Abstract
The brain is a dynamic system where complex behaviours emerge from interactions across distributed regions. Accurately linking brain function to cognition requires methods sensitive to these interactions. We introduce Feature Similarity (FS), which integrates a broad set of interpretable time-series features-such as covariance, temporal dependencies, and entropy -to move beyond traditional single-metric approaches. FS captured functional brain organization: regions within the same network showed greater similarity than those in different networks, and FS identified the principal gradient from unimodal to transmodal cortices. Compared with Pearson correlation-based functional connectivity (FC) and 46 out of 49 statistical pairwise interaction metrics (SPIs), FS demonstrated greater sensitivity to task modulation. Critically, FS revealed a task-dependent double dissociation in the Dorsal Attention Network, interacting more strongly with the Visual network during working memory but with the default mode network during long-term memory. FS thus provides a powerful tool for uncovering task-specific brain network interactions.