Predictions
Closed this issue · 2 comments
First n_init=7 and non-deterministic, second only non-deterministic. What do you think we should do with this figure? @aarmey
If what we're trying to show is how the weight affects performance, is in it acceptable to make it deterministic (third)? The fittings might not be optimal but they're comparable. It's weird that the prediction is so high with the 1M weight.
(1) It's definitely not ok to make it purely deterministic, because we don't know those results are anywhere close to the optimal clustering for that parameterization. (2) What does the 1M weight look like when non-deterministic? (3) Is there anything you don't like about the n_init=7 results?
The 1M weight is 1250, I switch it manually with illustrator after making the plot to not mess up the x axis. So to answer (2), you can find non-deterministic 1M in plots 1 and 2. I don't like the size of the error bars about the n_init=7 results. Do you think this is good enough? Also by just looking at the mean AUCs across weights, I'm not sure we can claim that increasing the weight helps improve prediction...