A recall stage dynamic hybrid recommendation method and system based on two-path model collaborative enhancement

Abstract

The application discloses a recall stage dynamic hybrid recommendation method based on two-way model collaborative enhancement, comprising the following steps: constructing a heterogeneous knowledge graph fusing explicit interaction, attribute association and implicit semantic relationship; taking an item-user interaction subgraph in the heterogeneous knowledge graph as input, and using a graph attention network to capture the structural relationship between the user and the item to generate an embedding vector; using the embedding vector to filter and recall and generate a score of a structural recall path, and screening out an item node with the highest predicted score as a query anchor point; taking the query anchor point as a starting point, performing semantic similarity retrieval in a vector space and multi-hop retrieval between physical entities on the heterogeneous knowledge graph in parallel, mining long-tail items and generating a score of a semantic recall path; introducing a ranking alignment normalization mechanism to map the two-way scores to a unified scale, and through adaptive weighted fusion, reordering and deduplication are performed on candidate items composed of the structural recall path and the semantic recall path to generate a unified recommendation list.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of recommendation systems and artificial intelligence. Specifically, it relates to a dynamic hybrid recommendation method and system for the recall stage based on dual-path model collaborative enhancement. It combines graph neural networks and graph representation learning enhanced by large language models to make full use of unstructured knowledge for recommendation. Background Technology

[0002] Recommendation systems analyze users' historical behavior to predict their preferences and generate personalized item recommendation lists. Mathematically, the recommendation task can be abstracted as a prediction function. Here, U and I represent the user set and item set, respectively, and the output is a score matrix R representing the strength of user preferences. The goal of a recommender system is to maximize the correlation between predicted preferences and users' actual preferences, thereby achieving accurate and stable recommendation results.

[0003] To address the key challenges in the field of recommender systems, researchers have developed a variety of mainstream technical approaches, but several unresolved issues remain.

[0004] Graph neural network-based collaborative learning models model user-item interactions through graph structures. LightGCN simplifies traditional graph convolutional networks, omitting feature transformations and nonlinear activation operations, retaining only neighbor aggregation, thus significantly improving the model's learning efficiency for ID interaction features. SimGCL builds upon this by introducing a graph contrastive learning mechanism, constructing positive and negative sample pairs to enhance the robustness of representations and further alleviate data sparsity and overfitting problems. However, such models rely entirely on explicit interaction data, lacking utilization of item content semantics and external knowledge. This not only limits their performance in cold-start scenarios but also makes them susceptible to noise-induced errors and reduced accuracy when faced with interaction noise due to the lack of semantic constraints.

[0005] Knowledge graph augmentation models improve recommendation performance by introducing external structured knowledge. For example, KGAT integrates a knowledge graph with a user-item interaction graph and uses graph neural networks for joint representation learning, thereby enriching the semantic information of items. However, the knowledge graphs constructed by these methods are usually static and structured, making it difficult to capture the fine-grained semantics and contextual relationships contained in unstructured text. Furthermore, the graph construction process is relatively complex, prone to problems such as entity alignment difficulties, and has relatively limited support for long-tail items not covered in the graph.

[0006] Generative recommendation methods based on large language models leverage the deep semantic understanding capabilities of pre-trained language models for recommendation generation. For example, Chat-Rec transforms candidate sets and user profiles into text prompts for input into a large model for inference; K-Rag Rec converts structured relationships in knowledge graphs into text fragments, using retrieval-enhanced generation mechanisms to improve recommendation accuracy. These methods effectively enhance recall and semantic understanding depth for long-tail items. However, large-model recommendation still faces three challenges: first, the illusion problem, which may generate false or irrelevant recommendation results; second, high inference latency, making it difficult to meet the response requirements of real-time recommendation systems; and third, the lack of effective ID representation mechanisms, making it impossible to accurately model the collaborative filtering strength between IDs mathematically like graph neural networks, and it is limited by the context window length when processing extremely long user behavior sequences.

[0007] In summary, existing recommendation technologies have three major limitations: First, the asymmetry between semantics and structure in information fusion makes it difficult for various methods to effectively coordinate ID-based structural retrieval results with pre-trained model-based semantic retrieval results; second, the failure problem in long-tail scenarios, where performance significantly degrades due to the lack of multi-dimensional cross-validation when data is sparse; and third, the contradiction between inference cost and recommendation quality, where while large models improve recommendation capabilities, they face dual implementation bottlenecks of response latency and illusion suppression in high-concurrency real-time scenarios. Summary of the Invention

[0008] The purpose of this invention is to overcome the challenges of semantic-structural asymmetry, long-tail failure, and inference cost contradictions in existing technologies. It proposes a dynamic hybrid recommendation method and system based on dual-path model collaborative enhancement for the recall stage. The core of this invention lies in constructing a recommendation framework capable of generating and utilizing knowledge graphs from unstructured text and structured interactive data. Through graph attention network training, and semantic recall enhanced by knowledge graphs constructed from large language models, a dual-pathway system of structural and semantic recall is formed. Finally, a recommendation list is generated through a dynamic fusion strategy. This invention, by collaboratively utilizing the structured reasoning capabilities of graph models and the semantic understanding capabilities of large language models, is of great significance for achieving more accurate, richer, and more stable intelligent recommendations.

