Revealing the Relationship between Task Types and Query Methods through a Comparative Analysis of Latent Space with KNN and Direct Queries Methods

Abstract

Research based on knowledge graphs has been a hot topic in various fields. Users can either perform direct queries using the rich semantic information provided by the knowledge graph or search through the latent space to complete their tasks. However, existing research lacks a comprehensive comparison between these two query methods. This thesis aims to clarify the relationships between these query methods and different task types in the paper domain. Therefore, we redefine a graph task classification for the paper domain, which enhances the understanding of user intent in task categorization. We present a Node2Vec or GraphSAGE with KNN model for queries in the latent space. We evaluate the recall of the direct query's link prediction algorithm and the Node2Vec or GraphSAGE with KNN models on the Cora and Movielens datasets. The experimental results show that Node2Vec or GraphSAGE with KNN outperforms the link prediction algorithm. Furthermore, the diversity of query results is assessed in both the paper and movie domains. The results reveal that direct querying performs better for Adjacency, Accessibility by Links, Common Connection, Nodes Attribute, and Hybrid tasks, while the latent space with KNN method is better for the Explore task. These findings fill a gap in current research and enhance the effectiveness of user queries. Additionally, we design a novel dynamically updated paper recommendation system, improving the explainability of recommendation results.