Personalized recommendation method based on semantic and structure double-path fusion and related equipment

Abstract

The embodiment of the invention provides a personalized recommendation method based on semantic and structure double-path fusion and related equipment, and belongs to the field of computer science and technology. The method comprises the following steps: constructing a structured interaction graph of user-article interaction and a semantic association graph based on text semantic similarity; utilizing a graph neural network and a graph attention network to respectively learn collaborative features and semantic features in two channels; a large language model is innovatively introduced as an external knowledge brain, and a personalized fusion supervision signal is generated based on node information; through a self-adaptive gating network supervised by the signal, two-way features are dynamically fused to obtain a final representation; and generating a recommendation list according to the information. According to the method, noise interference is avoided through two-way separation learning, the performance bottleneck of an existing recommendation system in a data sparse and cold start scene is effectively solved through a dynamic fusion mechanism achieved through LLM, and the recommendation accuracy, robustness and interpretability are remarkably improved.

Description

Technical Field

[0001] This application relates to the fields of computer science and technology, and in particular to a personalized recommendation method and related equipment based on semantic and structural dual-path fusion. Background Technology

[0002] With the widespread adoption of the internet and mobile devices, users are exposed to an explosive growth in information every day. Personalized recommendation systems have become a key tool to help users efficiently discover content of interest (such as products, news, videos, and music) from massive amounts of data. Among numerous recommendation algorithms, collaborative filtering is widely used because it does not rely on item content information and can discover user preferences solely through user-item interaction history. In recent years, collaborative filtering models based on graph neural networks have significantly improved recommendation performance by modeling user-item interaction relationships as a bipartite graph and aggregating multi-level neighbor information using operations such as graph convolution, and have become the mainstream paradigm in the current recommendation system field.

[0003] However, these collaborative filtering methods based on graph neural networks still face the following key challenges: 1) Data sparsity and cold start problem: When new users join the system or new items are launched, the lack of sufficient historical interaction records makes it difficult for the model to learn effective feature representations, leading to a significant drop in recommendation performance. 2) Insufficient utilization of semantic information: Traditional methods mainly focus on user behavior patterns, while ignoring the rich textual descriptive information of users and items (such as user profiles, product descriptions, etc.), which is crucial for understanding the user's true intentions and the core attributes of items. 3) Fixed information fusion methods and lack of dynamic adaptability: Existing fusion methods usually adopt simple feature concatenation or fixed attention mechanisms, which cannot adaptively adjust according to the characteristics of different users or items, limiting the model's expressive power and personalization level. 4) Poor noise robustness: Noise in user interaction data and text descriptions is easily contaminated and amplified during the early fusion process, thus affecting the final recommendation quality.

[0004] To address these challenges, the rise of Large Language Models (LLMs) in recent years has brought new opportunities to recommender systems. However, how to effectively integrate the capabilities of LLMs into recommender systems to overcome these challenges remains a research hotspot. Existing LLM-enhanced recommender methods either only use LLMs as feature extractors or employ static alignment strategies, failing to fully leverage the potential of LLMs in dynamic decision-making and knowledge guidance. Summary of the Invention

[0005] The main objective of this application is to propose a personalized recommendation method and related device based on semantic and structural dual-path fusion. This method effectively combines user interaction patterns and textual semantic information. Through an LLM-supervised adaptive gating fusion mechanism, it dynamically adjusts the fusion weights of the two information sources, thereby significantly improving recommendation performance in data-sparse and cold-start scenarios, and enhancing the model's robustness, dynamic adaptability, and interpretability.

[0006] To achieve the above objectives, one aspect of this application proposes a personalized recommendation method based on semantic and structural dual-path fusion, the method comprising: Construct a dual-path graph network: Based on the historical interaction data between users and items, construct a structured interaction graph; based on the text description data of users and items, generate semantic vectors using a pre-trained large language model, and construct a semantic association graph based on semantic similarity; Dual-channel feature learning: Information propagation and aggregation are performed on the structured interaction graph through a graph convolutional network model to learn collaborative feature vectors; weighted information propagation and aggregation are performed on the semantic association graph through a graph attention network model to learn semantic feature vectors based on the semantic associations and attention weights of nodes. Generate a fusion supervision signal: For the target node, construct a structured prompt word, which presents the semantic features and interaction structure features of the node in text form; input the prompt word into a pre-trained large language model, and obtain a value returned by the large language model indicating the relative reliability of the collaborative features and semantic features, and use this value as a supervision signal; Adaptive gated fusion: The collaborative feature vector and the semantic feature vector are concatenated and input into a gated network to obtain a gated vector; the collaborative feature vector and the semantic feature vector are weighted and fused using the gated vector to obtain the final fused feature vector; during training, the gated network is supervised learning using the supervision signal output by the large language model. Recommendation generation: Based on the final fused feature vector, a multi-task loss function is constructed to jointly optimize the model; using the optimized model, the matching score between the user and the candidate items is calculated, and a personalized recommendation list is generated based on the matching score.

[0007] In some embodiments, the specific steps for constructing the semantic association graph include: Calculate the semantic similarity of all user-item pairs to form a semantic similarity matrix; For each user, only the top K items with the highest semantic similarity to them are retained as semantic neighbors; for each item, only the top K users with the highest semantic similarity to them are retained as semantic neighbors. Based on the semantic neighbor relationships, a sparse semantic association graph is constructed.

