Problem
A feed recommender scores each item by the chance of a positive action, such as a save. A list built only from these scores fills with visually similar items. Users arrive at Pinterest Home Feed with low intent, so repetitive content cuts session length and return visits. The final layer must balance short- and long-term engagement and business rules, and be easy to change. Hard filtering of risky content is blunt and leaves gaps when nothing backfills.
Example from the article: with the feed-level diversity component removed, saves rose on day 1 and turned negative by the second week, with shorter sessions.
Solution in brief
Pinterest re-ranks the ranker output in a final multi-objective layer. First came a determinantal point process (DPP, 2021), then Sliding Spectrum Decomposition (SSD, early 2025) written in PyTorch and served from the shared model cluster. SSD is cheaper to run, so it takes more similarity signals (visual, text, graph embeddings, Semantic IDs) and a soft spacing penalty for risky content.
Questions this case answers
- How to compose the final ordered list of a feed so that it stays diverse over a long horizon while relevance scores favour near-duplicates
- How to penalize clustering of risky items in a ranked list without removing them and leaving gaps
- How to keep a recommendation feed diverse without losing engagement?
- How to replace a DPP diversification step with something cheaper to serve and easier to iterate on?
Tools and methods
- Determinantal point process (DPP) — first-generation slate diversification; kernel mixes item relevance and pairwise similarity
- Sliding Spectrum Decomposition (SSD) — position-adaptive greedy re-ranking over a sliding window
- Soft spacing penalty — pushes apart items from a sensitive set instead of filtering them
- GraphSAGE, PinCLIP, text embeddings, Semantic IDs — signals for pairwise Pin similarity
- PyTorch on CPU model servers — runs the blending logic
Google or a one-time email code. No password.