Currently, CAGRA does an assertion that graph degree is not less than intermediate graph degree, but even if graph degree is too close to intermediate graph degree, it's possible there's not enough neighborhood information encoded in the intermediate knn graph for the pruning to have a connected graph.
When this case happens, the user gets an error that's nebulus and doesn't provide enough information for the user to
- know what happened, and
- fix the issue.
In general, users aren't going to know the complex relationships between parameters like this, but we often make a suggestion that intermediate graph degree by 1.5 * graph degree.
Currently, CAGRA does an assertion that graph degree is not less than intermediate graph degree, but even if graph degree is too close to intermediate graph degree, it's possible there's not enough neighborhood information encoded in the intermediate knn graph for the pruning to have a connected graph.
When this case happens, the user gets an error that's nebulus and doesn't provide enough information for the user to
In general, users aren't going to know the complex relationships between parameters like this, but we often make a suggestion that intermediate graph degree by 1.5 * graph degree.