Forward classification on data streams

Abstract

In this paper, we explore a new research problem of predicting an incoming classifier on dynamic data streams, named as forward classification. The state-of-the-art classification models on data streams, such as the incremental and ensemble models, fall into the retrospective classification category where models used for classification are built from past observed stream data and constantly lag behind the incoming unobserved test data. As a result, the classification model and test data are temporally inconsistent, leading to severe performance deterioration when the concept (joint probability distribution) evolves rapidly. To this end, we propose a new forward classification method which aims to build the classification model which fits the current data. Specifically, forward classification first predicts the incoming classifier based on a line of recent classifiers, and then uses the predicted classifier to classify current data chunk. A learning framework which can adaptively switch between forward classification and retrospective classification is also proposed. Empirical studies on both synthetic and real-world data streams demonstrate the utility of the proposed method. © 2014 Springer International Publishing.