Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder with profound genetic basis. Machine learning approaches applied to genomic data analysis show promise for patient diagnosis and stratification, but they are highly sensitive to class imbalance. Several methods have been developed to address class imbalance. They vary in complexity, ranging from rudimentary oversampling to sophisticated generative models. However, their relative usefulness in genomic classification of ASD remains to be elucidated. This paper propose a first class imbalance benchmark ASD microarray data set, featuring 25 balancing techniques of varying complexity. Our research focuses on six on six categories of methods: (i) oversampling, (ii) undersampling, (iii) Hybrid approaches (iv) Ensemble techniques, (v) Generative methods and (vi) Algorithmic approaches. The pipelines utilized on the GEO Datasets with ANOVA and mRMR feature selection, stratified nested cross-validation, and Random Forest classifier. Our study proved hybrid approaches as most effective, particularly SMOTE + ENN, which attained the best ROC–AUC (0.90), PR–AUC (0.94) and Recall (0.94), which confirms their superiority over the rest of the approaches. Classical oversampling approaches retained their competitiveness, while lower-performing generative models were unstable within scenarios of limited sample sizes.