Semi-supervised FP-growth
Also known as: SS-FP-growth, constrained FP-growth, label-guided frequent pattern mining, semi-supervised frequent itemset mining
Semi-supervised FP-growth extends the classical Frequent Pattern growth algorithm by incorporating partial labels, user-defined constraints, or class-level information to guide frequent itemset discovery. Instead of mining all patterns indiscriminately, it focuses on patterns that are both statistically frequent and semantically meaningful given the available supervision signal.
Key highlights
- Targets patterns that are both frequent and aligned with known class structure, producing a more interpretable and actionable rule set.
- Inherits FP-growth's efficiency — no candidate itemset generation, compact tree representation — so it scales to large transaction databases.
- Constraint-based pruning reduces the search space, often cutting mining time compared to unconstrained FP-growth on the same data.
- Works with very few labels or soft constraints, making it practical when full annotation is expensive.
- Outputs human-readable association rules that domain experts can audit directly.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use semi-supervised FP-growth when you have transactional or categorical data where a portion of records carry class labels or domain constraints and you want to discover association rules or frequent patterns that respect that supervision. It is well suited for market-basket analysis with product-category labels, clinical records with partial diagnoses, or text mining with a few tagged documents. Avoid it when your data is fully labelled (use supervised classification instead), when you have no meaningful constraints to inject (plain FP-growth suffices), or when data is continuous and not naturally transactional.
Strengths & limitations
- Targets patterns that are both frequent and aligned with known class structure, producing a more interpretable and actionable rule set.
- Inherits FP-growth's efficiency — no candidate itemset generation, compact tree representation — so it scales to large transaction databases.
- Constraint-based pruning reduces the search space, often cutting mining time compared to unconstrained FP-growth on the same data.
- Works with very few labels or soft constraints, making it practical when full annotation is expensive.
- Outputs human-readable association rules that domain experts can audit directly.
- Requires at least partial labelling or well-specified constraints; without these it reduces to plain FP-growth with no benefit.
- Choosing meaningful constraint types and minimum support thresholds demands domain expertise and can require iterative tuning.
- The FP-tree can still become very large for high-dimensional, sparse transaction data even with semi-supervised pruning.
- Theoretical guarantees on pattern completeness weaken when hard constraints eliminate entire branches of the search space.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How is semi-supervised FP-growth different from constrained FP-growth?
Constrained FP-growth uses hard user-specified item or pattern constraints. Semi-supervised FP-growth additionally uses class labels attached to transactions to score pattern discriminativeness — it combines constraint pruning with label-based filtering, making it a broader framework.
How many labelled transactions do I need?
There is no strict minimum, but the supervision signal is useful only when labelled transactions are representative of the class structure. Even 5–10% labelled data can meaningfully bias the search if the labels are informative and evenly distributed across classes.
What minimum support threshold should I set?
Start high (e.g., 10–20% of transactions) and lower it gradually while monitoring the number of rules and their average confidence. With semi-supervised pruning, you can often afford a lower threshold than with plain FP-growth without drowning in noise.
Can I use this method with continuous features?
Not directly. FP-growth operates on categorical or binary items. Continuous features must first be discretised (e.g., binned into low/medium/high ranges) before the algorithm can treat them as items in a transaction.
Does it guarantee finding all frequent patterns?
Semi-supervised variants that apply anti-monotone constraints preserve the completeness guarantee for the constrained pattern set. However, if soft or label-based filtering is applied post-hoc, some frequent patterns may be excluded by design — this is intentional, not a bug.
Sources
- 1.Han, J., Pei, J., & Yin, Y. (2000). Mining frequent patterns without candidate generation. Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, 1–12.
- 2.FP-growth algorithm. Wikipedia.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Semi-supervised FP-growth. ScholarGate. https://scholargate.app/machine-learning/semi-supervised-fp-growth