Cross-border e-commerce intelligent recommendation system based on big data analysis

Abstract

The invention discloses a cross-border e-commerce intelligent recommendation system based on big data analysis, and relates to the technical field of cross-border e-commerce intelligent recommendation. In order to solve the problem that high-quality interaction edges are blindly deleted due to the uniform discarding proportion, the cross-border e-commerce heterogeneous graph intelligent recommendation system is provided. The method comprises the following steps: firstly, constructing a user-commodity-attribute multi-relation heterogeneous graph and obtaining a complete interaction edge set; then extracting characteristics such as transaction frequency and return dispute rate to calculate credibility and sparseness, setting control points according to quantiles, dividing interaction edges into high, medium and low quality segments, correspondingly endowing decreasing discarding probabilities, and training by using a Bernoulli sampling joint graph neural network; according to a model feedback dynamic correction threshold and edge classification, reconstructing a heterogeneous graph; and finally, carrying out online fine adjustment on the discarding probability through a UCB multi-arm Bandit algorithm, and carrying out hot updating on the model to realize noise reduction, long tail reservation and continuous improvement of recommendation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cross-border e-commerce intelligent recommendation, more particularly, to a cross-border e-commerce intelligent recommendation system based on big data analysis. BACKGROUND

[0002] In the field of cross-border e-commerce, the intelligent recommendation system as the core link connecting users and goods directly affects user experience and business conversion efficiency. In recent years, the recommendation technology based on graph neural network (GNN) has become a mainstream solution because it can effectively model the association relationship between users, goods and attributes. However, the existing cross-border e-commerce recommendation model based on graph neural network generally adopts a unified discard ratio (such as a fixed probability random discard of part of the interaction edges) for training. This way does not consider the quality difference of the interaction edges, resulting in frequent blind deletion, which seriously restricts the recommendation effect.

[0003] The prior art has the following disadvantages: There are a large number of high-value interaction behaviors in cross-border e-commerce, and such interaction edges contain user real preference information, which is crucial for the model to learn user demand. When the existing model adopts a unified discard ratio, it will randomly discard part of the high-quality interaction edges, resulting in the model losing key features and being unable to fully capture the core preferences of users, and the recommendation result is prone to lack of accuracy; There are a large number of low-quality noise in the interaction data, such as user mis-click, high return dispute rate of goods interaction, and frequent abnormal logistics order records. Such interaction edges cannot reflect the real demand of users, and even mislead the model learning. The existing model treats low-quality interaction edges and high-quality interaction edges equally, and does not reduce their influence through a discard strategy, resulting in the model being disturbed by noise data and the recommendation result being deviated.

[0004] In view of the above problems, the present application provides a solution. SUMMARY

[0005] The purpose of the present application is to provide a cross-border e-commerce intelligent recommendation system based on big data analysis, which is used to solve the problem that the unified discard ratio of the graph neural network in the existing cross-border e-commerce intelligent recommendation model treats different quality interaction edges equally and causes blind deletion.

[0006] The purpose of the present application can be achieved by the following technical solutions: The cross-border e-commerce intelligent recommendation system based on big data analysis comprises the following modules: a heterogeneous graph benchmark training module for constructing user, goods and attribute main data into a multi-relation heterogeneous graph and obtaining an initial model parameter set and a complete interaction edge set through graph neural network training under a fixed discard rate; the initial model parameter set comprises an initial embedding vector of each user, goods and attribute node; The edge quality differentiated training module extracts interaction features from each interaction edge based on the complete interaction edge set and cross-border e-commerce log data to obtain a credibility score and a sparsity index, and generates an edge quality label stream in combination with the edge category; determines control points based on the edge quality label stream distribution statistics, divides each interaction edge into a low-quality section, a medium-quality section, and a high-quality section based on the control points, and assigns decreasing discard probability parameters to each quality section respectively; then generates a differentiated mask using Bernoulli sampling, and jointly trains the graph neural network and outputs an updated model parameter set; The edge classification correction module dynamically adjusts the control points and corrects the interaction edge division based on the updated model parameter set and the quality sections to generate a new multi-relation heterogeneous graph. The strategy optimization and model deployment module fine-tunes the discard probability parameters by performing a multi-armed Bandit algorithm on real-time business indicators, outputs the optimal discard probability parameters, and updates the recommendation model.

[0007] As a further scheme of the present application: the heterogeneous graph benchmark training module includes: generating a commodity entity node set based on the commodity main data table extraction commodity identifier, category identifier and language identifier fields; generating a user entity node set based on the user main data table extraction user identifier and country identifier fields; generating attribute nodes based on foreign key relationships, and establishing attribute relationship edges between user entity nodes and attribute nodes, and between commodity entity nodes and attribute nodes; Assigning semantic weights to different interaction behaviors based on user commodity interaction logs and obtaining a comprehensive interaction score by weighted accumulation, the semantic weights include positive behavior weights and negative behavior weights; and comparing the comprehensive interaction score with a preset positive and negative interaction threshold, generating a plurality of matrices based on the interaction relationship mapping; Mapping the non-zero elements of the matrix to a classification interaction edge, and integrating the attribute relationship edges to form a multi-relation heterogeneous graph. Using a fixed discard rate graph neural network for embedding training, outputting an initial embedding vector and a complete interaction edge set.

