Research theme recommendation method and system based on heterogeneous information network and large language model

Abstract

The invention discloses a research theme recommendation method and system based on a heterogeneous information network and a large language model. According to the method, complex association among scholar historical research, a cooperation network and a thesis theme can be deeply mined, and scholar and theme representation learning fusing semantic content and a network structure is realized, so that reasonable association of a recommended theme and a scholar research background in a structural level is ensured. A large language model is further introduced for semantic reasoning and refining, intelligent evaluation and sorting can be conducted on candidate topics from multiple dimensions such as correlation, novelty and feasibility, and individuation, interpretability and academic perspectiveness of recommendation results are remarkably enhanced. Finally, a refined theme list which is highly matched with the research venation and has innovation potential and feasibility is provided for scholars, and the scholars are powerfully assisted in mastering the research dynamic state and expanding the research direction.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of scientific and technological intelligence and artificial intelligence, specifically to a research topic recommendation method and system based on heterogeneous information networks and large language models. Background Technology

[0002] Research topic selection is a crucial decision in a scholar's academic career. Faced with the explosion of knowledge and the trend of interdisciplinary integration, identifying research directions from a vast amount of academic resources that both align with one's own research foundation and possess innovative potential is a significant challenge. Traditional academic recommendation systems primarily focus on recommendations based on papers, patents, and collaborators, while paying insufficient attention to personalized and forward-looking recommendations for research topics.

[0003] Existing research topic recommendation methods suffer from the following limitations: First, content-based methods (such as topic models and word embeddings) rely excessively on surface-level semantic similarity of text, easily leading to information redundancy and "filter bubbles," making it difficult to discover novel interdisciplinary topics. Second, collaborative filtering methods are limited by data sparsity and the cold-start problem, making them unsuitable for scholars with unique research interests or those just starting out. Third, while graph-based methods (such as homogeneous graph neural networks) can model entity relationships, they typically ignore the heterogeneity of entity types such as scholars, papers, and topics in academic networks, failing to accurately depict diverse and complex academic relationships.

[0004] In recent years, heterogeneous graph neural networks (GNNs) have demonstrated powerful capabilities in modeling complex networks, effectively learning representations of heterogeneous nodes and relationships. Meanwhile, large language models (GLLMs) have made breakthroughs in deep semantic understanding, knowledge reasoning, and generation. However, how to organically combine the ability of GNNs to capture the connectivity of structured networks with the ability of LLMs to deeply understand unstructured semantics, and construct a topic recommendation system that can both understand the long-term research trajectory of scholars and perform expert-level logical reasoning, remains a pressing technical problem to be solved. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a method and system for recommending research topics for scholars that integrates graph neural networks and large-scale models. The core idea of ​​this invention is to utilize heterogeneous graph neural networks as "structure perceptrons" to mine potential topic diffusion patterns and long-term research trajectories from a heterogeneous network composed of scholars' historical data; simultaneously, it utilizes large-scale language models as "content insight devices" to perform in-depth evaluation and refinement of candidate topics generated by the GNN based on domain knowledge. Through this two-stage collaborative framework of "generation-refinement," an elevation from "connectivity" to "insight" is achieved, providing scholars with highly personalized research topic recommendations that are both relevant and novel.

[0006] To achieve the above objectives, this invention provides a research topic recommendation method based on heterogeneous information networks and large language models, comprising: S1. Based on the historical paper data of the target scholar group, obtain scholars, papers and topic entities, and construct a heterogeneous information network based on the association between the subject entities; S2. Based on the heterogeneous information network, obtain a low-dimensional dense vector representation of each node in the network that integrates semantic content and network structure; S3. Based on the low-dimensional dense vector representation, the scholar-topic recommendation task is formalized into a link prediction problem, and a preliminary set of candidate recommendation topics for each scholar is obtained based on the link prediction results. S4. Based on the preliminary candidate recommendation topic set and the historical research information of the corresponding scholars, semantic refinement and reordering are performed using a large language model to obtain the final personalized research topic recommendation list.

[0007] Preferably, S1 includes: Based on historical paper data obtained from academic databases, scholars, papers and their keywords are extracted as three types of entities: themes. Based on the extracted three types of entities, a heterogeneous information network of node sets is constructed, including scholar nodes, paper nodes, and topic nodes. Based on the associations between the three types of entities, an edge set is constructed in the heterogeneous information network. The edge set includes: scholar-paper edges representing scholars publishing papers, and paper-topic edges representing the topics contained in the papers.

