A review on cost-based feature selection algorithms in the various applications of machine learning

Knowledge acquisition is the most important challenge in building an expert system in any field, and one of the sources of knowledge will be the data collected in that field. Traditionally, the data collection process is assumed to have a symmetric cost. For example, this assumption will not be acceptable in the medical due to various expenses. Designing a cost-sensitive classification and a cost-sensitive feature selection method are two approaches to considering cost factors. Cost-effective feature selection improves financial return by significantly saving feature data cost as well as limiting credit losses and this can be used in different areas, for example, computer imaging and medical diagnosis which also have a large number of features that may be irrelevant or redundant. Analysis of the research reviewed in this study shows that cost-sensitive feature selection focuses on selecting a feature subset with minimum total cost while achieving a classification accuracy that is as high as possible. The review of selected studies showed a downward trend in using heuristic methods in this field, Wrapper methods are in the first rank regarding usage in evaluation criteria, and ۷۶\% of selected studies are in the single-objective category. Most of the studies were classified in the single-label category based on the number of determined labels.

کلیدواژه ها:

Cost-based approaches ، Cost-sensitive classification ، Feature selection ، single-label data