Good introductory review for LLM-based recommendation systems, I summarized the key points in the video.
LLMs for Recsys
Word2Vec(2013): Item embedding from co-occurrence in user interaction sequences
GRU4Rec(2015): Short-term, next-item indent from user interaction sequences
SASRec(2018), BERT4Rec(2019): Attention on long-range depedencies in user interaction sequences
LLMs(2024): Semantic IDs, Data Augementation, Unified Models
Challenge 1: Hash based item IDs don’t encode item content; thus struggle with cold-start and sparsity
Problem: Help users discover new items, faster (@Tiktok)
Solution: Trainable, multimodal, semantic IDs
Benefits: Address cold-start, Recs that understand content
Challenge 2: High-quality metadata is essential for search (and RecSys) but costly and high effort to get
General Solution -> using LLMs for synthetic data & labels
Problem 1: Poor user experience and lost trust due to low-quality job recommendation(@Indeed)
Solution 1: Lightweight classifier to filter bad recs
Problem 2: Help users search for new items (podcasts, audiobooks) in catalog of known items (song, artist) (@Spotify)
Solution 2: Query recommendation system
Benefits: Richer, higher-quality data at larger scale, Far lower cost and effort than human annotation
Challenge 3: Task-specific models duplicate engineering, increase maintenance cost, and don’t benefit from transfer learning
General Solution -> Unified models(it works for vision and language, so why not recsys? Even work for payments and fraud at Stripe, too!)
Problem 1: Teams deal with complexity from bespoke models for search, similar item recs, pre-query recs (@Netflix)
Solution 1: Unified ranker for all of the above
Problem 2: Help users get better results with highly specific or broad queries, on ever-changing inventory (@Etsy)
Solution 2: Unified Embedding + retrieval
Benefits: Simplifies systems, reduced maintenance overhead, transfer learning(but there may be the alignment tax), Gains to unified model benefit other use cases