-
Notifications
You must be signed in to change notification settings - Fork 748
Ideas list
Ann-Benchmarks contains scripts that produce the interactive plots available at http://ann-benchmarks.com. While interactive plots provide many advantages, such as displaying exact parameter settings used in individual runs, there are many ideas left unexplored.
A list of possible improvements:
- Add error bars to the individual plotting points
- Add contour maps to show uncertainty of individual average points
- Improve navigation in individual plots
- Introduce technique to compare different plots to each other
Working on these suggestions can also yield to improvements and added feature to existent open-source charting libraries (we use https://www.chartjs.org).
While ann-benchmarks contains some datasets already, many other datasets are used in the literature. This project explores inhowfar these datasets differ from the existing ones and includes possible new and interesting datasets. This project can also consist of a part where challenging, synthetic datasets are created.
The current code structure consist of some very tight couplings between different mechanisms, for example how datasets, algorithmic parameters and result files are coupled with each other. In this project we try to identify ways to make it more loosely coupled, improve reusability, improve documentation and add tests to the existing ones to improve code coverage.
Nearest neighbor search is a key primitive in many applications. For some applications, for example outlier detection, it has been observed that setting up the NN algorithm to return fewer actual nearest neighbors shows better performance. In this project, we add support to outlier detection applications to ann benchmarks and benchmark the quality of individual NN algorithms in this area.