A small recommender built on implicit feedback: every row of the input data is treated as an interaction between a user and a song, with no metadata, ratings, audio features, or learned embeddings involved.
The popularity model is the non-personalized baseline — songs ranked by interaction count, the same ten returned to everyone. It cannot adapt to individual taste, but it recommends to a completely new user, which the personalized model cannot.
The item-item collaborative filter compares songs by the sets of users who listened to them, scoring pairs with Jaccard similarity. A user’s recommendations are the mean similarity between each candidate song and every song already in their history, with heard songs and zero-score candidates removed. Using sets rather than counts means repeated plays by one user do not inflate similarity. It is memory-based: training stores the interaction data, and similarities are computed at request time.
Both models are evaluated on held-out interactions with precision@k and recall@k for k = 1..10, macro-averaged across a deterministically sampled set of users present in both splits — which is what makes the comparison between the personalized model and the baseline meaningful rather than anecdotal.