A multi-dimensional feature-driven B2B2C collaborative recommendation method and system

Abstract

This invention relates to the field of data processing, providing a multi-dimensional feature-driven B2B2C collaborative recommendation method and system. The method includes: collecting multi-dimensional features from B-end merchants, C-end users, and products through a heterogeneous data source interface to obtain standardized multi-dimensional features; dynamically learning weights on the standardized multi-dimensional feature dataset using a multi-head self-attention mechanism to obtain a fused feature vector; constructing a three-layer collaborative matrix on the fused feature vector based on tensor decomposition to obtain a multi-dimensional factor matrix; modeling causal relationships on the multi-dimensional factor matrix using a collaborative filtering algorithm based on a causal graph structure to obtain a deep collaborative network model; and performing real-time recommendation of items to be recommended using the deep collaborative network model to obtain a personalized B2B2C recommendation list. This invention can capture complex patterns in business scenarios and improve the accuracy of personalized recommendation results.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-dimensional feature-driven B2B2C collaborative recommendation method and system. Background Technology

[0002] With the rapid development of e-commerce, the B2B2C (Business to Business to Consumer) business model has matured. This model connects B-end merchants and C-end users through platforms, forming a complex ecosystem of interaction among the three parties. Existing B2B2C recommendation systems mainly use traditional collaborative filtering algorithms, calculating recommendations through a user-product rating matrix. Some advanced systems have begun to introduce deep learning technologies, such as neural network-based collaborative filtering and matrix factorization. When dealing with user preference prediction, these methods typically use a single user-product binary relationship model, utilizing historical interaction data for similarity calculation or latent factor learning.

[0003] However, existing technologies have significant shortcomings: First, traditional methods neglect the role of merchants as important intermediaries in B2B2C scenarios, failing to construct a user-product-merchant ternary relationship model, resulting in recommendation results lacking consideration of merchant characteristics and merchant-user suitability; second, most existing feature fusion methods employ simple linear weighting or shallow neural networks, which cannot effectively capture the complex interaction relationships and dynamic changes between features of different dimensions; third, traditional collaborative filtering algorithms generally suffer from causal confusion problems, failing to distinguish between genuine user preference causal relationships and spurious correlations, and are easily affected by confusion factors such as selection bias and popularity bias, reducing the accuracy and fairness of recommendations. Summary of the Invention

[0004] This invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation method and system to address the shortcomings of existing technologies.

[0005] This invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation method, comprising:

[0006] S1: Collect B-end merchant characteristics, C-end user characteristics and product characteristics from multiple dimensions through heterogeneous data source interfaces to obtain standardized multidimensional features;

[0007] S2: Dynamic weight learning is performed on the standardized multidimensional feature dataset through a multi-head self-attention mechanism to obtain a fused feature vector;

[0008] S3: Based on tensor decomposition, a three-layer collaborative matrix is ​​constructed on the fused feature vector to obtain a multidimensional factor matrix;

[0009] S4: By using a collaborative filtering algorithm based on a causal graph structure, the multidimensional factor matrix is ​​modeled for causal relationships to obtain a deep collaborative network model;

[0010] S5: Real-time recommendation of items to be recommended is performed through the deep collaborative network model to obtain a personalized B2B2C recommendation list.

[0011] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S1 further includes:

[0012] S11: A distributed feature acquisition system is obtained by deploying feature acquisition agents on multiple data nodes through a real-time stream processing architecture.

[0013] S12: Based on the distributed feature acquisition system, B-end merchant data, C-end user data and product data are collected through the data source interface to obtain multi-source heterogeneous raw data;

[0014] S13: Extract semantic features from the multi-source heterogeneous raw data using NLP technology to obtain standardized multidimensional features.

[0015] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, the multi-source heterogeneous raw data in step S12 includes:

[0016] B-end merchant data, which includes: business category data, geographical location data, user review data, transaction volume data, and inventory status data;

[0017] C-end user data, which includes: demographic data, historical purchase behavior data, browsing preference data, social network relationship data, and time preference data;

[0018] Product data includes: category attribute data, price range data, brand information data, user rating data, and sales data.

[0019] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S2 further includes:

[0020] S21: Nonlinear interaction modeling of different dimensions of features in standardized multidimensional features is performed by a feature interaction encoder to obtain feature interaction representation;

[0021] S22: Calculate the importance of features in different dimensions, and optimize the feature weight parameters online based on the feature contribution scores obtained from the importance calculation to obtain dynamic weight parameters;

[0022] S23: Based on the dynamic weight parameters and the feature interaction representation, the different dimensions of the standardized multidimensional features are reduced and fused to obtain a fused feature vector.

[0023] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S3 further includes:

[0024] S31: Based on the fused feature vector, tensor modeling is performed on the user-product-merchant relationship triple to obtain the collaborative tensor;

[0025] S32: Using the improved CP decomposition algorithm, latent factors are extracted from the fourth-order cooperative tensor to obtain initial factor results;

[0026] S33: Optimize the ranking objective of the relation triples based on the BPR loss function to obtain the preference strength score;

[0027] S34: Based on the preference intensity score, the initial factor results are decomposed through parallel matrix decomposition to obtain a multidimensional factor matrix including a user latent factor matrix, a product latent factor matrix, and a merchant latent factor matrix.

[0028] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S34 further includes:

[0029] S341: Select non-interactive relation triples in the cooperative tensor through a negative sampling mechanism to obtain a balanced training sample set;

[0030] S342: The user historical behavior in the balanced training sample set is weighted by a time decay factor to obtain an optimized training sample set.

[0031] S343: By optimizing the training sample set, sparsity control is applied to the preference intensity score to obtain interpretable latent factors;

[0032] S344: Based on the alternating least squares method, the interpretable latent factors are iteratively optimized to obtain a multidimensional factor matrix.

[0033] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S4 further includes:

[0034] S41: Based on the multidimensional factor matrix, causal structure identification is performed to obtain a relational causal graph;

[0035] S42: Identify confounding variables in the causal graph to obtain a set of confounding factors;

[0036] S43: By using a causal attention network, feature filtering is performed on the set of confusing factors to obtain a causal feature interaction representation;

[0037] S44: The weights of the causal feature interaction representation are adjusted by the counterfactual reasoning module to train the causal attention network and obtain a deep collaborative network model.

