Recommendation Systems — Interview Grill

45 questions on collab filtering, matrix factorization, two-tower, sequential models, ranking, evaluation. Drill until you can answer 30+ cold.

A. Foundations

1. What does a recommender output? Ranked list of top-$k$ items for a (user, context) input.

2. Common business targets? CTR, conversion, dwell time, retention, revenue. Pick one or weighted combination.

3. User-user vs item-item CF? User-user: similarity between users via shared items. Item-item: similarity between items via shared users. Item-item is Amazon's classic.

4. Why is CF data sparse? Most user-item pairs unobserved (a typical user interacts with $\ll 1\%$ of catalog).

5. What's the cold-start problem? New users / new items have no interaction data. CF can't help; need content features.

B. Matrix factorization

6. Matrix factorization core idea? $R\approx U{V}^{\top }$. Low-rank approximation of user-item interaction matrix.

7. Loss for explicit ratings? ${\sum }_{(u,i)}(R_{ui}-u_u^{\top }{v}_i{)}^2+\lambda (\Vert U{\Vert }^2+\Vert V{\Vert }^2)$.

8. How do you handle implicit feedback? Weighted ALS, BPR (pairwise ranking), or pointwise BCE on observed clicks vs sampled negatives.

9. What's BPR loss? Bayesian Personalized Ranking: $-\log \sigma (s(u,i^{+})-s(u,i^{-}))$. Pairwise; encourages observed > unobserved.

10. Strengths of MF? Simple, fast, scales, captures latent factors automatically.

11. Weaknesses? Cold start, no side features, static (no sequence).

C. Two-tower retrieval

12. Two-tower architecture? User encoder $f_u$ and item encoder $f_i$ → fixed-dim embeddings. Score = $u^{\top }v$.

13. Why do encoders need to be independent? To pre-compute item embeddings offline and serve via ANN. If user-item features mixed, every (user, item) pair would need explicit forward pass.

14. Training objective? Contrastive: positive pair (clicked item) → high score; negatives → low. In-batch negatives standard.

15. Hard negative mining? Include "almost-positive" negatives (e.g., items shown but not clicked). Better than random negatives at training discrimination.

16. Sampled softmax — why? Full softmax over millions of items intractable. Sample $K$ negatives from sampling distribution $Q$, then subtract $\log Q(j)$ from each sampled logit (the logQ correction) before applying softmax over sample. Yields an unbiased estimator of the full softmax gradient.

17. Inference flow? Embed user → ANN search over precomputed item embeddings → return top-$K$.

18. Why ANN, not exact? Exact KNN over millions is too slow. ANN (HNSW, IVF-PQ) trades small recall for huge speedup.

D. Sequential models

19. Why sequential models? User interests evolve. Recent clicks predict next click better than oldest clicks. Order matters.

20. GRU4Rec — what's the architecture? RNN over session of clicks. At each step, predict next item.

21. SASRec — improvement? Transformer self-attention instead of RNN. Better for long sequences.

22. BERT4Rec — twist? Masked-item prediction (BERT-style) on the sequence. Bidirectional context.

23. Modern industry sequential setup? Transformer over user history; produce user representation; combine with item features for ranking.

E. Two-stage retrieval + ranking

24. Why two-stage? Catalog ($N$) too large to rank exhaustively. Retrieval narrows to $K\ll N$. Then expensive ranking on $K$.

25. Latency budgets? Retrieval ~10ms. Ranking ~30-50ms. Total ~50ms.

26. Stage 1 methods? Two-tower ANN, item-item CF, popularity, rules. Often hybridized.

27. Stage 2 methods? GBDT, DeepFM, DLRM, transformer-based rankers.

28. When three stages? Pinterest, YouTube: retrieval (1M → 1000) → coarse ranking (1000 → 100) → final (100 → top-$K$).

F. Ranking models

29. Why does GBDT often win for ranking? Tabular interaction features; robust; fast; interpretable; handles missing data well.

30. DeepFM — what's the idea? Factorization Machines (low-order interactions) + Deep MLP (high-order). Combined.

31. DLRM architecture? Categorical embeddings → element-wise dot products + dense features → MLP.

32. LambdaMART — what is it? GBDT trained with pairwise/listwise ranking objectives (LambdaRank). Produces ranking-aware models.

33. Pointwise vs pairwise vs listwise? Pointwise: predict score per item. Pairwise: predict $i>j$. Listwise: optimize full list metric (e.g., NDCG).

G. Evaluation

34. NDCG@K formula intuition? $DCG@K={\sum }_{i=1}^K(2^{{\mathrm{rel}}_i}-1)/{\log }_2(i+1)$ (graded relevance form; for binary relevance simplifies to $\sum {\mathrm{rel}}_i/{\log }_2(i+1)$). Position-discounted relevance with $i$ 1-indexed (rank). Normalized by ideal ordering: $\mathrm{NDCG}=\mathrm{DCG}/\mathrm{IDCG}$. Higher = better.

35. MAP@K? Mean Average Precision. Average precision at each correct hit position.

36. MRR? Mean Reciprocal Rank: $1/k_1$ where $k_1$ is position of first hit. Captures "did we rank a correct item near top?"

37. Hit@K? 1 if any correct item in top-$K$, else 0. Simple recall measure.

38. AUC for ranking? Pairwise: probability that positive item ranks above negative. Threshold-free.

39. Why does offline often disagree with online? Position bias (top items click more), counterfactual (offline data from old policy), selection bias (only see clicks on shown), long-term effects.

40. Holdback test for recommenders? Permanent control population on old model. Catches long-term drift / degradation that A/B doesn't.

H. Cold start and exploration

41. New user — what do you do? Popularity, demographics, onboarding survey, exploration.

42. New item — what do you do? Content features (CLIP for images, text encoder for descriptions), forced exposure, similar-to-existing.

43. Echo chamber problem? Greedy ranking reinforces popular items; users see less diversity over time. Filter bubbles.

44. Thompson sampling for exploration? Maintain posterior over each item's reward. Sample, act greedily under the sample. Natural exploration-exploitation balance.

45. Diversity bonus? Add penalty for items similar to ones already in the recommended list. Encourages variety.

Quick fire

46. MF dimension typical? 64-256. 47. Two-tower scoring? Dot product. 48. In-batch negatives? Items from other queries in same batch. 49. NDCG decay? ${\log }_2(i+1)$. 50. Echo chamber fix? Exploration + diversity. 51. DLRM bottom layers? Embedding tables. 52. Cold-start tools? Content features + popularity. 53. GBDT for ranking strength? Tabular interactions. 54. Stage 1 latency? ~10ms. 55. Stage 2 latency? ~30-50ms.

Self-grading

If you can't answer 1-15, you can't talk about recommenders. If you can't answer 16-30, you'll struggle on architecture and ranking questions. If you can't answer 31-45, big-tech recommender system design will go past you.

Aim for 35+/55 cold.