Real-time Updating for Recommendation System Embeddings

Summary & Methodology Analysis

The paper tackles the challenge of embedding staleness, where user preference vectors remain frozen until a full model retrain occurs. To enable real-time adaptation, the authors replace static storage with a KP-tree, which is a sparse segment tree that uses sum aggregation. When a user provides a new rating, the system performs an O(log n) tree insertion to update the preference vector. This updated vector is then projected through a fixed low-rank basis, which is a set of primary vectors used to define the embedding space, to generate the current user embedding on-the-fly. To keep the sketch representative, the system employs norm-proportional sampling, a technique that selects items based on the magnitude of their influence, within the tree structure. This ensures that the sampling distribution remains consistent even as user preferences evolve. The approach was tested across various datasets, including KuaiRec, Amazon Electronics 2023, Amazon Video Games 2023, Amazon Music 2023, Book-Crossing, ML-1M, ML-25M, ML-10M, and Goodreads Comics, while comparing against methods like ALS, eALS, FunkSVD, RMFX, AutoEmb, DESS, Ekko, LiveUpdate, and FAISS. Despite the flexibility, the design has notable performance characteristics. The reliance on pointer-chasing tree traversals means the implementation is not optimized for GPU workloads. Furthermore, the approach degrades in efficacy when working with large catalogs where the item space lacks compactness relative to the sketch size. Finally, because the low-rank basis remains fixed, the model can still accumulate staleness as the underlying data distribution shifts over time.