[0038] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S44 further includes:

[0039] S441: Perform causal intervention calculations on the aforementioned causal graph using the Do-calculus algorithm to obtain an estimated causal effect value;

[0040] S442: Perform counterfactual reasoning on the estimated causal effect to obtain a counterfactual behavior pattern;

[0041] S443: The counterfactual behavior pattern and the causal feature interaction representation are fused to obtain the debiased causal features;

[0042] S444: Based on the debiased causal features, the model parameters of the causal attention network are adjusted to obtain a trained deep collaborative network model.

[0043] According to the multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention, step S5 further includes:

[0044] S51: Perform model quantization on the trained deep collaborative network model to obtain a lightweight recommendation model;

[0045] S52: Output the predicted score of the item to be recommended through the lightweight recommendation model;

[0046] S53: Generate a personalized B2B2C recommendation list based on the predicted scores.

[0047] This invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation system, comprising:

[0048] The data collection module is used to collect B-end merchant characteristics, C-end user characteristics, and product characteristics from heterogeneous data source interfaces to obtain standardized multi-dimensional features.

[0049] Fusion module: used to dynamically learn weights on the standardized multidimensional feature dataset through a multi-head self-attention mechanism to obtain a fused feature vector;

[0050] Construction module: used to construct a three-layer collaborative matrix from the fused feature vector based on tensor decomposition, to obtain a multidimensional factor matrix;

[0051] Modeling module: used to model the causal relationships of the multidimensional factor matrix using a collaborative filtering algorithm based on a causal graph structure, to obtain a deep collaborative network model;

[0052] The recommendation module is configured as a deep collaborative network model trained by the modeling module, and is used to make real-time recommendations through the deep collaborative network model to obtain a personalized B2B2C recommendation list.

[0053] This invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation method and system. First, through a distributed feature acquisition system and NLP semantic feature extraction technology, it efficiently acquires and standardizes multi-dimensional feature information from heterogeneous data sources, including B-end merchants, C-end users, and products. This solves the problems of data silos and feature inconsistencies in traditional systems, enabling the system to comprehensively capture the feature representations of various entities in the business ecosystem, laying a solid data foundation for subsequent intelligent recommendations. The subsequent introduction of a multi-head self-attention mechanism allows the system to adaptively learn the dynamic interaction relationships and importance weights between features of different dimensions, compared to traditional static feature fusion. The method significantly improves the richness and accuracy of feature representation. At the same time, the feature interaction encoder captures complex feature dependencies through nonlinear modeling, enabling the fused feature vectors to better reflect the complex patterns in actual business scenarios. Secondly, the three-layer collaborative matrix construction method based on tensor decomposition breaks through the limitations of traditional binary relation modeling. Through user-product-merchant triple tensor modeling and improved CP decomposition algorithm, it can simultaneously characterize the potential relationships between user preferences, product attributes, and merchant features, enabling the recommendation system to more comprehensively understand the complex interaction patterns in B2B2C scenarios, thereby generating more accurate personalized recommendation results.

[0054] It should be noted that the collaborative filtering algorithm based on causal graph structure and the application of the counterfactual reasoning module of this invention fundamentally solve the causal confusion problem in traditional recommendation systems. By performing causal intervention calculations and identifying confounding variables through the Do-calculus algorithm, the interference of confounding factors such as selection bias and popularity bias can be effectively eliminated, making the recommendation results more realistically reflect the user's intrinsic preferences and needs. This significantly improves the accuracy, fairness, and interpretability of recommendations, and has important theoretical value and practical significance in the practical application of artificial intelligence recommendation algorithms, providing an effective technical path for building more intelligent and trustworthy recommendation systems. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0056] Figure 1 A schematic diagram of a multi-dimensional feature-driven B2B2C collaborative recommendation method provided by the present invention;

[0057] Figure 2 This is a schematic diagram of a multi-dimensional feature-driven B2B2C collaborative recommendation system provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0059] The embodiments of the present invention are described below with reference to the figures.

[0060] like Figure 1 As shown, this invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation method, including:

[0061] S1: Collect B-end merchant characteristics, C-end user characteristics, and product characteristics from multiple dimensions through heterogeneous data source interfaces to obtain standardized multidimensional features.

[0062] Step S1 further includes:

[0063] S11: A distributed feature acquisition system is obtained by deploying feature acquisition agents on multiple data nodes through a real-time stream processing architecture.

[0064] Furthermore, the distributed feature acquisition system of the present invention establishes a real-time data transmission channel with various heterogeneous data sources through a pre-configured data source interface connector. The data source interface connector is an adapter component designed for different data formats and protocols, responsible for parsing and converting various data formats such as JSON, XML, CSV, and database query results.

[0065] S12: Based on the distributed feature acquisition system, B-end merchant data, C-end user data and product data are collected through the data source interface to obtain multi-source heterogeneous raw data.

[0066] The multi-source heterogeneous raw data in step S12 includes: B-end merchant data, which includes: business category data, geographical location data, user review data, transaction volume data, and inventory status data; C-end user data, which includes: demographic data, historical purchase behavior data, browsing preference data, social network relationship data, and time preference data; and product data, which includes: category attribute data, price range data, brand information data, user rating data, and sales data.

[0067] Specifically, for the collection of B-end merchant data, business category data is obtained through the API interface of the merchant management backend, including first-level category codes, second-level category codes, third-level category codes, and corresponding category name strings. Geographic location data is obtained through GPS coordinate collectors and address parsing services to obtain latitude and longitude coordinates and detailed address text information. User review data is extracted from the review system database, including structured and unstructured mixed data such as rating values, review text content, and review timestamps. Transaction volume data is collected through the financial system interface, including monthly transaction amount, order quantity, average order value, and other numerical indicators. Inventory status data is obtained in real time from the inventory management system, including the current inventory quantity, safety stock threshold, replenishment cycle, and other operational parameters for each SKU.

[0068] For the collection of C-end user data, demographic data includes basic attribute information such as age, gender, occupation, and income level filled in by users during registration. Historical purchase behavior data is obtained through the order system interface, which obtains time-series data such as the user's purchased product ID sequence, purchase time sequence, purchase amount sequence, and purchase frequency statistics. Browsing preference data comes from the user behavior log system, which records the user's page access path, product browsing duration, search keywords, filtering conditions, and other interactive behavior trajectories. Social network relationship data is collected through the social module interface, which collects network structured data such as the user's friend relationship graph, interaction frequency matrix, and influence score. Time preference data is extracted by analyzing the user's active time period distribution, purchase time pattern, seasonal preferences, and other time-dimensional characteristics.

