Sentiment classification over opinionated data streams through informed model adaptation

Abstract

Opinionated data streams are very popular data paradigms nowadays as more and more users share their opinions online about almost everything from products to persons, brands and ideas. One of the key challenges for opinionated stream mining is dealing with concept drifts in the underlying stream population by building learners that adapt to such concept changes. Ageing is a typical way of adapting to change in a stream environment as it potentially allows us to discard outdated information from the learning models and focus on the most recent information. Most of the existing approaches follow a fixed ageing strategy which remains the same over the whole stream; for example, a fixed window size in the sliding window model or a fixed ageing factor in the damped window model. This implies that we forget at the same rate over the whole course of the stream, which is counterintuitive given the volatile nature of the stream. What is more intuitive is to forget faster in times of change so as to adapt to new data and to forget slower, or in other words, to remember more, in times of stability. In this work, we propose an informative-adaptation-to-change approach where we first detect changes in the underlying data stream and then we tune the ageing factor of the ageing-based Multinomial Naive Bayes (MNB) classifier based on the detected change. Except for the up-to-date classifier our method also outputs the points of change in the stream, therefore offering more insights to the final users.