LIA: A Label-Independent Algorithm for Feature Selection for Supervised Learning

摘要

The current study considers an unconventional framework of unsupervised feature selection for supervised learning. We provide a new unsupervised algorithm which we call LIA, for Label-Independent Algorithm, which combines information theory and network science techniques. In addition, we present the results of an empirical study comparing LIA with a standard supervised algorithm (MRMR). The two algorithms are similar in that both minimize redundancy, but MRMR uses the labels of the instances in the input data set to maximize the relevance of the selected features. This is an advantage when such labels are available, but a shortcoming when they are not. We used cross-validation to evaluate the effectiveness of selected features for generating different well-known classifiers for a variety of publicly available data sets. The results show that the performance of classifiers using the features selected by our proposed label-independent algorithm is very close to or in some cases better than their performance when using the features selected by a common label-dependent algorithm (MRMR). Thus, LIA has potential to be useful for a wide variety of applications where dimension reduction is needed and the instances in the data set are unlabeled.