Enhancing Cross-Category Learning in Recommendation Systems with Multi-Layer Embedding Training
This addresses a key bottleneck in recommendation systems for improving performance on rare items, though it is an incremental advancement over existing embedding methods.
The paper tackles the problem of obtaining high-quality embeddings for rarely-occurring categories in DNN-based recommendation systems by introducing a multi-layer embedding training (MLET) technique, which reduces embedding dimension and model size by up to 16x and 5.8x on average while maintaining or improving model quality.
Modern DNN-based recommendation systems rely on training-derived embeddings of sparse features. Input sparsity makes obtaining high-quality embeddings for rarely-occurring categories harder as their representations are updated infrequently. We demonstrate a training-time technique to produce superior embeddings via effective cross-category learning and theoretically explain its surprising effectiveness. The scheme, termed the multi-layer embeddings training (MLET), trains embeddings using factorization of the embedding layer, with an inner dimension higher than the target embedding dimension. For inference efficiency, MLET converts the trained two-layer embedding into a single-layer one thus keeping inference-time model size unchanged. Empirical superiority of MLET is puzzling as its search space is not larger than that of the single-layer embedding. The strong dependence of MLET on the inner dimension is even more surprising. We develop a theory that explains both of these behaviors by showing that MLET creates an adaptive update mechanism modulated by the singular vectors of embeddings. When tested on multiple state-of-the-art recommendation models for click-through rate (CTR) prediction tasks, MLET consistently produces better models, especially for rare items. At constant model quality, MLET allows embedding dimension, and model size, reduction by up to 16x, and 5.8x on average, across the models.