Twitter Data Clustering on issues of Children with Special Needs using Hybrid Topic Models with Multi-viewpoints Similarity Metric

Abstract

Social networks are an excellent source for users to share or exchange information ontopics. Twitter is the most prioritized social network concerning the issues of children withspecial needs related topics of social users. Extracting good quality of topics from twittercorpus depends on the quality of text pre-processing and in finding optimal clustertendency. With traditional topic models, cluster tendency identification is difficult becausethey use less frequent words in tweets. In traditional topic models, k value (number ofclusters) decided manually and used Euclidean distance metric in most methods andcosine distance metrics in some methods. Proper Visualization of cluster tendency is alsoessential as corpus consists of a large number of documents and billions of words. In thispaper, hybrid topic models with multi-viewpoints based similarity metric proposed toVisualize topic clouds, to find cluster tendency of various topics related to issues ofchildren with special needs twitter datasets. Experimental evaluation and comparison ofthese proposed hybrid models done with other distance metrics. Empirical analysisperformed with convergence speed and computational complexities. Cluster validity ofproposed models done with external validity indices to quantify the quality of cluster andwith internal validity indices to evaluate clustering structure. Visual Non-MatrixFactorization (VIS NMF) under multi-viewpoints similarity metric performed well thanother models with a more informative assessment.