[0008] As a further scheme of the present application: mapping generates a matrix reflecting the interaction relationship, specifically, including: mapping the interaction relationship to three 0 / 1 matrix representations, namely membership matrix, non-membership matrix and hesitation matrix; If the comprehensive interaction score is greater than or equal to the preset positive interaction threshold, the corresponding position in the membership matrix is recorded as 1, otherwise it is recorded as 0; thereby forming the membership matrix; If the comprehensive interaction score is less than or equal to the preset negative interaction threshold, the corresponding position in the non-membership matrix is recorded as 1, otherwise it is recorded as 0; thereby forming the non-membership matrix; If the number of interactions of the user with the commodity is between the positive interaction threshold and the negative interaction threshold, 1 is recorded in the hesitation matrix, otherwise 0 is recorded; thereby forming the hesitation matrix; the non-zero elements of each matrix correspond to the initial interaction edge category label generated by them, that is, normal interaction edge, suspicious interaction edge and long-tail sparse interaction edge.

[0009] As a further scheme of the application: the interaction features in the edge quality differentiated training module include: transaction frequency, return dispute rate, logistics abnormality rate and language matching degree between the user and the commodity.

[0010] As a further scheme of the application: the credibility score in the edge quality differentiated training module is obtained by linearly weighting and summing the return dispute rate and the logistics abnormality rate based on their corresponding preset weights to obtain an abnormality coefficient, and using 1 minus the obtained abnormality coefficient to obtain the credibility score. The sparsity index is obtained by counting the interaction frequency of each interaction edge in the complete interaction edge set and calculating the quantile position of the interaction frequency in the whole interaction frequency distribution, and the reciprocal of the quantile position is defined as the sparsity index.

[0011] As a further scheme of the application: the edge quality label stream is generated in combination with the edge category; based on the quantile statistics of the edge quality label stream, each control point is determined, including: the credibility score, the sparsity index and the initial interaction edge category label of each interaction edge are combined to form a real-time updated edge quality label stream. The credibility scores in the edge quality label stream are sorted in full amount, and the value corresponding to the preset low quantile threshold of the credibility score is set as the first control point, which is the first control point for dividing the low quality section and the medium quality section. The value corresponding to the preset high quantile threshold of the credibility score is set as the second control point, which is the second control point for dividing the medium quality section and the high quality section.

[0012] As a further scheme of the application: the decreasing discard probability parameters are respectively allocated according to each quality section, including: when the credibility score of the corresponding interaction edge is less than or equal to the first control point, it is determined to be in the low quality section, and a first preset discard probability is allocated accordingly. When the credibility score of the corresponding interaction edge is greater than the first control point and less than or equal to the second control point, it is determined to be in the medium quality section, and a second preset discard probability is allocated accordingly. When the credibility score of the corresponding interaction edge is greater than the second control point, it is determined to be in the high quality section, and a third preset discard probability is allocated accordingly.

[0013] As a further scheme of the application: the edge classification correction module includes: the similarity of each interaction edge is calculated using the updated embedding vectors in the updated model parameter set as the model feedback edge weight. The model feedback edge weight, the credibility score and the sparsity index corresponding to each interaction edge are linearly weighted and summed to obtain a comprehensive quality score in combination with a preset trade-off coefficient;The preset positive interaction threshold and the preset reverse interaction threshold are updated based on the distribution of the comprehensive quality score, and each interaction relationship matrix is updated;Matrix completion is performed on specific sparse items, and weight suppression is performed on specific noise items.

[0014] As a further scheme of the application: the multi-arm Bandit algorithm includes generating a set of candidate strategy arms by preset rules according to the discard probability, that is, performing segmented threshold fine-tuning on the first control point and the second control point;Performing discard probability fine-tuning on the first preset discard probability parameter, the second preset discard probability and the third preset discard probability to form multiple groups of candidate strategy arms with small step differences; For a multi-relation heterogeneous graph, a directional candidate strategy arm is designed, including a candidate strategy arm for reducing the discard probability of a high-quality segment to which a long-tail sparse interaction edge belongs, and a candidate strategy arm for increasing the discard probability of a low-quality segment to which a suspicious interaction edge belongs, to form a directional arm set corresponding to the problem strategy;And a business index reward function is constructed, and a confidence upper bound algorithm is used to distribute traffic. The model hot update adopts weighted fusion of the initial embedding vector and the updated embedding vector.

[0015] As a further scheme of the application: the construction of the business index reward function and the distribution of traffic using the confidence upper bound algorithm include: extracting click-through rate, conversion rate and average browsing depth from real-time business index flow as basic indexes;Assigning a preset weight to the basic indexes according to the business target, and generating a comprehensive reward value by linear weighting;The comprehensive reward value is bound to the corresponding candidate strategy arm as the feedback basis for strategy evaluation; And distribute user traffic to each candidate strategy arm at a preset proportion, record the recommendation results corresponding to each candidate strategy arm and synchronously collect real-time business indexes; Based on the comprehensive reward value corresponding to each candidate strategy arm, the UCB upper confidence bound algorithm formula UCB(a) = ra+ clnNna is used, where a is the candidate strategy arm identifier and a∈M, M is the number of candidate strategy arms, ra is the average reward value of the candidate strategy arm a, N is the total number of trials, that is, the total user coverage of all candidate strategy arms, na is the user coverage of the candidate strategy arm a, and c is a preset exploration coefficient;The remaining traffic is distributed to the candidate strategy arm with the highest upper confidence bound;A preset proportion of exploration traffic is continuously retained to cope with business fluctuations; When the candidate strategy arm meets any of the following conditions in a continuous preset period, it is determined to be converged: its upper confidence bound is continuously higher than that of other strategy arms and the reward value difference is greater than or equal to a preset difference threshold, or its comprehensive reward value increment is lower than a preset fluctuation threshold;After convergence, the discard probability parameter corresponding to the optimal strategy arm is output as the current optimal configuration, which is denoted as the optimal discard probability parameter.

