BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Statistics
SUMMARY:On the Consistency of Supervised Learning with Mis
 sing Values - Julie Josse\, École Polytechnique
DTSTART;TZID=Europe/London:20190510T160000
DTEND;TZID=Europe/London:20190510T170000
UID:TALK115933AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/115933
DESCRIPTION:In many application settings\, the data have missi
 ng features which make data analysis challenging. 
 An abundant literature addresses missing data in a
 n inferential framework: estimating parameters and
  their variance from incomplete tables. Here\, we 
 consider supervised-learning settings: predicting 
 a target when missing values appear in both traini
 ng and testing data. We show the consistency of tw
 o approaches in prediction. A striking result is t
 hat the widely-used method of imputing with the me
 an prior to learning is consistent when missing va
 lues are not informative. This contrasts with infe
 rential settings where mean imputation is pointed 
 at for distorting the distribution of the data. Th
 at such a simple approach can be consistent is imp
 ortant in practice. We analyze further decision tr
 ees. These can naturally tackle empirical risk min
 imization with missing values\, due to their abili
 ty to handle the half-discrete nature of incomplet
 e variables. After comparing theoretically and emp
 irically different missing values strategies in tr
 ees\, we recommend using the "missing incorporated
  in attribute" method as it can handle both non-in
 formative and informative missing values.
LOCATION:MR12
CONTACT:Dr Sergio Bacallado
END:VEVENT
END:VCALENDAR