Google's SDF framework slashes stale recommendations by 54.9%

On large content platforms, stale recommendations remain a top source of user frustration. Google's new paper, accepted at CIKM 2026, identifies two distinct causes: supersession (new information making old items outdated) and relevance decay (inherent loss of value over time). Current heuristics like age cutoffs and engagement signals fail to capture these nuances. To solve this, Google built SDF (Supersession-Decay Filtering), a system fully deployed in Google Discover—a feed used by hundreds of millions daily. SDF pairs two learned models: a relational staleness model that compares item pairs to detect supersession, and a predicted traffic ratio (PTR) model that forecasts an item's traffic decline based on its content and lifetime visit patterns. Both filters run upstream of ranking, pruning stale candidates before they compete for placement, which also reduces downstream serving costs.

Online experiments confirm SDF's effectiveness: stale content prevalence fell significantly while user engagement metrics improved. Over two years of production, user-filed staleness reports dropped 54.9% relative to baseline, proving that targeted AI filtering outperforms crude proxies. The design is complementary—the relational model handles supersession while PTR handles decay, and applying them as a disjunction catches both failure modes. This work offers a scalable blueprint for any platform battling content freshness, from news feeds to e-commerce recommendations. The paper is available on arXiv (2608.15780) with an upcoming CIKM 2026 presentation.