Decision Trees Aren't Trained. They're Grown.

Short summary

An SRE explains decision trees as locally-grown, greedy algorithms with no global optimizer — each split minimizes impurity but the whole tree is unstable and can change completely with small data shifts. Trees assume the world is made of discrete boxes, making them intuitive for operational rule systems but risky when treated as explanations. The author recommends versioning tree structure like configuration and diffing it between retrains to catch when the model's 'explanation' of reality silently changes.

• •Decision trees are grown greedily, not optimized globally — each split is locally optimal but the overall tree is unstable
• •Fewer assumptions mean higher variance: small data changes can produce entirely different trees
• •Version and diff tree structure between retrains, especially when the tree is used as an explanation for stakeholders

Generated with AI, which can make mistakes.