of the product within 90 days to establish the conversion rate curve, and calculate the promotion effect index in combination with the current promotion discount depth;

[0141] Normalizing the price adaptation index, the diversity index and the promotion effect index to obtain normalized index values, constructing an initial population for commodity combination evaluation based on the normalized index values, and iteratively optimizing the initial population using a non-dominated sorting genetic algorithm, wherein in each iteration, the crowding distance of individuals in the population is calculated and a new population is generated through tournament selection, crossover and mutation operations until convergence to obtain a Pareto optimal solution set of the commodity combination;

[0142] Based on the Pareto optimal solution set, the state space and action space of the deep Q network are constructed, wherein the state space includes the normalized index value, the historical commodity conversion data and the market environment characteristics, and the action space includes the adjustment scheme of the evaluation parameters. The reward function that comprehensively considers the click rate, conversion rate and rate of return is used for parameter optimization training, and the ε-greedy strategy is used to select the parameter adjustment action during the training process;

[0143] Input the parameter optimization results obtained from the deep Q network training into the scene recognition model, map the current scene to a predefined scene category based on time attributes, user group characteristics and promotion activity attributes, assign differentiated weight initial values ​​to the normalized index values ​​according to the scene category, and dynamically adjust the weight coefficients of each index through an online learning method based on real-time conversion effect feedback data;

[0144] Differentiated sliding time windows are set according to the dynamically adjusted indicator weight coefficients, and the price adaptation index, diversity index and promotion effect index are incrementally updated using corresponding time window sizes. When the fluctuation of any indicator exceeds a preset threshold, the latest indicator value is re-input into the deep Q network for parameter optimization.

[0145] The product combination recommendation method aims to improve the user shopping experience and promote product sales. This method achieves dynamic evaluation and recommendation of product combinations by constructing multi-dimensional evaluation indicators, optimizing parameters using intelligent algorithms, and combining scene recognition.

[0146] First, construct a multi-dimensional evaluation index for product combinations. Obtain the user's historical purchase price data and the price data of the current product combination to be recommended. In order to measure the price adaptability, calculate the degree of difference between the two price distributions. Specifically, use the covariance matrix to capture the shape and direction differences of the two price distributions, and calculate the Mahalanobis distance as the price adaptability indicator. The smaller the distance, the higher the adaptability. For example, if the user's historical purchase prices are concentrated between 10-20 yuan, and the current product combination price is also within this range, the Mahalanobis distance is small and the price adaptability is high. On the contrary, if the user's historical purchase prices are concentrated between 50-60 yuan, and the current combination price is between 10-20 yuan, the Mahalanobis distance is large and the price adaptability is low.

[0147] Next, evaluate the category diversity of the product portfolio. Extract the category tree path of each product in the product portfolio, such as "apparel-men's clothing-T-shirts" and "electronics-mobile phones-smartphones". Convert these paths into vector representations and calculate information entropy based on these vectors to measure the uncertainty of category distribution. The higher the information entropy, the more dispersed the category distribution and the higher the diversity. For example, a portfolio containing T-shirts, mobile phones, and books has a higher category diversity than a portfolio containing only three different styles of T-shirts.

[0148] Then, evaluate the promotion effect of the product combination. Collect the historical promotion records of the product within 90 days, such as the conversion rate data under different discount levels. Based on this data, establish a conversion rate curve, and combine the promotion discount depth of the current product combination to predict its conversion rate as a promotion effect indicator. For example, if the conversion rate of a product with a 20% discount promotion is 20% and the conversion rate of a 30% discount promotion is 30% in the past 90 days, the conversion rate of the current 40% discount promotion can be predicted based on the curve.

[0149] The above price suitability, category diversity and promotion effect indicators are normalized and the indicator values ​​are mapped to between 0 and 1 for subsequent comparison and calculation. These normalized indicator values ​​are used to construct the initial population for product combination evaluation, where each individual represents a product combination and its corresponding indicator value.

