Schizophrenia has been considered as a dysconneciton syndrome, which means the disintegration, or over interaction between brain regions may underlie the pathophysiology of this disease. Noninvasive techniques like functional magnetic resonance imaging (fMRI) were utilized to test this hypothesis. However, there is no consensus on which brain areas and which functional network is related with it, mostly due to the small sample size of previous studies. Supervised machine learning techniques are able to examine fMRI connectivity data in a multivariate manner and extract features predictive of group membership. This technique requires large sample sizes and results from small sample study may not generalize well. By applying a multi-task classification framework to large size multi-site schizophrenia resting functional MRI (rsMRI) dataset, we were able to find consistent and robust features. We observed that schizophrenia patients had widespread deficits in the brain. The most informative and robustly selected functional connectivity (FC) features were between and within functional networks such as the default mode network (DMN), the fronto-parietal control network (FPN), the subcortical network, and the cingulo-opercular task control network (CON). Our finding validated the dysconnection hypothesis of schizophrenia and shed light on the details of the impaired functional connectivity.