This study provides managerial the utility of fairness-aware predictive analytics machine learning methods for predicting consumer preferences for treadmill products. Four models, multinomial logistics regression, Random Forest, XGBoost and LightGBM, were bench- marked on the Cardiogoodfitness dataset. Per-class accuracy, ROC-AUC, and cumulative gain analysis metrics were used to evaluate the model’s performance. SHAP and LIME address interpretability. The results of fairness analysis conducted across sex and age subgroups using true- positive and, false-positive rates, followed by subgroup specific threshold adjustments indicate Random Forest and XGBoost outperformed logistic regression, by achieving superior accuracy and class separability for TM195 and TM798, whereas TM498 remained challenging. XGBoost achieved balanced accuracies of 0.90 for TM 195 and TM498, and 1.00 for TM798. Annual income, miles, fitness, and age consistently turned out to be dominant predictors across the interpretability methods. Cumulative gain curves confirmed that by targeting the top ranked 20–30% of customers captured nearly all TM798 customers. Fairness diagnostics discovered sub-group disparities. These were reduced by post-processing adjustments. The study shows that integrating ensemble learning, interpretability, and fairness provides a reproducible framework for consumer choice analytics, highlighting the value of usage-based segmentation, equitable deployment, and transparent management trust predictions.