Explainable machine learning classification of \textit{Chandra} X-ray sources: SHAP analysis of multi-wavelength features

Machine learning can reliably classify astronomical objects when confidence levels are tracked: The study identified 54,770 robustly classified X-ray sources with high confidence thresholds (over 3σ), demonstrating that ML classification becomes trustworthy when systems quantify their certainty rather than treating all predictions equally.

Explainability techniques reveal which data features actually matter for decisions: Using SHAP analysis, researchers identified specific decision boundaries—such as infrared-optical features for distinguishing active galactic nuclei from stars—showing that understanding why an AI makes a classification is as important as the classification itself.

AI systems can work with incomplete information when trained on complete data: The study found that decision boundaries between object types could be estimated using limited multiwavelength data, suggesting ML models learn robust patterns that don't require all available information for each individual prediction.

Transparency in AI enables better science: By explaining which features the model relied on, researchers could validate whether the ML system was using scientifically meaningful relationships rather than spurious correlations, increasing confidence in discoveries made with automated classification.

Responsible AI requires investigating feature interactions, not just individual features: The research examined how multiple features work together in the model's decision-making, recognizing that understanding isolated feature importance is insufficient for ensuring reliable and interpretable AI systems.