Recommendation systems · End-to-end case study
PickMyMovie
An explainable hybrid recommender that turns a few liked films into a diverse shortlist, backed by a reproducible temporal evaluation rather than an untested similarity demo.
The problem
Recommendation demos often produce plausible-looking results without showing whether the model retrieves what users actually choose next. PickMyMovie needed to work as a useful product and withstand technical scrutiny.
The constraints were deliberate: no account system, no required third-party API, a serverless-friendly artifact, understandable explanations, and honest handling of cold-start and popularity bias.
System design
Method
- Content signal: cosine similarity over inverse-frequency weighted genre vectors.
- Collaborative signal: item-to-item cosine over positive ratings, with minimum support and shrinkage to reduce unreliable small-sample similarities.
- Quality signal: a small confidence-weighted rating term that cannot dominate personalization.
- Diversity layer: a genre-overlap penalty prevents near-identical cards from filling the visible shortlist.
The blend was selected on validation users. The held-out test set was used once for the comparison below.
Evaluation
For each eligible user, the latest positive rating became the test target, the second latest became validation, and earlier ratings formed training history. The recorded run scanned 10,000 users, evaluated a deterministic sample of 1,000, and ranked within the 5,000 most-observed candidate movies.
| Strategy | Recall@10 | NDCG@10 | MRR@10 | Coverage@10 |
|---|---|---|---|---|
| Popularity baseline | 5.0% | 0.0230 | 0.0149 | 0.28% |
| Genre content | 1.2% | 0.0060 | 0.0042 | 34.14% |
| Collaborative | 9.3% | 0.0431 | 0.0284 | 18.70% |
| Selected hybrid | 8.9% | 0.0414 | 0.0272 | 20.18% |
The collaborative model achieved the strongest held-out relevance in this run. The validation-selected hybrid traded 0.4 percentage point of Recall@10 for broader catalog coverage and averaged 37.9 ms per offline ranking. With one held-out target per user, Recall@10 is equivalent to hit rate. These results apply only to the documented sample and candidate universe.
Engineering decisions
- Two streaming passes avoid loading the 678 MB ratings file into memory.
- A compact JSON artifact keeps inference dependency-free and suitable for a Python serverless function.
- The browser stores picks locally; the server receives movie IDs but stores no visitor profile.
- API contract tests, model tests, syntax checks, Docker support, and CI cover the deployable path.
- The portfolio uses a same-domain rewrite while the recommender remains independently deployable.
Limitations
The snapshot ends in 2019, so this is not a current-release catalog. Sparse and new titles receive weaker collaborative evidence. Genre metadata cannot capture tone, cast, plot, or context, and offline implicit-feedback metrics do not prove user satisfaction.
Those limitations are kept visible instead of being hidden behind an unsupported performance claim. The next rigorous experiment would repeat the protocol across multiple user samples and confidence intervals before making a broader claim.