[0150] The non-dominated sorting genetic algorithm is used to iteratively optimize the initial population. In each iteration, the crowding distance of each individual in the population is first calculated, which reflects the density of solutions around the individual. Then, a new population is generated through tournament selection, crossover, and mutation operations. Tournament selection refers to the random selection of a part of individuals, among which individuals with better indicator values ​​are more likely to be selected. Crossover operation refers to exchanging some indicator values ​​of two selected individuals to generate new individuals. Mutation operation refers to randomly changing the indicator value of a certain individual. Repeat the iteration until the algorithm converges and obtains the Pareto optimal solution set of the product combination. The Pareto optimal solution set refers to a set of solutions that cannot improve a certain indicator without damaging other indicators.

[0151] Use deep Q network for parameter optimization. Take the indicator values ​​in the Pareto optimal solution set, the historical conversion data of the product, and the market environment characteristics as the state space, and take the adjustment scheme of the evaluation parameters as the action space. For example, the state space can contain normalized price adaptability, category diversity, and promotion effect indicators, as well as historical click-through rate, conversion rate, and sales data of the product. The market environment characteristics can include information such as seasons and holidays. The action space can contain different schemes for adjusting the weights of price adaptability, category diversity, and promotion effect indicators. Design a reward function that comprehensively considers click-through rate, conversion rate, and rate of return. For example, the reward function can be defined as click-through rate * conversion rate * rate of return. Use this reward function to train the deep Q network, and use the ε-greedy strategy to select the parameter adjustment action during the training process. The ε-greedy strategy is to randomly select an action with a certain probability, and select the action with the highest expected return in the current state with a certain probability.

[0152] The trained deep Q network is applied to scene recognition. Based on time attributes (such as weekends and holidays), user group characteristics (such as age and gender), and promotional activity attributes (such as full discounts and discounts), the current scene is mapped to a predefined scene category. Differentiated weight initial values ​​are assigned to normalized indicator values ​​according to the scene category. For example, for price-sensitive users, the price adaptation index has a higher weight; for users who pursue novelty, the category diversity index has a higher weight. Based on real-time conversion effect feedback data, the weight coefficients of each indicator are dynamically adjusted through online learning methods.

[0153] Finally, differentiated sliding time windows are set according to the dynamically adjusted indicator weight coefficients. The price fit, category diversity, and promotion effect indicators are incrementally updated using the corresponding time window sizes. For example, for products with large price fluctuations, the time window of the price fit indicator can be set shorter. When the fluctuation of any indicator exceeds the preset threshold, the latest indicator value is re-input into the deep Q network for parameter optimization.

[0154] This application can:

[0155] Personalized recommendation: This method can provide users with personalized product combination recommendations based on their historical purchasing behavior, current scenarios, and product characteristics, thereby improving their shopping experience and satisfaction. Improve conversion rate: By optimizing the multi-dimensional indicators of product combinations, such as price suitability, category diversity, and promotional effects, the attractiveness of product combinations can be improved, thereby increasing users' click-through rate and conversion rate. Increase sales: Through personalized recommendations and improved conversion rates, it is ultimately possible to promote product sales and increase merchants' sales and profits.

[0156] Figure 2 FIG. 1 is a schematic diagram of the structure of a collocation recommendation system based on association rule mining according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0157] The first unit is used to obtain the historical transaction data of the user, extract the user number, product number, product category number, product price, purchase time, and transaction score from the historical transaction data, and generate a user product transaction score matrix; calculate the average value and standard deviation of the user score, standardize the user product transaction score matrix, eliminate the score deviation, and obtain a standardized score matrix; based on the standardized score matrix, use cosine similarity to calculate the cosine similarity value between users to generate a user similarity matrix; predict the missing score of each product according to the user similarity matrix to obtain a score prediction matrix; filter the product combinations whose predicted scores are lower than a preset score threshold in the score prediction matrix to obtain an initial product candidate set;

[0158] The second unit is used to perform logarithmic normalization processing on the commodity prices in the initial commodity candidate set to generate a price feature vector; calculate the price distance between commodities based on the price feature vector to construct a price similarity matrix; obtain the commodity category hierarchical structure, calculate the category path distance between commodities, and generate a category similarity matrix; perform weighted fusion of the price similarity matrix and the category similarity matrix to obtain a comprehensive commodity similarity matrix; cluster and group the commodities in the initial commodity candidate set based on the comprehensive commodity similarity matrix to obtain a commodity combination set with similar price ranges and categories; calculate the support and confidence of each commodity combination in the commodity combination set, and select the combination that meets the minimum support threshold and the minimum confidence threshold as a frequent item set;

