The Seattle Single Decision Tree

1
The Seattle Single Decision Tree

• Toby BurnettUniversity of Washington

2
The single tree definition

• We originally copied the NN procedure of training
  on two subsets of the background, then using 2-D
  binned likelihood in the two variables.
• Get to use all the tools developed for this, just
  change a few file names
• Gordon presenting the results using this analysis
• But we asked, isnt this a bit of a kluge? Why
  not just apply the classification tree technology
  to all the background at once?

3
Classification tree training

• Use the gini criterion
• Quit when a node has less than 100 entries
• One pass, no boosting, averaging over subsets
• Train on ½ the data, test with the rest.

4
The variables used just copied from NN
Table of the gini improvement
5
Derived statistical limits, vs. cut on purity.

• Easy application of top_statistics.
• Summed over 1-tag and 2-tag.
• Make cut at 0.7 efficiency for following.

6
Now, Systematics!

• Very much harder, could not do in time.
• But easy to just look at the derivatives which
  are used as input to the procedure. I think that
  what it does is
• Assume a priori Gaussian distribution in the
  various systematic parameters, JES TRF, TRIG,
• Create new MC files with or one sigma change
  in value
• Recalculate new estimates of single top content
  using varied background and signal.
• Assume this is linear, and propagate the errors
  to the estimate fudge a little and take larger
  of two deviations from nominal
• A question is, how linear are the estimators as
  functions of the systematic parameters?

7
Linearity? Something weird with wjj JES
JES differences on all other background sources
look symmetric
8
OK, what is unique with JES and/or wjj?
A quick look at histograms of wjj (bad) and wbb
(good)
9
(Tentative) conclusions, plans

• This seems to invalidate the JES systematic
  calculation, since it badly fails the assumption
  used to compute its contribution to the limit
• Rerun the training on the systematic files?
• Need to optimize training
• Try all variables the classifier only selects
  the ones that help, ranks them
• Allow smaller final nodes
• Try boosting, etc.