Drs.NAS: Ultra-Efficient Neural Architecture Search for Recommendation Systems
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Claims an automated NAS that uses a superproxy metric (seven zero-cost proxies) to avoid training and finishes search in two minutes on a commodity CPU, not GPU hours. It frames search as differentiable subgraph sampling and optimizes a composite objective that trades AUC for FLOPs and size. Results show about 108.3× and 34.9× smaller models versus handcrafted and SOTA NAS, 88.8× and 14.7× fewer FLOPs, and average AUC gains of 0.0123 and 0.0056 on three CTR benchmarks.
AI-written summary · about
chapters (13)
- 0:00Introduction to DrSNAS
- 1:25Recommendation system challenges
- 3:34Limitations of NAS methods
- 4:39DrSNAS pipeline overview
- 5:42Super proxy design
- 7:30Graph-based search space
- 8:11Composite objective function
- 9:12Experimental setup
- 11:19Search cost and efficiency
- 12:24Inference cost results
- 13:47Predictive performance
- 14:29Robustness across models
- 15:32Conclusion
related talks
38mAutoML for Efficient Neural Architecture Design
30mNeural Architecture Search using Property Guided Synthesis
23mML-Powered Search & Recommendation 101
13mFast and Explainable Search on a Budget With OpenSearch Neural Sparse
26mNeuOS: A Latency-Predictable Multi-Dimensional Optimization Framework for DNN-drive
30mSemantic Search and Neural Information Retrieval
32mBudget Friendly Semantic Search With Neural Sparse Search
01h 52mNeural Network Accelerator Design