Market-basket analysis in Python (with a free POS dataset)

Get the dataset

Generic random transactions are useless here — if items are independent, no rules emerge. This dataset is built from affinity groups (chips + salsa + soda, diapers + wipes + baby food…), so association-rule mining returns genuine results.

Download market-basket dataset (CSV) → Customize in the generator

Each row is one item in a basket; group by transaction_id to reconstruct baskets.

Steps

1. Install & load

   pip install pandas mlxtend
   
   import pandas as pd
   df = pd.read_csv("retail_pos.csv")
   df.head()

2. Reshape into a basket matrix

   Turn the long table into one row per transaction, one column per product, with 1/0 for presence:

   basket = (df.groupby(["transaction_id", "product"])["quantity"]
               .sum().unstack().fillna(0))
   basket = (basket > 0).astype(int)
   basket.shape

3. Run Apriori for frequent itemsets

   from mlxtend.frequent_patterns import apriori, association_rules
   
   itemsets = apriori(basket, min_support=0.02, use_colnames=True)
   itemsets.sort_values("support", ascending=False).head(10)

4. Generate association rules

   rules = association_rules(itemsets, metric="lift", min_threshold=1.0)
   rules = rules.sort_values("lift", ascending=False)
   rules[["antecedents", "consequents", "support", "confidence", "lift"]].head(10)

5. Interpret the results

   Support = how often the combo appears. Confidence = P(consequent | antecedent). Lift > 1 = the items co-occur more than chance — a real association. You should see the seeded affinity groups rise to the top (e.g. salsa → tortilla chips with high lift). Use these for cross-sell, store layout, or recommendation demos.

Why it works on this data

Because baskets are assembled from real affinity groups with the occasional impulse buy, the joint distribution has genuine structure. Apriori and FP-Growth surface lift you can defend — unlike random transaction generators where every rule has lift ≈ 1 and the exercise falls flat.

Keep going