Search

CN-121981803-A - Commodity recommendation system and method based on deep learning

CN121981803ACN 121981803 ACN121981803 ACN 121981803ACN-121981803-A

Abstract

The invention relates to the technical field of electronic commerce recommendation systems and discloses a commodity recommendation system and method based on deep learning. The method comprises the steps of obtaining an interactive behavior sequence of a target user, carrying out framing slicing to generate a behavior period data frame set, extracting multidimensional interactive features, extracting entities and relations from a commodity knowledge graph, carrying out association matching between the features and the entities to construct a dynamic preference subgraph, analyzing the subgraph by utilizing a multi-mode fusion neural network, outputting an implicit intention vector, carrying out multi-hop reasoning in the knowledge graph based on the vector, retrieving to generate a candidate commodity pool, calculating potential association strength between commodities and the user through cross-domain migration learning, and sequencing to generate a final recommendation list. According to the method and the device, the short-term preference dynamics of the user can be described in fine granularity, and the implicit intention of the user is deeply excavated through dynamic knowledge association and fusion analysis, so that the accuracy and individuation degree of the recommendation result are effectively improved.

Inventors

  • ZHANG YING

Assignees

  • 深圳巨为科技开发有限公司

Dates

Publication Date
20260505
Application Date
20260123

Claims (10)

