Estimating Shortcut Learning and Inductive Bias in Multi-view Behavioural Trace Classification

Heterogeneous behavioural data admit multiple valid representations, but different representations may expose different statistical structure and favour different inductive biases. Despite this, behavioural classification systems are typically evaluated primarily through predictive performance, with limited attention to whether different models rely on the same explanatory evidence. This work studies the relationship between representation, inductive bias, and explanation stability in a multi-view behavioural classification benchmark. We compare heterogeneous model families under a shared nested-validation protocol and investigate whether their explanations converge towards a common discriminative structure. To address this question, we introduce Information Concordance (IC) and Selector-Adjusted Information Concordance (SAIC), two metrics designed to quantify whether cross-model explanation agreement reflects shared behavioural structure rather than artefacts of the learning pipeline. We evaluate the proposed framework on a multi-view ransomware benchmark comprising syscall telemetry, DLL/path artefacts, hardware-counter entropy summaries, and a combined representation. Experiments show that syscall telemetry provides the most stable evidence across model families, whereas the other views are more sensitive to representation and fusion choices. More generally, the results suggest that agreement across heterogeneous inductive biases provides stronger evidence of meaningful structure than predictive performance alone.