Open orelgueta opened 3 years ago
Below are the confusion matrices for the best model (MLP_small) for 3 types. Both 1D and 2D versions are shown. What I conclude from these matrices is that perhaps 3 types is too many, maybe 2 types is more realistic (plots for 2 types to come soon). We need to see the IRFs to decide perhaps. Also, I don't think the performance is better than regression. Definitely not significantly better.
Below are the same score comparison as above, but this time for 2 event types. This time the best model is not so clear. MLP small is probably still best overall, but it has a drop at one energy bin (1.2 < E < 1.7 TeV) and at higher energies even the linear models (Ridge), perform a bit better. Unclear why this is, but since we will anyway switch to the Prod5 dataset, add more variables and might anyway use regression, I won't investigate at the moment. I still take MLP small as the default for the rest of the plots.
Confusion matrices for two event types can be seen below. I think they are not significantly better than the regression confusion matrices.
Please see below updated plots for classification with the Prod5 sample. Additional details about the sample can be seen in #4. Below are the results for 2 event types, where we can see that MLP_small is clearly the best. I don't see a reason to assume it will not be the best also for other numbers of event types, so for 3 event types I only train MLP_small (results in the next comment).
Question: Aren't they way too good?! Getting 70-80% of the classifications right seems really good... Are we expecting such a performance?
With just two types? Yea that doesn't surprise me considering the regression performance we saw before. Soon I will get the 3-event-type results and then we will see if it's too good.
Below are the confusion matrices for 2 and 3 types MLP_small classification model. With 3 event types we get 60-65% classification accuracy, which is more reasonable I guess. The comparison with regression will be discussed in #2.
The plots below compare the scores of all classification models for 3 types. Surprisingly (or not!), MLP_small is the best model also for classification (apologies for it changing colours between plots...). The drop in performance in the two highest energy bins is interesting. Perhaps we need to split those bins to smaller bins?