[0159] The third unit is used to generate a dynamic price matching degree for the commodity combinations in the frequent item sets by first extracting the recent price change sequence of each commodity and calculating the correlation coefficient of the price change trend; reordering the frequent item sets based on the dynamic price matching degree to generate a commodity combination sequence that takes into account the dynamic price changes; obtaining the Pareto optimal solution set by constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect and combining the non-dominated sorting genetic algorithm; optimizing parameters using a deep Q network and dynamically adjusting the evaluation indicator weights based on scene recognition; and using a differentiated sliding window update mechanism to complete commodity combination evaluation and make commodity combination matching recommendations.

[0160] According to a third aspect of the embodiments of the present invention,

[0161] An electronic device is provided, comprising:

[0162] processor;

[0163] a memory for storing processor-executable instructions;

[0164] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0165] A fourth aspect of the embodiments of the present invention is:

[0166] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0167] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A collocation recommendation method based on association rule mining, characterized in that: include: Obtain the user's historical transaction data, extract the user number, product number, product category number, product price, purchase time, and transaction score from the historical transaction data, and generate a user product transaction score matrix; calculate the average and standard deviation of the user score, standardize the user product transaction score matrix, eliminate the score deviation, and obtain a standardized score matrix; based on the standardized score matrix, use cosine similarity to calculate the cosine similarity value between users to generate a user similarity matrix; predict the missing score of each product according to the user similarity matrix to obtain a score prediction matrix; filter the product combinations whose predicted scores are lower than a preset score threshold in the score prediction matrix to obtain an initial product candidate set; Performing logarithmic normalization processing on the commodity prices in the initial commodity candidate set to generate a price feature vector; Calculate the price distance between commodities based on the price feature vector and construct a price similarity matrix; Obtain the product category hierarchy structure, calculate the category path distance between products, and generate a category similarity matrix; Performing weighted fusion on the price similarity matrix and the category similarity matrix to obtain a comprehensive product similarity matrix; Based on the comprehensive similarity matrix of commodities, the commodities in the initial commodity candidate set are clustered and grouped to obtain a commodity combination set with similar price ranges and categories; the support and confidence of each commodity combination in the commodity combination set are calculated, and the combination that meets the minimum support threshold and the minimum confidence threshold is selected as a frequent item set; For the commodity combinations in the frequent item set, firstly, the recent price change sequence of each commodity is extracted, and the correlation coefficient of the price change trend is calculated to generate a dynamic price matching degree; based on the dynamic price matching degree, the frequent item set is reordered to generate a commodity combination sequence that takes into account the dynamic price changes; by constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect, and combining with a non-dominated sorting genetic algorithm, a Pareto optimal solution set is obtained; a deep Q network is used for parameter optimization, and the weight of the evaluation indicator is dynamically adjusted based on scene recognition; a differentiated sliding window update mechanism is used to complete the commodity combination evaluation and make commodity combination matching recommendations; Performing logarithmic normalization processing on the commodity prices in the initial commodity candidate set to generate a price feature vector; Calculate the price distance between commodities based on the price feature vector and construct a price similarity matrix; Obtain the product category hierarchy structure, calculate the category path distance between products, and generate a category similarity matrix; The price similarity matrix and the category similarity matrix are weighted and merged to obtain a comprehensive commodity similarity matrix including: Extracting price data of each commodity in the initial commodity candidate set, performing logarithmic transformation on the price data to obtain an initial logarithmic value; calculating the mean and standard deviation of the initial logarithmic value, normalizing the initial logarithmic value based on the mean and standard deviation, and obtaining a price feature vector corresponding to each commodity; calculating price distribution characteristics within the commodity category based on the price feature vector, the price distribution characteristics including the price variation coefficient within the category and the price range of the category; Based on the price feature vector, a commodity price distance matrix is ​​constructed, wherein each element of the price distance matrix is ​​obtained by calculating the square root of the sum of the differences between the components of the two commodity price feature vectors; a Gaussian kernel function is applied to the price distance matrix for nonlinear mapping, wherein the bandwidth parameter of the Gaussian kernel function is optimized and determined by a cross-validation method, and a price similarity matrix is ​​generated; Obtain the category label of each product in the initial product candidate set and construct a category hierarchical tree structure; calculate the maximum depth of the category hierarchical tree structure and assign a weight coefficient to each depth level, wherein the weight coefficient decays exponentially with increasing hierarchical depth; extract the category path between product pairs and calculate the depth of the nearest common ancestor node; calculate the weighted category path distance between product pairs based on the maximum depth, the weight coefficient and the depth of the nearest common ancestor node, and generate a category similarity matrix; Based on the price distribution characteristics, the price variation coefficient within the category is calculated, and the price variation coefficient is obtained by dividing the standard deviation of the commodity prices within the category by the average value; the degree of overlap of the price ranges between commodity pairs is calculated to generate the price range overlap rate; the feature fusion weight is calculated by combining the price variation coefficient and the price range overlap rate; the price similarity matrix and the category similarity matrix are weightedly combined according to the feature fusion weight to obtain a comprehensive commodity similarity matrix.

