ASC Algorithms Speed Up Swing Similarity for Large-Scale Recommender Retrieval by Orders of Magnitude

_reachsumit · x · 2026-09-16

• The paper targets Swing similarity, a key measure for item-to-item retrieval in industrial recommenders: existing solutions are either quadratic in item degree or rely on quality-losing truncation heuristics, impractical on graphs with billions of interactions.
• Authors propose ASC and K-ASC for approximate and top-K Swing queries with rigorous probabilistic guarantees on relative and additive errors.
• ASC adaptively combines two randomized algorithms, GNS and USS, to handle high- and low-degree query items with minimal runtime cost; K-ASC uses a filter-refinement paradigm with heuristics.
• Experiments on eight real datasets show orders-of-magnitude speedups over existing solutions; code is open-sourced.

Original post →

More from Research

• Beyond Jev: a guide to 15 new System One decision-model research approaches — TheTuringPost · 2026-10-26
• Paper: embeddings are robust to truncation without Matryoshka learning, except heavy truncation — srchvrs · 2026-10-26
• Is Matryoshka Representation Learning actually useful? A thread digs into the evidence — srchvrs · 2026-10-26
• NNP/CG-MM: embedding all-atom neural network potentials into coarse-grained mechanics — MoAlQuraishi · 2026-10-26
• Cardchoose: a 12-line, provably fair no-retry algorithm for sampling without replacement — NathanpmYoung · 2026-10-26
• Robots that read anticipatory brain signals for collaborative task planning — paper pick of the week — tomssilver · 2026-10-26

Research channel →