[0008] In some embodiments, the graph convolutional network model is a lightweight graph convolutional network, which obtains the collaborative feature vector by accumulating the node embedding vectors output by each layer.

[0009] In some embodiments, the gating network is a fully connected neural network, which calculates the gating vector using the following formula. :

[0010] in, For collaborative feature vectors, For semantic feature vectors, This represents a vector concatenation operation. and For learnable parameters, This is the Sigmoid activation function.

[0011] In some embodiments, the adaptive gating fusion yields the final fused feature vector using the following formula. :

[0012] in, This indicates element-wise multiplication.

[0013] In some embodiments, the multi-task loss function in the recommendation generation includes at least: Personalized ranking loss is used to optimize user-item ranking preferences; Gated supervision loss is used to minimize the difference between the gated vector and the supervision signal output by the large language model.

[0014] In some embodiments, the multi-task loss function further includes at least one of the following: Contrastive learning loss is used to encourage the alignment of collaborative and semantic features in the representation space; Knowledge distillation loss is used to distill the knowledge from the semantic feature vectors into the final fused feature vector; L2 regularization loss is used to prevent the model from overfitting.

[0015] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0018] The embodiments of this application include at least the following beneficial effects: This application provides a personalized recommendation method, electronic device, storage medium, and program product based on semantic and structural dual-path fusion. This solution avoids noise interference through dual-path separation learning and utilizes the dynamic fusion mechanism implemented by LLM, effectively solving the performance bottleneck of existing recommendation systems in data sparsity and cold start scenarios, and significantly improving the accuracy, robustness, and interpretability of recommendations. Attached Figure Description

[0019] Figure 1 This is a flowchart of the steps of the personalized recommendation method based on semantic and structural dual-path fusion in the embodiments of this application; Figure 2 This is a flowchart of the personalized recommendation method in the embodiments of this application; Figure 3 This is a flowchart illustrating the personalized recommendation method in an embodiment of this application; Figure 4 This is a schematic diagram of the hardware structure of the electronic device in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0023] 1) Large Language Models (LLMs) are deep learning models trained on large amounts of text data, enabling them to generate natural language text or understand the meaning of language text. These models can provide in-depth knowledge and language production on a wide range of topics by being trained on massive datasets. The core idea is to learn patterns and structures of natural language through large-scale unsupervised training, thus mimicking the human language cognition and generation process to some extent.

[0024] With the widespread adoption of the internet and mobile devices, users are bombarded with an explosive growth of information every day. Personalized recommendation systems have become a key tool to help users efficiently discover content of interest (such as products, news, videos, and music) from massive amounts of data. Among numerous recommendation algorithms, collaborative filtering (CF) is widely used because it does not rely on item content information and can discover user preferences solely through user-item interaction history. In recent years, collaborative filtering models based on graph neural networks (GNNs) have significantly improved recommendation performance by modeling user-item interaction relationships as a bipartite graph and aggregating multi-order neighbor information using operations such as graph convolution, and have become the mainstream paradigm in the current recommendation system field.

[0025] However, these collaborative filtering methods based on graph neural networks still face the following key challenges: 1) Data sparsity and cold start problem: In real-world recommendation scenarios, user-item interaction data is typically very sparse, with the vast majority of users interacting with only a small number of items. When new users join the system or new items are launched, the lack of sufficient historical interaction records makes it difficult for the model to learn effective feature representations, leading to a significant drop in recommendation performance. This is a fundamental flaw commonly found in existing methods.

[0026] 2) Insufficient utilization of semantic information: Traditional collaborative filtering methods primarily focus on user behavior patterns (i.e., structured interaction information), neglecting the rich textual descriptive information of users and items themselves (such as user profiles, interest tags, social media profiles, and product titles, detailed descriptions, categories, and user review summaries). This text contains deep semantic information about users' true intentions and the core attributes of items. For example, a user profile mentioning "likes science fiction movies," or a product description emphasizing "made of organic cotton," is crucial for understanding user preferences and item characteristics. However, existing methods have failed to effectively integrate and utilize this valuable unstructured semantic information.

[0027] 3) Fixed and Lack of Dynamic Adaptability in Information Fusion Methods: Existing methods for fusing structural and semantic information typically employ simple feature concatenation, weighted summation, or fixed attention mechanisms. This "one-size-fits-all" fusion strategy cannot adaptively adjust to the characteristics of different users or items. For example, for popular items with rich interaction records, their collaborative signals are usually more reliable; while for new users with sparse interactions, their semantic information, such as their personal profile, may be more valuable. Fixed fusion strategies limit the model's expressive power and personalization level, making it impossible to achieve personalized decision-making tailored to each individual.

[0028] 4) Poor noise robustness: User interaction data often contains noise, such as accidental clicks and abnormal behavior. Simultaneously, textual descriptions may be inaccurate, incomplete, or even false. In existing fusion frameworks, information from different modalities is mixed early on, which easily leads to noise contamination and amplification during information propagation, thus affecting the final recommendation quality.

