To mitigate the accuracy loss, if the query point is close to other partitions, you can not only search the most similar centroid but the neighbouring centroid of most similar cemtroid too. Usually the parameter is called we called "n_probs". But it comes as the expense of time complexity.
We have other similar options like Kd Tree and LSH for nearest neighbor search.
To mitigate the accuracy loss, if the query point is close to other partitions, you can not only search the most similar centroid but the neighbouring centroid of most similar cemtroid too. Usually the parameter is called we called "n_probs". But it comes as the expense of time complexity.
Absolutely, Sam :)
As I mentioned (and you too), there is a trade-off between time and accuracy so it really boils down to the latency you can afford to have.