  1. 1. The commodity recommendation method based on deep learning is characterized by comprising the following steps: Acquiring an interactive behavior sequence of a target user in a preset time period, framing and slicing the interactive behavior sequence to generate a behavior time period data frame set, and extracting features of the behavior time period data frame set to obtain multidimensional interactive features; extracting commodity entities and entity relations thereof from the commodity knowledge graph, carrying out association matching on multidimensional interaction features and the commodity entities, and constructing a dynamic preference subgraph; analyzing the dynamic preference subgraph by utilizing a multi-modal fusion neural network, and outputting an implicit intention vector; based on the implicit intention vector, performing multi-hop reasoning in the commodity knowledge graph, and searching to generate a candidate commodity pool; And performing cross-domain transfer learning on the commodities in the candidate commodity pool, calculating potential association strength between the commodities and the target user, and generating a final commodity recommendation list according to the potential association strength.
  2. 2. The deep learning-based commodity recommendation method according to claim 1, wherein the obtaining the interactive behavior sequence of the target user in the preset time period, framing and slicing the interactive behavior sequence, generating a behavior time period data frame set, and extracting features of the behavior time period data frame set to obtain multidimensional interactive features comprises: Continuously recording browsing, clicking, collecting and adding shopping carts and purchasing behaviors of a target user and a commodity system, and sequencing according to time stamps of behavior occurrence to form an initial interaction behavior sequence; Dividing an initial interactive behavior sequence by adopting a sliding time window, dividing continuous interactive behaviors into a plurality of non-overlapping behavior time period data frames, and forming a behavior time period data frame set; and respectively extracting the behavior frequency, the behavior type distribution, the behavior duration and the behavior interval characteristics of each behavior time period data frame in the behavior time period data frame set, and splicing and normalizing the characteristics to generate the multidimensional interaction characteristics.
  3. 3. The deep learning-based commodity recommendation method according to claim 2, wherein the extracting commodity entity and entity relation thereof from the commodity knowledge graph, and performing association matching on multi-dimensional interaction characteristics and commodity entity, and constructing a dynamic preference subgraph, comprises: Acquiring commodities, brands, categories, attributes and user entity nodes from a pre-constructed commodity knowledge graph, and acquiring attribution, synonymy, upper and lower positions and co-occurrence relation edges among the entity nodes; analyzing the commodity identification interacted by the target user from the multidimensional interaction characteristics, and positioning the corresponding commodity entity node in the commodity knowledge graph; Carrying out neighborhood expansion along entity relation sides in a knowledge graph by taking the positioned commodity entity nodes as cores to acquire all associated entities and relations in a one-hop range; And commonly constructing the associated entity, the relation, the target user entity node and the behavior pattern characteristics contained in the multidimensional interactive characteristics into the dynamic preference subgraph centering on the target user.
  4. 4. The deep learning-based commodity recommendation method according to claim 3, wherein said analyzing the dynamic preference subgraph by using the multi-modal fusion neural network, outputting the implicit intention vector, comprises: Constructing a multi-mode fusion neural network comprising graph convolution network branches and time sequence attention network branches, wherein the graph convolution network branches are used for processing topological structure information of dynamic preference subgraphs, and the time sequence attention network branches are used for processing behavior time sequence information contained in the dynamic preference subgraphs; Inputting the dynamic preference subgraph into a multi-modal fusion neural network, and carrying out feature propagation and aggregation on entity nodes and relationship edges in the subgraph by graph convolution network branches to generate a graph structure representation vector; the time sequence attention network branch models the time dependency relationship of the multidimensional interactive features in the behavior time period data frame set to generate a time sequence behavior characterization vector; and carrying out feature level fusion on the graph structure representation vector and the time sequence behavior representation vector, and obtaining the implicit intention vector through multi-layer perceptron coding.
  5. 5. The deep learning based commodity recommendation method according to claim 4, wherein said performing multi-hop reasoning in a commodity knowledge graph based on implicit intent vectors, retrieving and generating a candidate commodity pool comprises: mapping the implicit intention vector into a query vector in the commodity knowledge graph space; In the commodity knowledge graph, taking commodity entity nodes interacted by a target user as a starting point, guiding by query vectors, carrying out multi-jump walking along relation edges, and calculating semantic similarity between each commodity entity node and the query vectors in the graph; Setting a similarity threshold, and screening all commodity entity nodes with semantic similarity higher than the similarity threshold to form a preliminary candidate set; And filtering commodity entity nodes in the preliminary candidate set by combining the recent popularity and the inventory state of the commodity entity nodes to form the candidate commodity pool.
  6. 6. The deep learning-based commodity recommendation method according to claim 5, wherein the performing cross-domain transfer learning on the commodities in the candidate commodity pool, calculating the potential association strength between the commodities and the target user, includes: in the source field, pre-training a depth matching model by using large-scale user commodity interaction data, wherein the depth matching model learns a general mapping relation from user characteristics and commodity characteristics to matching scores; Acquiring long-term portrait features of a target user and multi-modal features of commodities in a candidate commodity pool, wherein the long-term portrait features are formed by historical stable preferences of the target user, and the multi-modal features of the commodities comprise text description features, visual image features and knowledge map embedding features; Inputting the long-term portrait features of the target user and the multi-modal features of the commodities in the candidate commodity pool into a pre-trained depth matching model; The depth matching model is finely adjusted by utilizing a small amount of real-time interaction data of a target user in the target field, so that the depth matching model is adapted to the characteristic distribution of the target field; and calculating the potential association strength between each commodity in the candidate commodity pool and the target user through the finely-adjusted depth matching model.
  7. 7. The deep learning based commodity recommendation method according to claim 6, wherein said generating a final commodity recommendation list according to the potential correlation strength ranking comprises: the potential association strengths obtained by calculation of all commodities in the candidate commodity pool are arranged in a descending order; Introducing a diversity control mechanism, and carrying out diversity rebalancing on the goods list with the top ranking, so as to avoid excessive homogenization of the recommendation result; and carrying out final sorting adjustment on the commodity list subjected to diversity rebalancing by combining with a commodity business rule strategy to generate the final commodity recommendation list.
  8. 8. The deep learning based commodity recommendation method according to claim 7, further comprising the incremental updating step of commodity knowledge maps: monitoring the change of online commodity information and the generation of new user interaction behaviors, and extracting newly-added commodity entity, attribute and user behavior triples in real time; fusing the newly added triples with the existing commodity knowledge graph, and performing representation learning on the updated graph by utilizing a knowledge graph embedding algorithm to update vector representation of the entity and the relation; And the commodity knowledge graph after incremental updating is used for a dynamic preference sub-graph construction step in the next round of recommendation tasks.
  9. 9. The deep learning based commodity recommendation method according to claim 1, further comprising the step of enhancing the interpretation of the implicit intent vector: after the implicit intention vector is output by utilizing the multi-mode fusion neural network, adopting an attention backtracking mechanism to analyze attention weight distribution of different characteristics in a graph convolution network branch and a time sequence attention network branch; According to the attention weight distribution, positioning the key behavior characteristic and the knowledge graph entity path which have the greatest contribution to generating the implicit intention vector; and converting the key behavior characteristics and the knowledge map entity path into readable natural language description, and associating the readable natural language description with the final commodity recommendation list as recommendation reasons.
  10. 10. Deep learning based commodity recommendation system comprising a memory, a processor and a computer program stored in the memory and running on the processor, characterized in that the processor, when executing the computer program, realizes the steps of the deep learning based commodity recommendation method according to any one of claims 1 to 9.

