Comparative Study of Clustering Algorithms for Nepali News

Abstract

Clustering is an important technique to separate data categories based on their feature similarity. Clustering belong to unsupervised type of machine learning algorithms. Among many clustering algorithms, three representative algorithms namely K-means, X-means and Expectation Maximization are experimented for the Nepali news clustering problem in this research work. News clustering is the task of categorizing news into groups that share similar interests. Clustering algorithms are evaluated for optimal performances based on cluster evaluation metrics and execution time. Evaluation metrics used are Dunn index, DB index and CH index. Execution time includes clustering time and training time. TF-IDF is used as a news embedding representation. Algorithms are also evaluated with reduced feature dimensions by applying PCA. To select the winner algorithm and setting the values of DB index, training time and clustering time must be lower and value of CH index and Dunn index must be higher. So, based upon the evaluation results, we conclude the winning algorithm and strategies in some states as follows. When feature dimension is high (>= 10000) K-Means perform better then others. When applied PCA to reduce feature space, EM algorithm better performs than others. With reduced feature space, K-Means still performs better then X-Means clustering algorithm.