Abstract Resting-state functional connectivity (rsFC) is increasingly used to develop biomarkers for psychiatric disorders. Despite progress, development of the reliable and practical FC biomarker remains an unmet goal, particularly one that is clinically predictive at the individual level with generalizability, robustness, and accuracy. In this study, we propose a new approach to profile each connectivity from diverse perspective, encompassing not only disorder-related differences but also disorder-unrelated variations attributed to individual difference, within-subject across-runs, imaging protocol, and scanner factors. By leveraging over 1500 runs of 10-minute resting-state data from 84 traveling-subjects across 29 sites and 900 participants of the case-control study with three psychiatric disorders, the disorder-related and disorder-unrelated FC variations were estimated for each individual FC. Using the FC profile information, we evaluated the effects of the disorder-related and disorder-unrelated variations on the output of the multi-connectivity biomarker trained with ensemble sparse classifiers and generalizable to the multicenter data. Our analysis revealed hierarchical variations in individual functional connectivity, ranging from within-subject across-run variations, individual differences, disease effects, inter-scanner discrepancies, and protocol differences, which were drastically inverted by the sparse machine-learning algorithm. We found this inversion mainly attributed to suppression of both individual difference and within-subject across-runs variations relative to the disorder-related difference by weighted-averaging of the selected FCs and ensemble computing. This comprehensive approach will provide an analytical tool to delineate future directions for developing reliable individual-level biomarkers.