Search-oriented Differentiable Product Quantization

Product quantization (PQ) is a popular approach for maximum inner product search (MIPS), which is widely used in ad-hoc retrieval. Recent studies propose differentiable PQ, where the embedding and quantization modules can be trained jointly. However, there is a lack of in-depth understanding of appropriate joint training objectives; and the improvements over non-differentiable baselines are not consistently positive in reality. In this work, we propose Search-oriented Product Quantization (SoPQ), where a novel training objective MCL is formulated. With the minimization of MCL, query and key's matching probability can be maximized for the differentiable PQ. Besides, VCS protocol is designed to facilitate the minimization of MCL, and SQL is leveraged to relax the dependency on labeled data. Extensive experiments on 4 real-world datasets validate the effectiveness of our proposed methods.