Query-Aware Quantization for Maximum Inner Product Search

Authors: Jin Zhang, Defu Lian, Haodi Zhang, Baoyun Wang, Enhong Chen

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed method is evaluated on three real-world datasets. The experimental results show that it outperforms the state-of-the-art baselines.
Researcher Affiliation Collaboration Jin Zhang1, Defu Lian1,2*, Haodi Zhang3, Baoyun Wang4, Enhong Chen1,2 1 University of Science and Technology of China 2 State Key Laboratory of Cognitive Intelligence, Hefei, China 3 Shenzhen University 4 Hisense
Pseudocode Yes Algorithm 1: Iterative optimization; Algorithm 2: Query-aware Quantization
Open Source Code No No explicit statement or link providing access to the source code for the methodology described in the paper was found.
Open Datasets Yes Our method is evaluated on three real datasets, Last FM, Glove, and Music100, which were used in the evaluation of multiple MIPS algorithms (Morozov and Babenko 2018; Guo et al. 2020; Krishnan and Liberty 2021; Zhang et al. 2022). The Last FM dataset is a real music recommendation dataset... as described in (Lian et al. 2015). The Glove dataset is a collection... as described in (Pennington, Socher, and Manning 2014). The Music100 dataset was introduced in IP-NSW (Morozov and Babenko 2018).
Dataset Splits Yes Input: Database X, validation set V , held-out set H, integer N, kv. ... Zbest, cbest evaluate on V , save best samples and best codebooks according to evaluation metric; The dataset splitting and experimental results on other datasets, including Echo Nest and Yahoo!Music, are attached in the Appendix.
Hardware Specification Yes The proposed algorithm is implemented in Julia, and all experiments are conducted in a Linux server with 3.00GHZ intel GPU and 300G main memory.
Software Dependencies No The paper states 'The proposed algorithm is implemented in Julia', but does not provide a specific version number for Julia or any other software dependencies with their versions.
Experiment Setup Yes For all quantization methods, we set the dimension of the subspace to 4, i.e., 25 codebooks, and use 16 codewords in each codebook, which is the same as the settings in Sca NN (Guo et al. 2020). In our method Alg.2, we set N to 500, #iter to 20, max iter to 2, and kv to 2000. We use Recall as the evaluation metric, and k-means as clustering algorithm. In the Appendix, we will explain in detail how these hyper-parameters are set.