[0069] The collection of product data covers category attribute data, including structured attribute information such as the multi-level classification system to which the product belongs, brand affiliation, specifications, and functional tags. Price range data is obtained from the product management system, including historical price change sequences, current selling price, promotional price, and price sensitivity indicators. Brand information data includes brand feature vectors such as brand awareness rating, brand level classification, and brand user group profile. User rating data is aggregated from the evaluation system to obtain statistical features such as average rating, rating distribution histogram, and rating trend curve. Sales data includes daily sales, monthly sales, quarterly sales, as well as sales performance indicators such as sales growth rate and market share.

[0070] S13: Extract semantic features from the multi-source heterogeneous raw data using NLP technology to obtain standardized multidimensional features.

[0071] In step S13, during NLP semantic feature extraction, the collected text data is first preprocessed using natural language processing, including text cleaning, HTML tag removal, special character filtering, and stop word removal. Then, a word segmentation algorithm is applied to segment the Chinese text according to semantic units. A hybrid method combining dictionary matching and statistical learning is used to identify word boundaries and perform part-of-speech tagging. For word vectorization, a pre-trained Word2Vec model is used to convert the segmentation results into high-dimensional dense vector representations. The Word2Vec model learns distributed representations of words through a Skip-gram architecture, mapping each word to a 300-dimensional real-valued vector, where each dimension represents the projection strength of the word in a latent semantic space. Sentence-level semantic representations are encoded using a BERT pre-trained language model. The BERT model employs a Transformer architecture bidirectional encoder, using a self-attention mechanism to calculate semantic association weights between words, encoding variable-length text sequences into fixed-dimensional semantic vector representations. The sentiment analysis module employs a deep learning-based sentiment classifier to determine the sentiment polarity of user review texts. The classifier uses a combination of convolutional neural networks and recurrent neural networks to extract sentiment features from the text, outputting positive, negative, and neutral sentiment labels and their corresponding confidence scores. The topic extraction algorithm uses a Latent Dirichlet Allocation (LDA) model to model the topics of merchant descriptions and product introductions. The LDA model assumes that each document is composed of multiple latent topics mixed in a certain proportion. The Gibbs sampling algorithm infers the topic distribution and vocabulary distribution of the topics, thereby extracting representative topic keywords and topic weights. The entity recognition module uses a Named Entity Recognition (NER) algorithm to extract structured entity information such as product names, brand names, and specifications from the text. The NER algorithm, based on a Conditional Random Field (CRF) model, combines character-level and vocabulary-level features for sequence labeling, using the BIO labeling system to identify entity boundaries and types. Finally, through the above semantic feature extraction process, the original unstructured text data is transformed into structured numerical feature vectors, providing high-quality semantic representation input for subsequent collaborative recommendation algorithms.

[0072] S2: Dynamic weight learning is performed on the standardized multidimensional feature dataset through a multi-head self-attention mechanism to obtain a fused feature vector.

[0073] Step S2 further includes:

[0074] S21: Nonlinear interaction modeling of different dimensions of features in standardized multidimensional features is performed by a feature interaction encoder to obtain feature interaction representation.

[0075] In step S21 of this invention, a standardized multidimensional feature dataset is first received as input through a feature interaction encoder. This dataset includes feature vectors of B-end merchants, C-end users, and product features. After input, a bilinear cross function is used to perform pairwise interaction calculations on features of different dimensions. Specifically, the interaction between B-end merchant features and C-end user features is calculated using a first weight matrix to obtain the interaction result. Similarly, the interaction between merchant features and product features is calculated using a second weight matrix, and the interaction between user features and product features is calculated using a third weight matrix. Subsequently, each interaction result undergoes a nonlinear transformation using a ReLU activation function, setting negative values ​​to zero while retaining positive values. Next, based on a gating mechanism, a gating weight is calculated using a sigmoid activation function. Finally, the gating weight is multiplied by the activated interaction result to obtain the output feature interaction representation.

[0076] S22: Calculate the importance of features in different dimensions, and optimize the feature weight parameters online based on the feature contribution scores obtained from the importance calculation to obtain dynamic weight parameters.

[0077] In step S22, during the importance calculation and dynamic weight optimization, the feature importance assessment uses a gradient-based method to calculate the contribution of each feature dimension to the final recommendation target. Specifically, for features... Its importance score It is calculated using the absolute value of the gradient, and the expression is: in It is the loss function for the recommendation task. The loss function represents the feature The partial derivatives of .

[0078] After obtaining the scores, the backpropagation algorithm is used to calculate the gradient, starting from the output layer and calculating the gradient layer by layer. The gradient of each layer is passed to the previous layer through the chain rule. The feature contribution score is further quantified using SHAP values. The importance of a feature is determined by calculating its marginal contribution in all feature subsets. The online optimization algorithm uses a stochastic gradient descent method with an adaptive learning rate to update the feature weight parameters. The learning rate adopts the adaptive adjustment strategy of the Adam optimizer. Finally, the dynamic weight parameters are obtained after smoothing by the exponential moving average method.

[0079] S23: Based on the dynamic weight parameters and the feature interaction representation, the different dimensions of the standardized multidimensional features are reduced and fused to obtain a fused feature vector.

[0080] In the dimensionality reduction and fusion process of step S23, the multi-head self-attention mechanism first processes the feature interaction representation and dynamic weight parameters as input in parallel. The multi-head self-attention mechanism contains multiple attention heads, each of which independently calculates the attention weights and output representation. The final output is obtained by concatenating the outputs of all attention heads and performing a linear transformation. Subsequently, the output of the multi-head attention is processed through residual connections and layer normalization to obtain the final output. The dimensionality reduction process uses principal component analysis to map high-dimensional feature vectors to low-dimensional space. The PCA algorithm first calculates the covariance matrix of the feature matrix, then obtains the eigenvectors and eigenvalues ​​through eigenvalue decomposition. The eigenvectors corresponding to the k largest eigenvalues ​​are selected to form the projection matrix. The dimensionality-reduced feature vectors are obtained by multiplying the covariance matrix and the projection matrix. Finally, a weighted average method is used for fusion to obtain the fused feature vector.

[0081] S3: Based on tensor decomposition, a three-layer collaborative matrix is ​​constructed from the fused feature vector to obtain a multidimensional factor matrix.

[0082] Step S3 further includes:

[0083] S31: Based on the fused feature vector, perform tensor modeling on the user-product-merchant relationship triples to obtain the collaborative tensor.

