Class imbalance learning (CIL), which aims to addressing the performance degradation problem of traditional supervised learning algorithms in the scenarios of skewed data distribution, has become one of research hotspots in fields of machine learning, data mining, and artificial intelligence. As a postprocessing CIL technique, the decision threshold moving (DTM) has been verified to be an effective strategy to address class imbalance problem. However, no matter adopting random or optimal threshold designation ways, the classification hyperplane could be only moved parallelly, but fails to vary its orientation, thus its performance is restricted, especially on some complex and density variable data. To further improve the performance of the existing DTM strategies, we propose an improved algorithm called CDTM by dividing majority training instances into multiple different density regions, and further conducting DTM procedure on each region independently. Specifically, we adopt the well-known DBSCAN clustering algorithm to split training set as it could adapt density variation well. In context of support vector machine (SVM) and extreme learning machine (ELM), we respectively verified the effectiveness and superiority of the proposed CDTM algorithm. The experimental results on 40 benchmark class imbalance datasets indicate that the proposed CDTM algorithm is superior to several other state-of-the-art DTM algorithms in term of G-mean performance metric.

Learning from imbalanced data is a challenging task in the machine learning field, as with this type of data, many traditional supervised learning algorithms tend to focus more on the majority class while damaging the interests of the minority class. Stacking ensemble, which formulates an ensemble by using a meta-learner to combine the predictions of multiple base classifiers, has been used for solving class imbalance learning issues. Specifically, in the context of class imbalance learning, a stacking ensemble learning algorithm is generally considered to combine with a specific sampling algorithm. Such an operation, however, might suffer from suboptimization problems as only using a sampling strategy may make it difficult to acquire diverse enough features. In addition, we also note that using all of these features may damage the meta-learner as there may exist noisy and redundant features. To address these problems, we have proposed a novel stacking ensemble learning algorithm named MSFSS, which divides the learning procedure into two phases. The first stage combined multiple sampling algorithms and multiple supervised learning approaches to construct meta feature space by means of cross combination. The adoption of this strategy satisfied the diversity of the stacking ensemble. The second phase adopted the whale optimization algorithm (WOA) to select the optimal sub-feature combination from the meta feature space, which further improved the quality of the features. Finally, a linear regression classifier was trained as the meta learner to conduct the final prediction. Experimental results on 40 benchmarked imbalanced datasets showed that the proposed MSFSS algorithm significantly outperformed several popular and state-of-the-art class imbalance ensemble learning algorithms. Specifically, the MSFSS acquired the best results in terms of the F-measure metric on 27 datasets and the best results in terms of the G-mean metric on 26 datasets, out of 40 datasets. Although it required consuming more time than several other competitors, the increment of the running time was acceptable. The experimental results indicated the effectiveness and superiority of the proposed MSFSS algorithm.

There have been many skewed cancer gene expression datasets in the post-genomic era. Extraction of differential expression genes or construction of decision rules using these skewed datasets by traditional algorithms will seriously underestimate the performance of the minority class, leading to inaccurate diagnosis in clinical trails. This paper presents a skewed gene selection algorithm that introduces a weighted metric into the gene selection procedure. The extracted genes are paired as decision rules to distinguish both classes, with these decision rules then integrated into an ensemble learning framework by majority voting to recognize test examples; thus avoiding tedious data normalization and classifier construction. The mining and integrating of a few reliable decision rules gave higher or at least comparable classification performance than many traditional class imbalance learning algorithms on four benchmark imbalanced cancer gene expression datasets.