Description

Commodity recommendation system and method based on deep learning Technical Field The invention relates to the technical field of electronic commerce recommendation systems, in particular to a commodity recommendation system and method based on deep learning. Background The existing commodity recommendation system widely adopts collaborative filtering, sequence modeling or knowledge graph combining methods. Collaborative filtering methods rely on co-occurrence patterns of user-item interaction matrices, which make it difficult to capture complex evolutions of user interests over time. Sequence models based on recurrent neural networks or self-attention mechanisms are able to handle sequences of user behavior, but often consider interaction recordings as equally spaced discrete event streams, or simply aggregated by sliding windows. This approach lacks efficient modeling capabilities for fine-grained temporal dynamics, such as natural period partitioning, periodic patterns, and short-term interest fluctuations, that exist in user behavior, resulting in inaccurate capture of the user's real-time intent. In the scheme of improving the interpretation and accuracy of recommendation by combining with a knowledge graph, a common practice is to learn static vector representation of commodities and relations by using a graph embedding technology, or to take a knowledge graph sub-graph corresponding to a user history interaction object as a fixed input. The user-knowledge graph correlation constructed by the methods is static or history dependent, and cannot be correlated and pruned in real time and dynamically according to the latest behavior characteristics of the user. The fixed graph structure is difficult to reflect the transfer of the current focus of the user, and the static embedded representation cannot encode the deep and implicit user intention triggered by the instant behavior, so that the capability of carrying out depth and personalized reasoning based on the knowledge graph is limited. A technique is needed that can extract features with more time resolution from a user behavior sequence to accurately characterize preference dynamics, and can construct a correlation with a knowledge graph in real time according to the dynamic features, so as to infer deep intentions that are not explicitly expressed by a user, thereby realizing more accurate personalized recommendation. Disclosure of Invention The invention aims to provide a commodity recommendation system and method based on deep learning, so as to solve the problems in the background technology. In order to achieve the above object, the present invention provides a commodity recommendation method based on deep learning, the method comprising: Acquiring an interactive behavior sequence of a target user in a preset time period, framing and slicing the interactive behavior sequence to generate a behavior time period data frame set, and extracting features of the behavior time period data frame set to obtain multidimensional interactive features; extracting commodity entities and entity relations thereof from the commodity knowledge graph, carrying out association matching on multidimensional interaction features and the commodity entities, and constructing a dynamic preference subgraph; analyzing the dynamic preference subgraph by utilizing a multi-modal fusion neural network, and outputting an implicit intention vector; based on the implicit intention vector, performing multi-hop reasoning in the commodity knowledge graph, and searching to generate a candidate commodity pool; And performing cross-domain transfer learning on the commodities in the candidate commodity pool, calculating potential association strength between the commodities and the target user, and generating a final commodity recommendation list according to the potential association strength. Preferably, the obtaining the interactive behavior sequence of the target user in the preset time period, framing and slicing the interactive behavior sequence, generating a behavior time period data frame set, and extracting features of the behavior time period data frame set to obtain multidimensional interactive features includes: Continuously recording browsing, clicking, collecting and adding shopping carts and purchasing behaviors of a target user and a commodity system, and sequencing according to time stamps of behavior occurrence to form an initial interaction behavior sequence; Dividing an initial interactive behavior sequence by adopting a sliding time window, dividing continuous interactive behaviors into a plurality of non-overlapping behavior time period data frames, and forming a behavior time period data frame set; and respectively extracting the behavior frequency, the behavior type distribution, the behavior duration and the behavior interval characteristics of each behavior time period data frame in the behavior time period data frame set, and splicing and normalizing the charact