HoeffdingTreeConfig

Hoeffding-bound confidence threshold. Lower means splits require more evidence.

This is the delta of the bound sqrt(-ln(delta) / 2n), so it enters logarithmically: an order of magnitude buys only a modest widening of the margin a candidate has to clear.

Multiplicative decay applied to delta per depth, which slows growth near the leaves.

Deeper leaves see less of the stream, so a fixed confidence level would let them split on proportionally thinner evidence. Shrinking delta with depth counteracts that.

Hard ceiling on tree depth.

Hard ceiling on internal plus leaf nodes.

Minimum weighted samples required on each side of a candidate split.

Minimum total weighted samples at a leaf before split evaluation runs at all.

Breiman-style random-subspace size: at every audit-leaf birth, draw a fresh random subset of this many candidates from the tree's full pool. null disables the trick and considers every candidate.

This is what decorrelates the trees in a forest. A single tree usually leaves it null.

Audit every Nth observation rather than every update.

Evaluating every candidate at every observation is the dominant cost in a VFDT, and the bound moves slowly, so checking periodically loses almost nothing.

If the Hoeffding bound itself shrinks below this, the leaf may split even when the runner-up is close: the classic VFDT tie-break parameter.

Without it, two candidates of genuinely equal merit deadlock forever, because the margin between them never exceeds any bound. This gives the tree permission to pick one once it has enough evidence to know the choice does not matter much.