[0016] Advantages of the present application: (1) The present application, by using the credibility-sparse driven three-stage differentiated discard strategy, applies a higher discard probability to low-quality interaction edges, effectively eliminating noise such as returns, logistics abnormalities, etc.; at the same time, dynamically reduces the discard probability of long-tail sparse and high-quality interactions, fully retains diversified information, so that the graph neural network can consider noise reduction and coverage in the training stage, and improve the accuracy and stability of user-goods embedding representation; (2) The present application, by means of the strategy optimization module, abstracts the three-stage discard probability as a multi-armed Bandit problem, combines click-through rate, conversion rate, and average browsing depth, etc. Comprehensive business feedback, uses upper confidence bound algorithm to fine-tune the discard curve online, and through hot updating, it acts on the recommendation model in real time, realizes the adaptive response to the changes of traffic and category structure, and further enhances the sustainability and robustness of the recommendation effect. BRIEF DESCRIPTION OF DRAWINGS

[0017] The present application will be further described below in conjunction with the accompanying drawings.

[0018] Figure 1 is the system framework structure diagram of the cross-border e-commerce intelligent recommendation system based on big data analysis of the present application; Figure 2 is the implementation logic diagram of the edge quality differentiated training module of the present application; Figure 3 is the implementation logic diagram of the edge classification correction module of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0020] Embodiment one Please refer to Figure 1 The present application is a cross-border e-commerce intelligent recommendation system based on big data analysis, which comprises: The heterogeneous graph benchmark training module is used to perform entity extraction on each business master data table of cross-border e-commerce to form a cross-border e-commerce heterogeneous entity set; based on user commodity interaction logs, a membership matrix, a non-membership matrix and a hesitation matrix are formed by mapping according to a scoring threshold; the non-zero elements of the three matrices are mapped by matrix adjacency to generate three types of interaction edges, namely normal interaction edges, suspicious interaction edges and long-tail sparse interaction edges, and attribute edges of various types are generated according to the master data foreign key; the three types of interaction edges and attribute edges are integrated and analyzed to form a multi-relation heterogeneous graph, and a benchmark graph embedding training is performed based on a fixed drop rate to obtain a complete interaction edge set and an initial model parameter set; Core entities are extracted from each business master data table of cross-border e-commerce to form a heterogeneous entity node set, specifically: the product identification field, the category identification field and the language identification field are extracted from the product master data table to generate a product entity node set; the user identification field and the country identification field are extracted from the user master data table to generate a user entity node set; meanwhile, according to the foreign key relationship in each business master data table of cross-border e-commerce, namely the country identification field in the user master data table and the country identification field in the country master data table; the language identification field in the product master data table and the language identification field in the language master data table; the category identification field in the product master data table and the category identification field in the category master data table; additional attribute nodes are generated, and attribute relationship edges are established for each user entity node and the corresponding country attribute node, each product entity node and the corresponding language attribute node and category attribute node respectively, and the initial weights of the attribute nodes and attribute edges are both set to 1; Finally, a heterogeneous entity node set of cross-border e-commerce is formed, which includes user entity nodes, product entity nodes and attribute nodes of various types such as country, language and category; Further, semantic weights are assigned to different interaction behaviors based on user commodity interaction logs, i.e. historical browsing, purchasing and rating interaction log data of users and commodities, and the comprehensive interaction score is obtained by weighted accumulation; the semantic weights are determined by business semantics, with positive weights for positive behaviors and negative weights for negative behaviors; and the comprehensive interaction score is compared with the preset positive and negative interaction threshold to map the interaction relationship to three 0 / 1 matrix representations, namely the membership matrix, the non-membership matrix and the hesitation matrix, specifically: if the comprehensive interaction score is greater than or equal to the preset positive interaction threshold, the corresponding position in the membership matrix is recorded as 1, otherwise as 0; thereby forming a membership matrix reflecting strong positive preference; If the comprehensive interaction score is less than or equal to the preset negative interaction threshold, the corresponding position in the non-membership matrix is recorded as 1, otherwise as 0; thereby forming a non-membership matrix representing explicit negative feedback; If the number of interactions between the user and the product is between the above-mentioned positive interaction threshold and the negative interaction threshold, 1 is recorded in the hesitation matrix, otherwise 0; thereby forming a hesitation matrix covering long-tail interactions of users; It should be noted that the above preset positive and negative interaction threshold is determined by the business target or initial statistics; for example, if the verified preference is focused on accurate recommendation, the positive interaction threshold can be set to 8, only strong positive behavior is passed, and the negative interaction threshold is -5, only strong negative behavior is passed; if potential preference needs to be mined, the threshold can be reduced, such as the positive interaction threshold is 3 and the negative interaction threshold is -2, and the range of the hesitation matrix is expanded.

