Digitalni repozitorijum
Univerziteta u Beogradu
{{ mc.sd.lang | uppercase }}
Comparison of Feature Selection Methods for Biomedical Data Classification
Đorđević, Nataša
Pljasković, Aldina
Marovac, Ulfeta
Dolićanin, Zana
Memić, Lejlija
Fetahović, Irfan
Abstract. Prior to applying machine learning techniques to huge datasets with large number of features, it is crucial to identify and select the most significant ones in order to reduce the size of the dataset and improve model’s performance. Feature selection offers numerous advantages, such as: lowering the price of classifier model training, reducing the model’s size and making classification models easier to understand. Filter, wrapper, and embedded methods are three general sets of techniques used for feature selection. The filter methods use a rating process to evaluate each feature’s importance before removing features with poor scores. It has been discovered that the filter approaches are quick, scalable, computationally easy, and independent of the classifier. Wrapper methods generate a set of candidate feature subsets and then employ a classification algorithm to evaluate them. Compared to filter methods, wrapped approaches usually provide more accurate results, but are computationally more expensive. There are, however, a number of alternative strategies as well as hybrid ones that do not fit into any of these three categories. In recent years, the application of metaheuristic algorithms has been proposed as a method for solving feature selection problem. In this research, we used both filter and wrapper-based metaheuristic approach for biomedical data classification. The obtained results demonstrate that applying feature selection improves the model’s performance, or at least provides the same results while reducing the size of the data set and making data collection easier.
engleski
2024
© All rights reserved
Keywords: feature selection, biomedical data classification, machine learning.