[0029] To address the aforementioned challenges, the rise of LLMs in recent years has brought new opportunities to recommender systems. LLMs, pre-trained on massive amounts of text, possess powerful semantic understanding, knowledge reasoning, and context awareness capabilities. However, effectively integrating the capabilities of LLMs into recommender systems to overcome these challenges remains a research hotspot. Existing LLM-enhanced recommender methods either use LLMs merely as feature extractors or employ static alignment strategies, failing to fully leverage the potential of LLMs in dynamic decision-making and knowledge guidance. Therefore, a new technical solution is urgently needed that can deeply integrate structured interactive information with unstructured semantic information. This solution should utilize a dynamic adaptive fusion mechanism supervised by an external knowledge engine (such as LLMs) to effectively address data sparsity and cold-start problems, and significantly improve the robustness and interpretability of the model.

[0030] In view of this, this application provides a personalized recommendation method, electronic device, storage medium, and program product based on semantic and structural dual-path fusion. This solution can effectively combine user interaction behavior patterns (structural information) and text semantic information (semantic information). Through an adaptive gating fusion mechanism supervised by LLM, the fusion weights of the two information sources are dynamically adjusted for each user or item node, thereby significantly improving recommendation performance in data sparsity and cold start scenarios, and enhancing the robustness, dynamic adaptability, and interpretability of the model.

[0031] This application provides a personalized recommendation method based on semantic and structural dual-path fusion, relating to the field of computer science and technology. This personalized recommendation method based on semantic and structural dual-path fusion can be applied to terminals, servers, or software running on either terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing a personalized recommendation method based on semantic and structural dual-path fusion, but is not limited to the above forms.

[0032] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0033] Figure 1 This is an optional flowchart of a personalized recommendation method based on semantic and structural dual-path fusion provided in an embodiment of this application. Figure 1The method may include, but is not limited to, steps S101 to S105.

[0034] Step S101: Construct a dual-path graph network.

[0035] In some embodiments, step S101 aims to prepare the data foundation for subsequent dual-channel learning. First, two types of core data are obtained from the recommender system database: user-item interaction data and text description data. User-item interaction data records the historical interaction behaviors of users and items, such as clicks, purchases, ratings, and browsing duration; this data constitutes a structured user-item relationship. Text description data includes users' personal profiles, interest tags, social information, as well as item titles, detailed descriptions, categories, and user review summaries; this data contains rich unstructured semantic information. Then, a structured interaction graph is constructed: all users and items are considered as nodes in the graph, and the set of nodes is denoted as . ,in For user collection, A collection of items. For users who interact with the content. and items An undirected edge is created between the corresponding nodes. This results in a bipartite graph. This is a structured interaction graph, where an element in the adjacency matrix is ​​1 indicating an interaction between the corresponding user and the item, and 0 otherwise. Next, a semantic association graph is constructed: first, a pre-trained LLM (such as BERT, GPT, etc.) is used to connect each user... and items The text description is encoded as a high-dimensional semantic vector. and Then calculate any user-item pair. The semantic similarity (e.g., cosine similarity) is used to obtain the semantic similarity matrix. To reduce computational complexity and noise interference, only the top-K items with the highest semantic similarity to each user are retained as semantic neighbors, and similarly, the top-K user neighbors are retained for each item. This constructs a sparse semantic association graph. Its adjacency matrix has non-zero values ​​(i.e., the corresponding similarity values) only at the Top-K connections, and the rest of the elements are 0.

[0036] Step S102: Dual-channel feature propagation and learning.

[0037] See Figure 2 and Figure 3 In some embodiments, step S102 designs two parallel and independent information processing channels to process the structured interaction graph and the semantic association graph respectively, in order to learn collaborative features and semantic features. Specifically: 1) Structured Collaboration Channel: Input structured interaction graph A lightweight Graph Convolutional Network (GCN) model is employed. GCN aggregates the features of a node and its first-order neighbors at each layer through a message-passing mechanism. Specifically, in the l-th layer of GCN, the embedding vector of node v is updated by a weighted sum of its own features and those of its first-order neighbors, with the adjacency matrix symmetrically normalized. This operation is concise and efficient, avoiding the overfitting risk associated with nonlinear transformations. To capture multi-order connectivity information, the node embedding vectors output from each layer are accumulated to obtain the final collaborative feature vector, which incorporates multi-order neighbor information.

[0038] 2) Semantic Analysis Channel: Input semantic association graph The algorithm employs a Graph Attention Network (GAT) model. GAT assigns different attention weights to neighboring nodes, thereby more accurately aggregating important information. Specifically, firstly, a multilayer perceptron (MLP) is used to map the text semantic vector to the same dimension as the collaborative features, serving as the initial embedding input to GAT. Then, in the l-th layer of GAT, attention weights are assigned to nodes... and its semantic neighbors Calculate the attention coefficient, and then aggregate the neighbor features based on the attention coefficient to obtain the node. The Layer embeddings are generated, with activation functions (such as LeakyReLU) used to produce nonlinear transformations. Finally, the output embeddings of each layer of GAT are aggregated (e.g., element-wise summation or taking the output of the last layer) to obtain the semantic feature vector H_S of the node. Through the attention mechanism, GAT can effectively mine heterogeneous information in semantic relationships and suppress noise.

[0039] Step S103: Generate LLM supervised fusion signal.

[0040] See Figure 2 and Figure 3 Step S103 is one of the core innovations of this embodiment, aiming to introduce an external "knowledge brain"—LLM—to provide dynamic and personalized monitoring signals for information fusion. Specifically, it is as follows: 1) Prompt Construction: For each node to be processed (user or item), design a structured prompt template to integrate two types of key information about the node: text features and interaction structure features. The text features are the node's original text description, while the interaction structure features can be summarized by statistically analyzing the node's historical interactions, for example: "This user has interacted with..." The items are mainly categorized as follows: and "or "the item was [Number] users interacted with the service, and the average rating was [Rating]. ".

