Network of Experts: Learning from Evolving Data Streams Through Network-Based Ensembles

Heitor Murilo Gomes, Albert Bifet, Philippe Fournier-Viger, Jones Granatyr, Jesse Read

Abstract: Ensemble classifiers are a promising approach for data stream classification. Though, diversity influences the performance of ensemble classifiers, current studies do not take advantage of relations between component classifiers to improve their performance. This paper addresses this issue by proposing a new kind of ensemble learner for data stream classification, which explicitly defines relations between component classifiers. These relations are then used in various ways, e.g., to combine the decisions of component models. The hypothesis is that an ensemble learner can yield accurate predictions in a streaming environment based on a structural analysis of a weighted network of its component models. Implications, limitations and benefits of this assumption, are discussed. A formal description of a network-based ensemble for data streams is presented, and an algorithm that implements it, named Network of Experts (NetEx). Empirical experiments show that NetEx’s accuracy and processing time are competitive with state-of-the-art ensembles.