[0008] Preferably, S2 includes: Based on the text content of nodes in the network, obtain the content feature vector of the nodes; Based on the structure of heterogeneous information networks, obtain the structural feature vectors of nodes; Based on content feature vectors and structural feature vectors, the final embedded representation of nodes is obtained through feature fusion and heterogeneous information aggregation.

[0009] Preferably, S3 includes: Based on the final embedding representations of scholars and topics, scholar-topic feature pairs are concatenated. Based on scholar-topic feature pairs, a pre-trained machine learning classifier is used to predict the probability that there is a link between scholars and topics; Based on the predicted probabilities, the Top-K unseen topics with the highest probabilities are selected for each scholar to form a preliminary set of candidate recommended topics.

[0010] Preferably, the steps for obtaining the content feature vector of a node include: Based on the title text and abstract text of the paper node, the title feature vector and abstract feature vector are obtained by encoding using a pre-trained language model, respectively. For scholar nodes and topic nodes, based on their association with paper nodes, the content feature vectors of scholar nodes and topic nodes are obtained by aggregating the content feature vectors of associated paper nodes.

[0011] Preferably, the step of obtaining the structural feature vector of a node includes: Based on heterogeneous information networks, node sequences are generated through random walks; Based on the node sequence, the structural embedding vector of each node is learned using a word embedding model.

[0012] Preferably, S4 includes: Based on the preliminary set of candidate recommended topics and the corresponding list of scholars' historical research topics, a structured prompt containing instructions for analysis, evaluation, and screening and ranking is constructed. Based on structured prompts, a large language model is driven to perform semantic reasoning and evaluation of candidate topics; Based on the output of the large language model, obtain the refined and sorted final personalized research topic recommendation list.

[0013] This invention also provides a research topic recommendation system based on heterogeneous information networks and large language models. The system is used to implement the above method and includes: a collection module, an extraction module, a filtering module, and a recommendation module. The acquisition module is used to acquire scholars, papers and topic entities based on the historical paper data of the target scholar group, and to construct a heterogeneous information network based on the association between the entities. The extraction module is used to obtain a low-dimensional dense vector representation of each node in the heterogeneous information network, which integrates semantic content and network structure. The filtering module is used to formalize the scholar-topic recommendation task into a link prediction problem based on the low-dimensional dense vector representation, and to obtain a preliminary set of candidate recommendation topics for each scholar based on the link prediction results. The recommendation module is used to perform semantic refinement and reordering based on the preliminary candidate recommendation topic set and the historical research information of the corresponding scholars, using a large language model to obtain the final personalized research topic recommendation list.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention significantly improves the accuracy and reliability of academic research topic recommendations. By combining heterogeneous information networks with deep learning models, it effectively captures the deep connections between scholars' research backgrounds and potential topics, ensuring the structural rationality of the recommendation results within the academic context. Furthermore, it introduces a large language model for semantic-level reasoning and refinement, enabling recommendations to go beyond co-occurrence relationships and comprehensively consider the relevance, novelty, and feasibility of topics, thus balancing academic continuity and innovative expansion. Ultimately, it provides scholars with highly personalized, interpretable, and forward-looking refined topic lists, assisting them in efficiently grasping disciplinary dynamics and clarifying research directions, demonstrating strong practical value. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 The following is an overall flowchart of the recommended method in the embodiments of the present invention; Figure 2 This is a schematic diagram of a heterogeneous information network of "scholar-paper-topic" according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the topic recommendation model architecture for collaborative enhancement of neural networks and large models in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the workflow of the semantic reordering module of a large language model based on LangChain and chained reasoning, as described in an embodiment of the present invention. Figure 5 A schematic diagram illustrating the extraction of feature vectors for the BERT and Longformer models. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 S1. Based on the historical paper data of the target scholar group, obtain scholars, papers and topic entities, and construct a heterogeneous information network based on the relationship between the subject entities.

[0020] Historical paper data of the target scholar group is obtained from academic databases (such as Scopus and Web of Science), and three types of entities are extracted: scholars, papers, and their keywords (as topics). A heterogeneous information network G=(V,E) is constructed, where the node set V contains the scholar node set A, the paper node set P, and the topic node set T. The edge set E contains two types of edges: if scholar a has published paper p, then there exists an edge (a, p); if paper p contains topic t, then there exists an edge (p, t). This network accurately models the triangular relationship of "scholar-paper-topic", such as... Figure 2 As shown.

