Bayesian latent variable models for the collaborative Web
Morgan Harvey · STAX · 2011
Since its creation in the early 1990s the Web has held the promise of allowing near-instantaneous communication, participation and sharing of resources and ideas between users across the globe. However up until fairly recently, the Web was predominantly a large collection of static documents providing no real scope for such interaction. The past decade has seen the arrival and rapid growth of the so called “Web 2.0” movement where sites have become increasingly more social with users able to share information with others. Due to the sheer volume of information available on the Internet and the massive number of products available on online shops, finding items which may be of interest can often be a very difficult task. Furthermore the continual expansion of the Web makes it impossible to manually evaluate each new item to determine if it might be of interest. In recent years the emergence of a more social web has resulted in the development of tools with the purpose of making this undertaking both easier and more enjoyable. This thesis explores two avenues of this new social web: social tagging and ratings-based collaborative filtering. Both of these methods rely on the users of the system to provide some information about the resources contained therein and then use this information to improve the user experience. The main hypothesis of this thesis is that these new social tools can be significantly improved by the use of machine learning methods to model and make sense of the data available. The work introduces a family of novel latent variable Bayesian models designed specifically to deal with this sparse and extremely noisy data. A series of experiments carried out on real-world data sets show that these models can overcome the inherent difficulties and provide significant improvements in performance over state of the art systems. Furthermore it is shown that the output of these models is more readily interpretable than from competing models and can therefore be utilised to gain a more complete understanding of the complex social and topical dynamics of such systems.