[0084] Furthermore, in step S31, the present invention constructs a three-dimensional relationship tensor of user-product-merchant based on the fused feature vector obtained in the previous steps. The fused feature vector includes user-dimensional feature vectors, product-dimensional feature vectors, and merchant-dimensional feature vectors. The tensor modeling process involves mapping these feature vectors into a three-dimensional space to form a collaborative tensor.

[0085] Specifically, in the tensor modeling process, user a's interaction records with product c at merchant b are first extracted, including explicit and implicit feedback information such as purchase behavior, browsing duration, and rating data. Then, these interaction data are associated with the corresponding fused feature vectors to generate triples. The feature representation is then processed through a feature mapping function. Calculate the tensor value of the triple, where For user-dimensional feature vectors, For product dimension feature vectors, For merchant-level feature vectors, This represents the tensor product operation, and the final constructed co-current tensor. The system records the intensity of user a's preference for product c at merchant b, forming a complete three-dimensional representation of the interaction relationship.

[0086] S32: Using the improved CP decomposition algorithm, latent factors are extracted from the fourth-order cooperative tensor to obtain the initial factor results.

[0087] In step S32 of this invention, the cooperative tensor is decomposed using an improved Canonical Polyadic Decomposition (CP) algorithm to extract latent factors. Specifically, the improved CP decomposition data processing procedure involves first initializing the user factor matrix. Commodity factor matrix Merchant factor matrix Initial values ​​are set using a Gaussian random initialization method, and then each factor matrix is ​​updated through alternating least squares iterations. The product factor matrix and merchant factor matrix are fixed, and the process is completed by minimizing... renew ,in, The cooperative tensor obtained in step S3, It is the Frobenius norm. This represents the expansion of the tensor in pattern 1. The Khatri-Rao product is represented by this matrix. The remaining factor matrix is ​​updated while fixing the other two matrices in sequence. It should be noted that the improvement in the CP decomposition algorithm of this invention lies in the introduction of a regularization term. To prevent overfitting, among which The regularization parameter is used, and non-negativity constraints are employed to ensure the interpretability of the factor. The iterative process continues until the convergence condition is met. If the initial factor result is obtained, the iteration conditions are as follows: To reconstruct the tensor, This is the convergence threshold.

[0088] S33: Optimize the ranking objective of the relation triples based on the BPR loss function to obtain the preference intensity score.

[0089] Bayesian Personalized Ranking (BPR) optimizes the ranking of relation triples to generate preference strength scores. BPR assumes that users prefer interacted items to uninterviewed items, and optimizes the ranking by maximizing posterior probabilities. The loss function is defined as follows:

[0090] ;

[0091] in, Represents the training sample set, This represents the sigmoid function. Indicates user For goods In merchants Predicted score For users For non-interactive products In merchants Predicted score Indicates model parameters.

[0092] In the data processing of BPR optimization, positive and negative sample pairs are first constructed for each user. Select the products they have interacted with. As a positive sample, randomly sampled non-interactive products. As negative samples, triples are formed, and then the predicted score is calculated. Next, the BPR loss gradient is calculated, and then the parameters are updated through stochastic gradient descent. The learning rate is set to α=0.01. After multiple rounds of iterative training, the model learns the user's preference ranking relationship for different products from different merchants and generates a preference intensity rating matrix. .

[0093] S34: Based on the preference intensity score, the initial factor results are decomposed through parallel matrix decomposition to obtain a multidimensional factor matrix including a user latent factor matrix, a product latent factor matrix, and a merchant latent factor matrix.

[0094] In step S34, based on the preference intensity score, the initial factor results are converted into a multi-dimensional factor matrix through parallel matrix decomposition. Parallel processing distributes the large-scale matrix decomposition task to multiple computing nodes to improve computational efficiency. Specifically, a block matrix decomposition strategy is adopted to divide the user-product-merchant matrix into multiple sub-matrix blocks according to the dimensions, and each computing node processes a specific matrix block.

[0095] Furthermore, in the parallel decomposition, the preference intensity rating matrix is ​​first... The matrix is ​​divided into multiple sub-matrices based on the user dimension. Each sub-matrix contains the user's rating data, and then each computing node independently performs matrix factorization. ,in For the first A matrix of decompositions, Represents the user latent factor submatrix, This represents the joint latent factor matrix of goods and merchants. Each node synchronously updates the factor matrix using an alternating least squares algorithm. Finally, the calculation results of each node are merged to obtain the complete user latent factor matrix. Product latent factor matrix Merchant potential factor matrix .

[0096] Step S34 further includes:

[0097] S341: Select non-interactive relation triples in the cooperative tensor through a negative sampling mechanism to obtain a balanced training sample set.

[0098] Furthermore, in step S341, this invention uses a negative sampling mechanism to select non-interactive relation triples in the collaborative tensor to construct a balanced training sample set. Since the user-product-merchant interaction data is extremely sparse, the number of positive samples (interacted) is far less than the number of negative samples (non-interacted), direct training would cause the model to be biased towards predicting the negative class. Negative sampling balances the ratio of positive to negative samples by intelligently selecting negative samples. During the negative sampling data processing, the number of interacted products for each user and the total number of products for each merchant are first counted. Then, for each positive sample, negative sample products are selected based on the sampling probability. The sampling process prioritizes selecting products similar to the positive samples but not interacted with by the user as negative samples to increase the information content of the negative samples. Five negative samples are generated for each positive sample to form training triples, and finally a balanced training sample set containing the positive sample set and the negative sample set is constructed.

[0099] S342: The user historical behavior in the balanced training sample set is weighted and decayed using a time decay factor to obtain an optimized training sample set.

[0100] The time decay factor applies weight decay to the historical user behavior in the balanced training sample set to generate an optimized training sample set. User preferences change over time, and recent behavior reflects current preferences better than historical behavior. The time decay factor highlights the importance of recent behavior by assigning different weights to interaction data at different times.

[0101] The data flow for time decay processing includes: first, obtaining the timestamp of each interaction record, calculating the difference between the current time and the interaction time, and then applying the exponential decay function to calculate the time weight. ,in The attenuation coefficient is set to 0.1. The time unit is used, and the present invention also considers the influence of interaction frequency. ,in This represents the number of times user a interacts with product b at merchant c. The final sample weight is the product of the time weight and the influence of the interaction frequency. After time decay processing, samples with recent high-frequency interactions receive higher weights, while samples with historical low-frequency interactions receive lower weights, forming an optimized training sample set.

