A Weighted Approach for Feature Selection in High-Dimensional and Incomplete Data

A recently encountered challenge in data science and more specifically in machine learning is growing amount of data. Exclusion of superfluous data and thereby focusing on essential variables, known as feature selection, proves vital for model performance optimization. This study undertakes a thorough investigation into pivotal feature selection strategies. The benefits and limitations of each method is clearly stated. An advanced methodology is also presented for tackling incomplete datasets, alongside introducing an innovative hybrid model that unites the Partial Mutual Information Criterion (PMIC), state-of-the-art null value completion strategies and neural network synergies to improve feature selection processes. Moreover, a newly defined benchmark, rational success weight is defined to certify the efficiency of our proposed algorithm in selecting relevant features. Finally, the suggested strategy is implemented in Python and numerical test results are reported on randomly generated data sets.