NearestNeighborClassifier API

User 2568 | 2/6/2016, 11:42:37 PM

I was experimenting with NearestNeighborClassifier for the Telstra Kaggle competition. This is a multi-class classification problem with log-loss metric

Unlike the other classifiers,

  1. NearestNeighborClassifier.evaluate does not support log-loss.
  2. NearestNeighborClassifier.predict does not support outputtype ‘probabilityvector’. Instead I had to use top_k and munge the results.

for convenience and simplicity when moving between models, it would be nice to have this be consistent across classifier APIs.


User 12 | 2/7/2016, 1:58:24 AM

Hi Kevin, thank you for the good feedback. I'll be working this month on polishing the NearestNeighborsClassifier. - Brian