E-commerce & content
AffinityGraph
Recommendations that surface the long tail
A hybrid recommendation engine combining collaborative filtering with graph signals, so suggestions lift basket size instead of repeating the bestseller list.
This example comes from the technical delivery experience behind AussieSync. It does not necessarily represent a project contracted directly through AussieSync.
Have a process like this one? Tell us about it.
01 Problem
Weak recommendations show everyone the same bestsellers, leaving the rest of the catalogue undiscovered.
- Popularity-ranked suggestions ignored what an individual customer was actually doing.
- New and niche inventory never surfaced, so it never sold.
- Pure collaborative filtering had nothing to say about a new product or a new customer.
- Recommendations had to be served inline, within the page render budget.
02 The solution
Collaborative filtering blended with graph signals over co-purchase and proximity.
- Behavioural similarity drives the collaborative half of the recommendation.
- A relationship graph adds co-purchase and shared-neighbour signals alongside it.
- Graph structure covers the cold-start cases collaborative filtering cannot reach.
- Batch precomputation plus online serving keeps responses inside the render budget.
03 Technical approach
Every choice earns its place.
- PythonModel training and the batch scoring pipeline.
- Graph storeCo-purchase and shared-neighbour relationships.
- ElasticSearchCandidate retrieval across a large catalogue.
- RedisOnline serving fast enough to sit inline in a page render.
04 Outcome
Higher basket size and engagement, and discovery of long-tail inventory.
- Higherbasket size and engagement
- Discoveredlong-tail inventory that never surfaced before
- Cold-starthandled via graph structure