Lightweight Alternatives for Hyper-parameter Tuning in Drifting Data Streams

Scenarios dealing with data streams often undergo changes in data distribution, which ultimately lead to a performance degradation of algorithms learning from such data flows (concept drift). This phenomenon calls for the adoption of adaptive learning strategies for algorithms to perform resiliently after a change occurs. A multiplicity of approaches have so far addressed this issue by assorted means, e.g. instances weighting, ensembling, instance selection, or parameter tuning, among others. This latter strategy is often neglected as it requires a hyper-parameter tuning process that stream learning scenarios cannot computationally afford in most practical settings. Processing times and memory space are usually severely constrained, thus making the tuning phase unfeasible. Consequently, the research community has largely opted for other adaptive strategies with lower computational demands. This work outlines a new perspective to alleviate the hyper-parameter tuning process in the context of concept drift adaptation. We propose two simple and lightweight mechanisms capable of discovering competitive configurations of learning algorithms used for data stream classification. We compare its performance to that of a modern hyper-parametric search method (Successive Halving) over extensive experiments with synthetic and real datasets. We conclude that our proposed methods perform competitively, while consuming less processing time and memory.