Textual Graph Contrastive Learning for Enhanced Dataset Recommendation

Ammar Qayyum, Bassamtiano Renaufalgi Irnawan, Sheng Xu, Zihao Hu, Fumiyo Fukumoto, Yoshimi Suzuki

Abstract: Given the current rapid expansion of available datasets, the necessity for a recommender system has arisen. This is because manually sifting through hundreds of datasets might be an inefficient use of time. Existing dataset recommender systems did not emphasize Graph Neural Networks (GNNs), which we believe can exploit academic paper-dataset relationships for the recommendation. In this paper, we propose a dataset recommender system that employs graph structures and textual features, utilizing contrastive learning to enhance performance and assess its efficacy with existing datasets. In our methodology, we conceptualize the academic article and the dataset as graph nodes, with their interrelationship represented as edges. We employed the bipartite graph of this relationship in conjunction with contrastive learning as the graph structure. Subsequently, it is integrated with the textual features enhanced by the SciBERT pre-trained language model (PLM), which serves as the fine-tuner for the model. The model is evaluated using 17,397 academic papers and 461 datasets accessible on Paper With Code websites. The results indicate that the integration of graph structure, textual attributes, and contrastive learning enhances the system's overall performance. The code is available at https://github.com/Maqif14/Textual_GCL.git