Breaking the Bubble: Semantic Patterns for Serendipity

The amount of information available online continues to grow exponentially. Techniques such as filtering and recommender systems help users manage this amount of information by filtering and organising information based on preferences to display more relevant content. To avoid phenomena such as filter bubbles and echo chambers, these algorithms display diversified content. Often, this diversification creates high levels of cognitive dissonance, pushing users further into polarised views. These algorithms are therefore indirectly responsible for exacerbating these problems. This thesis presents an approach designed to create manageable cognitive dissonance, enabling users to accept diverse content. Specifically, we focus on the concept of serendipity as a user-centric metric for evaluating recommendation results. The methodology introduced by this thesis combines two fields, i.e. Linked Open Data-based recommender systems and psychological theories of curiosity. Specifically, we use semantic patterns that emerge naturally in LOD sources, guided by curiosity-based user models, to identify serendipitous items. Central to this approach is the enrichment of data with LOD concepts. This includes user activities and interests as well as item metadata. The enriched data enable concept matching between the user profile and the item descriptions. The next major step in the approach is the selection of serendipitous items using the paths in these graphs (i.e., semantic patterns). To this end, we use a mixed evaluation approach: the algorithmic part was tested with offline methods, while the user-related part was tested with online experiments. Overall, our results show that it is possible to address serendipity in a traditional media context (e.g. music, television and books). Three conceptual contributions stand out from this result: an actionable definition of serendipity, a model for characterising users in detail, and a strategy for mining a rich knowledge base based on this characterisation. Having an actionable definition of serendipity is fundamental to operationalising such an abstract concept into a measurable metric. Therefore, we define serendipity as a balanced combination of diversity and relevance of recommended items with respect to the user's interests. We designed our system according to the principles of human-centred design research. In particular, we developed a persuasive technology by focusing on detailed user characterisation. To achieve this goal, our user profiles include what the user likes and a more personal characterisation of the user. A strategy for navigating a rich knowledge base guided by a user-centric profile is fundamental to achieving the serendipity effect. Semantic patterns allow us to view the large, rich knowledge base from the perspective of a particular user, bringing order to an otherwise complicated collection of information. Our three key contributions include an actionable definition of serendipity, a detailed user characterisation model, and a strategy for exploiting a knowledge base based on this characterisation. Nevertheless, breaking the filter bubble means accepting the intrinsic diversity in our society.