<p>Portable electronic noses are increasingly explored for food quality assessment; however, high-dimensional and redundant sensor features limit their deployment in resource-constrained environments. In this study, a compact multivariate sensing framework based on metal oxide semiconductor (MOS) sensors was developed as a beer-focused case study for rapid discrimination of fourteen commercial beer products under controlled variations in sample volume, standing time, and temperature handling. A unified preprocessing strategy combined with multiscale feature extraction was implemented, followed by mutual information–based feature selection to construct compact feature representations. Approximately 86% of the original features were eliminated without compromising discriminative capability. Support vector machine, random forest, and backpropagation neural network models were comparatively evaluated under Bayesian optimization and genetic algorithms to examine the interaction between classifier structure and hyperparameter optimization within reduced feature spaces. The results demonstrate that discrimination robustness in compact representations is strongly model-dependent, and optimization strategies exhibit distinct compatibility with different classifiers. A maximum test accuracy of 0.9048 with Macro-F1 above 0.90 was achieved under optimized conditions. The proposed portable electronic nose framework enables rapid qualitative discrimination without reliance on compound-specific analytical instruments, providing a practical and cost-effective solution for food quality screening and on-site monitoring.</p>