Pre-evaluation Strategy of Harmfulness Caused by Class Imbalance Based on Leave-one-out Cross Validation
Sen Xu · Journal of Chinese Computer Systems · 2012
In recent years,class imbalance problem has gradually evolved into one of the hotspots in several research fields,including artificial intelligence,machine learning and data mining.At present,many practical and effective methods have been proposed to solve this problem.However,the recent research indicated that not all of the imbalanced classification tasks are harmful and conducting specifically designed class imbalance learning algorithms on those unharmful classification tasks would hardly improve and even degenerate classification performance,meanwhile it is possible to increase training time to a large extent.To solve this problem,we propose a pre-evaluation strategy to estimate the harmfulness of skewed classification tasks.The strategy acquires the classification performance of training set by leave-one-out cross validation,and then uses the obtained performance to calculate a novel index named as Harmfulness Measure(HM) in order to assess the degree of damage.The index would provide helpful information to guide us to select appropriate learning algorithm.The experimental results on eight skewed datasets verified the effectiveness and feasibility of the presented strategy.