Online FP-growth

Standard FP-growth processes a fixed, complete dataset in a single pass and builds the FP-tree once. Online FP-growth updates an existing tree as new transactions arrive, avoiding repeated full scans. The core mining logic (conditional pattern bases, recursive sub-trees) is the same; the difference lies in the incremental maintenance of the tree structure.

A practical approach is to calibrate on a representative seed batch, then monitor whether the pattern count explodes or collapses as the stream progresses. Some implementations use a relative threshold defined as a fraction of transactions seen so far; others introduce a small error bound (as in the Lossy Counting scheme) to tolerate slight under-counting near the threshold boundary.

Basic Online FP-growth accumulates all transactions seen so far, so old patterns persist even if they are no longer relevant. To handle concept drift, use a sliding-window variant (which discards transactions outside a fixed window) or an exponential decay variant (which down-weights older transactions over time).

Generally yes, for the same reason standard FP-growth outperforms Apriori: no candidate generation means far fewer scans and much lower memory pressure. The advantage is most pronounced when transaction lengths are long and the item universe is large, precisely the conditions where online Apriori generates an exponential number of candidates.

Pure Online FP-growth is less common in mainstream ML libraries than static FP-growth (available in MLlib, mlxtend, and R's arules). Stream-specific frameworks such as MOA (Massive Online Analysis) and some custom Python implementations provide incremental FP-tree variants; in production settings it is often implemented as a custom operator in a stream-processing pipeline such as Apache Flink or Kafka Streams.