Recommendation Systems

Impression logs and time-based splits you can trust, baselines worth beating, retrieval and ranking models, diversity and business rules, and online tests that prove a recommender helps users and the business.

Free · Mac (Apple silicon) and Windows · v1.0.5 · Other platforms

Then open the Marketplace in the app, search for Recommendation Systems, and click Install. It goes into your own files, and works offline from then on.

What it installs.

1

Pillar

1

Area

50

Projects

201

Starter tasks

📊 Data & Analytics

Recommendation Systems

Building systems that suggest products, content or actions to users, balancing relevance, diversity and business goals, and testing their effect.

  1. Recommendation surface inventory and slot goals
  2. Impression and interaction logging for recommendations
  3. Item catalogue and metadata readiness check
  4. Popularity and co-occurrence baselines to beat
  5. Time-based offline evaluation split
  6. Offline ranking metric suite
  7. Online success metric and guardrails for recommendations
  8. Cold start inventory for new users and new items
  9. Business rules and exclusion filters for recommendations
  10. Weekly hand review of sampled recommendation lists
  11. Monthly offline benchmark rerun
  12. Candidate retrieval index health checks
  13. Catalogue coverage and long-tail exposure watch
  14. Weekly impression log health check
  15. Recommender test log and readout archive
  16. Merchandising boosts and pinned items register
  17. Not interested and complaint feedback triage
  18. Quarterly recommender roadmap review
  19. Serving latency budget and fallback lists
  20. Matrix factorisation for implicit feedback
  21. Item-to-item similarity for product page rows
  22. Two-tower retrieval model
  23. Learning to rank for the ranking stage
  24. Session-based and sequential recommendation
  25. Content-based recommendations from text and image embeddings
  26. Off-policy evaluation from logged recommendations
  27. Recommender systems reading programme
  28. Practice on public recommendation datasets
  29. Build versus buy a recommendation engine
  30. Retrieval and ranking architecture decision
  31. Diversity and novelty re-ranking
  32. Blending business objectives into the ranking score
  33. Calibrated recommendations matching user interests
  34. Explanations shown alongside recommendations
  35. Exploration slots with bandits for new items
  36. Nearest neighbour search library selection
  37. Negative sampling strategy comparison
  38. First online test of the new recommender
  39. Launch readiness review for a new recommendation surface
  40. Peak season recommendation readiness
  41. Post-mortem of a harmful recommendation incident
  42. Recommender research paper submission
  43. First recommender for a small catalogue shop
  44. Reproducing a published recommender result
  45. Inheriting a legacy recommendation pipeline
  46. Marketplace recommendations fair to buyers and sellers
  47. News and media recommendations with freshness
  48. Long-term value and slate optimisation with reinforcement learning
  49. Graph neural network recommenders
  50. Feedback loop and filter bubble simulation study

Where it starts.

Each project arrives with its purpose, milestones and first tasks written out. Here is how the first three read.

Recommendation surface inventory and slot goals

Before choosing an algorithm, list every place the product shows suggestions: the home feed, the product page carousel, the basket add-on row, the email digest and the empty search result. Each slot has a different job, such as discovery, completing a basket or bringing a lapsed user back, and one model tuned for clicks will serve some of them badly.

First tasks

  • Screenshot every place the product currently suggests items to users
  • Pull last month's impression counts for each recommendation slot
  • Write one goal and one guardrail metric per surface with its owner
  • Rank the surfaces by traffic and by the gap between goal and current logic

Impression and interaction logging for recommendations

Most teams log clicks and purchases but not what was shown, in which position and by which model version, and without impressions you cannot tell an ignored item from one never seen. Fixing the event schema now is far cheaper than discovering a year from now that none of the history supports training, position bias correction or off-policy evaluation.

First tasks

  • List the fields your current click events carry and which are missing
  • Add request ID, position and model version to the impression event
  • Join a day of clicks to impressions and measure the join rate
  • Document the event schema in the team's repository

Item catalogue and metadata readiness check

Content-based features, cold start fixes and business filters all depend on the item catalogue, and catalogues are rarely as clean as product teams believe. Measuring how many items have a category, description, image, price and live availability flag shows which approaches are possible now and which need catalogue work first.

First tasks

  • Export the item catalogue with every metadata field to a working table
  • Calculate fill rates for each field across live items
  • Decide how product variants group into one recommendable item
  • Send the ranked list of catalogue gaps to the catalogue owner

Make it yours.

Install Recommendation Systems, then tell Aurum the date you're aiming for. It checks whether the plan fits the weeks you actually have.

Free · Mac (Apple silicon) and Windows · v1.0.5 · Other platforms

Blueprints are general information to help you organise your own work. They are not medical, financial, legal or other professional advice. See the Terms.