Odor is a critical sensory indicator for evaluating the quality and freshness of aquatic products， with volatile organic compounds （VOCs） playing a central role in odor formation. To enhance the standardization and automation of odor molecule classification in Osteichthyes， this study proposes a machine learning-based framework integrating chemical structural descriptors and computational modeling. A total of 180 representative VOCs were selected from the FlavorDB database， and both two-dimensional and three-dimensional molecular descriptors were extracted. Hierarchical clustering using the Ward method was applied to group the compounds into three odor categories： plant-like （fresh）， complex aroma， and pungent/offensive. Subsequently， five conventional classification models were constructed， and a stacked ensemble model was developed using Gradient Boosting Decision Trees （GBDT） and Random Forest （RF） as base learners. The ensemble model achieved a classification accuracy of 89% on the test set and 82% on real tilapia samples collected at different storage stages， analyzed via Gas Chromatography–Ion Mobility Spectrometry （GC-IMS）. Confusion matrix analysis and Linear Discriminant Analysis （LDA） demonstrated the model's ability to distinguish between odor types， although the recall rate for complex aroma compounds was relatively low. Feature importance analysis revealed that molecular weight， charge distribution， and spatial configuration were the primary influencing factors. Overall， this study provides a structural descriptor-based modeling approach for odor classification in fish， and verifies its feasibility in practical applications. However， further improvements are needed in model robustness and coverage of diverse odor types.