[0021] S2. Based on heterogeneous information networks, obtain a low-dimensional dense vector representation of each node in the network that integrates semantic content and network structure.

[0022] This step aims to learn a low-dimensional dense vector representation that integrates semantic content and network structure for each type of node in the network.

[0023] (1) Content feature extraction For the paper node, the BERT model is used to encode its title, resulting in a title feature vector; the Longformer model is used to encode its abstract, resulting in an abstract feature vector. The detailed architecture of the two models is as follows: Figure 5 As shown, for scholar and topic nodes, their content features are obtained by aggregating the features of their associated paper nodes. This design can balance the precise semantics of short texts with the rich context of long texts.

[0024] (2) Structural feature extraction The DeepWalk algorithm is used to perform random walks on a heterogeneous network G=(V,E). Starting from each node v∈V, a truncated random walk of length L is performed to generate a node sequence: (1) The Skip-gram model is used to learn the structural embedding vector of each node, which captures the node's high-order proximity and functional role in the network. The goal is to: given a central node... vi In this case, maximize its context node vj Conditional probability: (2) in, Represents a node vi Context window in a random walk sequence.

[0025] (3) Feature fusion and heterogeneous information aggregation First, a bidirectional LSTM is used to fuse the content feature vector and structural feature vector of each node to obtain the initial representation of the node. For a node... v The content feature vector can be obtained through the above steps. and structural feature vectors Then, the Bi-LSTM algorithm is used to fuse the two types of features to generate the initial embedding of the node: (3) in, It is a node v The representation vector.

[0026] The Bi-LSTM model contains two LSTMs, one forward and one backward, enabling it to capture contextual information from two directions. Unlike simple feature concatenation or summation, the Bi-LSTM model can capture the deep relationship between content features and structural features, making it highly effective in heterogeneous feature fusion.

[0027] Next, a restarted random walk algorithm is used to sample a set of neighbors of various types for each central node. Then, a type-specific aggregator (such as Bi-LSTM pooling) is used to aggregate the information of neighbors of the same type. After sampling heterogeneous neighbor nodes, the node... v Neighbor set ,in T Represents a set of node types, denoted as Subsequently, the Bi-LSTM method is used to aggregate messages from neighbors of the same type. The calculation process is as follows: (4) in, Represents a node v Aggregate representation of type t; node Represents a node v's domain node, It is a node The representation vector; Given the varying strengths of relationships between different node types, an attention mechanism is further introduced to weighted aggregate the representations of different node types to obtain the node... v The final representation: (5) in, Represents a node v Self-weight; Indicates from type t Neighbors of nodes vThe weights are calculated as follows: (6) in, β For attention parameter vectors, LeakyReLU For activation function, f t For nodes t The representation is T, where T is the set of node types.

[0028] Finally, a node type-level attention mechanism is introduced to automatically learn the differences in importance of different types of neighbors (such as scholars' collaborators and published topics) to the central node's representation. The weighted summation yields the final embedding representation of the node. This process enables the model to fully understand the complex dependencies in heterogeneous networks.

[0029] S3. Based on low-dimensional dense vector representation, the scholar-topic recommendation task is formalized into a link prediction problem, and a preliminary candidate recommendation topic set for each scholar is obtained based on the link prediction results.

[0030] The scholar-topic recommendation task is formalized as a link prediction problem. For a scholar node 'a' and a topic node 't', their corresponding final embedding vectors are concatenated to form a feature pair. The machine learning model is trained on known scholar-topic association samples to learn the mapping function between input features and associations. Its prediction formula is shown below: (7) in, and These are the embedding representations of scholar nodes and topic nodes, respectively. This indicates a feature concatenation operation. The model function obtained through training. For model parameters, This represents the predicted association between scholar nodes and topic nodes.

[0031] During the recommendation phase, for each scholar, the probability of their links to all unseen topics is calculated, and the Top-K topics with the highest probabilities are selected to form an initial candidate recommendation set. This step ensures the structural rationality of the recommended topics based on the network structure.

[0032] S4. Based on the preliminary candidate recommendation topic set and the historical research information of the corresponding scholars, semantic refinement and re-ranking are performed using a large language model to obtain the final personalized research topic recommendation list.

