Integration Of Unsupervised Clustering Algorithm And Supervised Classifier For Pattern Recognition

• Main Author: Leong, Shi Xiang
• Format: Thesis
• Language: English
• Published: 2017
• Subjects: TK1-9971 Electrical engineering. Electronics. Nuclear engineering
• Online Access: http://eprints.usm.my/39416/1/Leong_Shi_Xiang_24_Pages.pdf; http://eprints.usm.my/39416/
• Institution: Universiti Sains Malaysia

In a real world, pattern recognition problems in diversified forms are ubiquitous and are critical in most human decision making tasks. In pattern recognition system, achieving high accuracy in pattern classification is crucial. There are two general paradigms for pattern recognition classification which are supervised and unsupervised learning. The problems in applying unsupervised learning/clustering is that this method requires teacher during the classification process and it has to learn independently which may lead to poor classification. Whereas for supervised learning method, it requires teacher or prior data (i.e. large, prohibitive and labelled training data) during classification process which in real life, the cost of obtaining sufficient labelled training data is high. In addition, the labelling is time consuming and done manually. To solve the problems mentioned, integration of unsupervised clustering algorithm and the supervised classifier is proposed. The objective of this research is to study the performance/capability of the integration between both unsupervised and supervised learning. In order to achieve the objective, this research is separated into two phases. Phase 1 is mainly to evaluate the performance of clustering algorithm (K-Means and FCM). Phase 2 is to study the performance of proposed integration system which using the data clustered to be used as train data for Naïve Bayes classifier. By adopting the proposed integration system, the limitation of the unsupervised clustering method can be overcome and for supervised learning, the labelling time can be reduced and more training examples are labelled which can be used to train for supervised classifier. As the result, the pattern classification accuracy is also xii increase. For examples, after applying the proposed integration system, the classification accuracy of Fisher’s Iris, Wine and Bacteria18Class has been increased from 88.67% to 96.00%, from 78.33% to 83.45% and from 93.33% to 94.67% respectively as compared to only used unsupervised clustering algorithm. The result has shown that the proposed integration system could be applied to increase the performance of the classification. However, further study is needed in the feature extraction and clustering algorithms part as the performance of the pattern classification is still depending on the data input.