Recap and final exam

Ten modules to move from an empty rating matrix to a two-tower engine evaluated by slice and A/B tested. Here is the course condensed, then the threads that run through it, then the exam.

The course at a glance

Module	The essential point
1. Framing	Recommendation is ranking, not rating prediction; explicit is rare, implicit is abundant and ambiguous
2. Neighborhood CF	Similarities on a sparse matrix; shrinkage by co-support is the cheapest big win, item-based scales farther
3. Matrix factorization	Latent factors on observed entries only; the three biases (global, user, item) explain most variance
4. Content-based	Sentence embeddings + weighted user profile; content produces a filter bubble unless mitigated
5. Hybrids	Weighted, switching, mixed, feature-enriched; normalize scores or the loudest wins
6. Two-tower	Independent user and item towers, in-batch negatives, retrieval then ranking with an ANN index
7. Cold start	Content for new items, context + short questionnaire for new users, controlled exploration
8. Metrics	recall@k, NDCG@k with the $\log_2(p+1)$ discount, coverage, diversity, novelty — always four numbers
9. Implicit feedback	Confidence weighting, inverse propensity on positions, popularity re-weighting; read curves by slice
10. Project	Temporal split, four models on one table, A/B with the novelty effect and winner's curse in mind

The threads running through the course

Sparsity is the defining constraint. Every design decision — factorization on observed entries only, shrinkage on similarities, content for cold items, in-batch negatives instead of full-catalog softmax — is downstream of the fact that a real user-item matrix is filled at well under 1 %. If you remember one sentence from ten modules, remember that one.

Ranking is the goal, not rating prediction. RMSE is a useful sanity check for a rating widget and a distraction everywhere else. Modules 1, 3, 8 and 10 all return to this. The wrong metric on the right task ships the wrong model.

Two towers, retrieval then ranking, everywhere. From module 3's matrix factorization (a linear two-tower) through module 5's LightFM (a feature-enriched two-tower) to module 6's neural two-tower, every modern architecture is a factorization of the score for reasons of serving cost. Understanding the pattern makes the rest deducible.

Implicit feedback is the training signal and it lies. Clicks correlate with position, popularity, mobile-vs-desktop, time of day and language. Modules 4, 7, 9 and 10 all name a bias to correct. A recommender trained on unadjusted implicit feedback slowly becomes its own past.

A decision grid to carry into your first project

Situation	First choice	Why
Zero interactions yet	Segmented popularity + content	Nothing else has data to model with
Under 5 interactions on this user	Content-based on onboarding topics	CF unstable, CB usable from day one
Warm user, small catalog (<10 k items)	LightFM feature-enriched, WARP loss	Best single-model on cold items
Warm user, large catalog	Two-tower + ANN retrieval + heavier ranker	Only pattern that scales to millions of items
Rating widget on a specific item	Matrix factorization with three biases, RMSE	The only surface where RMSE is the right metric
Cold new item	Any content-informed model, boosted with a decay	CF cannot score it

The final exam

The exam has 40 questions covering the ten modules: framing a ranking task, spotting shuffled or leaky splits, choosing a similarity and applying shrinkage, sizing $k$ and $\lambda$ in factorization, building a hybrid with score normalization, training a two-tower with in-batch negatives, mitigating the filter bubble, computing NDCG by hand, correcting for position and popularity bias, and reading the four-number metrics table by slice.

Many questions present a situation to diagnose: an offline metric that looks too good, a recommender that keeps proposing already-completed courses, a model whose coverage collapsed over training, a home page A/B whose first-week lift vanished by day 14. It is judgment that is assessed, not the recitation of formulas.

On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.

Before you start

For each row of the decision grid above, ask yourself "how would I see, in a metric, that the wrong choice was made?". If you can say why RMSE 0.87 can hide a bad recommender, why coverage below 30 % is a red flag even with strong recall, why a position-1 click is worth less than a position-8 click, and why the second week of an A/B is more trustworthy than the first, you are ready. Good luck.

Final exam

Ready to validate this course?

40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.

Start the exam

You need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.