Explainable Galaxy Interaction Prediction with Hybrid Attention Mechanisms
Sathwik Narkedimilli, Satvik Raghav, Om Mishra, Mohan Kumar, Aswath Babu H, Tereza Jerabkova, Manish M, Sai Prashanth Mallellu
TL;DR
This work tackles explainable galaxy interaction prediction by designing an attentive neural ensemble that fuses AG-XCaps, a Hybrid Self-Attention Network (H-SNN), and ResNet-GRU, trained on the Galaxy Zoo DESI dataset and augmented with LIME explanations. The approach achieves high predictive metrics ($P=0.95$, $R=1.00$, $F1=0.97$, $A=96 ext{%}$) and substantially reduces false positives compared with a Random Forest baseline, while maintaining a compact model footprint (~0.45 MB). By coupling capsule-based spatial features, self-attention, and sequential modeling on autoencoder-transformed representations, the method offers robust performance for large-scale surveys like Euclid and LSST and provides interpretable links to physical processes via LIME. The work also demonstrates strong ablation results and analyzes efficiency, scalability, and robustness, outlining practical deployment and avenues for future enhancements such as multimodal data fusion and SHAP-based explanations.
Abstract
Galaxy interaction classification remains challenging due to complex morphological patterns and the limited interpretability of deep learning models. We propose an attentive neural ensemble that combines AG-XCaps, H-SNN, and ResNet-GRU architectures, trained on the Galaxy Zoo DESI dataset and enhanced with LIME to enable explainable predictions. The model achieves Precision = 0.95, Recall = 1.00, F1 = 0.97, and Accuracy = 96%, outperforming a Random Forest baseline by significantly reducing false positives (23 vs. 70). This lightweight (0.45 MB) and scalable framework provides an interpretable and efficient solution for large-scale surveys such as Euclid and LSST, advancing data-driven studies of galaxy evolution.