[0102] S343: By optimizing the training sample set, sparsity control is applied to the preference intensity score to obtain an interpretable latent factor.

[0103] In step S343, the present invention optimizes the training sample set to control the sparsity of preference intensity scores and generates interpretable latent factors. The sparsity control reduces the influence of irrelevant features through L1 regularization and feature selection techniques, thereby improving the interpretability and generalization ability of the model. The sparse latent factors can clearly identify the key feature dimensions that affect user preferences.

[0104] The data processing flow for sparsity control is as follows: First, L1 regularization is applied to the preference intensity score matrix S, where the regularization coefficient is 0.01. L1 regularization causes some factor weights to be zero, thus achieving automatic feature selection. Then, a soft threshold operator is used to update the parameters. Next, the factor importance score is calculated, and the top-k factors with the highest scores are selected as key features. Finally, low-importance factors are removed by threshold filtering, and the top 10% of factors are retained. After sparsity control, an interpretable latent factor matrix is ​​generated.

[0105] S344: Based on the alternating least squares method, the interpretable latent factors are iteratively optimized to obtain a multidimensional factor matrix.

[0106] In step S344, the interpretable latent factors are iteratively optimized based on the alternating least squares method to obtain the final multidimensional factor matrix. The alternating least squares (ALS) method transforms the non-convex optimization problem into a convex optimization subproblem by alternately fixing one set of variables and optimizing another set of variables, thus ensuring convergence.

[0107] Specifically, the user factor matrix, product factor matrix, and merchant factor matrix are first initialized using a truncated normal distribution N(0,0.1) for random initialization. Then, an iterative optimization loop is entered. During the iteration, two factor matrices are fixed, and the optimal solution of the other factor matrix is ​​solved. Finally, a convergent multidimensional factor matrix is ​​obtained, including the user latent factor matrix, product latent factor matrix, and merchant latent factor matrix.

[0108] S4: Using a collaborative filtering algorithm based on a causal graph structure, the multidimensional factor matrix is ​​modeled for causal relationships to obtain a deep collaborative network model.

[0109] Step S4 further includes:

[0110] S41: Based on the multidimensional factor matrix, causal structure identification is performed to obtain a relational causal graph.

[0111] Furthermore, in step S41, this invention identifies causal structures based on a multidimensional factor matrix and constructs a causal relationship graph. The causal structure identification aims to discover causal relationships between variables from observed data, rather than merely correlations. Specifically, it first receives user latent factor matrices, product latent factor matrices, and merchant latent factor matrices from previous steps as input data. It extracts feature variables from the factor matrices, using each column of the user factor matrix as a user feature variable, each column of the product factor matrix as a product feature variable, and each column of the merchant factor matrix as a merchant feature variable. Then, it constructs a joint feature matrix by concatenating the three factor matrices row-wise. Next, it calculates the conditional independence test between feature variables using the partial correlation coefficient method, i.e., for variables... and In a given set of conditions In this case, the partial correlation coefficient is calculated using the following expression:

[0112] ;

[0113] in express and In a given set of conditions The partial correlation coefficient, for and The correlation coefficient, for and The correlation coefficient, for and The correlation coefficient was subsequently tested for conditional independence using Fisher's Z-transform.

[0114] ;

[0115] when At that time, it was believed and In a given Given conditional independence, the Peter-Clark algorithm (PC algorithm) is applied to construct the causal graph. This involves initializing a completely undirected graph, progressively removing conditionally independent edges, determining the directionality of the edges, and finally obtaining the relational causal graph. ,in This represents a set of nodes (containing user, product, and merchant feature variables). This represents a set of directed edges (representing causal relationships).

[0116] S42: Identify confounding variables in the causal graph to obtain a set of confounding factors.

[0117] Confounding variables are identified in the causal graph to obtain a set of confounding factors. Confounding variables are third-party variables that simultaneously affect both causal and outcome variables, leading to bias in causal inference. Identifying confounding variables requires finding variables on all possible backdoor and frontdoor paths in the causal graph.

[0118] In the identification of confusing variables, the first step is to traverse all node pairs in the causal graph. ,in Indicates the cause variable. This represents the outcome variable. Then, the backdoor criterion is applied to find confounding variables: for causal paths... → Search all from → Non-causal paths (backdoor paths), these paths cannot be... Starting with the arrow, calculate the backdoor path set BP = {P|P is from...} arrive The path, and P does not use →Start}, for each backdoor path P, identify all nodes on the path as potential confounding variables, and then apply the d-separation criterion to verify the confounding variables: given a set of confounding variables Z, check whether Z is separated. and All backdoor paths. The d-separation judgment rule is: a path P is blocked by the variable set Z if and only if there exists a node V on P such that V∈Z and (R is a chain node or branch node on P) or (R is a collision node on P and R and its descendants are not in Z). Then, the weight score of the confounding variables is calculated, which represents the frequency of the variable appearing in the backdoor path. Finally, important confounding variables are screened, and variables with a weight score greater than the threshold of 0.3 are selected to form the confounding factor set C.

[0119] S43: By using a causal attention network, feature filtering is performed on the set of confusing factors to obtain a causal feature interaction representation.

[0120] By using a causal attention network to filter features from a set of confusing factors, a causal feature interaction representation is obtained. The causal attention network is a neural network architecture that combines attention mechanisms and causal reasoning, and can automatically learn the importance weights of different features in causal relationships.

[0121] In causal attention networks, an input feature matrix is ​​first constructed, where each row represents the feature vector of the confounding variable. Then, a multi-head attention mechanism is designed to generate a query matrix, a key matrix, and a value matrix from the input features F through linear transformation, and the attention weights are calculated. Next, a causal mask matrix is ​​introduced, set according to the causal graph structure: if variable A has a causal influence on variable B, then =1; otherwise, =0. The causal mask is applied, and finally, the causal attention output is calculated. Then, feature transformation is performed through a multilayer perceptron, and finally, feature importance scores are calculated. The top k features with the highest scores are selected to form the causal feature interaction representation.

[0122] S44: The weights of the causal feature interaction representation are adjusted by the counterfactual reasoning module to train the causal attention network and obtain a deep collaborative network model.

[0123] In step S44, the weights of the causal feature interaction representation are adjusted through the counterfactual reasoning module, the causal attention network is trained, and a deep collaborative network model is obtained. Counterfactual reasoning is the core concept of causal reasoning. By comparing the actual events with the hypothetical counterfactual situations, the magnitude of the causal effect is estimated.

