3rd Place Solution to KDD Cup 2024 Task 2: Large Academic Graph Retrieval via Multi-Track Message Passing Neural Network

Yu Li, Taile Chen, Jingxin Hai, Hongbin Pei, Yixuan Zhou, Yong Fu, Tieying Zhang, Jing Tao

Keywords: academic graph mining, Information retrieval

Abstract: In the realm of academic research, accessing precise and useful information from huge literature is paramount for scholars. Traditional academic question answering (AQA) systems are based on dual-tower models, which, however, often falter when dealing with complex citation networks. To overcome these challenges, this paper proposes Multi-Track Textual Graph Retriever, a novel model that harnesses the power of a paper classifier, a multi-track textual graph retriever, and a dynamic retrieval task training strategy. Our model not only integrates text features and citation relationships into a cohesive framework but also employs a dynamic sampling strategy based on hard negative mining. This strategy dynamically refines the training process, leading to significant improvements in the accuracy and efficiency of academic paper retrieval. Finally, the proposed model achieved high-quality retrieval results and demonstrated competitive performance, the 3rd place, in the KDD CUP 2024 Task 2 competition. The code is available at https://github.com/liyu199809/PineappleHouse.

Submission Number: 30