Relational-View Ensemble Learning for Tabular Prediction

Abstract

Tabular prediction can benefit from relational information at multiple levels, including global feature relations, feature interactions within each sample, and relations among similar samples. Existing deep tabular methods exploit such information through attention-based, graph-based, and retrieval-based modeling. However, they often focus on only part of the relational structure or merge relational information into a shared representation, making it difficult to distinguish the predictive role of each relation type. We propose a relational-view ensemble framework for tabular prediction. Instead of simply integrating multiple relations into a single representation, the proposed model decomposes them into three distinct predictive views: global feature relations, intra-sample feature interactions, and inter-sample relations. Each view produces an independently supervised prediction logit, and the final prediction is obtained through learnable logit-level fusion. This design treats different relational axes as explicit predictive sources within a single architecture without training multiple model instances. Experiments on eight public tabular datasets show that the proposed framework achieves competitive performance against strong baselines, including gradient-boosted decision trees and recent deep tabular models, while achieving the best average rank across the evaluated methods. Ablation studies further show that the full three-view model achieves the best or tied-best performance on five of the six evaluated datasets, supporting the benefit of jointly modeling multiple relational views.