Association rule-based outlier detection flags categorical records that violate expected associations. It mines frequent item sets, forms rules X → Y, and identifies rows where X occurs without the expected Y. section 7.10
Two scoring perspectives are available:
- Count rules that hold; records satisfying fewer rules appear more outlierish. section 7.10
- Count exceptions where an antecedent is present but its consequent is absent; this approach produces interpretable rule-based explanations. section 7.10
- Combine it with FPOF when records containing few or zero frequent item sets also matter, because exception testing can skip them. section 7.10
Categorical preference data
One-hot-encode SpeedDating preference ranges, mine frequent item sets with minimum support 0.3, and generate rules with confidence at least 0.7. A row matching attractive [21–100] but not ambitious [0–15] violates an association rule.
