THE IMPACT OF FEATURE SELECTION AND DATA PRE-PROCESSING ON ML MODELS

This research aims to identify issues with datasets that affect the training of machine learning models. The paper demonstrates the impact of feature processing on the model evaluation metrics. Initially, the paper evaluates unprocessed categorical features based on customized rules. The results of the seven evaluated algorithms show very low metric values. The demonstration is conducted on a dataset with 16 features for predicting depression among students. The dataset consists of 7,022 records. The issue of depression was chosen because it involves many features, which facilitate the analysis of the correlations between them. Moreover, the large number of features and records allowed for the analysis of generalization capacity by training on different dataset scenarios for all seven algorithms. The demonstration shows that data pre-processing generates better results when the inputs are exclusively numerical. Subsequently, the research demonstrates the importance of individually analyzing the contribution of each parameter within the model. The study encompasses three categories of tests, which are implemented using the ML.NET framework in the C# programming language.