[0021] Convert the non-zero elements in the above three matrices into user-goods interaction edges in the graph, and classify the types of edges according to the source matrix, including: adding a normal interaction edge between the user entity node and the goods entity node from the non-zero element in the membership matrix as the initial interaction edge category label; adding a suspicious interaction edge as the initial interaction edge category label from the non-zero element in the non-membership matrix; adding a long-tail sparse interaction edge as the initial interaction edge category label from the non-zero element in the hesitation matrix; and there is at most one interaction edge between each pair of user entity nodes and goods entity nodes; all the interaction edges and attribute relationship edges generated in the above process are summarized to form a multi-relation heterogeneous graph containing multiple node types and multiple relation edge types; The graph neural network (GNN) is used for embedding training on the above multi-relation heterogeneous graph to obtain the initial vector representation of each node and the initial model parameters; to achieve this, a heterogeneous graph neural network model can be used to fuse node features and graph structure; the training process is as follows: A recommendation model containing graph convolution or graph attention layers is used to input user-goods interaction and attribute relationship as graph structure; the model parameters include node feature transformation weight, weight of aggregated neighbor information, etc. In this benchmark phase, the graph structure is discarded using a fixed dropout rate, that is, a portion of the edges is randomly discarded with a fixed probability in each training iteration to prevent overfitting; Common recommendation task losses are used, such as the error between the predicted user rating / click rate for a product and the real log data as a supervision signal, or a contrast learning framework (positive and negative samples) is used to optimize the embedding representation; in the benchmark training, no special processing is done for different category edges, and the overall performance is mainly focused on; After completing the benchmark training, the initial model parameter set is output, including the initial embedding vector of each user, goods, and attribute node; these vectors condense the graph structure information, which is used as the basic representation of the recommendation model and the weight parameters, bias, etc. of the graph neural network, and the complete interaction edge set of the normal / suspicious / long-tail three types of interaction edges is retained and constructed; It should be noted that the graph neural network (GNN) adopts a GraphSAGE architecture including three layers of graph convolution, including an input layer receiving a 128-dimensional feature vector, a hidden layer with ReLU and LeakyReLU activation functions, a Sigmoid output layer outputting a 16-dimensional embedding representation, and a two-order neighbor sampling mechanism. The specific construction process is an existing mature technology, which will not be described in detail in this embodiment. In this embodiment, various data of cross-border e-commerce business are fused into a multi-relation heterogeneous graph, and a graph neural network with a fixed dropout rate is used for benchmark embedding training to obtain initial node embedding representation and model parameters, which provide a base for subsequent modules.

[0022] Please refer to Figure 2 The edge quality differentiated training module is used to perform cross-border e-commerce interaction feature extraction on each interaction edge based on the complete interaction edge set and in combination with cross-border e-commerce log data, to obtain hierarchical cross-border e-commerce feature data; and the hierarchical cross-border e-commerce feature data is summarized in a sliding window to obtain a credibility score and a sparsity index; the credibility score, the sparsity index, and the edge category are integrated and analyzed to obtain an edge quality label stream and its quantile statistical cache; at the same time, the quantile statistical cache is read to pre-calculate and update three control points of the three-section dropout; as the dropout probability parameter of the current round, and the dropout probabilities of the three types of edges are calculated to generate a differentiated training batch; the edge domain entropy regularization and the topological consistency constraint are introduced in the same batch, and the joint anti-noise and long-tail gain strategy is used to complete the training; and the updated model parameter set and the dropout probability parameter used in the current round are output. Based on the cross-border e-commerce log data, within a preset sliding time window, transaction frequency, return dispute rate, logistics abnormality rate, and language matching degree indicators are extracted for each user-goods interaction edge; the indicators are used to evaluate the quality of each interaction edge, specifically including: Transaction frequency: the number of interactions or purchases of the user and the goods within the statistical window; Return dispute rate: the proportion of after-sales disputes in the user's order of the goods; Logistics abnormality rate: the proportion of logistics problems in the user's purchase of the goods; Language matching degree between the user and the goods: the similarity or matching degree of the user's language and the language of the goods; It should be noted that each of the above indicators is a conventional storage data of cross-border e-commerce log big data, and the logic of obtaining the data is an existing mature technology, which will not be described in detail in the embodiment. Within the preset sliding window, the following two key quality indicators are calculated according to the features: The credibility score: the abnormality coefficient is obtained by linearly weighting and summing the return dispute rate and the logistics abnormality rate based on their corresponding preset weights, and then the credibility is defined as 1 minus the abnormality coefficient, and is normalized to 0-1, the higher the score, the more credible the interaction; wherein the preset weight can be set based on business experience; The sparsity index: the interaction frequency of each interaction edge in the complete interaction edge set is counted, and its quantile position in the whole interaction frequency distribution is calculated, and then its reciprocal is obtained; the quantile is obtained based on the interaction frequency of the user and the commodity in the preset sliding window; The credibility score, sparsity index and initial interaction edge category label of each interaction edge, including normal interaction edge, suspicious interaction edge and long-tail sparse interaction edge, are combined to form a real-time updated edge quality label stream; The quantile statistical analysis is performed on the historical statistical data in the edge quality label stream to determine the control points of the three-section discard, specifically: The credibility score in the edge quality label stream is sorted in full, and the interval boundary is determined based on the preset quantile threshold, specifically including: The value corresponding to the preset low quantile threshold of the credibility score is set as the first control point Tlow, which is used to divide the low quality section and the medium quality section; The value corresponding to the preset high quantile threshold of the credibility score is set as the second control point Thigh, which is used to divide the medium quality section and the high quality section; Wherein, the preset quantile threshold such as the preset low quantile threshold can be 25% quantile, and the preset high quantile threshold can be 75% quantile, and the specific value can be determined according to the proportion demand of the system for low, medium and high quality edges and business targets, the control points divide the low quality section, the medium quality section and the high quality section, to give different edge discard probabilities respectively; According to the control points and the edge quality section, the credibility score of each interaction edge is determined in the interval and the corresponding discard probability is allocated, forming a three-section edge discard probability, including: When the credibility score of the corresponding interaction edge is less than or equal to the first control point, it is determined as the low quality section, and the first preset discard probability Plow is allocated correspondingly; the specific value of the first preset discard probability parameter Plow can be set according to the specific implementation requirement, which can be set in the interval of 0.6 to 0.9, to weaken the influence of low credibility edge; When the credibility score of the corresponding interaction edge is greater than the first control point and less than or equal to the second control point, it is determined as the medium quality section, and the second preset discard probability Pmid is allocated correspondingly; When the credibility score of the corresponding interaction edge is greater than the second control point, it is determined as a high-quality section, and a third preset discard probability Phigh is assigned accordingly; the values of the first, second and third preset discard probabilities decrease in turn, so as to realize high discard rate suppression of the low-quality section interaction edge and low discard rate reservation of the high-quality section interaction edge; Further, based on the above three-section edge discard probability, each interaction edge in the complete interaction edge set is subjected to differential mask processing, including: According to the edge discard probability parameter of the quality section corresponding to each interaction edge, Bernoulli random sampling is used to determine the retention or discard state of the interaction edge in the current training batch; Further, for the interaction edge belonging to the long-tail sparse interaction edge category and having a sparse degree index higher than a preset sparse degree threshold, the probability is adjusted downward on the basis of the edge discard probability of the original quality section, so as to increase the chance of retention and improve the attention of the model to long-tail interaction data; The subgraph structure of the current training batch is obtained through the above differential mask sampling method, and the edge domain entropy regular constraint and the topological consistency constraint of the prior art are introduced in the training process, as well as the noise suppression strategy based on adversarial perturbation, to jointly complete differential training; After training, the updated model parameter set and the three-section edge discard probability parameter used in the current round are output, including the updated embedding vector, the corresponding control point parameter, the specific edge discard probability parameter of each quality section and the sparse degree threshold parameter, so as to facilitate subsequent model iteration and strategy optimization; It should be noted that the above uses Bernoulli random sampling mechanism to introduce mask operation for each interaction edge; specifically, according to the edge quality section where the interaction edge is located, the corresponding edge discard probability is assigned as the Bernoulli distribution parameter, and the sampling result is used to determine whether the edge is retained in the current training round, so as to build a differential training adjacency structure. In addition, in combination with the existing graph neural network optimization method, two regular terms of edge domain entropy regular constraint and topological consistency constraint are introduced in the training process: the former is based on the attention entropy distribution of different interaction edge types, guiding the model to balance the use of normal, suspicious and long-tail sparse edges, preventing bias in the training set; the latter compares the node embedding adjacency relationship of the complete graph and the sampled graph, guiding the model to maintain the consistency of the overall graph structure expression under the condition that the edge is randomly discarded; the above two constraint mechanisms are known common methods for improving the robustness of edge distribution and the structure preservation ability in graph learning, which can be used to optimize the training effect in cooperation with the differential discard probability parameter, and the basic principles thereof will not be described in detail in this embodiment.

