BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Inferring application performance regardless of data completeness 
DTSTART:20140619T163000
DTSTAMP:20260509T234230Z
UID:80475096463ef11b89ee4e6685ae8034ad77163a49554395ddebdf99
CATEGORIES:Conferences - Seminars
DESCRIPTION:Alessandra Sala\, Bell Labs Ireland\nModern communication netw
 orks\, such as online social networks and call networks\, give us a unique
  opportunity of observing\, analyzing and better understanding human behav
 iors. In Telecommunication industry\, user’s data are considered a preci
 ous source of information to shed light into novel insights to drive the d
 esign of future communication platforms. Unfortunately\, in the presence o
 f factors such as increasing privacy awareness\, restrictions on applicati
 on programming interfaces (APIs) and constrained sampling strategies\, ana
 lyzing complete datasets is often unrealistic. For instance\, partial netw
 ork views are basically default in telco analytics\, as customers typicall
 y have frequent contacts with customers of other providers - which natural
 ly cannot be observed\; or\, accurately inferring user activity is the Hol
 y Grail of mobile advertisement and targeted service offering because priv
 acy restrictions usually do not allow the logging of complete URLs.\nThis 
 talk discusses the potential and risks of mining partial data with the ana
 lysis of two specific use cases. In the first use case\, we unveil the hid
 den effects in the evaluation of marketing campaign in social networks whe
 n the spread of information is estimated from a partial view of the networ
 k. The proposed methodology is able to quantify the error introduced due t
 o network partiality based on a theoretical oracle scenario and correct fo
 r the introduced error at large extent. In the second use case\, we show a
 n approach to mine mobile web traces form heavily truncated URLs and infer
 ring user activities with high accuracy. Truncated URLs are trimmed from i
 nformation like location or purchased products\, to mask possibly sensitiv
 e end user data. Furthermore\, URLs derived from real web traces are highl
 y noisy because dominated by unintentional web traffic like advertisement\
 , web analytics or third parties scripts. We have developed a statistical 
 model to segregate representative URLs characterizing the user activities 
 from unintentional web traffic and demonstrated that our approach classifi
 es user activities with 92% accuracy.
LOCATION:BC 420 https://plan.epfl.ch/?room==BC%20420
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR