Personalisation & ML · ML
Lift paying-rate on a games platform with collaborative filtering
+15%
A games platform already owns the strongest recommendation signal there is: what players do. This engagement turns behaviour logs into item embeddings and an offer-timing model, so the store stops showing everyone the same bundle.
The process
Assemble the behaviour log
Sessions, purchases, item interactions exported from your own database — no third-party data enters the model.
Clean and encode
Identifiers reconciled, bots and refunds stripped, events encoded into player and item vectors.
Train collaborative filtering
Similar players surface items they converge on; cold-start items fall back to content features.
Model the timing
A second model learns when a player is receptive — session depth, streak state, time since last purchase.
Serve and measure
Recommendations served through your existing store API behind a holdout split; paying-rate read against the holdout.
What we need from you
- Behaviour and purchase logs (any queryable form)
- Item catalogue with attributes
- A store surface we can vary per player
What you get
- Trained recommendation and timing models, owned by you
- Serving integration or an export your engineers wire in
- Holdout report on paying-rate movement
Timeline
Four to eight weeks depending on data readiness.