Commodity recommendation method, system and related device

Abstract

The invention provides a commodity recommendation method and system and a related device. Relates to the technical field of e-commerce and personalized recommendation, and comprises the following steps: S1, acquiring user-commodity data on an e-commerce platform in real time, and performing cleaning, time synchronization and feature normalization on the user-commodity data to form a user behavior feature set and a commodity attribute feature set; and S2, inputting the user behavior feature set into a fusion network, and generating a user comprehensive interest vector by using a graph neural network and an attention mechanism. According to the commodity recommendation method and system and the related device, hidden interests of the user can be depicted finely, preferences which are not explicitly exposed yet can be mined, the problem that a content-based method is not comprehensive in modeling of a user interest model is fundamentally solved, and therefore the recommendation accuracy and the recommendation individuation degree are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of e-commerce and personalized recommendation technology, specifically to a product recommendation method, system, and related apparatus. Background Technology

[0002] In recent years, the variety of goods and the scale of users on e-commerce platforms have expanded rapidly, and personalized recommendation technology has gradually become a core support for improving user experience and platform operational efficiency. Existing recommendation algorithms mainly include collaborative filtering-based methods, content-based methods, and hybrid recommendation methods that combine both. Collaborative filtering makes recommendations by mining behavioral similarities between users or products, and can effectively capture group preferences; content-based methods rely on product attributes and user profiles to make initial recommendations for new products; hybrid recommendation methods attempt to combine the advantages of both to alleviate their respective limitations. Furthermore, with the continuous emergence of multi-source heterogeneous data such as social media, location, and context, some advanced systems are attempting to integrate users' static preferences with dynamic behaviors and contextual information to improve the real-time performance and personalization of recommendations.

[0003] However, existing technologies still have many shortcomings: content-based methods do not fully characterize user interest models and are difficult to capture potential user preferences; recommendation systems struggle to achieve the best balance between accuracy, diversity, and real-time response, and cannot meet the comprehensive needs of e-commerce platforms for high precision, low latency, and high scalability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a product recommendation method, system, and related apparatus. It solves the problem of how to deeply mine potential user preferences by using multi-source heterogeneous data, fusing graph attention networks and causal Bayesian networks, and achieving high-precision, low-latency, and scalable personalized product recommendations through two-stage hierarchical inverted index retrieval and multi-objective online adaptive sorting.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a product recommendation method, comprising: S1. Acquire user-product data in real time on the e-commerce platform, and perform cleaning, time synchronization and feature normalization on the user-product data to form a user behavior feature set and a product attribute feature set; S2. Input the user behavior feature set into the fusion network, and use graph neural network and attention mechanism to generate a comprehensive user interest vector; S3. Construct a user behavior causal graph with behavioral events such as "browsing → adding to cart → purchasing" and key context variables as nodes. Use a causal Bayesian network to estimate the causal strength between nodes in the causal graph to obtain the potential causal effect vector of users on each product feature. Input the potential causal effect vector and the product attribute feature set into a multi-layer attention graph convolutional network to output a refined potential preference vector to represent the user's interests that have not yet been explicitly exposed. S4. Based on the refined potential preference vector, perform a two-stage hierarchical inverted index retrieval of the product library and generate a candidate product set according to diversity constraints to ensure that the overall retrieval latency is lower than a preset threshold. S5. Calculate a comprehensive score for the candidate product set and output a recommendation sequence. Based on user online feedback, use an incremental learning strategy to update the recommendation sequence in real time to achieve high-precision, low-latency, and scalable personalized recommendations.

[0006] Preferably, the user-product data includes user interaction logs, social relationship data, product metadata, and scenario context data.

[0007] Preferably, the fusion network is a graph attention network with L layers, each layer adopts a multi-head attention mechanism, and residual connections and layer normalization are implemented sequentially after the output of each layer to ensure numerical stability while modeling deep relationships.

[0008] Preferably, in the two-stage hierarchical inverted index retrieval, the first stage uses coarse retrieval based on quantization encoding to reduce the size of the candidate set, and the second stage performs vector nearest neighbor retrieval based on the refined potential preference vector, thereby ensuring that the retrieval latency is lower than a preset threshold.

[0009] Preferably, the candidate product set is scored and ranked using the following model formula: in, Candidate Products The initial correlation score, dimensionless. and Candidate Products With selected products The vector representation, of the same dimension, has been normalized to a unit vector. The diversity penalty coefficient is dimensionless and enhances the diversity of results by weighting and suppressing the high cosine similarity term.

