A collaboration with the data mining group at TU-Berlin and folks at Lancaster and Glasgow has seen a full paper accepted by ACM International Conference on Interactive Experiences for Television and Online Video (TVX 2017), Hilversum, The Netherlands, 06/2017. The acceptance rate is 31%, a competitive year for this conference series.

The paper describes our recent efforts in breaking the filter bubble, a term used to reflect the phenomenon that a recommendation algorithm guesstimates a user’s preference from limited contextual information (such as user clickstream data) and only provides the user with a very small selection of content based on the preference. A side-effect of such an approach is that it often ends up isolating a user from (a large amount) of content that the system does not believe would interest him or her. As a user selects from within the bubble, the bubble may also become smaller and more “specialised”, causing a negative cycle. We believe that the recommender should be smarter than it is and “talk” to its users as their friend. A friend who knows what you like and yet very often surprise you with new and cool things. We studied this contextual bias effect in an online IPTV system (to which I was a project lead for some years), and developed a novel approach to re-balance accuracy and diversity in live TV content recommendation using social media.

Yuan, J., Lorenz, F., Lommatzsch, A., Mu., M, Race, N., Hopfgartner, F., and Albayrak, S., Countering Contextual Bias in TV Watching Behavior: Introducing Social Trend as External Contextual Factor in TV Recommenders, to appear in 2017 ACM International Conference on Interactive Experiences for Television and Online Video (TVX 2017), Hilversum, The Netherlands, 06/2017

Abstract:
Context-awareness has become a critical factor in improving the predictions of user interest in modern online TV recommendation systems. In addition to individual user preferences, existing context-aware approaches such as tensor factorization incorporate system-level contextual bias to increase predicting accuracy. We analyzed a user interaction dataset from a WebTV platform, and identified that such contextual bias creates a skewed selection of recommended programs which ultimately leaves users in a filter bubble. To address this issue, we introduce a Twitter social stream as an external contextual factor to extend the choice with items related to social media events. We apply two trend indicators, Trend Momentum and SigniScore, to the Twitter histories of relevant programs.The evaluation reveals that Trend Momentum outperforms SigniScore and signalizes 96% of all peaks ahead of time regarding the selected candidate program titles.