Density estimation of missing data using OT
Malte Algren Geneva, On ATLAS
Interesting.
- Do OT between tag samples: 3tag -> 4 tag.
- However, dont have access to the full 4-tag distribution
- Instead "in fill" : replace 4b SR events with 3b SR events.
- Do OT from 3b → 4b infilled, not perfect… but better than before
- Now in fill with new SR prediction
- iterate
- (missed how they actually compute the maps with NNs,…) but it only works on 7d projected space
- Suggested that they add SvB to the list of variable they use to do the OT
- Very Sensitive to the CR boundary
- Maybe start with a better infill assumption … eg: the reweighted prediction
Notes
- Malte Algren
- "Density estimation of missing data using OT"
- PhD student in Geneva HEP
- Related to "In painting" problem in
- Setup the ABCD problem
- Crop out 4b SR replace removed events with 3b SR
- Fit OT from 3b -> 4b… Iterate
- How many steps ? …. How long for each step ? …
- What do you technical get in the end ? OT(map)
- How do you deal with point clouds ? … Yes, they use point clouds.
- 16D → 7D
- How do you predict another variable ?
- Suggest adding 8th SvB variable and then comparing that prediction to the rewieghting
- Very Sensitive to the CR boundry
- Maybe start with a better infill assumption … eg: the reweighted prediction
- Application generate synthetic data in the SR
Follow-ups
Links:
202504121742