[0124] Specifically, a counterfactual generation network is first constructed, with the original features and intervention variables as inputs and counterfactual features as outputs. The counterfactual generation network employs a variational autoencoder, which maps the inputs to a latent space, while the decoder generates counterfactual samples. Then, based on a causal effect estimator, the individual causal effect is calculated using a difference-in-differences method. A neural network parameterizes the expected values ​​to represent the expected outcome under intervention and the expected outcome under control. Next, balanced weights are constructed, and inverse probability weights are calculated for the treatment group and the control group. Finally, a weighted loss function is applied to train the model, resulting in a deep collaborative network model.

[0125] Step S44 further includes:

[0126] S441: The causal intervention calculation is performed on the causal graph of the relationship using the Do-calculus algorithm to obtain the estimated value of the causal effect.

[0127] In step S441, the Do-calculus algorithm is a graph-based causal inference algorithm used to calculate the causal effect under intervention conditions. Specifically, the algorithm inputs a relational causal graph and a causal feature interaction representation. First, it identifies the target causal effect, i.e., the impact of a specific intervention behavior on the outcome variable. For a user's preference prediction task, the target causal effect is the impact of the intervention user feature or product feature on the recommendation result. Further, the algorithm constructs an intervention distribution, fixing certain variables in the original probability distribution as intervention values. It then removes all incoming edges pointing to the intervened variable through graph cutting operations. Next, it applies the Do-calculus rule to transform the probability expression, converting the intervention distribution containing the do operator into a function of the observable distribution. In the specific calculation process, the algorithm traverses all directed paths from the intervention variable to the outcome variable in the relational causal graph, calculating the causal effect contribution of each path. The path causal effect is obtained by multiplying the causal strengths of each edge on the path. The causal strength of the edges is extracted from the causal feature interaction representation. Finally, the causal effects of all paths are summed to obtain an overall causal effect estimate. This causal effect estimate is represented as a scalar, quantifying the expected impact of the intervention on the outcome.

[0128] Specifically, starting with the original cause-effect graph Remove all edges pointing to the intervention variable to obtain the modified graph structure, thus obtaining the post-intervention causal graph. Then, three Do-calculus rules are applied for probability calculation: rule 1 is for inserting / deleting observations, rule 2 is for action / observation swapping, and rule 3 is for inserting / deleting actions. Next, conditional probabilities are calculated using the marginalization formula:

[0129] ;

[0130] in, As an intervention variable, For the outcome variable, Represents a set of confusing variables. The intervention operator is then used, and the adjustment formula is applied, i.e., when a valid set of adjustments exists. At that time, the causal effect is:

[0131] ;

[0132] Finally, the estimated causal effect was obtained through numerical calculation:

[0133] ;

[0134] ATE stands for average causal effect.

[0135] S442: Perform counterfactual reasoning on the estimated causal effect to obtain a counterfactual behavior pattern.

[0136] In step S442, the counterfactual reasoning process constructs a counterfactual scenario based on the causal effect estimate. Counterfactual reasoning answers the question "What would happen if a certain event had not occurred?" and understands the causal relationship by comparing the actual result with the counterfactual result.

[0137] Specifically, the counterfactual behavior patterns are first feature-encoded, with each pattern mapped to a fixed-dimensional vector through an embedding layer. Then, feature alignment is performed. Since the causal feature interaction representations and the counterfactual behavior pattern vectors come from different feature spaces, a unified representation is needed through an alignment network. The feature alignment module employs a shared encoder architecture, incorporating an attention fusion mechanism to calculate cross-attention weights and obtain fused features. After obtaining the fused features, adversarial training is introduced to remove bias, constructing a biased classifier that attempts to predict sensitive attributes (such as user gender, age group, etc.) from the fused features. The main network's goal is to minimize the accuracy of the biased classifier. Finally, adversarial training is implemented through a gradient inversion layer. During backpropagation, the gradient is multiplied by the negative of the adversarial weights and then fed back to the main network. After training, the bias-removed causal features are obtained.

[0138] S443: The counterfactual behavior pattern and the causal feature interaction representation are fused to obtain the debiased causal feature.

[0139] In step S443, the counterfactual behavior pattern and the causal feature interaction representation are first aligned in dimension. A linear transformation maps both features to the same vector space. The counterfactual behavior pattern is then converted into a dense vector representation through an embedding layer. Each user's counterfactual behavior pattern is encoded as a fixed-length feature vector. The gating network receives the aligned features as input and learns the optimal fusion weights. The gating network of this invention contains two parallel fully connected layers, which process the counterfactual behavior pattern features and the causal feature interaction representation respectively, outputting the corresponding gating weights. The gating weights are constrained between 0 and 1 by the sigmoid activation function, representing the contribution level of each feature. The fusion calculation is performed through weighted summation. The biased causal feature is equal to the counterfactual behavior pattern feature multiplied by its gating weight plus the causal feature interaction representation multiplied by its gating weight. Furthermore, the fusion process introduces residual connections, adding the original causal feature interaction representation as a residual term to the final result, enhancing feature expressiveness. The debiased causal features obtained in step S443 are standardized by layer normalization to ensure the stability of the feature distribution. The debiased causal features retain causal relationship information while reducing the impact of confounding variables and selection bias, providing a more reliable feature representation for subsequent model training.

[0140] S444: Based on the debiased causal features, the model parameters of the causal attention network are adjusted to obtain a trained deep collaborative network model.

[0141] The model parameter tuning process employs an end-to-end gradient descent optimization method. Partially causal features are used as training data input to the causal attention network, which outputs predicted user preference scores. During training, a multi-task loss function is first defined, including recommendation accuracy loss, causal consistency loss, and regularization loss. Recommendation accuracy loss uses mean squared error to calculate the difference between predicted and true scores. Causal consistency loss ensures that the model's predictions are consistent with causal inferences, calculated by the difference between predicted and true causal effects. Regularization loss includes L2 weight decay and dropout regularization to prevent overfitting. The optimization algorithm uses the Adam optimizer to dynamically adjust the learning rate and gradient update direction. During training, the algorithm monitors the loss function value on the validation set and stops training early when the loss no longer decreases. Model parameter tuning also includes hyperparameter optimization, selecting optimal learning rate, batch size, hidden layer dimensions, and other hyperparameters through grid search or Bayesian optimization methods. After training, the model's performance is evaluated using a test set, calculating metrics such as recommendation accuracy, recall, and causal effect estimation error. The resulting deep collaborative network model possesses causal reasoning capabilities and can generate bias-corrected personalized recommendation results. The model parameters are saved in binary file format, including the network structure definition and trained weight parameters.