[0009] The objective of this invention is achieved through the following technical solution: A dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement includes: Constructing a heterogeneous knowledge graph: Based on a predefined two-layer entity pattern, physical entities and semantic entities are extracted from unstructured text using a large language model, and then fused with structured physical entities composed of users and items. After entity alignment and noise cleanup, a heterogeneous knowledge graph that integrates explicit interactions, attribute associations and implicit semantic relationships is constructed. Execute graph attention network-driven structural recall: taking the item-user interaction subgraph in the heterogeneous knowledge graph as input, the graph attention network is used to capture the structural relationship between users and items to generate embedding vectors; the embedding vectors are used for filtering recall and generating scores for structural recall paths, and the item node with the highest prediction score is selected as the query anchor point. Perform knowledge graph-enhanced semantic recall: Starting from the query anchor point, perform semantic similarity retrieval in the vector space and multi-hop retrieval between physical entities on the heterogeneous knowledge graph in parallel, mine long-tail items and generate scores for semantic recall paths; Perform multi-path recall fusion: Receive ranking scores from structural recall paths and semantic recall paths, introduce a ranking alignment normalization mechanism to map the two scores to a unified scale, and then re-rank and deduplicate the candidate items composed of structural recall paths and semantic recall paths through adaptive weighted fusion to generate a unified recommendation list.

[0010] Furthermore, in the step of constructing the heterogeneous knowledge graph: The two-layer entity pattern includes an explicit entity layer and an implicit semantic layer. The explicit entity layer contains structured physical entities such as users and items; the implicit semantic layer contains semantic entities extracted from unstructured text. An adaptive text segmentation technique based on semantic coherence detection is used to divide unstructured text into text blocks using a sliding window; if the semantic similarity between adjacent sentences is less than a set threshold, the sentences are segmented. Calculate the TF-IDF weighted information entropy of the text block and use it to select representative text blocks; The large language model is invoked to extract physical entities and semantic entities, as well as semantic relationships and attribute associations, from representative text blocks, and then the alignment and disambiguation of physical entities and attribute associations are performed.

[0011] Furthermore, the alignment and disambiguation of entities and relations includes: For all entities in the heterogeneous knowledge graph, select any entity, use the Approximate Nearest Neighbor Search (ANN) to generate a set of candidate entities that are similar to it, and calculate the cosine similarity between the entity and the candidate entities. If the entity similarity in the explicit entity layer is greater than the preset explicit entity alignment threshold, then it is aligned after analysis by the large language model. If the entity similarity in the implicit semantic layer is greater than the preset implicit semantic layer alignment threshold, then they are directly merged. For all relations in the heterogeneous knowledge graph, the standard relation in the two-layer entity pattern is selected, and a set of candidate relations that are similar to it is generated by using the Approximate Nearest Neighbor Search (ANN). The semantic similarity between the standard relation and the candidate relations is calculated. If the similarity is greater than the relation normalization threshold, they are directly merged. Otherwise, the large language model determines whether to retain them as new relations.

[0012] Furthermore, in the step of performing graph attention network-driven structural recall: A two-layer graph attention network model is adopted. The first layer uses a 4-head attention mechanism to aggregate first-order neighbor information, and the second layer uses a single-head attention mechanism to aggregate second-hop neighbor information. The inter-layer activation function is ELU. The network model training uses a Bayesian personalized ranking loss function combined with L2 regularization; When providing online services, on the one hand, the item with the highest predicted score is selected from the positive sample items recently interacted with by the user, serving as the query anchor point for the semantic recall path; on the other hand, similar users are retrieved based on user vector similarity, and the historical interaction items of similar users are aggregated and sorted according to the scoring formula. The data are sorted and used as a candidate set for structure recall; among them, On behalf of users Embedded vector, The embedding vector representing item v. Pop(v) represents the popularity coefficient, which represents the number of times item v appears in the user-item interaction subgraph, indicating its popularity.

[0013] Furthermore, in the step of performing knowledge graph-enhanced semantic recall: Vector retrieval channel: Transforms query anchors into query vectors, performs approximate nearest neighbor retrieval to obtain scores. ; Graph propagation channel: On a heterogeneous knowledge graph, multi-hop retrieval between physical entities is performed starting from the query anchor point and following defined relationships. Node degree is used as a penalty to obtain a score. ; The scores from the two channels are linearly weighted and fused to obtain the semantic recall path score. .

