Model Selection Using Graph Neural Networks

This paper tackles the problem of selecting the optimal models (algorithms and their hyperparameters) for a structured classification problem using Graph Neural Networks (GNNs). Recent efforts in this direction associate statistical meta-features describing the problem with the performance of predefined models. However, the predictive power of these meta-features is insufficient while being expensive to compute. The approach presented in this paper encodes each problem as a granular knowledge graph where nodes denote prototypes, while edges capture their distance. Moreover, nodes are labeled with the most popular class in their neighborhood, and their quality is quantified with a purity score. The adjacency-based representations of these knowledge graphs establish positive arrows between close prototypes that belong to different decision classes. Therefore, solving the multilabel model selection problem consists of predicting the set of optimal models for a given dataset represented by its adjacency-based matrix knowledge graph. The results indicate that the proposed GNN-based meta-classifier can predict an optimal model for 92% of the datasets, suppressing the need to extract low-level features.