Effectiveness of Basic and Advanced Sampling Strategies on the Classification of Imbalanced Data. A Comparative Study Using Classical and Novel Metrics

摘要

The imbalanced class problem is noteworthy given its impact on the induction of predictive models and its constant presence in several application areas. It is a challenge in supervised classification, since most of classifiers are very sensitive to class distributions. Consequently, the predictive model is biased to the majority class, which leads to a low performance. In this paper, we analyze the reliability of resampling strategies through the influence of some factors such as dataset characteristics and the classifiers used for building the models, in order to improve the performance and determine which resampling method will be used according to these factors. Experiments over 24 real datasets with different imbalance ratio, using six different classifiers, seven resampling algorithms and six performance evaluation measures have been conducted aiming at showing which resampling method will be the most suitable depending on these factors.