[0014] Furthermore, in the step of performing multi-path recall fusion: First, calculate the corresponding normalized score based on the item's ranking in a certain recall list and the total number of items in the list; Next, weight parameters are used to merge the two-way recall, calculate the final recall score, and re-rank all candidate items composed of structural recall paths and semantic recall paths according to the final recall score. After removing duplicates, the final Top-K recommendation list is generated.

[0015] Preferably, the present invention also provides a dynamic hybrid recommendation system for the recall stage based on dual-path model collaborative enhancement, comprising: The heterogeneous knowledge graph construction module is used to extract physical and semantic entities from unstructured text based on a predefined two-layer entity pattern and a large language model. It then merges these entities with structured physical entities composed of users and items. After entity alignment and noise cleaning, it constructs a heterogeneous knowledge graph that integrates explicit interactions, attribute associations, and implicit semantic relationships. The structural recall module is used to take the item-user interaction subgraph in the heterogeneous knowledge graph as input, use a graph attention network to capture the structural relationship between users and items to generate an embedding vector; use the embedding vector to perform filtering recall and generate a score for the structural recall path, and select the item node with the highest prediction score as the query anchor point. The semantic recall module is used to perform semantic similarity retrieval in the vector space and multi-hop retrieval between physical entities on the heterogeneous knowledge graph in parallel, starting from the query anchor point, to mine long-tail items and generate a score for the semantic recall path. The multi-path recall fusion module receives ranking scores from structural recall paths and semantic recall paths. It introduces a ranking alignment normalization mechanism to map the two scores to a unified scale. Then, it uses adaptive weighted fusion to re-rank and de-duplicate the candidate items composed of structural recall paths and semantic recall paths to generate a unified recommendation list.

[0016] Preferably, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement.

[0017] Preferably, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the dynamic hybrid recommendation method for the recall phase based on dual-path model collaborative enhancement.

[0018] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows: 1. Based on a two-layer entity model, this invention utilizes a large language model to extract physical and semantic entities, as well as semantic and attribute relationships, from unstructured comments. It then aligns and disambiguates physical entities and attribute relationships, constructing a heterogeneous knowledge graph entirely offline. This invention pre-computes the high-cost computation of large models, effectively solving the high latency and illusion risks caused by direct reasoning. It provides lightweight, real-time queryable semantic knowledge support for the online stage, standardizes the unstructured text extraction process, and lays a high-quality knowledge foundation.

[0019] 2. Graph Attention Network (GAT) is used to capture the structural relationships between users and items within the item-user interaction subgraph, generating embedding vectors for online collaborative filtering. This achieves accurate modeling of pure collaborative signals, ensuring the core recommendation capabilities and recall efficiency of the recommendation system.

[0020] 3. This invention performs semantic similarity retrieval in the vector space and multi-hop retrieval between physical entities in the heterogeneous knowledge graph for semantic recall in parallel, and introduces node degree as a penalty in the graph propagation channel. This proactively reduces the weight of popular entities, greatly alleviating the long-tail item coverage dilemma faced by recommendation systems, and effectively uncovering long-tail items in heterogeneous knowledge graphs that are marginal but semantically highly matched with user intent.

[0021] 4. This invention introduces a ranking alignment normalization mechanism and an adaptive weighted fusion method (multi-path fusion strategy). This effectively solves the incomparability problem caused by the different sources and dimensions of scores in structural recall and semantic recall (asymmetric decision fusion of semantic and structural information), balancing the accuracy and diversity of recommendations.

[0022] 5. This invention combines the structured path of GAT with the semantic path based on heterogeneous knowledge graphs; it achieves complementary advantages between structural reasoning ability and the semantic understanding ability of large models, improves the model's understanding of semantic information, and effectively improves the accuracy of overall recall results. Attached Figure Description

[0023] Figure 1 The overall process framework diagram of the dynamic hybrid recommendation system in the recall phase of the embodiment.

[0024] Figure 2 This is a schematic diagram of a heterogeneous knowledge graph construction module driven by a large language model.

[0025] Figure 3 This is a schematic diagram of a structured recall module driven by a graph attention network.

[0026] Figure 4 A schematic diagram of a knowledge graph-enhanced semantic recall module.

[0027] Figure 5a This is a schematic diagram of the structure of the user-item interaction subgraph in a specific embodiment.

[0028] Figure 5b This is a schematic diagram of the complete heterogeneous knowledge graph constructed in a specific embodiment. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0030] Example 1 like Figure 1 As shown, this embodiment provides a dynamic hybrid recommendation system for the recall stage based on dual-path model collaborative enhancement, including: The heterogeneous knowledge graph construction module aims to address the high latency and illusion risks associated with direct reasoning from large models by building a unified knowledge foundation to offline perform high-cost computations. Its core is to extract semantic nodes from unstructured comment text using a large language model based on a predefined two-layer entity pattern, and then fuse these nodes with structured physical entities composed of users and items. This process strictly limits semantic extraction to supplementing and associating existing structured entities, thereby reducing the illusion risks associated with arbitrary generation. After entity alignment and noise cleansing, a heterogeneous knowledge graph integrating explicit interactions, attribute associations, and implicit semantic relationships is formed. This graph is built entirely offline, bringing forward high-cost steps such as large model computations, thus providing lightweight, real-time queryable semantic knowledge support for the online phase and resolving the contradiction between reasoning efficiency and quality.

