ABSTRACT

Document clustering is an automatic unsupervised machine learning technique that aimed at grouping related set of items into clusters or subsets. The target is to create clusters with high internal coherence, but different from each other substantially. Simply, items within the same cluster should be highly similar, while maintaining high dissimilarity with items within other clusters. Automatic clustering of documents has played a very significant role in many fields including data mining and information retrieval. This thesis aimed to improve the overall efficiency of a document clustering technique using N-grams and efficient similarity measure. The thesis improves the purity and accuracy of the obtained clusters. The preprocessing method is based on N-grams (sequence of N consecutive characters) which do not give consideration to stop-words or other special punctuations but creates and overlap among the content of a document which further gives room to ignore errors thereby increasing the quality of the clusters to a great extent. This approach clusters the news articles based on their N-grams representation, thereby reducing noise and increase the probability of occurrences of the sequences within the articles document. The proposed clustering technique has parameters which can be changed accordingly at the document representation level in order to improve the efficiency and quality of the generated clusters. The results from the experiment using R programming environment were carried out on real datasets of the Reuters21578 and 20Newsgropus proved the effectiveness of the proposed clustering technique at different levels of N-grams in terms of the accuracy and purity of the generated clusters. The results also showed that the proposed clustering technique perform averagely better than the baseline technique both in terms of accuracy and purity with a best results when the window of N-grams = 3.