Understanding and Improving Proximity Graph Based Maximum Inner Product Search
Authors: Jie Liu, Xiao Yan, Xinyan Dai, Zhirong Li, James Cheng, Ming-Chang Yang139-146
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiment results show that ip-NSW+ consistently and significantly outperforms ip-NSW and provides more robust performance under different data distributions. We used the four datasets listed in Table 1. |
| Researcher Affiliation | Academia | The Chinese University of Hong Kong {jliu, xyan, xydai, zrli6, jcheng, mcyang}@cse.cuhk.edu.hk |
| Pseudocode | Yes | Algorithm 1 NSW: Query Processing via Graph Walk (Fu et al. 2019) Algorithm 2 NSW: Graph Construction (Morozov and Babenko 2018) Algorithm 3 ip-NSW+: Query Processing via Graph Walk |
| Open Source Code | Yes | See https://github.com/jerry-liujie/ip-nsw/tree/Graph MIPS for all experiment code and data. |
| Open Datasets | Yes | We used the four datasets listed in Table 1. Yahoo!Music is obtained by conducting ALS-based matrix factorization (Yun et al. 2013)... Word Vector is sampled from the word2vec embeddings released in (Mikolov et al. 2013)... Image Net contains the visual descriptors of the Image Net images (Deng et al. 2009)... Tiny5M is sampled from the Tiny80M dataset... |
| Dataset Splits | No | The paper describes using '1,000 randomly selected queries' for evaluation and constructing graphs on the datasets but does not provide explicit train/validation/test splits for the data itself. |
| Hardware Specification | Yes | The experiments were conducted on a machine with Intel Xeon E5-2620 CPU and 48 GB memory in a single-thread mode. |
| Software Dependencies | No | The paper mentions 'We implemented ip-NSW+ by modifying the code of ip-NSW' but does not provide specific software versions or dependencies. |
| Experiment Setup | Yes | We used M = 10 and l = 10 for the angular graph As in ip-NSW+ in all experiments and the parameter configurations of Gs in ip-NSW+ is the same as the inner product graph in ip-NSW. |