Drs.NAS: Ultra-Efficient Neural Architecture Search for Recommendation Systems

Deep learning-based recommendation systems (DRS) have become a dominant workload in hyperscale data centers. However, designing DRS architectures that balance high predictive performance with computational efficiency remains a major challenge due to ever-increasing model complexity and scale. Neural architecture search (NAS) has recently emerged as a promising automated design approach and is now adopted in production by major hyperscalers. Yet, existing NAS methods face two critical limitations: (i) prohibitive search costs—often requiring several GPU hours to days—which hinder rapid iteration, and (ii) the resulting architectures are typically computation- and memory-intensive, limiting practical deployment. In this paper, we propose Drs.NAS, an ultra-efficient NAS framework for DRS. (i) Ultra-efficient search: We propose a novel metric, superproxy, which enables NAS without the costly training and validation required by existing NAS methods. Compared to SOTA NAS search times of 5!∼!18 GPU-hours, Drs.NAS completes the search within two minutes on a commodity CPU. (ii) Ultra-efficient results: The models discovered by Drs.NAS drastically reduce resource demands—achieving on average 108.3× and 34.9× smaller model sizes, and 88.8× and 14.7× fewer FLOPs, compared to handcrafted and SOTA NAS results, respectively. Crucially, these gains come without sacrificing predictive quality: Drs.NAS delivers on par or even superior predictive performance, surpassing handcrafted and NAS baselines by 0.0123 and 0.0056 in average AUC across three representative benchmarks, respectively.