[0033] The recommended model architecture for this step is as follows: Figure 3As shown, a large language model is introduced to refine the coarse-grained candidate set into fine-grained semantics. First, a structured suggestion engineering template based on the LangChain framework is designed, such as... Figure 4 As shown. This template explicitly assigns the LLM "Academic Research Assistant" system role and embeds chained reasoning instructions. The specific reasoning chain includes: (1) Analysis: In-depth interpretation of the scholar's list of historical research topics, summarizing his core areas, methodological preferences and potential expansion directions.

[0034] (2) Evaluation: Within the candidate topic set provided in step S3, each candidate topic is scored in a structured manner from three dimensions: "relevance", "novelty" and "feasibility".

[0035] (3) Screening and Ranking: Based on the comprehensive evaluation results, unsuitable topics are eliminated, and the remaining topics are ranked in descending order of their relevance to the scholars' research directions. By strictly limiting the generation of the LLM to the candidate set and guiding it through step-by-step reasoning, model illusion is effectively avoided, and the reliability and interpretability of the recommendations are improved. The final output is a concise and orderly list of personalized research topic recommendations.

[0036] The overall process framework of this embodiment is as follows: Figure 1 As shown.

[0037] Example 2 This embodiment uses a scholar from a university in my country as an example to illustrate the specific implementation process of the present invention.

[0038] (1) Data preparation A search of the Scopus database for English papers by RUC scholars published after 2000 yielded 21,166 records. After deduplication and cleaning of missing fields, 15,487 valid papers were obtained. The "Author ID," "EID," and "Author Keywords" of each paper were extracted to construct an entity list. The author keywords were then standardized and used as the "Topic."

[0039] (2) Network construction The training network was constructed using data from 2000 to 2020. The network contains 1498 scholar nodes, 7731 paper nodes, and 2998 topic nodes. The edge relationships include 13026 scholar-paper edges and 14155 paper-topic edges, as shown in Table 1.

[0040] Table 1 .

[0041] (3) Model training and recommendation GNN Stage: HetGNN was implemented using PyTorch. The paper title was BERT-based encoded, and the abstract was Longformer encoded. DeepWalk was set with a walk length of 80 and a window size of 10. Feature fusion used a 256-dimensional Bi-LSTM. An attention mechanism aggregated heterogeneous neighbor information. The final result was a 128-dimensional node embedding.

[0042] Link prediction: The 128-dimensional embeddings of scholars and topics are concatenated into a 256-dimensional vector, which is then input into a logistic regression classifier for training (positive and negative samples are 1:1). After training, the link probability between each scholar and all topics is predicted.

[0043] LLM Refinement: Use LangChain to call the GPT-3.5-turbo API. The prompt template is shown in Appendix A, explicitly instructing the model to analyze scholars' historical topics and filter from the Top-20 candidate list generated by GNN. Guide it through Chain-of-Thought evaluation for relevance and novelty.

[0044] The recommendation results are shown in Tables 2-4, where Table 2 shows the recommendation results based on graph neural networks and machine learning; Table 3 shows the recommendation results based on large models; and Table 4 shows the recommendation results of the method of this invention.

[0045] Table 2 Table 3 Table 4 The results, evaluated on test data from 2021 to 2024, show that the F1 score of the pure GNN+LR model is 63.15%. After introducing GPT-3.5-turbo for re-ranking, the F1 score of the GNN-LLM collaborative model improves to 71.46%, representing an 8.31% improvement, and significantly enhances the novelty of recommended topics and scholar satisfaction.

[0046] Example 3 This embodiment also provides a research topic recommendation system based on heterogeneous information networks and a large language model, including: a collection module, an extraction module, a filtering module, and a recommendation module; the collection module is used to obtain scholars, papers, and topic entities based on the historical paper data of the target scholar group, and construct a heterogeneous information network based on the associations between entities; the extraction module is used to obtain a low-dimensional dense vector representation of each node in the network that integrates semantic content and network structure based on the heterogeneous information network; the filtering module is used to formalize the scholar-topic recommendation task into a link prediction problem based on the low-dimensional dense vector representation, and obtain a preliminary set of candidate recommended topics for each scholar based on the link prediction results; the recommendation module is used to perform semantic refinement and re-ranking using a large language model based on the preliminary set of candidate recommended topics and the historical research information of the corresponding scholars, and obtain a final personalized research topic recommendation list.

