Pulkitg64 commented on PR #15003: URL: https://github.com/apache/lucene/pull/15003#issuecomment-3497672783
Ran another experiment where I keep deleting 10% nodes from the graph. Started from 100k docs and ended with 49963 docs at the end of 5th iteration. This benchmark run is similar to [this one](https://github.com/apache/lucene/pull/15003#issuecomment-3412537982), where we saw 10% degradation in recall till 5th iteration but that's not the case anymore. For maxCon=8,16, we are seeing little degradation but maxConn=32,64 is performing as good as baseline. | MaxConn | Experiment | Iteration 1 | Iteration 2 | Iteration 3 | Iteration 4 | Iteration 5 | |---------|------------|-------------|-------------|-------------|-------------|-------------| | 8 | Baseline | 0.796 | 0.800 | 0.806 | 0.808 | 0.812 | | | Candidate | 0.783 | 0.767 | 0.743 | 0.753 | 0.753 | | 16 | Baseline | 0.89 |0.895 | 0.898 | 0.902 | 0.906 | | | Candidate | 0.884 | 0.877 | 0.871 | 0.868 | 0.868 | | 32 | Baseline | 0.926 | 0.930 |0.933 | 0.936 | 0.939 | | | Candidate | 0.923 | 0.923 | 0.923 | 0.926 | 0.929 | | 64 | Baseline | 0.935 | 0.937 |0.94 | 0.945 | 0.946 | | | Candidate | 0.935 | 0.936 | 0.94 | 0.943 | 0.943 | -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
