Search

CN-122022952-A - Commodity recommendation method based on fusion of multiple user representations

CN122022952ACN 122022952 ACN122022952 ACN 122022952ACN-122022952-A

Abstract

The invention relates to the technical field of recommendation systems, in particular to a commodity recommendation method based on fusion of multiple user representations. According to the method, multiple heterogeneous user representations such as behavior representations, content representations and social representations of users are firstly obtained, and then representation alignment is carried out through an antagonism fusion network, wherein a multi-view fusion device generates unified fusion representations by using a cross-view attention mechanism, and a view discriminator discriminates view specific features through a three-layer neural network and realizes antagonism training through a gradient inversion layer. Based on the aligned fusion representation, calculating the matching degree of the user and the object by adopting a multi-granularity matching mechanism, projecting the representation to three subspaces, respectively calculating the similarity, and weighting and fusing. And finally introducing diversity constraint when generating a recommendation list, performing weight reduction treatment on the articles in the same category, and controlling category distribution. The method and the device effectively solve the problem of distribution difference in the fusion of the multi-source user representations, and promote recommendation accuracy and diversity.

Inventors

  • WANG XIAOYE
  • XU CHEN

Assignees

  • 天津理工大学

Dates

Publication Date
20260512
Application Date
20260202

Claims (10)

  1. 1. A commodity recommendation method based on fusion of multiple user representations is characterized by comprising the following steps: S1, acquiring various heterogeneous user representations and article representations of a user, wherein the user representations comprise behavior representations based on a user history behavior sequence, content representations based on user generated content and social representations based on a user social network, and the article representations are extracted through an embedded layer based on article attribute characteristics; s2, inputting the multiple heterogeneous user representations into an antagonism fusion network, wherein the network comprises a multi-view fusion device and a view discriminator, the multi-view fusion device calculates the correlation weights among different user representations through a cross-view attention mechanism and generates a unified fusion representation, and meanwhile, the view discriminator discriminates the residual view specific features in the fusion representation through a 3-layer neural network and reversely propagates discrimination loss to the multi-view fusion device through a gradient inversion layer; s3, calculating the matching degree between the user and the article by adopting a multi-granularity matching mechanism based on the fusion user representation and the article representation which are aligned by the antagonism; and S4, generating a personalized commodity recommendation list according to the matching degree sequencing result, and introducing diversity constraint to ensure coverage of the recommendation result.
  2. 2. The commodity recommendation method according to claim 1, wherein the cross-view attention mechanism is specifically implemented by assigning a learnable query vector to each of the behavior representation, the content representation and the social representation, wherein each query vector has a dimension of 64, calculating attention scores between each query vector and key vectors of all user representations, wherein the key vectors are obtained from the original user representations through linear transformation, the attention scores are normalized by a softmax function to obtain weight distribution, and finally weighting and summing the original user representations based on the weight distribution to generate a fused representation having a dimension of 256.
  3. 3. A commodity recommendation method based on multiple user representation fusion according to claim 1, wherein said gradient inversion layer performs a reversal operation on the gradient values during back propagation and multiplies the counter strength coefficient which increases linearly from 0.1 to 1.0 during training, and after the gradient inversion layer is located at the input layer of the view discriminator, forces the fusion representation generated by the multi-view fusion device to gradually lose view-specific features while maintaining discrimination capability of the fusion representation.
  4. 4. The commodity recommendation method based on multiple user representation fusion according to claim 1, wherein the behavior representations are extracted by a time sequence encoder that introduces time interval aware position coding when processing the user behavior sequence, the position coding is calculated based on time differences of adjacent behaviors, a 4-layer transducer module is used to capture time dependence in the behavior sequence, each layer contains 8 attention heads, the hidden layer dimension is 512, and the output is pooled by averaging to obtain the behavior representation.
  5. 5. The commodity recommendation method based on multiple user representation fusion according to claim 1, wherein the content representations are extracted by a text encoder that processes user-generated text content using a 12-layer transform module based on a pre-trained BERT model, inputs text cut-off lengths of 128 tokens, selects text features most relevant to the recommendation task through an attention pooling layer that calculates attention weights using a learnable query vector, and outputs a content representation with dimensions 768.
  6. 6. The commodity recommendation method based on multiple user representation fusion according to claim 1, wherein the social representation is extracted by a graph encoder that performs a 2-hop graph convolution operation on the user's social graph, each hop using an independent graph attention layer that calculates the attention weight of neighbor nodes, aggregates the social features of direct neighbors and second order neighbors, outputs a vector representation with dimension 256, and performs a nonlinear transformation using a ReLU activation function after the graph convolution operation.
  7. 7. The commodity recommendation method based on multi-user representation fusion according to claim 1, wherein the multi-granularity matching mechanism is specifically implemented by projecting the fusion user representation and the object representation into 3 different subspaces respectively, each subspace dimension is 128, projection is implemented through a full-connection layer, cosine similarity is calculated in each subspace, the similarity score is normalized to be between 0 and 1 through a sigmoid function, the final matching degree is a weighted sum of the similarity of each subspace, and the weight is learned from the fusion user representation through another full-connection layer.
  8. 8. The method of claim 1, wherein the specific implementation of the diversity constraint comprises classifying candidate items when generating a recommendation list, calculating an initial duty ratio of each category in the recommendation list, applying a weight reduction coefficient of 0.7 to candidate items belonging to the same item category, and further adjusting a matching degree ranking if the duty ratio of a single category in the recommendation list exceeds 40%, so as to ensure that the recommendation list covers at least 5 different categories.
  9. 9. The commodity recommendation method based on multiple user representation fusion according to claim 1, further comprising a model training stage, wherein a multi-view fusion device, a view discriminator and a matching degree calculation module are simultaneously optimized by using an end-to-end training mode, and the training target comprises 3 loss items, wherein a main loss of a recommendation task uses a cross entropy loss function, a discrimination loss of resistance alignment uses a two-class cross entropy loss function, and a contrast loss of multi-granularity matching uses an interval loss function.
  10. 10. The commodity recommendation method based on multiple user representation fusion according to claim 1, wherein a progressive update strategy is adopted in a deployment stage, user representations are updated every 24 hours in increment, incremental update is calculated in real time through a time sequence encoder based on user latest behavior data, fusion model parameters are updated every 7 days in full quantity, the full quantity update uses data retraining models of the past 30 days, the architecture of an resistance fusion network is kept unchanged in the updating process, and only model weights are adjusted to adapt to changes of data distribution.