[0041] 2) LLM Inference and Signal Generation: The filled-in prompts are input into the pre-trained LLM, which evaluates the relative reliability of the collaborative and semantic information of the node based on the prompt content and outputs a value y between 0 and 1. The closer the value is to 1, the more reliable the collaborative information; the closer it is to 0, the more reliable the semantic information. To reduce computational costs, this step can be performed offline in batches before model training, and the generated supervision signal y can be cached for subsequent training. For example, the LLM can be asked the following question: "Given a user profile and interaction summary, assign a weight between 0 and 1 to the importance of the collaborative information (0 indicates complete reliance on semantic information, 1 indicates complete reliance on collaborative information)." The LLM's output value serves as the supervision signal y.

[0042] Step S104: Adaptive gating fusion.

[0043] In some embodiments, step S104 utilizes the LLM supervision signal generated in step S103 to guide the gated network in learning how to adaptively fuse features from both channels. Specifically: 1) Gated Network Design: A simple fully connected neural network is used as the gated network, with the input being the collaborative feature vector. With semantic feature vector spliced ​​vector The output is a gated vector with the same dimension as the features. Specifically, gating vector ,in and For learnable parameters, For the Sigmoid function, ensure Each element is in Within the range.

[0044] 2) Adaptive fusion: utilizing gated vectors The dual-channel features are weighted and fused to obtain the final fused feature vector. ,Right now , where ⊙ denotes element-wise multiplication. This fusion formula allows the model to independently determine the contribution ratio of collaborative and semantic information in each feature dimension.

[0045] 3) LLM supervised learning: During training, the gated vector output by the gated network... It will be related to the supervision signal generated by LLM. By comparing the two, the learning of the gating network is guided by minimizing the mean square error between them, so that it gradually approaches the judgment criteria of LLM.

[0046] Step S105: Joint optimization of models and generation of recommendations.

[0047] In some embodiments, step S105 defines a loss function for multi-task joint optimization and trains the entire model in an end-to-end manner. Specifically: 1) Loss function design: Total loss function It consists of several parts, including: personalized ranking loss (such as Bayesian personalized ranking loss, BPR), used to optimize user-item ranking preferences; and gated supervision loss, used to guide the gated network to learn the supervision signals provided by the LLM, which can take the form of a gated vector. With signal Mean squared error between them; contrastive learning loss, used to encourage collaborative features. and semantic features Alignment in the representation space ensures consistency in dual-channel learning; knowledge distillation loss distills the knowledge of high-dimensional semantic embeddings into the final fused embedding, which is particularly beneficial for cold-start scenarios; and L2 regularization loss prevents model overfitting.

[0048] 2) Joint optimization: The total loss function is obtained by weighting the losses of each part proportionally, and the total loss is minimized by the backpropagation algorithm, thereby optimizing all parameters of the model end-to-end.

[0049] 3) Recommendation generation: After the model is trained, for a given user and candidate items, calculate the inner product score of the final fused features of each user-item pair, and generate a Top-N recommendation list for the user based on the score.

[0050] The following three specific embodiments illustrate in detail the implementation of the personalized recommendation method based on semantic and structural dual-path fusion in different application scenarios. Each embodiment addresses an e-commerce platform recommendation scenario, a news application distribution scenario, and an enterprise SaaS function guidance scenario, respectively, and sequentially introduces the business background and challenges, data acquisition and preprocessing, graph construction methods (collaborative graph and semantic graph), model training and optimization process, and online deployment and effect verification.

[0051] Example 1: Recommendation by an e-commerce platform (1) Business background and challenges This embodiment addresses a personalized product recommendation scenario on a large e-commerce platform. In this scenario, the platform possesses massive amounts of user and product data, with frequent user behaviors such as browsing, clicking, favorites, and purchases. Products also possess rich category, attribute, and textual description information. While traditional collaborative filtering recommendation methods can uncover potential preferences based on user-product interaction history, they face challenges such as data sparsity, long-tail effects, cold start, and interest drift. For example, newly registered users or newly launched products are often difficult to recommend due to a lack of behavioral data; user behavior data is diverse and high-dimensional, making it difficult for single collaborative filtering to capture the semantic or multimodal features of product content. To address these challenges, this embodiment proposes a dual-channel fusion method: fully utilizing user behavior structure information in the user-product collaborative graph channel, extracting semantic features of product content or attributes in the semantic graph channel, and adaptively fusing the two types of information through an LLM gating mechanism to comprehensively characterize user preferences and improve recommendation accuracy.

[0052] (2) Data acquisition and preprocessing In this e-commerce scenario, the first step is to collect historical user behavior logs, including records of user clicks, browsing, purchases, and favorites related to products, as well as attribute information and textual descriptions of each product (e.g., name, category, tags, detailed descriptions, etc.). The acquired data may contain missing information, noise, and inconsistencies, making preprocessing crucial. Data preprocessing includes: cleaning dirty data, such as removing abnormal or duplicate behavior records; unifying the mapping between user and product IDs; converting user-product interactions into rating or implicit feedback formats; performing text cleaning on product descriptions, such as word segmentation, stop word removal, and part-of-speech tagging; encoding product text descriptions or tags into initial semantic vectors using pre-trained models (e.g., Word2Vec, BERT); and extracting and normalizing user profile information (e.g., age, region, interest tags). Furthermore, behavioral data can be reduced, for example, by generating user preference tags based on statistical features (number of clicks, number of purchases). After these processes, structured user-product interaction data and product semantic feature data are obtained, preparing for subsequent graph construction and model training.