[0031] The structured recall module is a structured representation engine that ensures recommendation capabilities. It takes the user-item interaction subgraph from a heterogeneous knowledge graph as input and utilizes a Graph Attention Network (GAT) to capture high-order structural relationships between users and items, generating more powerful embedding vectors. Offline, GAT training, embedding pre-computation, and indexing are completed. Online, the physical entity embedding vectors generated by GAT are directly used for efficient collaborative filtering recall, selecting the items with the highest prediction scores (i.e., the "feature items" of the user's core preferences) as query anchors to guide semantic recall.

[0032] The semantic recall module is the core of improving long-tail coverage. Starting with the query anchors provided by the structural recall module, it retrieves relevant items in the vector space based on semantic similarity, and performs multi-hop retrieval between physical entities on a heterogeneous knowledge graph. During the retrieval process, a penalty based on node degree is introduced to increase the probability of discovering long-tail items with sparse interactions. The combination of these two methods effectively uncovers long-tail items that are semantically relevant to the user's intent but may be located on the edge of the knowledge graph, generating a semantic recall score.

[0033] The multi-path recall fusion module is crucial for addressing the asymmetry in the fusion of semantic and structural information at the decision-making level, responsible for unifying and balancing the two signals. It simultaneously receives rankings from both the structural recall and semantic recall modules. Addressing the fundamental problem of poor comparability between the two scores due to differences in their sources and scales, this module first introduces a ranking alignment and normalization mechanism, mapping them to a unified scale for fair comparison. Subsequently, through an adaptive weighted fusion method, candidate items composed of structural and semantic recall paths are re-ranked and deduplicated, ultimately generating a unified recommendation list that considers both semantic and structural information.

[0034] Preferably, this embodiment also provides a recommendation method based on the above-described dynamic hybrid recommendation system for the recall phase, including: Offline model building and training: 101. Constructing a Heterogeneous Knowledge Graph: First, the system collects unstructured text and structured data. Second, the system constructs a heterogeneous knowledge graph based on these two types of data: For structured data, physical entities are constructed based on item and user information, and relationships are built using historical interaction records between them, resulting in a user-item interaction subgraph. For unstructured text data, adaptive text segmentation technology based on semantic coherence detection is used to divide long unstructured texts into semantically complete text blocks. Subsequently, the TF-IDF weighted information entropy standard is introduced to select representative text blocks with the largest information content. Based on this, the Large Language Model (LLM) is invoked to extract physical and semantic entities contained in the unstructured text according to the entity requirements in the predefined two-layer entity model. Relationships between entities are constructed based on the preset standard relationships in the two-layer entity model, resulting in an implicit semantic subgraph. Finally, the physical entities extracted from the implicit semantic subgraph are aligned and disambiguated with the physical entities in the user-item interaction subgraph, constructing a unified heterogeneous knowledge graph that integrates explicit interactions, attribute associations, and implicit semantic relationships. After the heterogeneous knowledge graph is constructed, all physical entities are vectorized to build a vector library.

[0035] 102. Training the Graph Attention Network: A multi-layer, multi-head graph attention network (GAT) is built and trained on the user-item interaction subgraph of a heterogeneous knowledge graph. A Bayesian Personalized Ranking (BPR) loss function with L2 regularization is adopted, and the Adam optimizer is used for iterative updates to learn high-quality embedding vectors for all users and items.

[0036] 103. Configure semantic recall path: Utilize a pre-trained text embedding model to vectorize all physical entities in the heterogeneous knowledge graph and construct an efficient Approximate Nearest Neighbor (ANN) index. Simultaneously, pre-define multi-hop retrieval logic between physical entities based on the heterogeneous knowledge graph, ensuring dual preparation of vectors and the knowledge graph for online semantic recall.

[0037] Model predictions during the online phase: 201. Identifying User Profiles and Query Anchor Points: When the online system receives a user request, it first obtains the user's embedding vector based on the user information using an offline-trained graph attention network. The system then calculates the matching score between this embedding vector and all item physical nodes in the user's historical interaction records to identify the current user's core interests and preferences, and selects the item physical nodes with the highest scores as query anchor points for subsequent semantic deep mining.

