Research On Knowledge Graph Citation Recommendation Combined With BERT And Graph Attention Networks

In the face of massive literature resources,the establishment of citation recommendation models to assist researchers in improving literature retrieval efficiency is currently a research hotspot in the field of recommendation algorithms.Traditional citation recommendation algorithms suffer from limited content analysis and cold start issues.Leveraging the powerful semantic representation and structural processing capabilities of knowledge graphs can break through the bottlenecks of traditional methods at the technical level,effectively capturing researchers’ preferences and needs,and rapidly recommending authoritative,cutting-edge,and highly relevant literature resources.Based on this,the main contributions of the paper are threefold,as follows:(1)A knowledge graph model for citation domains is established.Literature resources in the field of computer science are taken as an illustration in this paper,with citation resources as the main body and citation reference as the relational.An enhanced TF-IDF model is employed to generate candidates set of keywords,which is subsequently converted into word vectors for clustering purposes.By fusing computations obtained from multiple convolutional neural networks with varying kernel sizes,multi-scale contextual features of the paper are extracted.The attention mechanism is incorporated to enhance the node features from a global perspective selectively and accomplish the task of entity extraction.A self-attention mechanism is utilized to capture the semantic relationships between words.The obtained vectors,together with position vectors,are subsequently input into a convolutional neural network for joint computation,resulting in the final features required for classification and accomplishing the task of relation extraction.Knowledge storage is implemented and data preparation for subsequent algorithms is completed using the high-performance graph database Neo4 j.(2)A citation recommendation algorithm fusing knowledge graph and graph attention network is proposed.First,the Trans R algorithm is used to map the knowledge graph information into low-dimensional dense vectors;secondly,the graph attention network is used to aggregate the neighbor node information through the multi-channel fusion mechanism to enrich the semantics of the target node.Then,a dynamic convolution layer is introduced to dynamically aggregate neighbor node information to enhance the expression ability of the model;finally,the interaction probability between the user and the citation is calculated through the prediction layer.Through the comparative experiment analysis,the performance of the proposed algorithm is better than that of the comparative model,and the evaluation index MRR is 6.0 percentage points and 3.4 percentage points higher than the sub-optimal model NNSelect in the comparative model,and the Precision and Recall metrics are also improved to different degrees,which verifies the effectiveness of the algorithm.(3)A novel hybrid citation recommendation algorithm is proposed,which is based on the Siamese BERT and dynamic graph attention network.Firstly,to address the issue of information loss and confusion caused by BERT model’s simultaneous learning of contrasting sentences,Siamese networks and BERT model are combined to separately learn the textual input information of the citing paper and the candidate cited paper,which can extract their textual features to enrich the contextual information of citations.In order to address the issue of graph-based citation recommendation algorithms neglecting semantic features of paper texts,the high-order embedding representation of the knowledge graph is concatenated with it to obtain the final citation features,which are then fed into the prediction module to compute the interaction probability between them.Through the comparative experiment analysis,the proposed hybrid recommendation algorithm outperforms the baseline models in all evaluation metrics,which verifies the superiority of this algorithm over single-technique recommendation algorithms,and indicates a significant improvement in recommendation performance.