An innovative ‘Cluster-then-predict’ Approach for improved sentiment prediction

Abstract

Sentiment analysis is a field related to data mining in which subjective information is extracted from source materials such as Twitter, blogs, newspaper articles, etc. Twitter data presents an opportunity for companies to analyze the sentiment the customers or potential users have towards its products. This paper presents an innovative sentiment prediction approach in which the data is first clustered using K-means clustering and then CART algorithm is applied to each cluster to classify the tweets as positive or negative. The results of innovative ‘cluster-then-predict’ approach directs towards an improved overall prediction accuracy with an increased collected sample data size leading to better clustering and improved classification of each cluster. Also, clustering of data provides useful insights, which helps the companies to gauge consumer sentiment. For the purpose of this paper, tweets related to ‘Windows 10’ which was launched by Microsoft on July 29, 2015; have been extracted.