Local Optimization Policy for Link Prediction via Reinforcement Learning

Link prediction effectively recovers missing and undiscovered link structures in a graph, enhancing researchers' ability to comprehend the generation mechanisms and evolutionary processes of the graph. Graph Neural Networks (GNNs) address link prediction tasks by aggregating complex structures and features within a specified scope. However, determining the optimal aggregation scopes for nodes in different graph-structured data poses challenges in terms of complexity and time consumption. Handcrafted or expert-based aggregation scopes require significant computational resources and involve high complexity. To address these challenges, in this paper, we propose exploring diverse information aggregation scopes for individual nodes to enhance the performance of GNNs. We introduce the Local Optimization Policy (LOP) to jointly learn the creation of the GNNs and the link prediction task. LOP adaptively learns the aggregation scope of each node through deep reinforcement learning and utilizes the learned aggregation scopes to construct the GNNs. Furthermore, we introduce the virtual node and edge features to enhance the performance of link prediction. Experimental results on multiple datasets demonstrate the superior performance of LOP compared to baselines, providing evidence for the feasibility, effectiveness, and reliability of combining GNNs and deep reinforcement learning.