A simple yet effective feature completion model for relation-level learning in missing data

Junyang Chen, Liang-Jie Zhang, Zhelin Huang, Han Liu, Yirui Wu, Huan Wang, Wei Liu, Victor C. M. Leung

Abstract: Relation-level learning through network embedding plays a pivotal role in network analysis. However, current embedding models often assume the availability of all associated node features during both training and testing, a strong assumption not always met in practical scenarios. Nodes possess diverse relation levels when interacting with neighbors, demanding the incorporation of node-associated features for effective relation-level learning. Yet, many existing models struggle with nodes containing missing data, be it edges or features, as they tend to either ignore or statically assign values to such nodes. Consequently, embedding models relying on incomplete data often yield suboptimal results. To address this challenge, we propose the Feature Completion with Attention-based Aggregation (FCAA) model for relation-level learning. FCAA explicitly models edges and features by leveraging an attention-based aggregation mechanism to learn node relations. Specifically, we adopt an inductive learning approach to capture feature interactions between node pairs, thus enhancing the handling of missing data. Experimental findings underscore the efficacy of FCAA, showcasing substantial improvements over state-of-the-art methods in link prediction and vertex classification tasks.

External IDs:dblp:journals/inffus/ChenZHLWWLL26