Subspace constraint clustering for semi- supervised sparse data

The complexity of many scientific fields has brought about its impact on not only the amount of generated data but also the nature of the data itself. Nowadays, it is common to deal with high-dimensional datasets. Processing and analyzing this type of datasets require extreme attention as several persistent problems can be faced. Grouping or partitioning data into a number of several sets is one of the applications widely used in data mining and machine learning. This process is referred to by clustering and it provides an idea on the underlying knowledge presented in the dataset. One of the most commonly used clustering algorithms is k-means, due to its simplicity and effectiveness in normal circumstances. However, considering the nature of high-dimensional data, k-means seems to fail in obtaining the desired results. This is mainly caused by the curse of dimensionality and possible noise within the data, which hinder the outliers of the produced clusters. As a possible solution, dimensionality reduction for sparse data has been proposed. Dimensionality reduction can be obtained on both space and subspace bases for clustering applications. However, subspace dimensionality reduction has attracted more attention in recent studies due to its robustness and clusterability improvement. Thus, in this research, it has been decided to investigate soft subspace clustering for sparse data. Four algorithms have been chosen, based on their mechanism orientation, to compare their performances with k-means in clustering hard-to-cluster datasets. A Graphical User Interface (GUI) has been implemented using MATLAB 2018b to run all clustering processes and to produce both numerical and visual results. It has been found that the performance of k-means can be dramatically improved through incorporating soft subspace computing. Particularly, ReliefF algorithm, in this study, could improve the clustering performance to more than 50% in some cases.