Rethinking search and recommendation for human multimedia
Dr. Cynthia Liem, Technical University of Delft
In today's digital information society, we daily rely on information from very large archives and collections. Therefore, we increasingly have to rely on automatic filtering and recommendation algorithms to help us navigating these spaces. These days, automatic filtering procedures are frequently acknowledged for giving more empowerment and oversight over large-scale information resources than humans can have on their own. The procedures are also typically set up to operate in 'frictionless' ways, trying to minimize the amount of required user effort in indicating what is relevant, and trying to avoid that a user will disagree too often with the system's suggestions.
In this talk, I will focus on several use cases dealing with larger collections of multimedia, in which items are about (or intended for) humans. For these cases, I will argue that empowerment and oversight cannot exist without multidisciplinary domain and explicit purpose understanding, that the frictionless paradigm may not always be the right one, and that we generally need to rethink carefully how and where the success of a system should be assessed.
First, I will discuss ongoing debates and current challenges regarding computer-assisted decision-making in personnel recruitment and selection, as currently investigated under an ERASMUS+ 'Big Data for Psychological Assessment' Strategic Partnership. The introduction of pattern recognition and machine learning techniques for multimodal job candidate applications has promises (e.g. higher scalability, more systematic analyses), but also evokes threats (e.g. susceptibility to data bias) that call for transparency and explainability. At the same time, 'transparency' and 'explainability' may be understood differently by different fields, and even counteract 'acceptability' for end users. I will discuss where different disciplines and stakeholders will have different perspectives on the problem, and how clear scoping of concepts and focus areas before tackling transparency and explainability will be critical.
Secondly, I will focus on multimedia recommendation, discussing how common data representation and evaluation paradigms in the field, as well as the 'frictionless' concept, work against consumption diversity and information accessibility. I propose that consciously introducing friction, and purposefully going *against* a user's explicitly evidenced taste---even in domains that are dominated by personal and subjective taste, such as music recommendation---may be much more beneficial and rewarding in the long run, provided that this will be done in strategic and respectful ways.
Dr. Cynthia C. S. Liem MMus is an Assistant Professor in the Multimedia Computing Group of Delft University of Technology, The Netherlands, and a Visiting Researcher at the FLOWERS team of Inria Bordeaux Sud-Ouest. She holds degrees in both Computer Science and Classical Piano Performance, and gained industrial experience interning at Bell Labs Europe Netherlands, Philips Research, Google UK and Google Research, Mountain View, USA. Her research interests have strongly been motivated by her background in both engineering and music, and increasingly shift towards making people discover new interests and content which would not trivially be retrieved in search and recommendation scenarios. As a computer scientist, she is recipient of several prestigious awards, including the Lucent Global Science (2005) and Google Anita Borg Europe Memorial (2008) scholarships and the Google European Doctoral Fellowship 2010 in Multimedia. She also was a finalist of the Dutch-Belgian New Scientist Science Talent Award 2016 for young scientists committed to public outreach, and will be Researcher-in-Residence at the National Library of The Netherlands in the second semester of 2018. As a performing musician, her Magma Duo (with Emmy Storms, violin) has also been award-winning both nationally and internationally.
Location: JRC Sevilla.
Registration required before April 13th. Click here to follow registration link.