A Hybrid Feature Subset Selection Method Based On Grasp And Relief

Official URL: https://risc01.sabanciuniv.edu/record=b3205825

The abundance of complex, high-dimensional data in various fields has amplified the necessity for effective feature selection strategies. This thesis proposes an innovative hybrid Feature Subset Selection technique to identify and retain the most valuable features, thereby enhancing model interpretability and combating the curse of dimensionality. Our method uniquely merges the computational efficiency of filter techniques and the precision of wrapper methods, explicitly combining the metaheuristic algorithm Greedy Randomized Adaptive Search Procedure (GRASP) and a reputable feature filtering algorithm Relief. The process initiates with a comprehensive exploration of feature combinations, subsequently applying various filter techniques, amongst which Relief exhibited superior performance. Additionally, Relief was integrated into the construction and improvement stages of GRASP. Experiments are conducted by checking the average 30 runs of K-Nearest Neighbors scores and time. The results underscore the potency of the hybrid approach, significantly improving model performance and demonstrating the potential of integrating filter and wrapper methods for efficient feature selection in high-dimensional datasets. This contribution allows us to maximize the accuracy of the machinelearning model while minimizing the time dedicated to feature selection.