[0142] S5: Real-time recommendation of items to be recommended is performed through the deep collaborative network model to obtain a personalized B2B2C recommendation list.

[0143] Step S5 further includes:

[0144] S51: Perform model quantization on the trained deep collaborative network model to obtain a lightweight recommendation model.

[0145] In step S51, the floating-point parameters of the trained deep collaborative network model are first converted to a low-precision representation. The deep collaborative network model obtained in the preceding steps contains a multi-layer neural network structure, with each layer containing a weight matrix and a bias vector. The original model parameters are stored as 32-bit floating-point numbers. This invention uses a quantization algorithm to compress the precision of each parameter layer in the model, converting the 32-bit floating-point parameters into 8-bit integer parameters. Specifically, the quantization process uses a linear quantization method. First, the maximum and minimum values ​​of the parameters in each layer are counted, and the quantization range and quantization step size are calculated. For each floating-point parameter value, the minimum value is subtracted, and then divided by the quantization step size. The integer quantized value is then obtained. Second, the quantization process also includes quantization-aware training. After quantization, the model is fine-tuned to compensate for the precision loss caused by quantization. Quantization-aware training uses a pass-through estimator gradient propagation method. Forward propagation uses the quantized parameters, and backward propagation uses the original floating-point gradient. Subsequently, this invention also employs structural pruning techniques to delete neuron connections with smaller weights, further compressing the model size. The pruning process calculates the importance score for each connection. The importance score is calculated based on a combination of weight magnitude and gradient information. Connections with importance scores below a threshold are marked as redundant and deleted. The remaining connections are renumbered to form a sparse network structure. Ultimately, the storage space for the quantized model parameters is reduced, and memory usage is significantly decreased.

[0146] S52: Output the predicted score of the item to be recommended through the lightweight recommendation model.

[0147] Furthermore, in step S52, a user query request and a candidate product set are first received as input. The user query request includes a user identifier, query timestamp, geographic location information, and contextual features, while the corresponding candidate product set includes all products to be recommended and their associated merchant information. In the specific prediction process, the lightweight recommendation model obtained in step S51 first encodes the input data by converting user identifiers into user feature vectors, product identifiers into product feature vectors, and merchant identifiers into merchant feature vectors. Subsequently, feature vector extraction is performed. First, a pre-trained feature embedding table is queried, which stores the dense vector representation of each entity. The obtained user feature vectors contain potential user preference factors, product feature vectors contain potential product attribute factors, and merchant feature vectors contain potential merchant business factors. After obtaining vectors with preference factors, the lightweight recommendation model inputs the three types of feature vectors into a causal attention network for interaction modeling. The attention weights between users and products are calculated using a dot product attention mechanism, reflecting the user's level of attention to different products. In the calculation, the user feature vector and product feature vector are multiplied to obtain a similarity score. The similarity score is then normalized using softmax to obtain the attention weight. The merchant feature vector serves as contextual information to adjust the interaction intensity between users and products. A gating mechanism controls the influence of merchant factors. Finally, the model output layer maps the weighted feature representations to prediction scores. Prediction scores constrained between 0 and 1 represent the user's preference for the product; scores closer to 1 indicate a higher degree of preference.

[0148] S53: Generate a personalized B2B2C recommendation list based on the predicted scores.

[0149] In step S53, the predicted scores of all candidate products are first sorted. This invention uses a quick sorting method, arranging the products in descending order of predicted scores. The sorted product list reflects the user's preference order for different products, with products ranking higher. Simultaneously, the personalized B2B2C recommendation list generation of this invention also needs to consider business constraints and diversity requirements. Business constraints include product inventory status, merchant service capabilities, and geographical location restrictions. The inventory constraint filtering process checks the real-time inventory quantity of each product; products with zero inventory are removed from the recommendation list. Geographical location constraints calculate the delivery distance based on the user's and merchant's locations; products outside the delivery range are filtered out. Furthermore, this invention employs a diversity control algorithm to ensure the recommendation list includes products of different categories, avoiding excessive concentration of recommendations in a single category. The diversity algorithm uses the maximum marginal relevance method to increase the diversity between products while maintaining high relevance. Specifically, it first maintains a set of selected products and a set of candidate products. In each iteration, it selects products from the candidate set that have the lowest similarity to the selected products and the highest predicted scores to add to the recommendation list. The similarity calculation is based on product category tags, brand information, price range, and other attribute features. The length of the recommendation list is dynamically adjusted based on the user interface display capabilities and user behavior patterns. The mobile recommendation list is set to 10 products, and the PC recommendation list is set to 20 products. A personalized list is then generated, and the personalized B2B2C recommendation list is output in JSON format, containing fields such as product identifier, product name, merchant information, predicted score, and recommendation reason.

[0150] like Figure 2 As shown, this invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation system, comprising:

[0151] Data collection module 100: Used to collect B-end merchant features, C-end user features and product features from multiple dimensions through heterogeneous data source interfaces to obtain standardized multi-dimensional features;

[0152] Fusion module 200: used to perform dynamic weight learning on the standardized multidimensional feature dataset through a multi-head self-attention mechanism to obtain a fused feature vector;

[0153] Construction module 300: used to construct a three-layer collaborative matrix based on tensor decomposition of the fused feature vector to obtain a multidimensional factor matrix;

[0154] Modeling module 400: used to model the causal relationships of the multidimensional factor matrix using a collaborative filtering algorithm based on a causal graph structure, and obtain a deep collaborative network model;

[0155] The recommendation module 500 is configured as a deep collaborative network model trained by the modeling module 400, and is used to make real-time recommendations through the deep collaborative network model to obtain a personalized B2B2C recommendation list.

[0156] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the multi-dimensional feature-driven B2B2C collaborative recommendation method described in various embodiments or some parts of embodiments.

