LightFM recommendations in production
Personalised menu sequencing
LightFM over sparse customer-item data with six months of order history, recency boosting, and scheduled reruns that keep live recommendations fresh.
6 mo
order history
(Python · LightFM · pandas · SQL · Linux cron)
# Problem
The ordering app showed every customer the same menu order. Sparse order histories and stale models tend to collapse back to “most popular”.
# Approach
Productionised a LightFM recommender over sparse customer-item matrices built from item codes, purchase frequency, and six months of history. Recency boosting lifts recent orders, and scheduled pipeline reruns refresh the live recommendations.
# Result
Personalised, regularly refreshed menu order in the app, driven by collaborative filtering instead of a hand-ranked list.