[0053] (3) Graph construction method Regarding the collaborative graph channel, this embodiment constructs a user-product interaction bipartite graph. The graph nodes include user nodes and product nodes, and edges represent the interaction relationships between users and products (such as clicks or purchases). This collaborative graph can be viewed as a user-product association network, implicitly reflecting the preference relationships between users and products. Simultaneously, interactions can be categorized into different types (such as browsing, adding to cart, and purchasing) according to business needs, and different weights can be assigned to edges of different types. This embodiment uses a graph neural network with first-order and higher-order neighbor aggregation to represent the graph structure, where each user node and product node corresponds to a vector representation, and information is aggregated from neighbor nodes through graph convolution or graph attention mechanisms.

[0054] Regarding the semantic graph channel, this embodiment constructs a product semantic relation graph or knowledge graph. Nodes can include the product itself and attribute nodes associated with the product (such as brand, category, keyword nodes, etc.). For example, a heterogeneous graph containing various relationships such as product-category, product-tag, and tag-tag can be constructed, allowing the semantic graph to reflect the semantic connections between products through attributes or tags. Furthermore, natural language processing is performed on product text information (such as product descriptions) to extract keywords or topics, and these are connected as nodes to corresponding product nodes, forming a product-semantic node graph structure. This semantic graph channel can use content-based graph neural networks or pre-trained language models to learn node representations and capture the implicit semantic features of product content. For example, relational knowledge graph embedding methods can be used, or subgraph neural networks can be used to embed the semantic graph. Through the above graph construction, this embodiment can extract user and product features from the collaborative structure level and the content semantic level, respectively.

[0055] (4) Model training and optimization process The core model of this embodiment includes dual inputs: a collaborative graph channel and a semantic graph channel, along with an LLM-gated fusion module in between. First, the collaborative graph obtained in the graph construction step is input into the collaborative channel network to learn the vector representations of users and items in the graph; the semantic graph is then input into the semantic channel network to learn the semantic representations of item nodes and (if applicable) user preference representations. In the collaborative channel, a multilayer graph convolutional network (GCN) or a graph attention network (GAT) can be used for message passing, with each layer updating node vectors to absorb neighbor information. In the semantic channel, a similar approach to GCN, or a combination of a pre-trained language model (such as BERT), can be used to encode the item text, and then a multilayer perceptron extracts deep semantic features. After obtaining the two types of embeddings, they enter the LLM-gated fusion module: this module uses a set of gating coefficients generated by a pre-trained LLM to dynamically control the weights of the collaborative embeddings and semantic embeddings in the final fusion. For example, the semantic descriptions of users and candidate items can be used as prompts input to the LLM, which outputs fusion weights. Then, the two embeddings are weighted and fused to obtain the final user-item joint representation. Finally, the fused representation is used for similarity calculation or rating prediction (such as inner product or MLP) to obtain the user's predicted rating for the product. The model training process employs cross-entropy loss or Bayesian Personalized Ranking (BPR) loss, and uses stochastic gradient descent algorithms (such as Adam) for end-to-end optimization. To improve training efficiency, a negative sampling strategy can be used to dynamically construct negative examples; simultaneously, model parameters (learning rate, regularization coefficient, etc.) are tuned during training to prevent overfitting and accelerate convergence.

[0056] (5) Deployment and effect verification After training, this embodiment is deployed online in a production environment. The model is typically divided into two phases: offline training and online recommendation. In the offline phase, the model parameters are refreshed periodically (e.g., daily) using the latest user behavior data, and the semantic and structural embeddings of products are pre-calculated. In the online phase, the system receives real-time user requests, aggregates the current user's historical behavior, and calculates recommendation scores sequentially with the candidate product list. The system calls the model service to return a TOP-N recommendation list based on different interfaces such as portals and mobile applications. In the initial stage of deployment, A / B testing and other experiments were used to verify system performance: monitoring changes in metrics such as click-through rate (CTR), conversion rate (CVR), user dwell time, and product purchase rate. Experimental results show that the accuracy and coverage of recommendations are significantly improved after introducing the dual-channel fusion model, especially in scenarios involving new user cold starts and long-tail product recommendations. This model has significant advantages over traditional collaborative filtering methods in offline evaluation metrics such as MAP and NDCG, while also showing a significant improvement in online CTR, verifying the effectiveness and feasibility of this embodiment in e-commerce recommendations.

[0057] Example 2: News App Distribution (1) Business background and challenges This embodiment is applicable to personalized content distribution scenarios in news and information applications. In this scenario, news content is updated frequently and contains a large amount of information. User attention changes rapidly with current events, so the recommendation system needs to balance timeliness and semantic understanding. News data has obvious textual features and thematic attributes. Users' reading records are often short text sequences, and user interests are prone to drift with news events. Traditional collaborative filtering faces challenges in this scenario. For example, newly released news lacks sufficient user interaction data, and simply recommending based on the user-news click matrix can lead to cold start problems. At the same time, the similarity between news items is mainly reflected at the semantic level, such as news on the same topic or related events, requiring in-depth semantic analysis of the content. To address the characteristics of news recommendation, this embodiment adopts a dual-path fusion method: extracting structural preference information from the collaborative graph channel of user-news interaction and extracting textual semantic information from the semantic graph channel of news content, and then using LLM gating fusion to more accurately match user interests with news content.

