Subject fingerprinting and task classification rely on distinct functional connectivity features.

Functional connectivity (FC) measured by functional magnetic resonance imaging (fMRI) has been shown to be a marker of individual brain characteristics, and also to reflect cognitive tasks. However, it remains unclear how the choice of FC measures affects the encoding of both subject and task properties. We address this question using a high-quality deep-phenotyping dataset consisting of multiple naturalistic tasks (listening to a story, watching three different movies and playing a video game) and resting-state, while working on two classification problems: subject fingerprinting and task classification. We compare the performance of a combination of two FC measures and three covariance estimation methods. We then examine the similarity and subject specificity of FC across tasks and with structural connectivity (SC). We find that sparse partial correlation, obtained from the Graphical-Lasso estimator, performs best in subject identification tasks and is most similar to SC; it stands out as a marker of identity. In contrast, task information is better captured by Pearson correlation measures, as they account for distributed brain activity. Overall, we find that pairwise interactions captured by partial correlation are optimal for fingerprinting, while multi-way relationships underlying full correlation are an accurate marker of cognitive function.