Most Convolutional Neural Network (CNN) interpretation techniques visualize only the dominant cues that the model relies on, but there is no guarantee that these represent all the evidence the model uses for classification. This limitation becomes critical when hidden secondary cues—potentially more meaningful than the visualized ones—remain undiscovered. This study introduces CasCAM (Cascaded Class Activation Mapping) to address this fundamental limitation through counterfactual reasoning. By asking “if this dominant cue were absent, what other evidence would the model use?”, CasCAM progressively masks the most salient features and systematically uncovers the hierarchy of classification evidence hidden beneath them. Experimental results demonstrate that CasCAM effectively discovers the full spectrum of reasoning evidence and can be universally applied with nine existing interpretation methods.

This study proposes a novel forecasting framework that simultaneously captures the strong periodicity and irregular meteorological fluctuations inherent in solar radiation time series. Existing approaches typically define inter-regional correlations using either simple correlation coefficients or distance-based measures when applying spatio-temporal graph neural networks (STGNNs). However, such definitions are prone to generating spurious correlations due to the dominance of periodic structures. To address this limitation, we adopt the Elastic-Band Transform (EBT) to decompose solar radiation into periodic and amplitude-modulated components, which are then modeled independently with separate graph neural networks. The periodic component, characterized by strong nationwide correlations, is learned with a relatively simple architecture, whereas the amplitude-modulated component is modeled with more complex STGNNs that capture climatological similarities between regions. The predictions from the two components are subsequently recombined to yield final forecasts that integrate both periodic patterns and aperiodic variability. The proposed framework is validated with multiple STGNN architectures, and experimental results demonstrate improved predictive accuracy and interpretability compared to conventional methods.