Assessing training data reduction in dry bean classification: performance and efficiency

Reducing the amount of training data can decrease computational requirements, but it may also affect classification ‎performance. ‎This study evaluates the effect of reducing training data on predictive performance and measured training time for Logistic Regression and Random Forest, using the dry bean dataset. The dataset comprises 13,611 samples, ‎‎16 numerical features and 7 classes. The training proportion (20%, 40%, 60%, 80%, and 100%) of the remaining training pool was tested, and a fixed 20% of the dataset was reserved for testing. The accuracy and Macro-F1 were used for the performance assessment, while the training time was measured. ‎Logistic Regression showed only slight performance variation across the evaluated training levels. At 20%, it achieved 91.99% Accuracy and 93.24% Macro-F1, 0.116 percentage points lower than the full-data case, with a 76.17% reduction in mean training time. The performance changes were larger for the Random Forest model at the lower training levels, ‎from 93.43% with the full training set to 92.45% at the 20% training level, as seen for the Macro-F1 model. The results indicate that the impact of lower training ratios depends on the chosen classifier, and that training times and predictive performance differ significantly between the two classifiers used in this study.