An arXiv preprint reports a method called DICS that is designed to speed classification-tree training by shrinking the set of candidate splits. In synthetic tests, the DICS-based tree model CGCT was reported as 8–22 times faster than a classical decision tree, while DICS-based ensembles also showed large reported speedups.
A smaller set of choices at each split
Rather than search every possible threshold, DICS forms candidate classification splits from clustering boundaries and adjusts those boundaries using the spread around the cluster centers. The aim is to shrink the candidate split set before the tree chooses among them.
In the paper’s operation-count analysis, DICS has a reported split-finding cost of O(mPU), compared with O(NPU) for exhaustive search and O(HPU) for histogram search. That is a theoretical comparison of the amount of work, not a guarantee that every implementation or computer will deliver the same wall-clock speed.
Synthetic tests delivered the biggest gains
On synthetic data, CGCT was reported 8–22 times faster than a classical decision tree, with a marginal accuracy trade-off. CGRF was 8–30 times faster than a random forest; FastC-GBM was 3.6–27 times faster than LightGBM and 5–50 times faster than XGBoost, with reported accuracy differences within 0.02.
The synthetic simulations were repeated 10 times, except for BDTKS, which was run once because of its high training cost.
Real datasets were less uniform
The real-data benchmark table covered 4,601 to 200,000 samples, 27 to 3,072 features and 2 to 100 classes.
On those datasets, CGCT was approximately 2–21 times faster than a classical decision tree, but its accuracy decrease was slightly larger than in the synthetic tests. DICS-based random-forest training was approximately 2.3–12.8 times faster than standard random-forest training, with an accuracy drop of up to 0.02.
FastC-GBM was up to 9.5 times faster than LightGBM and 18.75 times faster than XGBoost; Santander was the only stated exception. BDTKS was slower and unstable in the real-data tree comparisons.
Speed and split quality are different measures
At maximum depth 8 on CIFAR-10, CGCT’s cumulative gain was 0.123 versus 0.127 for exhaustive decision-tree search, retaining 97.18% of the exhaustive gain. That score is a model-internal split-quality measure, not a direct test-accuracy result.
Under the paper’s stated assumptions about node size and data density, its theorem writes the gap between the optimal DICS and decision-tree gains as O(1/N) and describes the gap as vanishing as N increases. Across 40 independent runs, the root-level gain gap fell from 0.0039 to 0.0011 as N grew from 1,000 to 20,000. The result does not establish that full trees will make identical predictions.
What the preprint leaves open
The method was evaluated for classification, with regression left as future work. The authors attributed the Santander boosting exception to implementation-level memory management issues, adding uncertainty to that comparison.
Taken together, these results support a narrower conclusion: DICS may reduce training cost in the tested classification settings, while its reported speedups still need to be checked across other implementations and hardware.
Paper data and sources
Original title: DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
Authors: MD Saifur Rahman Mazumder, Feng Yu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text