[0038] 202. Perform Dual-Path Collaborative Recall: The system performs inference in parallel on both recall paths. In the structured path, based on the similarity metric of user embedding vectors, it retrieves the neighboring user group with the closest interests to the current user from the vector library, and extracts items that these neighboring users have historically interacted with but that the current user has not yet encountered as the candidate set for the structured path output. In the semantic path, starting from the query anchor point determined in the previous step, it performs multi-hop retrieval between physical entities based on heterogeneous knowledge graphs and semantic similarity retrieval of item embedding vector space in parallel, mining semantically relevant long-tail items as the candidate set for the semantic path output.

[0039] 203. Dynamic Fusion and Ranking Output: The system summarizes the candidate sets generated by the structured path and semantic path in real time, calls the preset ranking alignment normalization function and fusion weight to perform comprehensive scoring and re-ranking, removes duplicates, generates the final recall recommendation list and outputs it.

[0040] Example 2 This embodiment provides the underlying implementation logic of each module involved in the method of the present invention, and supplements the explanation of the dynamic hybrid recommendation system and method in the recall stage in conjunction with specific applications and data, as follows: 1. A heterogeneous knowledge graph construction module driven by a large language model, such as... Figure 2 As shown.

[0041] 1.1 Data schema definition and preprocessing: First, a two-layer entity schema is defined: an explicit entity layer (containing structured physical entities such as users and items) and an implicit semantic layer (semantic entities extracted from text). Relationships are categorized into three types: interaction relationships between explicit entities, semantic relationships between implicit entities, and attribute associations connecting the two layers of entities. For structured data, physical entities are constructed based on the item and user information, and relationships are built using historical interaction records between them, resulting in a user-item interaction subgraph. For unstructured text, an adaptive text segmentation technique is used to cut the text into an ordered sequence of sentences. Where 's' represents unstructured text, Represents adjacent sentence sequences. k This represents the total number of sentence sequences. A sliding window is used to handle sequences exceeding a length threshold. Adaptive text segmentation is performed on the text. During the segmentation process, if the semantic similarity between adjacent sentences is high... In this case, the similarity is determined using cosine similarity, and the text is finally divided into several text chunks. This represents the threshold for semantic similarity between sentences.

[0042] 1.2 Key Text Block Selection and Knowledge Extraction: The segmented text blocks are deduplicated and merged, and the representative text block with the highest information content is selected from each cluster. The selection criterion for the representative text block is TF-IDF weighted information entropy. The maximum value is calculated using the following formula: ; Here, chunk represents a text block, and w represents a word within that text block. The frequency of word w in a text chunk is represented by TF-IDF, which is a word frequency-inverse document frequency weighting method, reflecting the discriminative power across the entire corpus.

[0043] Subsequently, a large language model is invoked to extract physical and semantic entities, as well as semantic and attribute relationships, from representative text blocks based on a predefined two-layer entity pattern. Simultaneously, a text embedding model is used to convert all physical entities in the graph into vectors and create an index, forming a vector library.

[0044] 1.3 Alignment and Fusion of Entities and Relationships: For all entities in the graph, select any entity, generate a set of candidate entities that are close to it using Approximate Nearest Neighbor Search (ANN), and calculate the cosine similarity between the entity and the candidate entities. If the entity similarity of the explicit entity layer... Then, alignment is performed after analysis by a large language model; if the similarity of entities within the implicit semantic layer... If so, then merge directly. Indicates the explicit entity alignment threshold. Indicates the implicit semantic layer alignment threshold; For all attribute relationships in the graph, a preset standard relationship is selected from the two-layer entity pattern. Approximate nearest neighbor search (ANN) is used to generate a set of candidate relationships that are similar to it. The semantic similarity between the preset standard relationship and the candidate relationships is calculated. If the similarity is... If the relationship is not found, it will be merged directly; otherwise, the large language model will determine whether the relationship has sufficient uniqueness to be retained as a new relationship. Indicates the relation alignment threshold.

[0045] 2. A structure recall module driven by graph attention networks, such as... Figure 3 As shown, this module is implemented based on the user-item interaction subgraph in the constructed heterogeneous knowledge graph.

[0046] 2.1 Network Structure Design and Training: A two-layer graph attention network model is adopted: the first layer uses a 4-head attention mechanism to aggregate first-order neighbor information, and the second layer uses a single-head attention mechanism to aggregate second-hop neighbor information. The inter-layer activation function is ELU. Model training uses the Bayesian personalized ranking loss function. Combined with L2 regularization loss function The formula is as follows: ; ; ; Where u represents a user, i represents a positive sample (item that user u has interacted with), and j represents a negative sample (item that user u has not interacted with). This represents the prediction score for user u for positive sample item i. Let B be the predicted score for user u on negative sample item j, and let B represent the training batch. Let N be the initial embedding matrix for all nodes in the user-item interaction subgraph, where N is the total number of nodes in the user-item interaction subgraph, and D is the preset dimension. The regularization coefficients for graph attention networks; This represents the sigmoid function.

