Multilevel Association Rule Mining

What is it?

Multilevel association rule mining discovers association rules at multiple levels of abstraction, using a concept hierarchy over the items (e.g., Item → Subcategory → Category). Mining only at the lowest (raw item) level often misses valuable patterns, because individual items may be too specific to have high support, while their generalized categories do.

                All Items
                /        \
          Beverages      Bakery
          /      \        /    \
      Milk      Juice  Bread  Cookies
      /   \
 Amul Milk  Nestle Milk

Rules can be mined at any level:

• Low level (too specific, low support): {Amul Milk} ⇒ {Nestle Cookies}
• High level (general, higher support): {Beverages} ⇒ {Bakery}

Approaches to Multilevel Mining

1. Uniform Minimum Support
The same min_sup threshold is used across all levels of the hierarchy.

Level 1 (General):   min_sup = 5%  ──▶  applied
Level 2 (Specific):  min_sup = 5%  ──▶  same threshold

• Pros: simple, single threshold to set.
• Cons: high-level (general) items naturally have more support, while low-level (specific) items may fail the same threshold even if they represent genuinely important niche patterns — some interesting low-level rules are missed.

2. Reduced (Level-Specific) Minimum Support
Each level of the hierarchy has its own, typically lower, minimum support threshold as you go deeper.

Level 1 (General):   min_sup = 5%
Level 2 (Category):  min_sup = 3%
Level 3 (Specific):  min_sup = 1%

• Pros: accounts for the fact that specific items are naturally rarer; lets genuinely interesting low-level patterns surface.
• Cons: more thresholds to tune; poor tuning can either flood results with noise or still miss patterns.

Within reduced support, there are two search strategies:

• Level-by-level independent — each level is mined independently regardless of results at other levels.
• Level-cross filtering by single/k-itemset — an itemset at a lower level is examined only if its corresponding ancestor itemset at the level above was found frequent (this prunes the search space significantly using apriori-style reasoning across levels).

3. Group-Based (Item-Specific) Minimum Support
Different groups of items (rather than levels) get their own custom support thresholds, letting users assign lower thresholds to specific item groups they know to be rare-but-important (e.g., luxury goods) and higher thresholds elsewhere.

Summary

Approach	Threshold granularity	Trade-off
Uniform min_sup	One threshold, all levels	Simple, but biased toward general-level rules
Reduced min_sup	One threshold per level	Captures specific-level rules, more tuning needed
Group-based min_sup	Custom threshold per item group	Most flexible, most tuning effort

In short: multilevel mining lets analysts see both broad, general trends and narrow, specific patterns by mining across a concept hierarchy — with reduced/level-specific support being the most common practical approach to avoid missing important niche rules.