CORRELATION-BASED FEATURE SELECTION WITH BAG-BASED FUSION SCHEME FOR MULTI-INSTANCE LEARNING APPLICATION

Multi-Instance Learning (MIL) classifies a bag of instances rather than an individual instance. There is a lack of consideration of feature selection in MIL. The large number of features that are irrelevant and redundant in MIL affects the classification performance. Besides that, a genuine label of instance is unknown in MI data and evaluation of relevancy using bag class label cannot be done directly. To address this gap, this paper proposes a Fusion Bag-based Correlation Feature Selection (FBC-FS) technique using multiple bag summarization to accommodate MI data in an effort to increase performance in MI classification. The proposed technique consists of three steps: feature transformation, feature evaluation using the bag correlation and fusion of candidate features. The FBSFS is evaluated based on the MI dataset (Breast Cancer and Tiger image) with a standard Support Vector Machine, K-Nearest Neighbour (KNN) and Decision Tree. The superior result achieves up to 91.5% AUC when using KNN for the Breast Cancer dataset and the improvement achieves up to 16% with proposed FS compared without performing FS task. The results also proved that correlation measures in evaluating relevance and redundancy criteria with extended parameters to find optimal features contribute highly to the improvement of the classification performance.