Description

Commodity recommendation method based on fusion of multiple user representations Technical Field The invention relates to the technical field of recommendation systems, in particular to a commodity recommendation method based on fusion of multiple user representations. Background With the rapid development of electronic commerce and content platforms, personalized recommendation systems have become a key technology for improving user experience and platform value. The system constructs comprehensive user representation by analyzing various information sources such as historical behaviors, content preferences, social relations and the like of the user, so that the user interests are predicted more accurately and related commodities or contents are recommended. The recommendation method based on the multi-source user representation can understand user preference from different dimensions, has important significance for solving the traditional problems of data sparseness, cold start and the like, and is an important research direction in the field of recommendation systems at present. In the prior art, a recommendation method based on multi-source user representation usually adopts a simple characteristic splicing or weighted average mode to perform representation fusion, and the method ignores the distribution difference and semantic gap between different source user representations. For example, representations extracted from a behavior sequence reflect mainly the immediate interests of the user, while representations extracted from a social network reflect the long-term stability preferences of the user, with significant differences in the distribution characteristics of the two representations in the feature space. Direct simple fusion can lead to feature conflict and semantic confusion in the representation space, and the accuracy of subsequent matching calculation is reduced. In addition, the existing method lacks an effective alignment mechanism in the fusion process, and semantic consistency of different source representations in a unified space cannot be guaranteed. Such defects are particularly evident in scenes where the user behavior data are sparse or the social relationship is absent, and may cause deviation of the recommendation result. In addition, the traditional method generally adopts a single matching degree calculation mode, so that complex association relations between users and articles in different granularities are difficult to capture, and the performance improvement of a recommendation system is limited. Therefore, the commodity recommendation method based on the fusion of multiple user representations is provided for the problems, and the technical problems to be solved are how to overcome distribution differences and semantic inconsistencies among multiple user representations, achieve more effective representation fusion, how to keep recommendation accuracy under the condition of sparse data or partial information loss, and how to understand the matching relationship between the user and the object from a multi-granularity level, so that the overall performance of a recommendation system is improved. Disclosure of Invention In order to overcome the above-mentioned drawbacks of the prior art, an embodiment of the present invention provides a commodity recommendation method based on fusion of multiple user representations, so as to solve the problems set forth in the above-mentioned background art. In order to achieve the purpose, the invention provides a commodity recommendation method based on fusion of multiple user representations, which comprises the following steps: S1, acquiring various heterogeneous user representations and article representations of a user, wherein the user representations comprise behavior representations based on a user history behavior sequence, content representations based on user generated content and social representations based on a user social network, and the article representations are extracted through an embedded layer based on article attribute characteristics; s2, inputting the multiple heterogeneous user representations into an antagonism fusion network, wherein the network comprises a multi-view fusion device and a view discriminator, the multi-view fusion device calculates the correlation weights among different user representations through a cross-view attention mechanism and generates a unified fusion representation, and meanwhile, the view discriminator discriminates the residual view specific features in the fusion representation through a 3-layer neural network and reversely propagates discrimination loss to the multi-view fusion device through a gradient inversion layer; s3, calculating the matching degree between the user and the article by adopting a multi-granularity matching mechanism based on the fusion user representation and the article representation which are aligned by the antagonism; and S4, generating a personalized commo