[0047] The Adam optimizer is continuously iterated and updated to learn the embedding vectors of all users and items; 2.2 Online Recall and Anchor Point Generation: When providing online services: On the one hand, select predicted scores from positive sample items that the user has recently interacted with. The highest-value item serves as the query anchor for the semantic recall path. On the other hand, similar users are retrieved based on user embedding vector similarity, and the historical interaction items of these similar users are aggregated, excluding items that the current user has already accessed, and then sorted by score. The items are sorted and used as a candidate set for structure-based recall. On behalf of users Embedded vector, The embedding vector representing item v. Pop(v) represents the popularity coefficient, which indicates the number of times item v appears in the user-item interaction subgraph.

[0048] 3. Knowledge graph-based semantic recall module: such as Figure 4As shown, this module uses query anchor points. Starting with this, semantic recall is implemented.

[0049] 3.1 Dual-channel recall strategy: Vector retrieval channel: Utilizing a pre-trained text embedding model, the query anchor point... The data is converted into vectors, and then approximate nearest neighbor retrieval is performed in a vector database based on heterogeneous knowledge graphs to obtain a score based on the cosine similarity of the item semantic vectors. : On heterogeneous knowledge graphs, from query anchor points Starting from a predetermined point, a multi-hop retrieval of physical entities is performed along established relationships. During this retrieval process, physical entities, based on preset semantic and attribute relationships, travel through 2 to 3 hops, ultimately reaching neighboring physical entities via semantic entities, thereby calculating the corresponding graph propagation channel score. Definition For query anchor points The set reachable by k jumps, with a decay factor of , The ) indicates the degree of the anchor point being queried in the graph. The degree of a physical entity in the graph, the score obtained through the graph propagation channel. The calculation is as follows: ; ; 3.2 Score Fusion: The scores from the two channels are linearly weighted and fused to obtain the final score for the semantic recall path. The dual-channel fusion weight is defined as follows: ; .

[0050] 4. Multi-channel recall fusion module: 4.1 Ranking Normalization: To address the issue of inconsistent dimensions between the two component lists, a ranking-based normalization method is employed. For an item v in a certain recall list, let its ranking in that list be... The list has a total of N elements, and the ranking mapping function parameter weights are... Then the normalized fraction The calculation is as follows: ; Therefore, for item v in the candidate set of the structural recall path, by The scores are sorted, and then the ranking is normalized to obtain the corresponding normalized scores. Similarly, for an item v in a candidate set of paths to be retrieved, by The scores are sorted, and then the ranking is normalized to obtain the corresponding normalized scores. ; 4.2 Weighted Fusion and Output: After ranking and normalizing the two recall lists separately, the fusion weight parameter is used. The two recall routes are merged to obtain the final recall score. : ; according to All candidate items composed of structural recall paths and semantic recall paths are reordered to generate the final Top-K recommendation list.

[0051] Specifically, in the data preprocessing stage, this embodiment uses deepseek-v3.2 as the base for the large model. In the heterogeneous knowledge graph construction stage, the size and length threshold of the sliding window are... Each sentence is set to 5 sentences, with a semantic similarity threshold between sentences. The threshold was set to 0.7, using the BGE-large-en model as the text embedding model, Faiss as the vector database, and HNSW for type retrieval. An explicit entity alignment threshold was applied during the text block merging and graph construction stages. Set to 0.8, implicit semantic layer alignment threshold Set to 0.7, relation alignment threshold It is also set to 0.7. In the structured recall path stage driven by the graph attention network, for the embedding dimension parameter D, 128 dimensions are chosen, and the regularization coefficient is... Choose 0.0005. The popularity coefficient is 0.05. The Adam optimizer was used when training the GAT model with a learning rate of 0.0005. The recall count for this path was set to 100. The decay factor of the graph propagation channel is considered during the semantic recall path phase of knowledge graph augmentation. Choose 0.6 for dual-channel fusion weight. Choose 0.5. Select 150 for this path's recall count. During the multi-path fusion phase, the ranking mapping function parameters... The fusion weight parameter is 0.7. The value was 0.3. During the evaluation process, Recall@50 and NDCG@50 were selected for further evaluation, so the final merged recall candidate set Top-K was 50. The final results are shown in Table 1, and the results of this invention show improvements in all metrics.

[0052] Table 1 Experimental Results In summary, this invention standardizes the process of extracting semantic knowledge from unstructured text and fusing it with structured data, laying a high-quality knowledge foundation for semantic recall. It also designs a dual-path collaborative enhancement mechanism for structured and semantic recall, using GAT to model pure collaborative signals and knowledge graphs and large-scale models to mine semantic signals, allowing both to complement each other's strengths. Finally, a multi-path fusion strategy based on dynamic weights is implemented, effectively solving the problem of inconsistent results from different recall pathways and balancing the accuracy and diversity of recommendations.