[0023] Please refer to Figure 3, the edge classification correction and new graph module is used to adjust and correct the score thresholds and confidence intervals corresponding to the membership matrix, the non-membership matrix and the hesitation matrix according to the edge weight distribution and the edge quality label stream of the updated model parameter set, to perform low-rank completion on high sparse items and weight suppression on high noise items; the corrected three matrices and the divided interaction edges are formed to generate a new multi-relation heterogeneous graph; Using the updated embedding vectors of each entity node in the updated model parameter set output by the edge quality differential training module, the cosine similarity between the embedding vectors of the corresponding user entity node and the embedding vectors of the corresponding commodity entity node is calculated for any interaction edge, and the obtained cosine similarity score is taken as the model feedback weight of the interaction edge, which is used to reflect the judgment of the contribution of the model to the edge; then, the edge feedback weight is compared and analyzed with the credibility score and the sparsity index in the aforementioned edge quality label stream to determine which interaction edges are underestimated or overestimated in the initial classification, and the membership threshold, the non-membership threshold and the hesitation interval are dynamically adjusted accordingly to complete the adaptive correction of the membership matrix, the non-membership matrix and the hesitation matrix, so that the classification of the interaction edges is more in line with the learning results of the model; Using the edge feedback weight and the credibility score and the sparsity index calculated by the edge quality differential training module, the comprehensive quality score Q of the user-commodity edge is calculated, which can be defined as: Comprehensive quality score Q = α × model feedback edge weight + β × credibility score - γ × sparsity index; Wherein, α, β, γ are preset weighting coefficients determined by verification data or business requirements to balance the influence of model feedback and historical statistics; According to the statistical analysis of the real-time distribution of the comprehensive quality score Q of all interaction edges, the corresponding quantile is calculated based on the preset high-interaction quantile threshold and the preset low-interaction quantile threshold determined by business requirements, which are respectively taken as the new positive interaction threshold and the new negative interaction threshold; wherein, the preset high-interaction quantile threshold and the preset low-interaction quantile threshold can be any quantile level between 0% and 100%, and the specific value can be the quartile point but is not limited to the quartile point; specifically: The value corresponding to the preset high-interaction quantile threshold of the comprehensive quality score Q is calculated, and the value is taken as the new positive interaction threshold; The value corresponding to the preset low-interaction quantile threshold of the comprehensive quality score Q is calculated, and the value is taken as the new negative interaction threshold; Based on the dynamically updated positive interaction threshold and negative interaction threshold, the user-commodity interaction relationship is re-labeled, and the membership matrix, the non-membership matrix and the hesitation matrix are updated, and the specific rules are as follows: If the comprehensive quality score Q of the interaction edge is greater than or equal to the new positive interaction threshold, the corresponding position of the membership matrix is recorded as 1, indicating normal interaction; If the comprehensive quality score Q of the interaction edge is less than or equal to the new negative interaction threshold, the non-affiliation matrix at the corresponding position is marked 1, indicating a suspicious interaction; If the comprehensive quality score Q of the interaction edge is between the new positive interaction threshold and the new negative interaction threshold, the hesitation matrix at the corresponding position is marked 1, indicating a long-tail sparse interaction; Interactions that do not belong to any of the above cases are marked as no interaction, and all matrix corresponding positions are marked 0.

