Exploiting Social Networks for Recommendation in Online Image Sharing Systems
This thesis aims to demonstrate the distinct and so far little explored value of knowledge derived from social interaction data within large web-scale image sharing systems like Flickr, Picasa Web, Facebook and others for image recommendation. I have shown how such systems can be significantly improved through personalisation that takes into account the social context of users by modelling their interactions by mining data, building and evaluating systems that incorporate this information.
... improvements allow users to search and browse large online image collections more quickly and to find results that more accurately match their personal information needs when compared to existing methods. Traditional information retrieval and recommendation datasets are contrived to provide stable baselines for researchers to compare against but they rarely accurately reflect the media systems users tend to encounter online. The online photo sharing site Flickr provides rich and varied data that can be used by researchers to analyse and understand users' interactions with images and with each other. I analyse such data by modelling the connections between users as multigraphs and exploiting the resultant topologies to produce features that can be used to train recommender systems based on machine learnt classifiers. The core contributions of this work include insight into the nature of very large-scale on- line photo collections and the communities that form around them, as well as the dynamic nature of the interactions users have with their media. I do this through the rigorous evaluation of both a probabilistic tag recommendation system and a machine learnt classifier trained to mimic user decisions regarding image preference. These implementations focus on treating the user as both a unique individual and as a member of potentially many explicit and implicit communities. I also explore the validity of the Flickr 'Favourite' feedback label as proxy for user preference, which is particularly important when considering ot [...]