Integrating the optimal classifier set for sentiment analysis

This work proposed a three phase solution of multiple classifiers for sentiment analysis, in which an optimal set of classifiers is selected and integrated automatically and outperforms not only the best single classifier methods, but also the state-of-art competitors of ensemble learning.

Abstract

Automatic identification of users' sentiment is important for many Web applications, such as recommender systems and business intelligence. Sentiment analysis can be treated as a classification task, which tries to identify the user's overall sentiment expressed in documents. But it is difficult for users to select a classifier for a special analyzed domain, since each classifier would achieve various performances in different domains. Thus, we proposed a three phase solution of multiple classifiers for sentiment analysis, in which an optimal set of classifiers is selected and integrated automatically. An approximate algorithm is designed to tackle the Combinatorial Explosion Problem of classifier set selection, which can be proven to be 2-approximation. At last, extensive experiments carried out on real-world datasets show that the proposed solution outperforms not only the best single classifier methods, but also the state-of-art competitors of ensemble learning.