[0058] (2) Data acquisition and preprocessing In news distribution scenarios, data sources primarily include user reading behavior logs (such as clicks, likes, comment timestamps, etc.) and news content features (including titles, body text, category, keywords, and publication time). First, the data undergoes cleaning and structuring: invalid or abnormal reading events are removed, and low-quality news (such as news with repetitive content or ambiguous titles) is filtered out. User reading sequences are segmented to avoid interference from rapidly changing preferences. Preprocessing operations such as word segmentation and HTML tag removal are performed on the news text. Next, natural language processing techniques are applied to the news text content. Models such as TF-IDF, LDA, or BERT are used to generate semantic feature vectors for each news article, extracting topic distribution or keyword vectors. One-hot encoding is performed on news tags or category information. Statistical analysis and encoding are conducted on user-side features (such as region, device type, subscription preferences, etc.). Finally, a user-news interaction record matrix is ​​obtained, along with various features used to construct the graph: news text semantic vectors, tag relationships, and timestamps.

[0059] (3) Graph construction method In the collaborative graph channel, this embodiment constructs a user-news interaction graph. Nodes include user nodes and news nodes, and edges represent user-news interactions (different types such as first click and repeated clicks can be distinguished and assigned different weights). Since news recommendation needs to focus on temporal information, this embodiment can introduce time attributes into the graph structure (such as adding timestamp features to nodes or edges) and use a temporal graph neural network for dynamic updates. In the semantic graph channel, a news semantic relevance graph or knowledge graph is constructed. For example, news nodes can be connected to their topic tag nodes, or a news-keyword bipartite graph can be constructed based on keyword nodes extracted from news titles / text; or entity recognition can be used to connect news with corresponding entity knowledge graph nodes to form a cross-news topic association network. This semantic graph can also include semantic similarity edges between news items (connected based on content cosine similarity or topic similarity). Through the above construction, the semantic graph captures high-level semantic relationships between news content. In the collaborative graph channel, a graph convolutional network or graph attention network is used to calculate the collaborative preference representation between users and news; in the semantic graph channel, a heterogeneous graph neural network or knowledge graph embedding algorithm is used to learn the semantic representation of news. By combining the two channels, this embodiment can both capture users' interests in different news types by leveraging user behavior and discover related articles that users may be interested in by utilizing the semantic similarity of news content.

[0060] (4) Model training and optimization process The recommendation model structure in this embodiment is similar to that in Embodiment 1, including dual-input and gating fusion mechanisms. The input to the collaborative channel is a user-news interaction graph, which uses a multi-layer temporal GCN to learn the collaborative embeddings of users and news. The input to the semantic channel is a news semantic relevance graph, which uses a heterogeneous graph network or a pre-trained semantic model to extract the semantic representation of the news. The two are then fused through an LLM gating module: for example, the user's interest description (historical reading topics) and the news summary are input into the LLM, allowing the LLM to generate fusion strategies or weights based on user preferences and news content, thereby adaptively weighting and fusing structured preferences and semantic information. During training, a click-through rate prediction loss function is used, with actual user clicks as positive examples and other randomly sampled news items as negative examples for training. To address the data sparsity and hot recall issues in news streaming scenarios, an online learning strategy can be incorporated: the model is periodically updated online incrementally to quickly learn new news information and the latest user feedback. During large-scale training, multi-machine distributed parallel technology and graph sampling algorithms (such as neighbor sampling) can be used to improve training efficiency and to fine-tune model hyperparameters (learning rate, number of layers, negative sampling rate, etc.).

[0061] (5) Deployment and effect verification During deployment, this embodiment typically employs a streaming online recommendation architecture: In the offline phase, user-news graphs and news semantic graphs are maintained, and the training model and news embeddings are updated periodically; in the online phase, user recommendation requests are responded to in real-time. Specifically, when a user uses the news app, the system converts the user's recent reading history into interest feature vectors, calls the model to score and rank the candidate news set, and returns a personalized news list. Online validation combines online A / B testing with offline metric evaluation: by comparing the new system with benchmark systems (such as simple collaborative recommendation or traditional algorithms), metrics such as click-through rate (CTR), browsing depth, user dwell time, and user retention rate are monitored. Experimental results show that the introduction of semantic and structural dual-path fusion improves the relevance and diversity of the recommendation list; especially for newly published news and low-exposure news, the semantic graph channel can compensate for insufficient behavioral signals, significantly increasing the probability of such news being recommended. After multiple iterations of optimization, this model outperforms traditional models in offline evaluation metrics such as recommendation accuracy and nDCG, while user engagement and satisfaction metrics also show significant improvement in online experiments.