2. The method according to claim 1, characterized in that Based on the standardized rating matrix, the cosine similarity values ​​between users are calculated using cosine similarity to generate a user similarity matrix; the missing rating of each product is predicted according to the user similarity matrix to obtain a rating prediction matrix including: Obtain user product rating data, calculate the average rating and the standard deviation of the rating of each user in the user product rating data; based on the average rating and the standard deviation of the rating, perform standard processing on the user product rating data to obtain a rating standardization matrix; according to the rating vectors of the commonly rated products between the user pairs in the rating standardization matrix, use the cosine similarity formula to calculate the cosine similarity value between the user pairs, the cosine similarity value is calculated by dividing the inner product of the rating vectors between the user pairs by the product of the modulus of the rating vectors, and generate a user similarity matrix; Identify the missing rating position in the user's product rating data, where the missing rating position corresponds to the position where the user has not rated the product; for each missing rating position, extract the rating similarity value between the current user and other users from the user similarity matrix; sort other users in descending order according to the rating similarity value, and select K users with the highest rating similarity as the neighbor user set; extract a subset of users from the neighbor user set who have rated the product corresponding to the current missing rating position, and the rating of each user in the user subset needs to be standardized; For the users in the user subset, a scoring weight is calculated based on their cosine similarity values ​​in the user similarity matrix, and the scoring weight is obtained by normalizing the similarity value; the scoring weight is weighted and summed with the corresponding standardized score in the user subset to obtain the predicted score of the current missing score position; the predicted score is de-normalized by the average score and the score standard deviation of the current user to obtain the predicted score value under the original score scale; Repeat the above prediction process for all missing rating positions, combine the predicted rating values ​​with the original rating data, and construct a complete rating prediction matrix; calculate the confidence of each predicted rating value in the rating prediction matrix, and the confidence is calculated based on the number of neighboring users participating in the prediction, the discreteness of the neighboring user ratings, and the average similarity of the neighboring users; filter the predicted rating values ​​in the rating prediction matrix according to the confidence, mark the predicted rating values ​​with a confidence lower than a preset threshold as invalid predictions, and obtain the final rating prediction matrix.

3. The method according to claim 1, characterized in that Based on the comprehensive similarity matrix of commodities, the commodities in the initial commodity candidate set are clustered and grouped to obtain a commodity combination set with similar price ranges and categories; the support and confidence of each commodity combination in the commodity combination set are calculated, and the combinations that meet the minimum support threshold and the minimum confidence threshold are selected as frequent item sets, including: Clustering and grouping the commodities in the initial commodity candidate set based on the commodity comprehensive similarity matrix; the clustering and grouping process comprises: pairing the commodities in the initial commodity candidate set with each other to calculate commodity similarity values, and dividing the commodities into different initial commodity combinations based on the commodity similarity values; Obtaining price range information and category information of the commodities in the initial commodity combination, calculating the price mean and price variance of the commodities in each initial commodity combination based on the price range information, and calculating the category similarity of the commodities in each initial commodity combination based on the category information; splitting and reorganizing the initial commodity combination whose price variance is greater than a first price preset threshold or whose category similarity is less than a second category similarity preset threshold, to obtain a set of commodity combinations that meet the price range and have similar categories; Counting the support of each commodity combination in the commodity combination set based on the historical transaction data, where the support is the ratio of the number of co-occurrences of the commodity combination in the historical transaction data to the total number of transactions; calculating the confidence of each commodity combination in the commodity combination set, where the confidence is the co-occurrence probability of the commodity combination in the historical transactions; A minimum support threshold and a minimum confidence threshold are set, the support and confidence of each commodity combination in the commodity combination set are calculated, and commodity combinations whose support is greater than the minimum support threshold and whose confidence is greater than the minimum confidence threshold are selected as frequent item sets.

