You are asked to compare Pin similarity in Python. Begin by clarifying how a Pin is represented and what “similar” should mean. Then assume the interviewer provides one query Pin embedding and a collection of candidate Pin embeddings. Explain how you would return the top `k` candidates by cosine similarity and how you would evaluate whether the similarity is useful. ### Constraints & Assumptions - Each valid embedding is a one-dimensional numeric vector with the same dimension. - Candidate identifiers are unique, and the query itself should be excluded if it appears among the candidates. - Define behavior for zero vectors, missing values, invalid `k`, and tied scores. - Treat this as a DS Python and evaluation question, not a production retrieval-system design exercise. ### Clarifying Questions to Ask - Are Pins represented by dense embeddings, tags, user co-engagement, or another feature set? - Is similarity intended to capture visual, topical, or behavioral relatedness? - How large is the candidate set, and is exact retrieval required? - What labeled or behavioral outcome will define a good neighbor? ### What a Strong Answer Covers - Why the representation and similarity measure must match the product meaning of similarity. - Correct cosine-similarity logic with validation, deterministic ranking, and explicit edge-case behavior. - Computational cost and when vectorization or approximate nearest-neighbor retrieval becomes necessary. - Offline ranking metrics, slice analysis, leakage checks, and an online decision metric. ### Follow-up Questions 1. How would your solution change if each Pin were represented by a set of tags? 2. How would you support millions of candidates with a latency requirement? 3. How would you prevent near-duplicate results from reducing recommendation diversity?