Skip to main content

Rank recommendations

Goal: for each customer, the top 3 products they are most likely to order in the next 30 days ("buy it again").

The query​

LIST_DISTINCT predicts a set of linked IDs; RANK TOP K turns it into a ranking task:

PREDICT LIST_DISTINCT(orders.product_id, 0, 30, days) RANK TOP 3
FOR EACH customers.customer_id

Run it​

ds = relativedb.from_dataframes(
    {"customers": customers, "products": products, "orders": orders},
    links=[("orders", "customer_id", "customers"),
           ("orders", "product_id", "products")])

df = ds.predict(query, anchor_time=t0)

The result contains a ranked list of product IDs per customer. Note orders.product_id is an FK — the ranking works over graph edges (Row.parents), never over ID feature values.

Notes​

  • K bounds the returned list, not the candidate set.
  • Use CLASSIFY instead of RANK TOP K for a multilabel-style yes/no per item.
  • A complete self-checking version (habitual staple ranked #1 per customer) lives at examples/industry/pzn_buy_it_again.py.