Catalog Home Page

Document clustering with optimized unsupervised feature selection and centroid allocation

Al-Jadir, Ibraheem (2018) Document clustering with optimized unsupervised feature selection and centroid allocation. PhD thesis, Murdoch University.

[img]
Preview
PDF - Whole Thesis
Download (2MB) | Preview

Abstract

An effective document clustering system can significantly improve the tasks of document analysis, grouping, and retrieval. The performance of a document clustering system mainly depends on document preparation and allocation of cluster positions. As achieving optimal document clustering is a combinatorial NP-hard optimization problem, it becomes essential to utilize non-traditional methods to look for optimal or near-optimal solutions. During the allocation of cluster positions or the centroids allocation process, the extra text features that represent keywords in each document have an effect on the clustering results. A large number of features need to be reduced using dimensionality reduction techniques. Feature selection is an important step that can be used to reduce the redundant and inconsistent features. Due to a large number of the potential feature combinations, text feature selection is considered a complicated process.

The persistent drawbacks of the current text feature selection methods such as local optima and absence of class labels of features were addressed in this thesis. The supervised and unsupervised feature selection methods were investigated. To address the problems of optimizing the supervised feature selection methods so as to improve document clustering, memetic hybridization between filter and wrapper feature selection, known as Memetic Algorithm Feature Selection, was presented first. In order to deal with the unlabelled features, unsupervised feature selection method was also proposed. The proposed unsupervised feature selection method integrates Simulated Annealing to the global search using Differential Evolution. This combination also aims to combine the advantages of both the wrapper and filter methods in a memetic scheme but on an unsupervised basis. Two versions of this hybridization were proposed. The first was named Differential Evolution Simulated Annealing, which uses the standard mutation of Differential Evolution, and the second was named Dichotomous Differential Evolution Simulated Annealing, which used the dichotomous mutation of the differential evolution.

After feature selection two centroid allocation methods were proposed; the first is the combination of Chaotic Logistic Search and Discrete Differential Evolution global search, which was named Differential Evolution Memetic Clustering (DEMC) and the second was based on using the Gradient search using the k-means as a local search with a modified Differential Harmony global Search. The resulting method was named Memetic Differential Harmony Search (MDHS). In order to intensify the exploitation aspect of MDHS, a binomial crossover was used with it. Finally, the improved method is named Crossover Memetic Differential Harmony Search (CMDHS). The test results using the F-measure, Average Distance of Document to Cluster (ADDC) and the nonparametric statistical tests showed the superiority of the CMDHS over the baseline methods, namely the HS, DHS, k-means and the MDHS.

The tests also show that CMDHS is better than the DEMC proposed earlier. Finally the proposed CMDHS was compared with two current state-of-the-art methods, namely a Krill Herd (KH) based centroid allocation method and an Artifice Bee Colony (ABC) based method, and found to outperform these two methods in most cases.

Item Type: Thesis (PhD)
Murdoch Affiliation: School of Engineering and Information Technology
Supervisor(s): Wong, Kevin, Fung, Lance and Xie, Hong
URI: http://researchrepository.murdoch.edu.au/id/eprint/49947
Item Control Page Item Control Page

Downloads

Downloads per month over past year