BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pydata-amsterdam2026//speaker//LPTRML
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-pydata-amsterdam2026-WBBVBF@pretalx.com
DTSTART;TZID=CET:20260910T103500
DTEND;TZID=CET:20260910T110500
DESCRIPTION:Large scale transactional platforms processing millions of dail
 y events require high-performance algorithmic fraud intervention. However\
 , deploying models that block bad actors introduces a structural blind spo
 t: the censorship of positive labels needed to train the next generation o
 f models. To generate unbiased data\, systems typically rely on a holdout 
 group\, a subsample where the model doesn't act. This creates an operation
 al tradeoff: a large holdout caps the model's ability to intervene\, while
  a small holdout reduces the sample size of the training pipeline\, especi
 ally when the labels are heavily unbalanced.\n\nThis talk details the meth
 odology developed to mitigate this reject inference using LightGBM and Opt
 una. We will walk through the mathematical theory of Inverse Propensity We
 ighting (IPW) to reconstruct missing data\, and examine why applying stand
 ard IPW directly to GBDTs often causes gradient explosion and overfitting.
  To bridge the gap between statistical theory and tree-based mechanics\, w
 e will introduce Asymmetrical IPW: a hybrid weighting architecture that st
 abilizes tree gradients using optimizable\, asymmetrical weights. This app
 roach yielded a model recovering 99% of the benchmark's partial AUC (@20% 
 FPR)\, where the benchmark is a model trained on fully uncensored data.
DTSTAMP:20260710T150433Z
LOCATION:Room 1 (170)
SUMMARY:Beyond the Holdout: Mitigating Censorship Bias with Asymmetric IPW 
 - Leonardo Amorim
URL:https://pretalx.com/pydata-amsterdam2026/talk/WBBVBF/
END:VEVENT
END:VCALENDAR