Effective Multi-Label Classification Using Data Preprocessing

Background

The related work about multi-label classification and label imbalance is presented here. For multi-label classification, there exist methods that use the transformation approach. It changes multi-label data such that methods for single-label classification can be used. Sometimes multi-label data is not modified. Thus adaptation methods modify the process of dealing with such data. There also exists an approach that ensembles multiple existing methods. CC (Read, 2009), MLkNN (Zhang & Zhou, 2007) and RAkEL (Tsoumakas et al., 2011) are examples of these three approaches respectively.

For few decades, many researchers have worked in the field of multi-label classification (Tsoumakas & Katakis, 2007) (Tsoumakas et al., 2009) (Trohidis et al., 2008) (Tsoumakas et al., 2010) (Madjarov et al., 2012) (Zhang & Zhou, 2014) (Tidake & Sane, 2018). K nearest neighbor has also been the choice of many researchers for multi-label classification. From the study, it is noticed that neighbors are obtained using only features always. In contrast, the scenario is different for data that is multi-label. Each instance belongs to a predefined set of labels. Hence it is possible to consider labels along with features for obtaining neighbors.

Zhang and Zhou discuss an approach in (Zhang & Zhou, 2007). It follows an algorithm adaptation approach. It is an improved version of the well-known nearest neighbor algorithm. Several researchers use it to perform multi-label classification. It utilizes feature similarity to determine nearest neighbors (Zhang & Zhou, 2005) (Zhang & Zhou, 2007) (Spyromitros-Xioufis et al., 2008). In the case of multi-label classification, since the instances are associated with multiple labels, label dissimilarity may also help determine a set of nearest neighbors.