A probabilistic reciprocal recommender with temporal dynamics

Recommender systems are methods of personalisation that provide users of online services with suggestions for further interaction with those services. Most recommender systems are for product-to-consumer recommendation, suggesting items or products to users, but there is a growing need for reciprocal recommender systems, where the goal is to suggest users to other users of the system with whom they might like to interact. Unlike product-to-consumer recommendation, for reciprocal recommendation to be successful both parties must agree to the interaction. Reciprocal recommendation is needed, for example, in matching users in online social networks for friendship or dating, or matching users to employers in online recruitment. Since online social network websites typically have many millions of users, the problem of finding successful interactions is not a trivial task, and automating the process becomes essential.Most existing reciprocal recommender systems use either profile similarity or interaction similarity to recommend new matches, assuming that user preferences are static and ignoring temporal aspects of user behaviour. This thesis takes a different approach, and addresses the issue of representing user preferences as dynamic. We introduce a new representation for changes in user preferences and use that representation in creating a reciprocal recommender system applied to online dating.We first show by analysing a large real-world dataset on people-to-people interactions from a dating website that user behaviours and activity levels do change over time and thus require a dynamic model to represent them. The analysis includes a detailed specification of behaviours captured in the data and exploratory data analysis to understand the underlying features of the data. In addition we include experiments on the discovery of sub-groups of users based on time-varying activity levels, and the use of these subgroups in a measure of similarity which is applied in a collaborative filtering recommender system to improve the ranking of recommendations.Next, we introduce a novel method of representing observations in the dataset as a sequence of interaction events instead of the traditional approach in which each observation is considered as an isolated event. Then we propose, design and implement a Hidden Markov model (HMM) content-based reciprocal recommender system that uses this new representation to generate recommendations. This HMM content-based reciprocal recommender system was evaluated on the real-world dating site dataset and our results show that it generates improved recommendations compared to existing systems that do not use an explicit temporal representation.Finally, we develop a general framework for combining the HMM content-based reciprocal recommender system with collaborative filtering techniques to create a unified hybrid recommender. Additionally, a new similarity measure is introduced to rank the recommendations generated by this hybrid recommender. Moreover, we propose, design and implement a reciprocal recommender system using the suggested framework and the new similarity measure. Evaluation of this system shows that it generates better recommendations than existing systems in a time-efficient manner