A Novel Ensemble Learning Approach Based on Choquet Integral Using Multi Assessment of Classifiers for Breast Cancer Prediction
Soumita Guria, Chandrima Debnath, Debashree Guha, Aneek Adhya · 2024
Breast cancer has emerged as the principal cause of death among women across the globe. Early prediction of cancer remains a significant challenge for both doctors and researchers. Recently, there has been a development of automated diagnostic systems for the early diagnosis of breast cancer. For this purpose, a number of classifier fusion (classifier ensemble) approaches have been proposed, one of which is the Choquet integral-based classifier fusion. These techniques utilized accuracy as a fuzzy measure to assess the importance of each classifier combination, to improve the effectiveness of the diagnostic process. In medical decision making, other parameters such as specificity, sensitivity and area under the curve (AUC) also play an important role. This observation has motivated us to develop a novel method of classifier ensemble approach by incorporating the Choquet integral, in which the fuzzy measure values for classifiers are computed based on multiple criteria: accuracy, specificity, sen-sitivity, and AUC by leveraging the principles of multi-criteria decision-making (MCDM). This process effectively handles un-certainty and imprecision inherent in classifier outputs. In this framework, a heterogeneous classifier set is considered to train the model, and consequently, an individual fuzzy measure value is generated for each classifier during this training phase. These fuzzy measure values are then used to combine the classifiers using the Choquet integral in the test phase. This classifier fusion method incorporates the uncertainty inherent in decision scores, which is normally absent in traditional ensemble methods. The effectiveness of the proposed Choquet-based ensemble algorithm is evaluated utilizing the Breast Cancer Wisconsin dataset. Then the outcomes derived from the experimental analysis are compared with traditional methodologies, as well as the other methods available in existing literature.