[0010] Preferably, the online incremental update uses a multi-armed gambling machine model to adaptively adjust the weights of the three objectives of accuracy, diversity, and time delay.

[0011] A product recommendation system, comprising: The data acquisition and preprocessing module is used to acquire user-product data in real time and output cleaned, time-synchronized and normalized user behavior feature sets and product attribute feature sets; The feature fusion module is used to input the user behavior feature set into a graph attention network to obtain a comprehensive user interest vector; The latent preference mining module is used to estimate the causal strength based on the user behavior causal graph and the causal Bayesian network, generate a potential causal effect vector, and output a refined latent preference vector by passing it together with the product attribute feature set through an attention graph convolutional network. The candidate retrieval module is used to perform a two-stage hierarchical inverted index retrieval based on the refined potential preference vector to obtain a set of candidate products, and the retrieval latency is not higher than a set threshold. The online ranking update module is used to rank the candidate product set according to a multi-objective weighted model to generate a recommendation sequence, and to update the ranking parameters of the recommendation sequence in real time based on user feedback using a multi-armed gambling machine algorithm, so as to achieve high-precision and low-latency personalized recommendations.

[0012] A product recommendation device, comprising: The interest inference unit is used to collect and normalize user-product data in real time, obtain the comprehensive user interest vector through the graph attention network, construct the user behavior causal graph with events such as "browse → add to cart → purchase", obtain the potential causal effect vector through the causal Bayesian network, and output the refined potential preference vector by combining it with the product attribute features through the attention graph convolutional network. The recommendation generation unit is used to perform a two-stage hierarchical inverted index retrieval based on the refined potential preference vector to obtain a set of candidate products, sort them according to a multi-objective weighted model and output a recommendation sequence, and adjust the sorting weights in real time based on user feedback using a multi-armed gambling machine algorithm to achieve high-precision and low-latency recommendations.

[0013] This invention provides a product recommendation method, system, and related apparatus. It has the following beneficial effects: This product recommendation method, system, and related devices introduce a causal Bayesian network into the user behavior sequence to estimate the causal strength of events such as "browsing → adding to cart → purchasing" and key contextual variables. Then, by combining a multi-layer attention graph convolutional network to deeply integrate the potential causal effect vector with product attribute features, it can accurately characterize users' implicit interests and uncover preferences that have not yet been explicitly revealed. This fundamentally makes up for the problem of incomplete modeling of user interest models in content-based methods, thereby significantly improving the accuracy and personalization of recommendations.

[0014] This invention employs a two-stage hierarchical inverted index retrieval combined with diversity constraints to generate candidate sets in parallel. It also incorporates a multi-objective weighted ranking model and online adaptive weight updates based on a multi-armed gambling machine to achieve a dynamic balance between high accuracy, diversity, and real-time response in the recommendation results. Simultaneously, by distributing the retrieval and online updates to edge computing nodes and introducing quantization encoding technology, it ensures low latency and high concurrency system performance, meeting the comprehensive requirements of e-commerce platforms for scalability and real-time performance. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the process of realizing the invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, this embodiment of the invention provides a product recommendation method, including: S1. Acquiring user-product data in real time on an e-commerce platform, and performing cleaning, time synchronization, and feature normalization on the user-product data to form a user behavior feature set and a product attribute feature set. The user-product data includes user interaction logs, social relationship data, product metadata, and scenario context data.

[0018] S2. Input the user behavior feature set into the fusion network, and generate a comprehensive user interest vector using a graph neural network and attention mechanism. The fusion network is a graph attention network with L layers. Each layer adopts a multi-head attention mechanism, and residual connections and layer normalization are implemented sequentially after the output of each layer to ensure numerical stability while modeling deep relationships.

[0019] The specific implementation method is as follows: In this embodiment, it is assumed that there are 1,000 active users and 500 products on the platform, and approximately 10,000 user-product interaction records are collected. The initial feature of each user node and product node is a 64-dimensional real number vector. The user vector is derived from its historical behavior encoding, and the product vector is derived from the product's attribute and description encoding.

[0020] Network structure settings: The fusion network is set as a 3-layer graph attention network.

[0021] Each layer of the graph attention network has four parallel attention heads.

[0022] Each attention head receives a 64-dimensional vector from the previous layer's output, maps it to a 16-dimensional vector, and then participates in neighbor message aggregation.

[0023] After aggregation, the 16-dimensional outputs of the four heads are spliced ​​together in sequence to restore the 64-dimensional output.