[0062] Example 3: Enterprise SaaS Function Guide (1) Business background and challenges This embodiment focuses on the design of feature recommendation and user guidance scenarios in enterprise-level SaaS (Software as a Service) products. Unlike general e-commerce or news recommendations, SaaS recommendations aim to help enterprise users understand and use software functions. For example, in enterprise management systems or collaborative office software, it is desirable to promptly recommend functional modules that have not yet been used but may be needed (such as report generation, data analysis, collaborative approval, etc.) based on user behavior. The main challenges in enterprise SaaS scenarios include: on the one hand, new enterprises or new users have no usage history and need to quickly familiarize themselves with the software functions; on the other hand, product functions usually have a hierarchical structure (such as main functions and sub-functions) and are accompanied by rich semantic descriptions (help documents, usage scenario descriptions, etc.), and recommendations need to take into account user roles and industry characteristics. To this end, this embodiment adopts a dual-path fusion strategy: in the collaboration graph channel, user-function interaction data is used to mine group preferences, and in the semantic channel, the textual description of functions and business semantics are combined to construct recommendation logic, and the two information paths are intelligently fused through LLM (Local Management Model).

[0063] (2) Data acquisition and preprocessing The data collected by enterprise SaaS platforms includes user identity information (such as job title, department, role, etc.), user function usage logs (such as clicks, API call counts, session logs, etc.), and metadata and documentation for each functional module (function name, description, tags, related business concepts, etc.). The preprocessing stage first cleans the function usage logs, for example, by deleting data generated by test accounts and filtering for abnormal usage frequencies; standardizes user information (using unified role or department tags); and maps function operation logs to user-function usage counts or implicit ratings. Natural language processing is then applied to the text descriptions and tags of the functions, using knowledge extraction methods to build relationships between functions (such as discovering related functions or functions that depend on the same business metric). Simultaneously, a preliminary function relationship network can be established using domain knowledge bases or product manuals. The output of preprocessing includes a user-function interaction matrix, function text semantic vectors, and necessary user and function attribute features.

[0064] (3) Graph construction method In the collaborative graph channel, this embodiment constructs a user-function interaction graph. Nodes include user nodes and function nodes, and edges represent relationships where users have used corresponding functions, which can be weighted according to usage frequency. This graph can reflect the association patterns between users and functions, such as users in the same department frequently using certain functions together, or a user continuously using multiple function modules. To address the function cold start problem, a virtual "newbie" node can be introduced into the graph to connect new users, or collaborative filtering based on similar roles can be used to supplement information for new users during recommendations. In the semantic graph channel, this embodiment focuses on constructing a function semantic association graph. For example, function nodes with the same business domain can be connected with label nodes, or functions can be connected with their key business concept nodes (such as "approval process" or "financial analysis") to form a function-business concept knowledge graph. User roles can also be mapped to business concepts, forming multi-hop semantic associations across user functions. Through the above construction, the semantic graph contains both semantic relationships between functions and matching information between user roles and functions. The collaborative channel learns the structural embeddings of users and functions in the graph neural network, while the semantic channel learns the semantic embeddings of roles and functions from the knowledge graph.

[0065] (4) Model training and optimization process The model structure is the same as the previous two examples, including a dual-channel approach and an LLM fusion module. During training, structural embeddings of users and functions are first extracted from the collaboration graph to capture user habits and group effects within the enterprise. Semantic embeddings of functions and business concepts are extracted from the semantic graph to help the model understand the relationships between functions and their matching degree with user roles or job responsibilities. Then, an LLM gating module is used to fuse the two embeddings: for example, user role descriptions and textual descriptions of candidate functions are input into the LLM, and a fusion strategy is generated based on the model's inference results to adaptively adjust the weights of structural and semantic information. The training loss uses implicit feedback from user function usage, optimized using supervised learning methods. For the sampling strategy, frequently used functions can be used as positive examples, and unused functions as negative examples; alternatively, a threshold can be used to divide samples into approximate positive and negative. During training, model performance is evaluated through cross-validation and cold-start testing. Hyperparameters such as the number of model layers and attention heads are adjusted, and regularization techniques such as Dropout are attempted to improve generalization ability.

[0066] (5) Deployment and effect verification Upon deployment, the trained model is integrated into the user guidance system of the SaaS platform. As users interact within the system, the recommendation service obtains the user's current role and historical function usage records in real time. Using the user's role description and behavior sequence as input, it calculates recommendation scores for candidate functions and pushes function usage suggestions to the user interface according to the score ranking. The platform can recommend relevant modules in the function menu or help center and guide users to quickly access the corresponding function page when they are interested. Effectiveness is validated through enterprise user feedback and usage data: metrics such as increased usage rate of recommended functions, reduced time for users to complete tasks, and user satisfaction scores can be statistically analyzed. After deployment, it was observed that, through the dual-path fusion model, the system can accurately recommend functions that users have not yet used but should prioritize based on their roles. For example, the accounting management module is recommended when finance personnel log in for the first time; the attendance approval module is recommended when HR managers log in. Compared to traditional rule-based or single-collaboration methods, this method is significantly more effective in guiding function usage coverage and reducing user training costs, such as improved recommendation click-through rates and function usage conversion rates, fully demonstrating the effectiveness of this embodiment in enterprise SaaS function guidance scenarios.

[0067] In summary, compared with the prior art, the method of this embodiment has at least the following advantages and beneficial effects: 1) Dual-path information complementarity effectively solves the cold start problem: By processing structured interaction graphs and semantic association graphs in parallel, this invention fully utilizes two heterogeneous information sources. In cold start scenarios where user interaction data is extremely sparse, rich semantic information serves as a supplement, providing the model with additional prior knowledge, thereby significantly improving recommendation performance in cold start and sparse scenarios.

