Researchers have proposed two new decision tree models designed to make ensemble learning better at handling shifting data streams.
Most streaming classifiers use Hoeffding Trees, which split nodes on a fixed schedule based on a statistical bound. That predictable rhythm keeps individual trees in an ensemble different from each other, which is what makes ensembles outperform single trees. A newer approach, adaptive splitting, instead waits for a change detector to flag performance decay before splitting, which reacts faster to concept drift but tends to make every tree in the ensemble split at similar points, eroding diversity. The new paper introduces two variants, both called Hoeffding Adaptive Splitting Trees, that blend the two approaches: they keep periodic splits to preserve diversity while adding change-detection triggers to catch drift early.
Concept drift is the reason streaming models degrade in production, whether that's fraud patterns, sensor readings, or user behavior shifting under a model's feet. Ensemble diversity is the main lever for keeping accuracy stable through that drift, so a base learner that improves both responsiveness and diversity at once addresses a real bottleneck rather than a cosmetic one. The paper reports state-of-the-art results across benchmark comparisons, computational cost analysis, and drift-adaptation tests.
Take "state-of-the-art" with the usual asterisk: it's a self-reported benchmark from a preprint that hasn't been peer reviewed yet.