Hierarchical classification of data streams: a systematic literature review

The classification task usually works with flat and batch learners, assuming problems as stationary and without relations between class labels. Nevertheless, several real-world problems do not assume these premises, i.e., data have labels organized hierarchically and are made available in streaming fashion, meaning that their behavior can drift over time. Existing studies on hierarchical classification do not consider data streams as input of their process, and thus, data is assumed as stationary and handled through batch learners. The same can be said about works on streaming data, as the hierarchical classification is overlooked. Studies concerning each area individually are promising, yet, do not tackle their intersection. This study analyzes the main characteristics of the state-of-the-art works on hierarchical classification for streaming data concerning five aspects: (i) problems tackled, (ii) datasets, (iii) algorithms, (iv) evaluation metrics, and (v) research gaps in the area. We performed a systematic literature review of primary studies and retrieved 3,722 papers, of which 42 were identified as relevant and used to answer the aforementioned research questions. We found that the problems handled by hierarchical classification of data streams include mainly classification of images, human activities, texts, and audio; the datasets are mostly created or synthetic data; the algorithms and evaluation metrics are well-known techniques or based on those; and research gaps are related to dynamic context, data complexity, and computational resources constraints. We also provide implications for future research and experiments to consider common characteristics shared amongst hierarchical classification and data stream classification.

     

Adaptive random forests for evolving data stream classification

Random forests is currently one of the most used machine learning algorithms in the non-streaming (batch) setting. This preference is attributable to its high learning performance and low demands with respect to input preparation and hyper-parameter tuning. However, in the challenging context of evolving data streams, there is no random forests algorithm that can be considered state-of-the-art in comparison to bagging and boosting based algorithms. In this work, we present the adaptive random forest (ARF) algorithm for classification of evolving data streams. In contrast to previous attempts of replicating random forests for data stream learning, ARF includes an effective resampling method and adaptive operators that can cope with different types of concept drifts without complex optimizations for different data sets. We present experiments with a parallel implementation of ARF which has no degradation in terms of classification performance in comparison to a serial implementation, since trees and adaptive operators are independent from one another. Finally, we compare ARF with state-of-the-art algorithms in a traditional test-then-train evaluation and a novel delayed labelling evaluation, and show that ARF is accurate and uses a feasible amount of resources.