Hybrid Feature Selection using Shapley Value and ReliefF for Medical Datasets

The medical databases are composed of vast amount of data. Increment in data volume has led to a massive amount of high-dimensional medical data made available to the public on the Internet. These large amounts of medical data can be put into good use through knowledge discovery by identifying knowledge that is useful via data mining. These high-dimensional data are often associated with redundant features removal. A range of information theoretic methods have been deployed in selecting the most viable and relevant feature sets, which have led to reduction in the size of data. Nonetheless, these methods have mostly failed in identifying the significance of each feature derived from the sets of features. An exceptional feature set not only decreases computational time and cost, but also enhances classifier accuracy in classification. As such, this study proposes a feature selection technique based on filter-wrapper technique using the ReliefF-Shapley Value hybrid. The ReliefF filter method was applied in the early stage stage to determine the accuracy of a feature in discriminating among classes. Next, the reduced set of features yielded from ReliefF was passed to the wrapper-based Shapley Value. In the wrapper method, Shapley Value was employed to add weight, and later, to assess each attribute based on the assessment standards. The outcomes were assessed using UCI-derived five medical datasets. The proposed method was able to yield competitive outcomes for most datasets.