[0047] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A research topic recommendation method based on heterogeneous information networks and large language models, characterized in that, include: S1. Based on the historical paper data of the target scholar group, obtain scholars, papers and topic entities, and construct a heterogeneous information network based on the association between the subject entities; S2. Based on the heterogeneous information network, obtain a low-dimensional dense vector representation of each node in the network that integrates semantic content and network structure; S3. Based on the low-dimensional dense vector representation, the scholar-topic recommendation task is formalized into a link prediction problem, and a preliminary set of candidate recommendation topics for each scholar is obtained based on the link prediction results. S4. Based on the preliminary candidate recommendation topic set and the historical research information of the corresponding scholars, semantic refinement and reordering are performed using a large language model to obtain the final personalized research topic recommendation list.

2. The research topic recommendation method based on heterogeneous information networks and large language models according to claim 1, characterized in that, S1 includes: Based on historical paper data obtained from academic databases, scholars, papers and their keywords are extracted as three types of entities: themes. Based on the extracted three types of entities, a heterogeneous information network of node sets is constructed, including scholar nodes, paper nodes, and topic nodes. Based on the associations between the three types of entities, an edge set is constructed in the heterogeneous information network. The edge set includes: scholar-paper edges representing scholars publishing papers, and paper-topic edges representing the topics contained in the papers.

3. The research topic recommendation method based on heterogeneous information networks and large language models according to claim 1, characterized in that, S2 includes: Based on the text content of nodes in the network, obtain the content feature vector of the nodes; Based on the structure of heterogeneous information networks, obtain the structural feature vectors of nodes; Based on content feature vectors and structural feature vectors, the final embedded representation of nodes is obtained through feature fusion and heterogeneous information aggregation.

4. The research topic recommendation method based on heterogeneous information networks and large language models according to claim 1, characterized in that, S3 includes: Based on the final embedding representations of scholars and topics, scholar-topic feature pairs are concatenated. Based on scholar-topic feature pairs, a pre-trained machine learning classifier is used to predict the probability that there is a link between scholars and topics; Based on the predicted probabilities, the Top-K unseen topics with the highest probabilities are selected for each scholar to form a preliminary set of candidate recommended topics.

5. The research topic recommendation method based on heterogeneous information networks and large language models according to claim 3, characterized in that, The steps to obtain the content feature vector of a node include: Based on the title text and abstract text of the paper node, the title feature vector and abstract feature vector are obtained by encoding using a pre-trained language model, respectively. For scholar nodes and topic nodes, based on their association with paper nodes, the content feature vectors of scholar nodes and topic nodes are obtained by aggregating the content feature vectors of associated paper nodes.

6. The research topic recommendation method based on heterogeneous information networks and large language models according to claim 3, characterized in that, The steps for obtaining the structural feature vector of a node include: Based on heterogeneous information networks, node sequences are generated through random walks; Based on the node sequence, the structural embedding vector of each node is learned using a word embedding model.

7. The research topic recommendation method based on heterogeneous information networks and large language models according to claim 1, characterized in that, S4 includes: Based on the preliminary set of candidate recommended topics and the corresponding list of scholars' historical research topics, a structured prompt containing instructions for analysis, evaluation, and screening and ranking is constructed. Based on structured prompts, a large language model is driven to perform semantic reasoning and evaluation of candidate topics; Based on the output of the large language model, obtain the refined and sorted final personalized research topic recommendation list.

8. A research topic recommendation system based on heterogeneous information networks and large language models, said system being used to implement the method described in any one of claims 1-7, characterized in that, include: The module includes a data collection module, an extraction module, a filtering module, and a recommendation module. The acquisition module is used to acquire scholars, papers and topic entities based on the historical paper data of the target scholar group, and to construct a heterogeneous information network based on the association between the entities. The extraction module is used to obtain a low-dimensional dense vector representation of each node in the heterogeneous information network, which integrates semantic content and network structure. The filtering module is used to formalize the scholar-topic recommendation task into a link prediction problem based on the low-dimensional dense vector representation, and to obtain a preliminary set of candidate recommendation topics for each scholar based on the link prediction results. The recommendation module is used to perform semantic refinement and reordering based on the preliminary candidate recommendation topic set and the historical research information of the corresponding scholars, using a large language model to obtain the final personalized research topic recommendation list.

Images