jmazanec15 commented on issue #12533: URL: https://github.com/apache/lucene/issues/12533#issuecomment-1717981259
Additionally, the [FreshDiskANN](https://arxiv.org/pdf/2105.09613.pdf) paper did some work in this space. They ran a test for NSG where they iteratively repeat the following process a certain number of cycles and track the recall: 1. delete 5% of the index 2. patch the incident nodes that were impacted via local neighborhoods (similiar to @zhaih (1)) 3. reinsert the deleted nodes 4. measure recall They ran a similar one for HNSW where they do not patch the edges. In both cases, they saw some degradation: . Their intuition for this happening is because of the graphs become sparser as this process happens, leading to less navigability. The graphs become sparser because the pruning policy is more strict. In their system, they do employ a similar algorithm to @zhaih (1), where they connect the incident edges and prune based on some criteria that shows promise. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
