CLIO Core

Docs · Machine learning

fim: frequent itemsets and association rules.

Mines every itemset whose support is at least min_support, plus association rules, with the same output as Spark MLlib's FPGrowth. Candidates of size L are generated on the host from the frequent (L-1)-itemsets and counted in one paged pass per level.

Module
ml_fim
Header
clio_ml/frequent_itemsets.h
Rows
one 0/1 column per item

What it does

Counts are integers, so the result is exact and independent of paging or block order. Itemsets are returned sorted by size and then by items. Rules have a single consequent item.

Rows

One transaction per row with one 0/1 column per item. clio_ml.COLUMNS["fim"] is 0.

Python

python
import numpy as np, clio_ml as cm
T = (np.random.default_rng(1).random((4096, 8)) < 0.4).astype(np.float32)

r = cm.fit_fim(T, min_support=0.1, min_confidence=0.6, max_size=4)
for s in r.itemsets[:5]:
    print(s.items, s.count)
for rule in r.rules[:5]:
    print(rule.antecedent, "->", rule.consequent, rule.confidence, rule.lift)
  • fit_fim(X, **options) calls MineFrequentItemsets.
  • Options (FrequentItemsetOptions): min_support (0.05, fraction of transactions), min_confidence (0.6, for rules), max_size (4), threads (256).
  • Result FrequentItemsetResult: itemsets (list of Itemset with ascending items and count), rules (list of AssociationRule with antecedent, consequent, confidence, lift), stats.

C++

cpp
struct FrequentItemsetOptions {
  float min_support = 0.05f;
  float min_confidence = 0.6f;
  u32 max_size = 4;
  u32 threads = 256;
  ClusterExchange cluster;
};
struct Itemset { std::vector<u32> items; u64 count = 0; };
struct AssociationRule {
  std::vector<u32> antecedent;
  u32 consequent = 0;
  double confidence = 0, lift = 0;
};
struct FrequentItemsetResult {
  std::vector<Itemset> itemsets;       // sorted by (size, items)
  std::vector<AssociationRule> rules;
  VecStats stats;
  std::vector<float> Flatten() const;
};
FrequentItemsetResult MineFrequentItemsets(Dataset &ds, const FrequentItemsetOptions &opt);
FrequentItemsetResult MineFrequentItemsetsPlain(const float *data, u64 n_rows, u32 n_items,
                                                u64 elems_per_page, u32 nblocks,
                                                const FrequentItemsetOptions &opt);
FrequentItemsetResult MineFrequentItemsetsReference(/* same arguments */);

The ChiMod is module ml_fim.

Related: sequential patterns, LDA.