Two-Stage Hashing for Fast Document Retrieval

This work fulfills sublinear time Nearest Neighbor Search (NNS) in massive-scale document collections. The primary contribution is to propose a two-stage unsupervised hashing framework which harmoniously integrates two state-of-the-art hashing algorithms Locality Sensitive Hashing (LSH) and Iterative Quantization (ITQ). LSH accounts for neighbor candidate pruning, while ITQ provides an efficient and effective reranking over the neighbor pool captured by LSH. Furthermore, the proposed hashing framework capitalizes on both term and topic similarity among documents, leading to precise document retrieval. The experimental results convincingly show that our hashing based document retrieval approach well approximates the conventional Information Retrieval (IR) method in terms of retrieving semantically similar documents, and meanwhile achieves a speedup of over one order of magnitude in query time.