A Novel Graph-based Framework for Explainable Image Classification: Features That Matter

The efficacy of any machine learning model is largely contingent on the quality of the features used for training. Hence, the extraction of robust and discriminative features from raw data is a critical step. This task, however, presents significant challenges. Although modern deep learning models are very advanced, they are often criticized for their black-box nature. The predictions made by these models are not readily interpretable in terms of the features that influenced them. In our research, we propose a novel framework that innovatively combines the prowess of convolutional layers for feature extraction with robustness of Graph Neural Networks (GNNs) to model the relationships of neuron activations for better interpretability. The proposed architecture initially generates features to produce class-based neuron activations, these activations are then incorporated into graph structure. The GNN incorporates the relationship between neuron activations to produce final classifications. The proposed model provides explainability as in the predictions can be traced back to the specific neurons that contributed to them. The proposed model not only matches the accuracy of state-of-the-art models but also provides explainability through target class-specific feature importance.