Öneri Sistemleri

Framing

1. The Recommendation ProblemPredicting preferences from sparse implicit feedback; retrieval-then-ranking pipelines.1–2 sa

Methods

1. Collaborative FilteringRecommending from the behaviour of similar users or items, without content features.1–2 sa
2. Matrix Factorization for RecommendationUsers and items as vectors in a shared latent space; ratings as dot products.1–2 sa
3. Two-Tower Retrieval ModelsSeparate user and item encoders trained contrastively so candidates can be retrieved with nearest-neighbour search.1–2 sa
4. Cold Start and Content-Based RecommendationNew users and items have no interaction history; content features and hybrids fill the gap.1–2 sa
5. Learning to RankPointwise, pairwise, and listwise objectives for ordering candidates.1–2 sa

Dynamics

1. Bandits for RecommendationTreating recommendation as exploration: contextual bandits and off-policy evaluation from logs.1–2 sa
2. Feedback Loops and Popularity BiasRecommenders shape the data they learn from, so exposure bias compounds into popularity bias.1–2 sa