4. The method according to claim 1, characterized in that For the commodity combinations in the frequent item set, extract the recent price change sequence of each commodity, calculate the correlation coefficient of the price change trend, and generate a dynamic price matching degree; The frequent item sets are reordered based on the dynamic price matching degree, and the product combination sequence considering the dynamic price change is obtained, including: For each commodity combination in the frequent item set, extract the recent price change sequence of each commodity in the commodity combination from the historical transaction data, and the price change sequence includes the daily price of the past thirty days; use the maximum and minimum value normalization method to map the price change sequence into a standardized price sequence, and the value range of the standardized price sequence is zero to one; Calculating the price correlation coefficients of the price change trends for the standardized price series of each commodity in the commodity combination, and constructing a price correlation coefficient matrix based on the price correlation coefficients, wherein the price correlation coefficient matrix is ​​used to characterize the synchronization of price changes of each pair of commodities in the commodity combination; calculating the statistical eigenvalues ​​of the price correlation coefficient matrix, including the mean, variance and maximum eigenvalue; The dynamic price matching degree of the commodity combination is generated based on the statistical eigenvalues ​​of the price correlation coefficient matrix, and the dynamic price matching degree is obtained by weighted calculation of the mean, variance and maximum eigenvalue; a re-ranking scoring function is constructed, and the re-ranking scoring function integrates the dynamic price matching degree with the support and confidence of the commodity combination; and the commodity combinations in the frequent item set are sorted in descending order according to the re-ranking scoring function to obtain a commodity combination sequence that takes into account the dynamic changes in prices.

5. The method according to claim 4, characterized in that The method further comprises: Collect the competitive product data of each product in the product combination, convert the product title into a word vector to obtain a semantic feature vector, calculate the cosine similarity of the semantic feature vector to obtain the product similarity, and screen the competitive products based on the preset similarity threshold to form a competitive product price monitoring pool; Performing kernel density estimation on the competitive product price data in the competitive product price monitoring pool to obtain a competitive product price distribution curve, calculating the mean and quantile of the competitive product price distribution curve to obtain market price distribution characteristics, and calculating the standard deviation of the competitive product price to obtain market price dispersion; Calculate the price deviation rate based on the actual price of the commodity and the market price distribution characteristics, perform weighted combination of the price deviation rate and the market price dispersion to obtain a single product price competitiveness index, and use the historical sales volume of the commodity to weight the single product price competitiveness index to obtain a commodity combination price competitiveness index; Calculate the price change sequence of the product and the competing product, use a sliding time window to calculate the sliding coefficient of the price change sequence to obtain the price change trend correlation, and calculate the price linkage index based on the price change trend correlation; Inputting the commodity combination price competitiveness index and the price linkage index into a pre-trained multi-layer perceptron network, the multi-layer perceptron network outputs a feature weight coefficient, and using the feature weight coefficient to perform weighted fusion on the commodity combination price competitiveness index and the price linkage index to obtain a comprehensive score for the commodity combination; The commodity combination sequence is reordered based on the comprehensive score of the commodity combination to generate an optimized commodity combination recommendation sequence, and the commodity combination recommendation sequence is used for commodity combination recommendation.