[0024] Further, in the present embodiment, special processing is also included for the entries in the hesitation matrix and the non-affiliation matrix that are of particular interest: For high sparse entries in the hesitation matrix, i.e., interaction edges corresponding to a sparsity index exceeding the high sparsity threshold, the interaction score is predicted and filled using a low-rank matrix completion method to compensate for the lack of data; For high noise entries in the non-affiliation matrix, i.e., interaction edges corresponding to a credibility score significantly lower than the average credibility level, the influence on model training is further reduced by reducing the edge weight or directly removing the interaction edge; The non-zero elements in the modified affiliation matrix, non-affiliation matrix and hesitation matrix are remapped to normal interaction edges, suspicious interaction edges and long-tail sparse interaction edges, the existing attribute relationship edges are integrated, and a new multi-relation heterogeneous graph is constructed to prepare for the next round of iteration training.

[0025] The strategy optimization and model deployment module is used to input the real-time business indicator stream and the updated model parameter set, as well as the new multi-relation heterogeneous graph and the discarded probability parameter of the current round; perform a multi-armed Bandit algorithm to evaluate and fine-tune the discarded probability parameter; form a fine-tuned three-section discarded control point, generate the optimal discarded probability parameter for the next round and write the new control point back to the quantile statistics cache; output the final merged base embedding and the latest model parameter set of the recommendation model, and hot update to the online service; The three-section discarded control point (Tlow, Thigh) and the corresponding parameter combination of each section discarded probability (Plow, Pmid, Phigh) are abstracted as a strategy arm in the multi-armed Bandit problem; each complete parameter configuration, i.e., a set of values of (Tlow, Thigh, Plow, Pmid, Phigh), corresponds to an independent arm, representing a differentiated discarded strategy; Based on the benchmark control point of the current quantile statistics cache, such as the historical stable section threshold and the initial discarded probability, the candidate strategy arm set is generated according to the preset rules, including: On the basis of the benchmark parameters, segment threshold fine-tuning is performed on Tlow such as ±5% quantile offset, Thigh such as ±5% quantile offset; discard probability fine-tuning is performed on Plow such as ±0.1 adjustment, Pmid such as ±0.1 adjustment, Phigh such as ±0.1 adjustment, forming multiple groups of small-step-amplitude-difference alternative strategy arms; In view of the known problems of the interactive edges in the multi-relation heterogeneous graph, such as the long-tail edge discard leading to insufficient diversity, the suspicious edge retention causing noise, and the like, a directional alternative strategy arm is designed, including a long-tail sparse interactive edge belonging to a high-quality segment discard probability alternative strategy arm, i.e., Phigh down-regulation, an alternative strategy arm for improving the discard probability of a suspicious interactive edge belonging to a low-quality segment, i.e., Plow up-regulation, and the like, forming a directional arm set corresponding to the problem-strategy pair; From the real-time business indicator stream, core indicators are screened, and a reward function is constructed, including: extracting indicators strongly related to the recommendation effect from the real-time business indicator stream, including: click rate (CTR), conversion rate (i.e., the number of times a user purchases a recommended product / the number of clicks), and average browsing depth as basic indicators; assigning weights to the basic indicators according to the business target, with the weight being 1; generating a comprehensive reward value through linear weighting; binding the calculated comprehensive reward value to the corresponding alternative strategy arm, i.e., associating the strategy arm identifier and the effective timestamp, as feedback for subsequent strategy evaluation; Strategy iteration is implemented using the UCB upper confidence bound algorithm: A fixed proportion, such as 10% of the total traffic, is used to randomly allocate user requests to each alternative strategy arm, such as alternative strategy arm 1 covering 3% of users and alternative strategy arm 2 covering 3% of users, to ensure that each alternative strategy arm obtains sufficient interactive samples, such as the single-arm cumulative user coverage being greater than or equal to a preset threshold; in this stage, the multi-relation heterogeneous graph recommendation results corresponding to each alternative strategy arm, such as the influence of different edge types on the recommendation link, are recorded, real-time business basic indicators are synchronously collected, corresponding comprehensive reward values are calculated, and are associated to the alternative strategy arm; Based on the comprehensive reward values corresponding to each alternative strategy arm, the UCB upper confidence bound algorithm formula UCB(a) = ra+ clnNna is used, where a is the alternative strategy arm identifier and a ∈ M, M is the number of alternative strategy arms, ra is the average reward value of the alternative strategy arm a, N is the total number of trials, i.e., the total user coverage of all alternative strategy arms, na is the number of users covered by the alternative strategy arm a, and c is a preset exploration coefficient used to control the exploration intensity, which can be taken as c = 2; the confidence upper bound of each alternative strategy arm is dynamically calculated, the remaining traffic is allocated to the alternative strategy arm with the highest upper confidence bound, and high-potential configurations are preferentially verified; at the same time, a constant exploration ratio, such as a small traffic exploration triggered once an hour, is maintained to cope with business indicator fluctuations and strategy decay; When a certain alternative strategy arm satisfies any of the following conditions within a continuous K statistical periods, such as K = 72 hours, the algorithm is determined to have converged: The upper confidence limit of the alternative strategy arm is continuously higher than all other strategy arms, that is, the reward value difference is greater than or equal to a preset difference threshold; The integrated reward value increment of the alternative strategy arm is lower than a preset fluctuation threshold, such as a reward value fluctuation of less than 0.5% for consecutive K periods; after convergence, the alternative strategy arm is locked as the optimal strategy arm, and the corresponding discard probability parameters T* low, T* high and P* low, P* mid, P* high are output as the current optimal configuration, which is denoted as the optimal discard probability parameter.

