My work focuses on modelling user behaviour from internet clickstream data. During my PhD at the University of Amsterdam I used temporal probabilistic models to perform activity recognition for health monitoring elderly, you can find my PhD thesis here. I did my postdoc with the Activity Recognition and Ambient Sensing (ARAS) group at the Boğaziçi University in Istanbul, Turkey. And I worked as a senior researcher at AGT International, a smart city solutions provider.
My publications can best be seen as answers to the following questions:
- Which model can be used to recognize human activities from audio? [HHMM ->WASPAA 2013]
- Does the use of a hierarchy help recognition of human activities? [HHMM -> AmI 2011]
- Can we limit annotation efforts by asking the user for labels? [Active learning -> AmI 2011]
- How well do discriminative and generative models perform? [HMM, CRF -> Ubicomp 2008]
- Does modeling the duration of an activity help recognition? [Semi-Markov models -> JAISE 2010]
- How can we apply activity recognition on a large scale? [Transfer learning -> Pervasive 2010]
- How should we evaluate models for activity recognition? [F-measure, error metrics -> COSDEO 2011]
A full list of my publications can be found here.
We participated in the IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events. Our submission was for the Office Live, Event detection task and we finished second place using a hierarchical hidden Markov model for sound event recognition.
The results, can be found at: http://c4dm.eecs.qmul.ac.uk/sceneseventschallenge/Together with several members of the Data Analytics team at AGT R&D we finished first place in the IEEE GRSS Data Fusion competition. Our classification map ranked first after evaluation against an undisclosed validation set. More than 50 international teams entered the Best Classification Challenge.
The top-ten results, ranked based on their accuracy, are summarized at: http://hyperspectral.ee.uh.edu/?page_id=695