[0024] Residual connectivity and layer normalization: After each layer is aggregated and spliced, the 64-dimensional result is added to the 64-dimensional vector of the input of that layer item by item. Then, the added result is normalized to suppress numerical offset and ensure the stable distribution of the output of each layer.

[0025] Anti-overfitting settings: After calculating the attention at each layer and summing the residuals, 20% of the dimensions are randomly discarded to prevent overfitting and improve the robustness of the model.

[0026] Forward computation process: The first layer receives the original 64-dimensional features of all 1000 users and 500 items, and performs four parallel attention aggregations with their connected neighbor nodes to obtain 1000+500 new 64-dimensional vectors.

[0027] The second and third layers repeat the above operations, using the same number of heads, splicing method, residual and normalization strategies, and always performing 20% ​​discarding after output.

[0028] Finally, 64-dimensional vectors corresponding to 1000 user nodes are extracted from the output of the third layer as the "comprehensive interest vectors" of these 1000 users.

[0029] S3. Construct a user behavior causal graph with behavioral events such as "browsing → adding to cart → purchasing" and key context variables as nodes. Use a causal Bayesian network to estimate the causal strength between nodes in the causal graph to obtain the potential causal effect vector of users on each product feature. Input the potential causal effect vector and the product attribute feature set into a multi-layer attention graph convolutional network to output a refined potential preference vector to represent the user's interests that have not yet been explicitly revealed.

[0030] The specific implementation method is as follows: In a batch of nearly 50,000 complete session data, each session records the sequential events of user browsing, adding to cart, purchasing, etc., as well as their contextual information. The main implementation process is as follows.

[0031] Node and sample preparation: Seven discrete nodes are extracted for each session: browsing, adding to cart, purchasing, dwell time level, discount level, access time segment, and terminal category.

[0032] Dwell time is categorized into three levels: 0-3 seconds, 3-15 seconds, and over 15 seconds. Discounts are categorized into three levels: no discount, ≤10%, and >10%. Visit times are categorized into three levels: 0-8 AM, 8 AM-4 PM, and 4 PM-12 AM. This gives each node 3-4 finite possible values.

[0033] Learning about cause-effect graph structures: A constraint-based PC algorithm is adopted, which restricts possible parent nodes to dwell time levels and access time segments, starting with browsing.

[0034] The final result is a sparse directed acyclic graph with 7 nodes and 14 directed edges.

[0035] Causal Bayesian network parameter estimation: Dirichlet priors are used, with the all-zero prior strength set to 0.1.

[0036] The conditional probability table for each edge in the graph converges after 20 iterations using expectation maximization. The convergence criterion is that the increase in log-likelihood is less than 1 × 10⁻. 4 .

[0037] After convergence, the average treatment effect of the edge "discount level → purchase" was quantified as 0.18, and the average treatment effect of the edge "stay time level → purchase" was quantified as 0.23. The effect share corresponding to each product was extracted in turn and summarized into a 32-dimensional potential causal effect vector.

[0038] Refinement of potential preferences: The 32-dimensional potential causal effect vector is concatenated with the 64-dimensional attribute vector of the product to form a 96-dimensional joint vector.

[0039] The input is fed into a two-layer attention map convolutional network: the first layer has 64 hidden units, the second layer has an output dimension of 64, the drop-out rate is 0.3, and residual and layer normalization are both enabled.

[0040] The network was trained for 10 rounds on 50,000 samples with an Adam learning rate of 0.001. Finally, it outputs a 64-dimensional "refined latent preference vector" for each user-item pair, which serves as the matching basis for subsequent candidate searches.

[0041] S4. Based on the refined potential preference vector, perform a two-stage hierarchical inverted index retrieval of the product database and generate a candidate product set according to diversity constraints to ensure that the overall retrieval latency is lower than the preset threshold.

[0042] In the two-stage hierarchical inverted index retrieval, the first stage uses coarse retrieval based on quantization encoding to reduce the size of the candidate set, and the second stage performs vector nearest neighbor retrieval based on refined latent preference vectors, thereby ensuring that the retrieval latency is lower than the preset threshold.

[0043] The specific implementation method is as follows: In this embodiment, the platform's product database contains approximately 500,000 items. For the aforementioned "refined potential preference vector," a two-stage hierarchical inverted index retrieval method is employed: Phase 1: Coarse retrieval based on quantization encoding.

[0044] Quantization parameters: Each 64-dimensional refined potential preference vector and each commodity vector are divided into 8 subspaces, and each subspace is clustered into 256 centroids to generate 8-byte codes.

