Cross-Dataset Evaluation of Compression Strategies for Fuel Efficiency Prediction

Deploying machine learning models in resource-constrained environments requires balancing predictive performance with computational and memory costs. This study investigates how different model compression strategies affect fuel-efficiency prediction across datasets with substantially different sizes and feature dimensionalities. We evaluate five approaches: principal component analysis (PCA) for feature reduction, structural pruning of Random Forests, coreset-based training-data reduction, neural-network quantization, and weight pruning. The experiments use three vehicle fuel-efficiency datasets: Auto MPG, EPA fuel-economy data, and a large Vehicles dataset, covering different sample sizes and feature representations. Performance is evaluated using root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R²), while efficiency is assessed through training time and measures of model, feature, or data reduction. Results show that compression effectiveness depends strongly on the dataset and method. Coreset-based reduction generally provides a favorable accuracy-efficiency trade-off, while PCA can substantially reduce the feature space at the cost of predictive performance. Random Forest models also provide a strong predictive baseline across the datasets. The study provides a cross-dataset comparison of compression strategies for tabular regression and highlights practical considerations for deploying machine learning systems under computational constraints.