[0053] Preferably, in order to further illustrate the specific process of the present invention, a case analysis is conducted here, and representative cases are selected from the MovieLens-100K dataset to reflect the gains of the invention in the recall stage.

[0054] In the offline phase, during the heterogeneous knowledge graph construction module phase, for the films *The Horde* (La Horde) (2009), *Resident Evil* (2002), and *Cloverfield* (2008), three film physical entity nodes and their corresponding user physical entity nodes are constructed on the user-item interaction subgraph, and the corresponding interaction relationships are established. The user-item interaction subgraph is as follows: Figure 5a As shown in the diagram. Blue nodes represent user physical nodes, red nodes represent movie physical nodes, and the relationship between them represents the interaction relationship. Figure 5a As can be seen, The Horde (La Horde) (2009) belongs to long-tail data.

[0055] For unstructured text, after text segmentation and feature text selection; extracting physical and semantic entities, as well as semantic and attribute relationships from representative text blocks; using a vector library; and aligning and fusing all entity and attribute relationships, a complete heterogeneous knowledge graph is obtained. For example... Figure 5b As shown in the diagram. Yellow nodes represent semantic nodes, the relationships between yellow nodes represent semantic relationships, and the relationships between yellow and red nodes represent attribute associations.

[0056] The graph attention network-driven structural recall module is trained on the user-item interaction subgraph in a heterogeneous knowledge graph to obtain the embedding vectors of users and items.

[0057] During the online phase, this embodiment performs recall for user 610. The structure recall module calculates the vector similarity and graph structure of the item vector library to obtain a candidate set of structure recall paths. The movie The Horde (La Horde) (2009) ranked 93rd; Cloverfield (2008) ranked 23rd, and Resident Evil (2002) was selected as the query anchor for the semantic path. .

[0058] The semantic recall module calculated and obtained a candidate set of semantic recall paths. The movie *The Horde (La Horde)* (2009) ranked 10th in this set, and *Cloverfield* ranked 9th overall. In the graph propagation recall process within the semantic recall module, semantic nodes played a crucial role in propagation. Specifically, for the query anchor *Resident Evil* (2002), the search propagated to the movie *Cloverfield* (2008) in a two-hop scenario via (Resident_Evil, has_feature, Bloody) and (Cloverfield, has_feature, Bloody). The graph propagation score was calculated by statistically analyzing the nodes at the two and three hops. After weighted fusion and ranking normalization, *The Horde (La Horde)* (2009) and *Cloverfield* (2008) ranked 35th and 13th overall, respectively. This reflects a significant enhancement in the retrieval capability for long-tail data, while maintaining relatively good performance for popular data.

[0059] Preferably, embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps in the dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement as described in the above embodiments. The electronic device specifically includes the following: Processor, memory, communications interface, and bus; The processor, memory, and communication interface communicate with each other via a bus; the communication interface is used to realize information transmission between server-side devices, metering devices, and user-side devices.

[0060] The processor is used to call the computer program in the memory. When the processor executes the computer program, it implements all the steps in the dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement in the above embodiments.

[0061] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement in the above embodiments.

[0062] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0063] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0064] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set, the instruction means being implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement, characterized in that, include: Constructing a heterogeneous knowledge graph: Based on a predefined two-layer entity pattern, physical entities and semantic entities are extracted from unstructured text using a large language model, and then fused with structured physical entities composed of users and items. After entity alignment and noise cleanup, a heterogeneous knowledge graph that integrates explicit interactions, attribute associations and implicit semantic relationships is constructed. Execute graph attention network-driven structural recall: using the item-user interaction subgraph in the heterogeneous knowledge graph as input, the graph attention network is used to capture the structural relationship between users and items to generate embedding vectors; The embedding vector is used for filtering and recall, and the score of the structural recall path is generated. The item node with the highest predicted score is selected as the query anchor point. Perform knowledge graph-enhanced semantic recall: Starting from the query anchor point, perform semantic similarity retrieval in the vector space and multi-hop retrieval between physical entities on the heterogeneous knowledge graph in parallel, mine long-tail items and generate scores for semantic recall paths; Perform multi-path recall fusion: Receive ranking scores from structural recall paths and semantic recall paths, introduce a ranking alignment normalization mechanism to map the two scores to a unified scale, and then re-rank and deduplicate the candidate items composed of structural recall paths and semantic recall paths through adaptive weighted fusion to generate a unified recommendation list.

2. The dynamic hybrid recommendation method for the recall phase according to claim 1, characterized in that, In the steps of constructing a heterogeneous knowledge graph: The two-layer entity pattern includes an explicit entity layer and an implicit semantic layer. The explicit entity layer contains structured physical entities such as users and items; the implicit semantic layer contains semantic entities extracted from unstructured text. An adaptive text segmentation technique based on semantic coherence detection is used to divide unstructured text into text blocks using a sliding window; if the semantic similarity between adjacent sentences is less than a set threshold, the sentences are segmented. Calculate the TF-IDF weighted information entropy of the text block and use it to select representative text blocks; The large language model is invoked to extract physical entities and semantic entities, as well as semantic relationships and attribute associations, from representative text blocks, and then the alignment and disambiguation of physical entities and attribute associations are performed.