[0158] This invention provides a multi-dimensional feature-driven B2B2C collaborative recommendation method and system, which significantly improves recommendation accuracy. Through multi-dimensional feature fusion and causal reasoning techniques, the recommendation accuracy is improved compared to traditional collaborative filtering methods, and the cold start problem in recommendation systems is effectively solved. New users and new products can quickly obtain accurate recommendations. In addition, this invention greatly reduces recommendation bias. Through counterfactual reasoning and bias removal processing, the impact of selection bias and popularity bias on recommendation results is reduced, and the interpretability of recommendation results is enhanced. Users can understand the reasons for the recommendations, improving user trust and satisfaction. At the same time, it improves merchant conversion rates. Accurate personalized recommendations increase merchant sales conversion rates by an average of 20-30%, and also improve user experience. The diversity and novelty of recommendations are significantly enhanced, and user activity and retention rates are significantly improved.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-dimensional feature driven B2B2C collaborative recommendation method, characterized in that, The method comprises the following steps: S1: Collecting B-end merchant features, C-end user features and product features through a heterogeneous data source interface to obtain standardized multi-dimensional features; S2: Learning dynamic weights of the standardized multi-dimensional features through a multi-head self-attention mechanism to obtain a fusion feature vector; S3: Constructing three-layer collaborative matrices based on tensor decomposition of the fusion feature vector to obtain a multi-dimensional factor matrix; S4: Modeling causal relationships of the multi-dimensional factor matrix through a collaborative filtering algorithm based on a causal graph structure to obtain a deep collaborative network model; S5: Real-time recommending items to be recommended through the deep collaborative network model to obtain an individual B2B2C recommendation list. 2.The multi-dimensional feature-driven B2B2C collaborative recommendation method of claim 1, wherein, Step S1 further comprises: S11: Deploying feature collection agents on multiple data nodes through a real-time stream processing architecture to obtain a distributed feature collection system; S12: Collecting B-end merchant data, C-end user data and product data through data source interfaces based on the distributed feature collection system to obtain multi-source heterogeneous raw data; S13: Extracting semantic features of the multi-source heterogeneous raw data through NLP technology to obtain standardized multi-dimensional features. 3.The multi-dimensional feature-driven B2B2C collaborative recommendation method of claim 2, wherein, The multi-source heterogeneous raw data in step S12 comprises: B-end merchant data, including business category data, geographic location data, user evaluation data, transaction size data and inventory status data; C-end user data, including demographic data, historical purchase behavior data, browsing preference data, social network relationship data and time preference data; Product data, including category attribute data, price interval data, brand information data, user rating data and sales data.

4. The multi-dimensional feature-driven B2B2C collaborative recommendation method according to claim 1, characterized in that, Step S2 further comprises: S21: Nonlinearly interacting different dimensional features in the standardized multi-dimensional features through a feature interaction encoder to obtain feature interaction representations; S22: Calculating the importance of different dimensional features and obtaining feature contribution scores based on the importance calculation, and online optimizing feature weight parameters to obtain dynamic weight parameters; S23: Reducing and fusing different dimensional features in the standardized multi-dimensional features based on the dynamic weight parameters and the feature interaction representations to obtain a fusion feature vector.

5. The multi-dimensional feature-driven B2B2C collaborative recommendation method according to claim 1, characterized in that, Step S3 further comprises: S31: Tensor modeling of user-product-merchant relationship triples based on the fusion feature vector to obtain a collaborative tensor; S32: Extracting latent factors from the collaborative tensor through an improved CP decomposition algorithm to obtain initial factor results; S33: Optimizing the ranking target of the relationship triples based on a BPR loss function to obtain preference intensity scores; S34: Parallelizing matrix decomposition of the initial factor results based on the preference intensity scores to obtain a multi-dimensional factor matrix comprising user latent factor matrices, product latent factor matrices and merchant latent factor matrices.

6. The multi-dimensional feature-driven B2B2C collaborative recommendation method according to claim 5, characterized in that, Step S34 further comprises: S341: Selecting non-interaction relationship triples in the collaborative tensor through a negative sampling mechanism to obtain a balanced training sample set; S342: weight-decay processing the user historical behavior in the balance training sample set by a time decay factor, to obtain an optimized training sample set; S343: sparse control of the preference intensity score by the optimized training sample set, to obtain an interpretable latent factor; S344: iterative optimization of the interpretable latent factor based on an alternating least squares method, to obtain a multi-dimensional factor matrix.

7. The multi-dimensional feature-driven B2B2C collaborative recommendation method according to claim 1, characterized in that, Step S4 further includes: S41: causal structure identification based on the multi-dimensional factor matrix, to obtain a relationship causal graph; S42: confusion variable identification of the relationship causal graph, to obtain a confusion factor set; S43: feature screening of the confusion factor set by a causal attention network, to obtain a causal feature interaction representation; S44: weight adjustment of the causal feature interaction representation by a counterfactual reasoning module, to train the causal attention network, to obtain a deep collaborative network model.

8. The multi-dimensional feature-driven B2B2C collaborative recommendation method according to claim 7, characterized in that, Step S44 further includes: S441: causal intervention calculation of the relationship causal graph by a Do-calculus algorithm, to obtain a causal effect estimate value; S442: counterfactual reasoning of the causal effect estimate value, to obtain a counterfactual behavior pattern; S443: fusion of the counterfactual behavior pattern and the causal feature interaction representation, to obtain a debiased causal feature; S444: model parameter adjustment of the causal attention network based on the debiased causal feature, to obtain a trained deep collaborative network model. 9.The multi-dimensional feature-driven B2B2C collaborative recommendation method of claim 1, wherein, Step S5 further includes: S51: model quantization of the deep collaborative network model, to obtain a lightweight recommendation model; S52: output of a prediction score of a to-be-recommended item by the lightweight recommendation model; S53: generation of an individualized B2B2C recommendation list according to the prediction score. 10.A multi-dimensional feature driven B2B2C collaborative recommendation system, characterized in that, Comprise: The acquisition module is used for multi-dimensional acquisition of B-end merchant features, C-end user features and commodity features through a heterogeneous data source interface, to obtain standardized multi-dimensional features; The fusion module is used for dynamic weight learning of the standardized multi-dimensional features by a multi-head self-attention mechanism, to obtain a fusion feature vector; The construction module is used for three-layer collaborative matrix construction of the fusion feature vector based on tensor decomposition, to obtain a multi-dimensional factor matrix; The modeling module is used for causal relationship modeling of the multi-dimensional factor matrix by a collaborative filtering algorithm based on a causal graph structure, to obtain a deep collaborative network model; The recommendation module is configured as the deep collaborative network model trained by the modeling module, and is used for real-time recommendation, to obtain an individualized B2B2C recommendation list.

Images