Researchers have found a faster way to grow decision trees that stay both deep and accurate.
The method, described in a paper posted to arXiv, tackles a well-known trade-off in building decision trees. Methods that guarantee the mathematically best tree tend to work only on small, shallow trees with features chopped into yes-or-no splits. Faster heuristic methods can handle bigger, deeper trees and continuous-valued features, but they settle for good-enough predictions rather than the best possible one. This approach splits the difference: it solves the top of the tree with an exact search technique called branch-and-reduce, approximates the rest with quick greedy heuristics similar to a lookahead rollout in reinforcement learning, then loops back over the whole tree in a low-cost "moving-horizon" pass to sharpen accuracy.
Decision trees matter because, unlike neural networks, a person can actually read the rules a tree uses to make a decision, which counts for something in fields like healthcare or lending where regulators want an explanation, not a black box. The bottleneck has always been that the most interpretable, provably-optimal trees don't scale past small, shallow problems, forcing practitioners to choose between trustworthy-but-shallow models and accurate-but-opaque ones. If this approach holds up outside the paper's own benchmarks, it narrows that gap without demanding a bigger compute budget.
It's still an arXiv preprint, not a peer-reviewed benchmark war, so treat "near-optimal" as a claim worth retesting on your own data before swapping out your gradient-boosted forest.