Single Path Embedding Dimension Search in Recommender System
Gaofeng Lu, Jie Zhang · 2025
In practical large scale recommender systems, there are typically thousands of feature fields derived from users, items, contextual information, and their interactions. An embedding layer is widely used in deep learning recommendation models to map high-dimension sparse features into low-dimension dense embeddings. Traditionally, a uniform dimension is assigned to all feature embeddings, which can result in inefficient memory usage and sub-optimal performance due to inappropriate dimension assignments. Therefore, it is highly beneficial to allocate different embedding dimensions to various feature fields based on their importance. To address this problem, we introduce Single Path Embedding Dimension Search, a differentiable neural architecture search-based method that can automatically choose suitable dimensions for every feature field. Specifically, we construct a super-net which can calculate the weights across diverse dimensions for each feature field. Notably, only a single path within all fields is activated once a time, thereby mitigating the issue of weight mutual interference. Extensive experiments on two widely used public datasets show that the proposed algorithm achieves significant improvements in several recommendation metrics compared to the fixed dimension embedding algorithm, demonstrate the efficiency and superiority of the proposed method.