[0068] 2) LLM-Supervised Fusion for Dynamic Personalized Decision Making: This invention innovatively introduces LLM as an "external knowledge brain," providing dynamic and personalized supervisory signals for information fusion. The model can autonomously determine whether to rely more on behavioral information or textual descriptions for each specific user or item node, achieving personalized fusion tailored to each individual. Compared to traditional static fusion methods, the LLM-guided fusion strategy of this invention offers greater flexibility and superior effectiveness.

[0069] 3) Channel separation design significantly enhances noise robustness: Before information fusion, structural and semantic information propagate in independent channels, effectively avoiding mutual interference between different modal noises. Under the guidance of the LLM supervision signal, the gated fusion mechanism reduces the weights of channels deemed to have high noise levels, further improving the overall model's robustness. Theory and practice show that, compared to the strategy of fusion before propagation, the separation-then-fusion approach of this invention exhibits superior noise suppression performance.

[0070] 4) Complete technical solution, balancing performance and efficiency: The technical solution proposed in this invention is complete and highly feasible. By performing Top-K sparsification on the semantic association graph and adopting a lightweight graph network model (such as LightGCN), recommendation performance is significantly improved while training and inference efficiency are also considered, demonstrating strong practical application value.

[0071] 5) Strong Model Interpretability: The various designs in this invention enhance the model's interpretability. The attention weights of GAT reveal which semantic neighbors are most important to the current node; the supervision signals generated by LLM provide a basis for the gating network's decisions; and the final gating vector intuitively reflects the model's preference for collaborative and semantic information during fusion. These factors collectively make the model's decision-making process more transparent, no longer a black box operation.

[0072] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0073] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0074] Please see Figure 4 , Figure 4 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 401 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 402 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 402 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 402 and is called and executed by the processor 401 using the methods described in the embodiments of this application. Input / output interface 403 is used to implement information input and output; The communication interface 404 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 405 transmits information between various components of the device (e.g., processor 401, memory 402, input / output interface 403, and communication interface 404); The processor 401, memory 402, input / output interface 403 and communication interface 404 are connected to each other within the device via bus 405.

[0075] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0076] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0077] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0078] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0079] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented in the embodiments of this program product are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. The executable computer program code or "code" used to perform the various embodiments can be written in high-level programming languages ​​such as C, C++, Python, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0080] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0081] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0082] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0083] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0084] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0085] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0087] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0088] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0089] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A personalized recommendation method based on semantic and structural dual-path fusion, characterized in that, The method comprises the following steps: Constructing a two-path graph network: constructing a structured interaction graph according to historical interaction data of users and items; generating semantic vectors by using a pre-trained large language model according to text description data of users and items, and constructing a semantic association graph based on semantic similarity; Dual-channel feature learning: information propagation and aggregation of the structured interaction graph are performed by a graph convolution network model to learn a collaborative feature vector; weighted information propagation and aggregation of the semantic association graph are performed by a graph attention network model to learn a semantic feature vector according to the semantic association and attention weight of the node; Generating a fusion supervision signal: for a target node, a structured prompt word is constructed, which presents the semantic features and interaction structure features of the node in the form of text; the prompt word is input into a pre-trained large language model to obtain a value indicating the relative reliability of the collaborative feature and the semantic feature returned by the large language model, which is used as a supervision signal; Adaptive gating fusion: the collaborative feature vector and the semantic feature vector are spliced and input into a gating network to obtain a gating vector; the collaborative feature vector and the semantic feature vector are weighted and fused by using the gating vector to obtain a final fusion feature vector; in the training process, the gating network is supervised and learned by using the supervision signal output by the large language model; Recommendation generation: based on the final fusion feature vector, a multi-task loss function is constructed to jointly optimize the model; the optimized model is used to calculate the matching score between the user and the candidate item, and an individualized recommendation list is generated according to the matching score.

2. The method of claim 1, wherein, The specific steps of constructing the semantic association graph comprise: Calculate the semantic similarity of all user-item pairs to form a semantic similarity matrix; For each user, only the top K items with the highest semantic similarity are retained as semantic neighbors; for each item, only the top K users with the highest semantic similarity are retained as semantic neighbors; Based on the semantic neighbor relationship, a sparse semantic association graph is constructed.

3. The method of claim 1, wherein, The graph convolution network model is a lightweight graph convolution network, which obtains the collaborative feature vector by accumulating the node embedding vectors output by each layer.

4. The method of claim 1, wherein, The gating network is a fully connected neural network which computes the gating vector by the following formula : where, is a collaborative feature vector, is a semantic feature vector, denotes a vector concatenation operation, and is a learnable parameter, is a Sigmoid activation function.

5. The method of claim 4, wherein, The adaptive gated fusion obtains the final fusion feature vector by the following formula : wherein denotes element-wise multiplication.

6. The method of claim 1, wherein, In the recommendation generation, the multi-task loss function at least includes: An individualized ranking loss for optimizing the ranking preference of user-item pairs; A gating supervision loss for minimizing the difference between the gating vector and the supervision signal output by the large language model.

7. The method of claim 6, wherein, The multi-task loss function further includes at least one of the following: A contrastive learning loss for encouraging the alignment of collaborative features and semantic features in the representation space; A knowledge distillation loss for distilling the knowledge of the semantic feature vector to the final fusion feature vector; An L2 regularization loss for preventing model overfitting.

8. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method of any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the method of any one of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that, The computer program, which is executed by a processor, implements the method of any one of claims 1 to 7.

Images