Reproduce “SHARK: A Lightweight Model Compression Approach for Large-scale Recommender Systems” on OpenResearch

About this paper

Beichuan Zhang, Chenggen Sun, Jianchao Tan, Xinjun Cai, Jun Zhao, Mengqi Miao, Kang Yin, Chengru Song and 2 more

Kuaishou Technology developed SHARK, a lightweight model compression framework, to reduce the computational and storage demands of large-scale recommender systems. It achieves up to 70% storage savings and a 30% improvement in Queries Per Second (QPS) in production environments by combining efficient feature selection and adaptive mixed-precision quantization, all while maintaining or slightly improving model performance.

Key insights

• A first-order Taylor expansion provides a sufficiently accurate and dramatically faster approximation of feature importance, making feature selection feasible for large industrial datasets without altering model architecture.
• Row-wise mixed-precision quantization, guided by dynamically updated priority scores that consider both access frequency and the greater importance of positive samples, is crucial for minimizing performance degradation when reducing embedding precision.
• Iterative pruning combined with rapid fine-tuning, rather than full retraining, effectively recovers model performance and accounts for feature coupling after the removal of less important features.

Summary by alphaXiv.