A comparison of unsupervised feature selection algorithms and a new entropy-based method proposal

Feature selection task is essential for Machine Learning algorithms not to be influenced by the curse of dimensionality. In this regard, feature selection methods try to address this trouble. However, feature selection methods have some deficiencies: (i) the performance of each machine learning method can be remarkably different on the selected features (ii) significant changes can also be followed in the performance of the classifiers by depending on differences in the subset of selected feature (iii) they spend a long time on huge data sets. In this thesis, to cope with the aforementioned problems, we propose a fast unsupervised feature selection algorithm, which is based on a univariate and filter approach. The proposed method jointly regards both the cumulative entropy of the distribution and the Shannon entropy calculated by the symmetry of the distribution for each feature. As a result of comparisons with cutting-edge works, the experimental results demonstrate that the presented algorithm better overcomes these problems compared to other methods.