6. The method according to claim 1, characterized in that By constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect, and combining them with the non-dominated sorting genetic algorithm, the Pareto optimal solution set is obtained; the deep Q network is used for parameter optimization, and the evaluation indicator weights are dynamically adjusted based on scene recognition; The differentiated sliding window update mechanism is used to complete the product combination evaluation and make product combination recommendations, including: Construct multi-dimensional evaluation indicators, obtain the user's historical purchase price and the current product combination price, calculate the Mahalanobis distance between the two price distributions through the covariance matrix to obtain the price adaptation index; extract the category tree path of each product in the product combination to construct the category distribution vector, and calculate the category distribution uncertainty based on information entropy to obtain the diversity index; collect the historical promotion records of the product within 90 days to establish the conversion rate curve, and calculate the promotion effect index in combination with the current promotion discount depth; Normalizing the price adaptation index, the diversity index and the promotion effect index to obtain normalized index values, constructing an initial population for commodity combination evaluation based on the normalized index values, and iteratively optimizing the initial population using a non-dominated sorting genetic algorithm, wherein in each iteration, the crowding distance of individuals in the population is calculated and a new population is generated through tournament selection, crossover and mutation operations until convergence to obtain a Pareto optimal solution set of the commodity combination; Based on the Pareto optimal solution set, the state space and action space of the deep Q network are constructed, wherein the state space includes the normalized index value, the historical commodity conversion data and the market environment characteristics, and the action space includes the adjustment scheme of the evaluation parameters. The reward function that comprehensively considers the click rate, conversion rate and rate of return is used for parameter optimization training, and the ε-greedy strategy is used to select the parameter adjustment action during the training process; Input the parameter optimization results obtained from the deep Q network training into the scene recognition model, map the current scene to a predefined scene category based on time attributes, user group characteristics and promotion activity attributes, assign differentiated weight initial values ​​to the normalized index values ​​according to the scene category, and dynamically adjust the weight coefficients of each index through an online learning method based on real-time conversion effect feedback data; Differentiated sliding time windows are set according to the dynamically adjusted indicator weight coefficients, and the price adaptation index, diversity index and promotion effect index are incrementally updated using corresponding time window sizes. When the fluctuation of any indicator exceeds a preset threshold, the latest indicator value is re-input into the deep Q network for parameter optimization.

7. A collocation recommendation system based on association rule mining, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the historical transaction data of the user, extract the user number, product number, product category number, product price, purchase time, and transaction score from the historical transaction data, and generate a user product transaction score matrix; calculate the average value and standard deviation of the user score, standardize the user product transaction score matrix, eliminate the score deviation, and obtain a standardized score matrix; based on the standardized score matrix, use cosine similarity to calculate the cosine similarity value between users to generate a user similarity matrix; predict the missing score of each product according to the user similarity matrix to obtain a score prediction matrix; filter the product combinations whose predicted scores are lower than a preset score threshold in the score prediction matrix to obtain an initial product candidate set; The second unit is used to perform logarithmic normalization processing on the commodity prices in the initial commodity candidate set to generate a price feature vector; Calculate the price distance between commodities based on the price feature vector and construct a price similarity matrix; Obtain the product category hierarchy structure, calculate the category path distance between products, and generate a category similarity matrix; Performing weighted fusion on the price similarity matrix and the category similarity matrix to obtain a comprehensive product similarity matrix; Based on the comprehensive similarity matrix of commodities, the commodities in the initial commodity candidate set are clustered and grouped to obtain a commodity combination set with similar price ranges and categories; the support and confidence of each commodity combination in the commodity combination set are calculated, and the combination that meets the minimum support threshold and the minimum confidence threshold is selected as a frequent item set; The third unit is used to generate a dynamic price matching degree for the commodity combinations in the frequent item sets by first extracting the recent price change sequence of each commodity and calculating the correlation coefficient of the price change trend; reordering the frequent item sets based on the dynamic price matching degree to generate a commodity combination sequence that takes into account the dynamic price changes; obtaining the Pareto optimal solution set by constructing multi-dimensional evaluation indicators of price adaptability, category diversity and promotion effect and combining the non-dominated sorting genetic algorithm; optimizing parameters using a deep Q network and dynamically adjusting the evaluation indicator weights based on scene recognition; and using a differentiated sliding window update mechanism to complete commodity combination evaluation and make commodity combination matching recommendations.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images