Reciprocal Recommender Systems

Recommender systems are algorithms that use information about a user to predict how much they are likely to like an item. They are used on a variety of sites, including online shopping sites, streaming sites and social networks. Two classes of algorithms are particularly widespread: content-based filtering and collaborative filtering. Content-based filtering makes recommendations based on matches between user-defined preferences and item descriptions. Collaborative filtering uses similarities between the behaviour of different users to make recommendations. Collaborative filtering, popularized by the shopping site Amazon.com, has had the most success, and most modern algorithms use the principle to some degree. Traditionally, recommender systems are unidirectional: a single user is recommended an inanimate item, and makes a decision about the item. In the case of online dating, users are recommended other users instead of items. These recommendations are only useful if they are reciprocal: the user seeing the recommendations and the user being recommended must like each other, adding complexities to traditional models. There are many public datasets for unidirectional recommender systems, including the Netflix dataset of approximately 100 million movie ratings from users. Based on these, standard evaluation metrics for recommendation algorithms have been developed, including different metrics for the quality of individual recommendations, the quality of the ordering of the list and the reliability of the system
This project falls within the EPSRC Engineering research area'