> Speculative Recommendation Serving: From LLM Decoding to Caching

Recommender Systems 中级 2026-04-01 06:41 2026-04-01
#recommendation #caching #speculative decoding #latency #matrix factorization #regret bounds

RecCache transfers the draft-and-verify paradigm from speculative LLM decoding to recommendation caching: user clusters as draft models, score-ratio acceptance criterion, achieving 1.0-3.8x speedup training-free with O(T*epsilon) regret.

Speculative Recommendation Serving

The Structural Analogy

LLM Speculative Decoding Speculative Rec Serving
Draft model q(x) Cluster cache C[c]
Target model p(x) Full recommender R(u)
min(1, p/q) acceptance Score-ratio acceptance
Multi-token speculation Multi-cluster speculation

Results

Dataset NDCG@10 Accept Rate Speedup
ML-1M .091 91% 2.2x
A-Electronics .010 59% 3.1x
MIND .014 32% 3.8x

MIND: NDCG exactly matches fresh recommendations at 3.8x speedup.

Multi-Cluster Gain

49-68% of accepted recommendations come from non-nearest clusters. Single-cluster caching would substantially reduce coverage.

Training-Free Regret

\(O(T \cdot \epsilon)\) regret without training vs \(O(\sqrt{T})\) for RL-based caching (CARL). RecCache matches CARL's NDCG at zero training cost.

Methodology Fix

Critical correction: early experiments didn't exclude training history items from recommendations, inflating NDCG. The fix reduces absolute numbers but ensures genuine generalization — a common pitfall in recommendation evaluation.