3. The dynamic hybrid recommendation method for the recall phase according to claim 2, characterized in that, The alignment and disambiguation of entities and relations include: For all entities in the heterogeneous knowledge graph, select any entity, use the Approximate Nearest Neighbor Search (ANN) to generate a set of candidate entities that are similar to it, and calculate the cosine similarity between the entity and the candidate entities. If the entity similarity in the explicit entity layer is greater than the preset explicit entity alignment threshold, then it is aligned after analysis by the large language model. If the entity similarity in the implicit semantic layer is greater than the preset implicit semantic layer alignment threshold, then they are directly merged. For all relations in the heterogeneous knowledge graph, the standard relation in the two-layer entity pattern is selected, and a set of candidate relations that are similar to it is generated by using the Approximate Nearest Neighbor Search (ANN). The semantic similarity between the standard relation and the candidate relations is calculated. If the similarity is greater than the relation normalization threshold, they are directly merged. Otherwise, the large language model determines whether to retain them as new relations.

4. The dynamic hybrid recommendation method for the recall phase according to claim 1, characterized in that, In the step of structural recall driven by the execution graph attention network: A two-layer graph attention network model is adopted. The first layer uses a 4-head attention mechanism to aggregate first-order neighbor information, and the second layer uses a single-head attention mechanism to aggregate second-hop neighbor information. The inter-layer activation function is ELU. The network model training uses a Bayesian personalized ranking loss function combined with L2 regularization; When providing online services, on the one hand, the item with the highest prediction score is selected from the positive sample items that the user has recently interacted with, and used as the query anchor point for the semantic recall path; On the other hand, similar users are retrieved based on user vector similarity, and their historical interaction items are aggregated according to a scoring formula. The data are sorted and used as a candidate set for structure recall; among them, On behalf of users Embedded vector, The embedding vector representing item v. Pop(v) represents the popularity coefficient, which indicates the number of times item v appears in the user-item interaction subgraph.

5. The dynamic hybrid recommendation method for the recall phase according to claim 1, characterized in that, In the step of performing knowledge graph-enhanced semantic recall: Vector retrieval channel: Transforms query anchors into query vectors, performs approximate nearest neighbor retrieval to obtain scores. ; Graph propagation channel: On a heterogeneous knowledge graph, multi-hop retrieval between physical entities is performed starting from the query anchor point and following defined relationships. Node degree is used as a penalty to obtain a score. ; The two channel scores are linearly weighted and fused to obtain the semantic recall path score. .

6. The dynamic hybrid recommendation method for the recall phase according to claim 1, characterized in that, In the step of performing multi-path recall fusion: First, calculate the corresponding normalized score based on the item's ranking in a certain recall list and the total number of items in the list; Next, weight parameters are used to merge the two-way recall, calculate the final recall score, and re-rank all candidate items composed of structural recall paths and semantic recall paths according to the final recall score. After removing duplicates, the final Top-K recommendation list is generated.

7. A dynamic hybrid recommendation system for the recall stage based on dual-path model collaborative enhancement, characterized in that, include: The heterogeneous knowledge graph construction module is used to extract physical and semantic entities from unstructured text based on a predefined two-layer entity pattern and a large language model. It then merges these entities with structured physical entities composed of users and items. After entity alignment and noise cleaning, it constructs a heterogeneous knowledge graph that integrates explicit interactions, attribute associations, and implicit semantic relationships. The structural recall module is used to capture the structural relationship between users and items using a graph attention network to generate an embedding vector, taking the item-user interaction subgraph in the heterogeneous knowledge graph as input. The embedding vector is used for filtering and recall, and the score of the structural recall path is generated. The item node with the highest predicted score is selected as the query anchor point. The semantic recall module is used to perform semantic similarity retrieval in the vector space and multi-hop retrieval between physical entities on the heterogeneous knowledge graph in parallel, starting from the query anchor point, to mine long-tail items and generate a score for the semantic recall path. The multi-path recall fusion module receives ranking scores from structural recall paths and semantic recall paths. It introduces a ranking alignment normalization mechanism to map the two scores to a unified scale. Then, it uses adaptive weighted fusion to re-rank and de-duplicate the candidate items composed of structural recall paths and semantic recall paths to generate a unified recommendation list.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the dynamic hybrid recommendation method for the recall stage based on dual-path model collaborative enhancement as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the dynamic hybrid recommendation method for the recall phase based on dual-path model collaborative enhancement as described in any one of claims 1 to 6.

Images