Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling (Google, CIKM 2024)
This entry from the zhichai.net Recommender Engines collection indexes a Google Research paper published at CIKM 2024: *Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling*.
Metadata
| Field | Value | |-------|-------| | Title | Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling | | Venue | CIKM 2024 | | Organization | Google Research | | Source | https://research.google/pubs/improved-estimation-of-ranks-for-learning-item-recommenders-with-negative-sampling/ | | Resource type | Academic paper | | Section | Recommender Engines |
Context
Rank-based metrics (e.g., Recall@K, Hit@K) are standard offline evaluation tools for item recommenders. However, when a retrieval model is trained with negative sampling, directly computing rank metrics over the full item catalog can be problematic: the sampled negatives used during training bias naive rank estimates. This Google CIKM 2024 paper addresses improved estimation of ranks for item recommenders trained under negative sampling.
> Note: This post is an index entry based on public metadata. For the paper's exact methodology, theoretical results, and experiments, please consult the original publication via the link above.
Related entries in the collection
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
- Augmenting Netflix Search with In-Session Adapted Recommendations, RecSys
- Bridging Language and Items for Retrieval and Recommendation, Mar 2024
- Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
- DiffKG: Knowledge Graph Diffusion Model for Recommendation, WSDM 2024
- Original paper: *Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling*, Google Research, CIKM 2024. See the official publication page for the PDF and citation details.
Why it matters
For industrial-scale recommenders with catalogs of millions to billions of items, full-catalog evaluation is expensive, and sampled evaluation introduces bias. Techniques for unbiased or corrected rank estimation under negative sampling are directly relevant to anyone benchmarking two-tower retrieval models, candidate generators, or personalized ranking systems where evaluation consistency matters as much as training efficiency.