An Experimental Study on Supervised Learning and Random Forest Feature Selection for Edible and Toxic Mushroom Classification

This paper presents an experimental study on data-driven classification using machine learning techniques with a focus on feature selection for improving predictive performance. The study utilizes a structured benchmark dataset consisting of 8124 instances with 23 categorical features to classify samples into edible and poisonous categories. To enhance model efficiency and accuracy, a random forest (RF)-based feature selection method was applied to identify and retain the most informative attributes. Three supervised learning algorithms, logistic regression (LR), naïve Bayes (NB), and decision tree (DT), were implemented and evaluated both before and after applying RF-based feature selection. Accuracy, precision, and recall were the standard criteria used to evaluate performance. The experimental results show that all classifiers achieved improved performance after feature selection, with the decision tree model obtaining the highest classification accuracy. The findings demonstrate that integrating ensemble-based feature selection with traditional classifiers significantly enhances classification performance and model robustness, particularly for high-dimensional categorical datasets. This study highlights the effectiveness of feature-driven learning frameworks for reliable and efficient data-driven classification systems.