Winnowing Wrapper Method: An Efficient Feature Selection Technique

Feature selection is a critical task in machine learning, big data, and data mining, aiming to enhance model performance by identifying the most relevant features. It is widely used in data- and information-driven domains. Working with irrelevant and redundant features, however, reduces model performance and can lead to incorrect predictions. This paper introduces a novel Winnowing Wrapper Method (WWM) inspired by traditional winnowing techniques used in grain preparation. In this method, a window from a sliding window of features is selected, and unique pairs are generated for all combinations. The performance of these pairs is then evaluated to identify low-, medium-, and high-performing pairs and finally eliminates features with high occurrences on the irrelevant feature list based on a specified threshold. This approach enables us to examine the interactions and dependencies between each feature pair within the dataset. Thus, without using any optimization or hybridization methods, WWM ensures efficient identification of important features, reduces computational complexity compared to existing wrapper methods, and provides balanced performance on both high-dimensional and large datasets. The efficiency of this method is demonstrated through its application to benchmark datasets, including “Stroke Prediction” and “Heart Disease,” showcasing improved execution time with balanced performance for various performance metrics.