We propose a feature-refinement classification head, called FCH,  that builds graphs from transfer learning–based embeddings and employs a composite loss for joint classification and relational learning. Its modular design allows for seamless integration into existing architectures, showing strong potential for performance improvement.

We analyze relational structure learning per batch under an inductive setting, showing that GNN-based subgraph representations can capture meaningful relationships. The proposed classification head enables efficient and robust representation learning with batchwise training.

We assess how adjacency matrix design affects performance and efficiency in graph-based deep networks in terms of evaluation and complexity. The results show that RBF kernel–based matrices offer structural simplicity and strong performance, making them well-suited for image classification.

Experimental results show that the adaptation of an FCH module is highly competitive compared to existing approaches in terms of accuracy, loss distribution, and visual distribution of latent vectors. The adopted GNN can effectively learn feature representations by utilizing the latent structure and the relationship between samples.