Ensemble of Heterogeneous Classifiers for Improving Automated Tweet Classification

Twitter is widely used by businesses to communicate with and obtain feedback from their customers, almost in real time. Automated analysis is necessary to deal with the large volumes of tweets in a timely manner, and an insightful classification is a first step in this analysis. This paper presents an ensemble method that combines classifier outputs by including the individual probability distributions with an additional learning layer, as opposed to an arithmetic combination of weighted probabilities. Besides using probabilities generated by individual classifiers, we also show that the use of tweet vectors helps with the ensemble learning step. In addition, we have described a mapping method to employ a cloud taxonomy service as an additional classifier. We include LLDA, Naive Bayes, and the cloud taxonomy service for the ensemble method, and have applied our methods on a real industry tweet dataset. The proposed ensemble model is able to outperform several widely used probability based ensemble methods, i.e., Weighted Sum and Product of Experts (PoE). The ensemble model is also more adaptive than previous models in handling the variations in the probability distributions output from the individual classifiers. In addition to the algorithms used and the improved results, this paper's contribution is in insights from the application of models on a unique, real dataset.