Project · Two-stage baseline recommender vs a popularity baseline

Using a small real or synthetic interaction dataset (MovieLens-100k works well), build a two-stage baseline: a candidate generator (popularity or item-item co-occurrence) feeding a simple ranker, and measure recall@10 against a pure popularity baseline on a held-out time-based split.

• A cleaned interaction table built from raw events or a public dataset, with a documented train/test split by time.
• A popularity-baseline recommender and its recall@10 on the test split.
• A two-stage (retrieve + rank) recommender and its recall@10 on the same split.
• A short writeup naming your true objective vs the proxy metric you actually measured, and one selection-bias risk in your data.

• The train/test split is time-based, not random, so the ranker cannot see future interactions.
• The two-stage recommender beats the popularity baseline on recall@10 by a reported, non-trivial margin.
• Retrieval and ranking are implemented as clearly separate stages, not fused into one function.
• The writeup explicitly distinguishes the true objective from the proxy metric measured.