Schizophrenia (SCZ) is a severe mental disorder that impairs brain function and daily life, while its early and objective diagnosis remains a major clinical challenge due to the reliance on subjective assessments. This study aims to develop a machine learning-based framework for the auxiliary diagnosis of SCZ using multi-dimensional electroencephalogram (EEG) features and to investigate the underlying neural alterations. Resting-state EEG data were obtained from 45 male patients with pediatric SCZ and 39 age-and gender-matched healthy controls. Three types of EEG features (relative power (RP), fuzzy entropy (FuzEn), and functional connectivity (FC)) were extracted under various time window lengths and fed into four ensemble learning models. A data-driven feature selection approach (Recursive Feature Elimination) was applied to identify the most informative features, resulting in 212 most discriminative features (48 RP, 40 FuzEn, and 124 FC) out of the initial 760. Leveraging the selected features, the Categorical Boosting model achieved the highest classification accuracy of 99.60% at the 4-s window. Further analysis of the discriminative features revealed that the altered EEG characteristics were mainly in the alpha, beta, and gamma bands. Particularly, altered FCs exhibited a fronto-increase-parieto-decrease pattern mainly in the right hemisphere along with spectral-dependent RP alterations and a universally reduced FuzEn in the pediatric SCZ group. In summary, this study not only showcases the potential of advanced ensemble learning algorithms in precisely identifying pediatric SCZ, but also provides new insights into the altered brain functions in pediatric SCZ patients, which may benefit the future development of automatic diagnosis systems.