Anisotropic Additive Quantization for Fast Inner Product Search

Maximum Inner Product Search (MIPS) plays an important role in many applications ranging from information retrieval, recommender systems to natural language processing and machine learning. However, exhaustive MIPS is often expensive impractical when there are a large number of candidate items. The state-of-the-art approximated product quantization with score-aware loss, which weighs more heavily on items larger inner scores. it challenging extend the loss for additive due parallel-orthogonal decomposition residual error. Learning respect this since can achieve lower approximation error than quantization. To end, we propose method called Anisotropic Additive Quantization combine anisotropic efficiently update codebooks algorithm, develop new alternating optimization algorithm. proposed algorithm extensively evaluated three real-world datasets. experimental results show that outperforms baselines approximate search accuracy while guaranteeing similar retrieval efficiency.