[0045] Search process: The distance approximation is calculated for all 500,000 items using PQ codes.

[0046] Result size: Select the top 2000 items with the smallest approximate distance as coarse candidates.

[0047] Performance metrics: This stage takes an average of about 3ms, which is far below the preset 10ms threshold.

[0048] The second stage: fine retrieval based on vector nearest neighbors.

[0049] Search Index: Load the original 64-dimensional vectors corresponding to the 2000 coarse candidate products into the nearest neighbor graph index.

[0050] Search process: The index is searched using the user's 64-dimensional refined latent preference vector, and the top 300 most similar products are retrieved.

[0051] Performance metrics: This stage takes an average of about 4ms, with a total retrieval latency of about 7ms, which is still lower than the preset 10ms.

[0052] Diversity constraints and candidate set generation: Category coverage: The 300 refined search results are grouped into 20 major categories, ensuring that at least 3 products are retained in each major category.

[0053] Final candidate set: Select the top few items that meet the coverage criteria in each group in descending order of similarity, and summarize them to form a candidate set of about 300 items.

[0054] S5. The candidate product set is comprehensively scored and a recommendation sequence is output. Based on online user feedback, the recommendation sequence is updated in real time using an incremental learning strategy to achieve high-precision, low-latency, and scalable personalized recommendations.

[0055] The candidate product set is scored and ranked using the following model formula: in, Candidate Products The initial correlation score, dimensionless. and Candidate Products With selected products The vector representation, of the same dimension, has been normalized to a unit vector. The diversity penalty coefficient is dimensionless and enhances the diversity of results by weighting and suppressing the high cosine similarity term.

[0056] The online incremental update uses a multi-armed gambling machine model to adaptively adjust the weights of the three objectives of accuracy, diversity, and time delay.

[0057] The specific implementation method is as follows: After obtaining 300 candidate products in the previous stage, this embodiment completes the final sorting and online update according to the following process: Initialize correlation: The offline-trained matching model outputs an initial relevance score in the range of 0–1 for each candidate item; for example, the scores for the first five items are 0.92, 0.90, 0.88, 0.86, and 0.85, respectively.

[0058] Diversity penalty rearrangement: Let the penalty coefficient γ = 0.25.

[0059] Greedy iterative rearrangement: In the first round, the product A with the highest score is selected and added to the results list.

[0060] In the second round, the cosine similarity (normalized vector) between the remaining products and A is calculated. For example, if product B has a similarity of 0.8 with A, the corresponding penalty is 0.25 × 0.64 = 0.16, and the overall score of B drops to 0.74.

[0061] Select the highest new overall score (C) and add it to the results list.

[0062] Repeat this process until 40 final recommended products are generated.

[0063] The reordering phase takes an average of 2.1ms on a single 16-core CPU.

[0064] Online incremental updates: Define the three arms as corresponding to three sets of weight vectors: Arm 1: Precision 0.70 / Variety 0.20 / Latency 0.10.

[0065] Arm 2: Precision 0.60 / Variety 0.30 / Latency 0.10.

[0066] Arm 3: Precision 0.55 / Variety 0.35 / Latency 0.10.

[0067] After showing 40 products to the user each time, the revenue per round is calculated based on clicks, category coverage, and actual response time.

[0068] After approximately 10,000 online interactions, the device prioritizes arm 2, with its weights automatically approximating 0.60 / 0.30 / 0.10.

[0069] Compared to the fixed weights before the update (0.70 / 0.20 / 0.10), the overall click-through rate increased by 14%, the category coverage of the output list increased by 9%, and the average response latency remained at 38ms.

[0070] A product recommendation system, comprising: The data acquisition and preprocessing module is used to acquire user-product data in real time and output cleaned, time-synchronized, and normalized user behavior feature sets and product attribute feature sets.

[0071] The feature fusion module is used to input the user behavior feature set into the graph attention network to obtain the user's comprehensive interest vector.

[0072] The latent preference mining module is used to estimate the causal strength based on the causal graph of user behavior and a causal Bayesian network, generate a latent causal effect vector, and output a refined latent preference vector by passing it together with the product attribute feature set through an attention graph convolutional network.

[0073] The candidate retrieval module is used to perform a two-stage hierarchical inverted index retrieval based on the refined potential preference vector to obtain a set of candidate products, and the retrieval latency is not higher than a set threshold.

