BEGIN:VCALENDAR VERSION:2.0 X-WR-CALNAME:EventsCalendar PRODID:-//hacksw/handcal//NONSGML v1.0//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York LAST-MODIFIED:20240422T053451Z TZURL:https://www.tzurl.org/zoneinfo-outlook/America/New_York X-LIC-LOCATION:America/New_York BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT CATEGORIES:College of Engineering,Graduate Studies,Lectures and Seminars,Th esis/Dissertations DESCRIPTION:Thesis Advisor: Dr. Gökhan Kul, Computer and Information Scien ce Committee Members: Dr. Joshua Carberry, Computer and Information Scienc e and Dr. Yuchou Chang, Computer and Information Science Abstract: Machine learning models deployed in consequential domains can become unfair towar d protected subgroups as the data they receive drifts over time, yet the p rotected attributes needed to measure fairness directly are often unavaila ble at runtime due to privacy regulation and operational constraints. This creates a gap: existing fairness toolkits require protected labels and pe rform one-time audits, while generic drift detectors monitor continuously but cannot localize which subgroup a shift harms. This thesis develops a n on-invasive fairness drift monitor that addresses this gap by repurposing Conformance Constraints, a data-profiling primitive, as a temporal fairnes s signal. The monitor learns per-subgroup distributional profiles at basel ine, using protected attributes only once, and thereafter tracks violation of those profiles over incoming data batches without any runtime access t o protected attributes. Across three fairness benchmarks, two classifiers, and nineteen controlled drift scenarios, the conformance-constraint viola tion signals track fairness degradation more closely than KS and KL detect ors under global drift, and they remain competitive with them under group- targeted drift. Subgroup localization provides its clearest advantage unde r global drift, where a minority-subgroup signal substantially outperforms both aggregate signals and the baselines. The correlations are modest in absolute terms, indicating that the monitor functions as a screening instr ument that flags fairness degradation for closer investigation rather than as a precise estimator. The approach offers privacy-preserving, subgroup- aware fairness monitoring suited to regulated deployment settings. For fur ther information please contact Dr. Gokhan Kul at gkul@umassd.edu. \nEv ent page: https://www.umassd.edu/events/cms/20260803-non-invasive-fairness -drift-monitor-for-machine-learning.php\nEvent link: https://teams.microso ft.com/meet/245292205330763?p=xBhMejmhVBcZHQGSk6 X-ALT-DESC;FMTTYPE=text/html:

Thesis Advisor: Dr. Gökhan Kul \, Computer and Information Science

Committee Members: Dr. Joshua Carberry\, Computer and Information Science and Dr. Yuchou Chang\, Compute r and Information Science

Abstract: Machine learning models deploy ed in consequential domains can become unfair toward protected subgroups a s the data they receive drifts over time\, yet the protected attributes ne eded to measure fairness directly are often unavailable at runtime due to privacy regulation and operational constraints. This creates a gap: existi ng fairness toolkits require protected labels and perform one-time audits\ , while generic drift detectors monitor continuously but cannot localize w hich subgroup a shift harms. This thesis develops a non-invasive fairness drift monitor that addresses this gap by repurposing Conformance Constrain ts\, a data-profiling primitive\, as a temporal fairness signal. The monit or learns per-subgroup distributional profiles at baseline\, using protect ed attributes only once\, and thereafter tracks violation of those profile s over incoming data batches without any runtime access to protected attri butes. Across three fairness benchmarks\, two classifiers\, and nineteen c ontrolled drift scenarios\, the conformance-constraint violation signals t rack fairness degradation more closely than KS and KL detectors under glob al drift\, and they remain competitive with them under group-targeted drif t. Subgroup localization provides its clearest advantage under global drif t\, where a minority-subgroup signal substantially outperforms both aggreg ate signals and the baselines. The correlations are modest in absolute ter ms\, indicating that the monitor functions as a screening instrument that flags fairness degradation for closer investigation rather than as a preci se estimator. The approach offers privacy-preserving\, subgroup-aware fair ness monitoring suited to regulated deployment settings.

For furth er information please contact Dr. Gokhan Kul at gkul@umassd.edu.

Event page: https://www.umassd.edu/events/cms/20260803-non-invasive-fa irness-drift-monitor-for-machine-learning.php
Event link: htt ps://teams.microsoft.com/meet/245292205330763?p=xBhMejmhVBcZHQGSk6

DTSTAMP:20260708T235306 DTSTART;TZID=America/New_York:20260803T110000 DTEND;TZID=America/New_York:20260803T120000 LOCATION:Microsoft Teams SUMMARY;LANGUAGE=en-us:Non-Invasive Fairness Drift Monitor for Machine Lear ning Models Using Conformance Constraints UID:4565056d86a29cee67fd5068d8d15b19@www.umassd.edu END:VEVENT END:VCALENDAR