[0026] The optimal discard probability parameter of the optimal strategy arm after convergence is written into the quantile statistical cache to overwrite the original baseline parameter; and subsequent modules directly call the parameters verified online to realize a strategy iteration closed loop; The model parameters of historical training and the latest iteration are fused to generate a final recommendation model and perform hot updating online, and the steps are as follows: The base embedding generated by the heterogeneous graph baseline training module and the updated embedding vector generated by the edge are fused by weighting to obtain the fused node embedding Hfinal as follows: Hfinal = ε * H0 + (1-ε) * Ht), wherein H0 is the base embedding output by the edge quality differentiated training module, Ht is the updated embedding vector output by the edge quality differentiated training module, and ε is a weight such as ε = 0.3, which retains 30% of the global structure information. The fused node embedding Hfinal, the updated model parameter set and the optimal discard probability parameter are uploaded to the inference server together for seamless hot updating.

[0027] The present application proposes a cross-border e-commerce heterogeneous graph intelligent recommendation system to solve the problem of blind deletion of high-quality interaction edges caused by uniform discard ratio. First, a user-goods-attribute multi-relation heterogeneous graph is constructed to obtain a complete interaction edge set; then, features such as transaction frequency and return dispute rate are extracted to calculate credibility and sparsity, control points are set according to quantiles, interaction edges are divided into high, medium and low quality sections, and corresponding decreasing discard probabilities are assigned, and Bernoulli sampling is combined with graph neural network training; then, the threshold and edge classification are dynamically corrected according to the model feedback, and the heterogeneous graph is reconstructed; finally, the UCB multi-arm Bandit algorithm is used to online fine-tune the discard probability and hot update the model to realize noise reduction, long tail preservation and continuous improvement of recommendation accuracy.

[0028] The above formulas are dimensionless and the numerical values are calculated, the formula is obtained by software simulation of a large amount of data to obtain the latest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0029] The above embodiments can be realized by software, hardware, firmware or any combination thereof, in whole or in part. When realized by software, the above embodiments can be realized in the form of a computer program product in whole or in part.

[0030] Those skilled in the art can understand that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application of the technical solution and the constraints of the invention. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0031] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0032] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0033] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A cross-border e-commerce intelligent recommendation system based on big data analysis, characterized in that: The system includes the following modules: a heterogeneous graph benchmark training module, which is used to construct a multi-relationship heterogeneous graph from user, product, and attribute master data and train it through a graph neural network at a fixed dropout rate to obtain an initial model parameter set and a complete set of interaction edges; the initial model parameter set includes the initial embedding vector of each user, product, and attribute node; The edge quality differentiation training module, based on the complete set of interactive edges and cross-border e-commerce log data, extracts interactive features from each interactive edge to obtain a credibility score and sparsity index, and generates an edge quality label stream by combining the edge category. Based on the quantile statistics of the edge quality label stream, each control point is determined. Based on each control point, each interactive edge is divided into low-quality segment, medium-quality segment, and high-quality segment, and a decreasing drop probability parameter is assigned to each quality segment. Subsequently, a differentiation mask is generated using Bernoulli sampling, which is jointly trained with the graph neural network and outputs an updated set of model parameters. The edge classification correction module dynamically adjusts each control point and corrects the interaction edge division based on the updated model parameter set and each quality segment, generating a new multi-relationship heterogeneous graph; The strategy optimization and model deployment module fine-tunes each drop probability parameter by executing the multi-armed Bandit algorithm on real-time business metrics, outputs the optimal drop probability parameters, and updates the recommendation model.

2. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 1, characterized in that, The heterogeneous graph benchmark training module includes: generating a set of product entity nodes by extracting product identifier, category identifier, and language identifier fields from the product master data table; generating a set of user entity nodes by extracting user identifier and country identifier fields from the user master data table; generating attribute nodes based on foreign key relationships, and establishing attribute relationship edges between user entity nodes and attribute nodes, and between product entity nodes and attribute nodes; Based on user product interaction logs, semantic weights are assigned to different interaction behaviors and then weighted and accumulated to obtain a comprehensive interaction score. The semantic weights include positive behavior weights and negative behavior weights. The comprehensive interaction score is then compared with preset positive and negative interaction thresholds, and multiple matrices are generated based on the interaction relationship mapping. Map the non-zero elements of the matrix to classification interaction edges, and integrate attribute relationship edges to form a multi-relationship heterogeneous graph; An embedding training is performed using a graph neural network with a fixed dropout rate, which outputs an initial embedding vector and a complete set of interaction edges.

3. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 2, characterized in that, The mapping generates a matrix that reflects the interaction relationship. Specifically, it includes mapping the interaction relationship to three 0 / 1 matrix representations: a membership matrix, a non-membership matrix, and a hesitation matrix. If the overall interaction score is greater than or equal to the preset positive interaction threshold, then the corresponding position in the membership matrix is ​​recorded as 1; otherwise, it is recorded as 0; thus forming the membership matrix. If the overall interaction score is less than or equal to the preset negative interaction threshold, then the corresponding position in the non-membership matrix is ​​recorded as 1; otherwise, it is recorded as 0; thus forming the non-membership matrix. If the number of times a user interacts with a product falls between the positive and negative interaction thresholds mentioned above, it is recorded as 1 in the hesitation matrix; otherwise, it is recorded as 0. This forms the hesitation matrix. The non-zero elements of each matrix are used to generate their respective initial interaction edge category labels, namely, normal interaction edge, suspicious interaction edge, and long-tail sparse interaction edge.

4. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 1, characterized in that, Interactive features in the edge quality differentiation training module include: transaction frequency, return dispute rate, logistics anomaly rate, and user-product language matching index.

5. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 4, characterized in that, The credibility score in the edge quality differentiation training module is obtained by linearly weighting the return dispute rate and logistics anomaly rate based on their corresponding preset weights to obtain the anomaly coefficient. The anomaly coefficient is then subtracted from 1 and the result is defined as the credibility score. The sparsity index is defined by statistically analyzing the interaction frequency of each interaction edge in the complete set of interaction edges, calculating its quantile position in the distribution of all interaction frequencies, and defining its reciprocal as the sparsity index.

6. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 1, characterized in that, The edge quality label stream is generated by combining the edge categories; each control point is determined based on the quantile statistics of the edge quality label stream, including: combining the credibility score and sparsity index of each interaction edge with the initial interaction edge category label to form a real-time updated edge quality label stream; The credibility scores in the edge quality label stream are sorted in full. The value corresponding to the credibility score at the preset low percentile threshold is set as the first control point, which serves as the first control point for delineating the boundary between the low quality segment and the medium quality segment. The value corresponding to the confidence score being at the preset high percentile threshold is set as the second control point, which serves as the second control point for defining the medium quality segment and the high quality segment.

7. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 6, characterized in that, Each quality segment is assigned a decreasing discard probability parameter, including: when the credibility score of the corresponding interaction edge is less than or equal to the first control point, it is determined to be a low quality segment and a first preset discard probability is assigned accordingly. When the credibility score of the corresponding interaction edge is greater than the first control point and less than or equal to the second control point, it is judged as a medium quality segment and a second preset discard probability is assigned accordingly. When the credibility score of the corresponding interaction edge is greater than that of the second control point, it is judged as a high-quality segment and a third preset discard probability is assigned accordingly.

8. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 1, characterized in that, The edge classification correction module includes: calculating the similarity of each interaction edge using the updated embedding vector in the updated model parameter set as the model feedback edge weight; The comprehensive quality score is obtained by linearly weighting and summing the model feedback edge weights, credibility scores, and sparsity indices corresponding to each interaction edge by combining preset trade-off coefficients; the preset positive interaction threshold and preset negative interaction threshold are updated based on the distribution of the comprehensive quality score, and each interaction relationship matrix is ​​updated; matrix completion is performed on specific sparse entries, and weight suppression is performed on specific noise entries.

9. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 1, characterized in that, The multi-arm Bandit algorithm includes generating a set of candidate strategy arms according to preset rules for each drop probability, namely, performing segmented threshold fine-tuning on the first control point and the second control point; and performing drop probability fine-tuning on the first preset drop probability parameter, the second preset drop probability, and the third preset drop probability to form multiple sets of candidate strategy arms with small step differences. For heterogeneous graphs with multiple relationships, we design targeted alternative strategy arms, including alternative strategy arms that reduce the probability of discarding high-quality segments to which long-tail sparse interaction edges belong, and alternative strategy arms that increase the probability of discarding low-quality segments to which suspicious interaction edges belong, forming a set of targeted arms corresponding to the problem strategy; and we construct a business indicator reward function and use a confidence upper bound algorithm to allocate traffic. The model hot update uses a weighted fusion of the initial embedding vector and the updated embedding vector.

10. The intelligent recommendation system for cross-border e-commerce based on big data analysis according to claim 9, characterized in that, The construction of the business indicator reward function, which uses a confidence upper bound algorithm to allocate traffic, includes: extracting click-through rate, conversion rate, and average browsing depth per user from the real-time business indicator stream as basic indicators; assigning preset weights to the basic indicators according to business objectives, and generating a comprehensive reward value through linear weighting; and binding the comprehensive reward value with the corresponding alternative strategy arm as a feedback basis for strategy evaluation. User traffic is allocated to each alternative strategy arm according to a preset ratio, the recommendation results corresponding to each alternative strategy arm are recorded, and real-time business metrics are collected synchronously. Based on the comprehensive reward value corresponding to each candidate strategy arm, the UCB upper confidence boundary algorithm formula is applied: UCB(a) = ra + clnNna; where a is the candidate strategy arm identifier and a∈M, M is the number of candidate strategy arms, ra is the average reward value of candidate strategy arm a, N is the total number of trials, i.e., the total user coverage of all candidate strategy arms, na is the number of users covered by candidate strategy arm a, and c is the preset exploration coefficient; the remaining traffic is allocated to the candidate strategy arm with the highest upper confidence boundary; a preset proportion of exploration traffic is continuously reserved to cope with business fluctuations; Convergence is determined when the candidate strategy arm meets any of the following conditions within a continuous preset period: its upper confidence boundary is continuously higher than that of other strategy arms and the difference in reward value is greater than or equal to the preset difference threshold, or its comprehensive reward value increase is lower than the preset fluctuation threshold; after convergence, the drop probability parameter corresponding to the optimal strategy arm is output as the current optimal configuration and recorded as the optimal drop probability parameter.

Images