[0074] The online ranking update module is used to sort the candidate product set according to a multi-objective weighted model to generate a recommendation sequence, and to update the ranking parameters of the recommendation sequence in real time based on user feedback using a multi-armed gambling machine algorithm to achieve high-precision, low-latency personalized recommendations.

[0075] A product recommendation device, comprising: The interest inference unit is used to collect and normalize user-product data in real time, obtain the comprehensive user interest vector through a graph attention network, construct a causal graph of user behavior based on events such as "browsing → adding to cart → purchasing", obtain the potential causal effect vector using a causal Bayesian network, and output a refined potential preference vector by combining it with product attribute features through an attention graph convolutional network.

[0076] The recommendation generation unit is used to perform a two-stage hierarchical inverted index retrieval based on the refined potential preference vector to obtain a set of candidate products, sort them according to a multi-objective weighted model and output a recommendation sequence, and adjust the sorting weights in real time based on user feedback using a multi-armed gambler algorithm to achieve high-precision and low-latency recommendations.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A product recommendation method characterized by comprising: include: S1. Acquire user-product data in real time on the e-commerce platform, and perform cleaning, time synchronization and feature normalization on the user-product data to form a user behavior feature set and a product attribute feature set; S2. Input the user behavior feature set into the fusion network, and use the graph neural network and attention mechanism to generate a comprehensive user interest vector; S3. Construct a user behavior causal graph with behavioral events such as "browsing → adding to cart → purchasing" and key context variables as nodes. Use a causal Bayesian network to estimate the causal strength between nodes in the causal graph to obtain the potential causal effect vector of users on each product feature. Input the potential causal effect vector and the product attribute feature set into a multi-layer attention graph convolutional network to output a refined potential preference vector. S4. Based on the refined potential preference vector, perform a two-stage hierarchical inverted index retrieval of the product library and generate a candidate product set according to diversity constraints; S5. Calculate a comprehensive score for the candidate product set and output a recommendation sequence. Update the recommendation sequence in real time using an incremental learning strategy based on user online feedback.

2. The method of claim 1, wherein: The user-product data includes user interaction logs, social relationship data, product metadata, and scenario context data.

3. The method of claim 1, wherein: The fusion network is a graph attention network with L layers. Each layer adopts a multi-head attention mechanism, and residual connections and layer normalization are performed sequentially after the output of each layer.

4. The method of claim 1, wherein: In the two-stage hierarchical inverted index retrieval, the first stage uses coarse retrieval based on quantization encoding to reduce the size of the candidate set, and the second stage performs vector nearest neighbor retrieval based on the refined potential preference vector.

5. The method of claim 1, wherein: The candidate product set is scored and ranked using the following model formula: wherein, is the initial relevance score of the candidate item, is the initial relevance score of the candidate item, is the vector representation of the selected item, is the diversity penalty coefficient.​​ 6. The method of claim 1, wherein: The online incremental update uses a multi-armed gambling machine model to adaptively adjust the weights of the three objectives of accuracy, diversity, and time delay.

7. A product recommendation system characterized by, include: The data acquisition and preprocessing module is used to acquire user-product data in real time and output cleaned, time-synchronized and normalized user behavior feature sets and product attribute feature sets; The feature fusion module is used to input the user behavior feature set into a graph attention network to obtain a comprehensive user interest vector; The latent preference mining module is used to estimate the causal strength based on the user behavior causal graph and the causal Bayesian network, generate a potential causal effect vector, and output a refined latent preference vector by passing it together with the product attribute feature set through an attention graph convolutional network. The candidate retrieval module is used to perform a two-stage hierarchical inverted index retrieval based on the refined potential preference vector to obtain a set of candidate products, and the retrieval latency is not higher than a set threshold. The online ranking update module is used to rank the candidate product set according to a multi-objective weighted model to generate a recommendation sequence, and to update the ranking parameters of the recommendation sequence in real time based on user feedback using a multi-armed gambling machine algorithm.

8. A commodity recommendation related device characterized by comprising: include: The interest inference unit is used to collect and normalize user-product data in real time, obtain the comprehensive user interest vector through the graph attention network, construct the user behavior causal graph with events such as "browse → add to cart → purchase", obtain the potential causal effect vector through the causal Bayesian network, and output the refined potential preference vector by combining it with the product attribute features through the attention graph convolutional network. The recommendation generation unit is used to perform a two-stage hierarchical inverted index retrieval based on the refined potential preference vector to obtain a set of candidate products, sort them according to a multi-objective weighted model and output a recommendation sequence, and adjust the sorting weights in real time based on user feedback using a multi-armed gambling machine algorithm.

Images