Context Boosting Collaborative Recommendations Conor Hayes, Pádraig Cunningham Computer Science Department, Trinity College Dublin Abstract This paper describes the operation of and research behind a networked application for the delivery of personalised streams of music at Trinity College Dublin. Smart Radio is a web based client-server application that uses streaming audio technology and recommendation techniques to allow users build, manage and share music programmes. Since good content descriptors are difficult to obtain in the audio domain, we originally used automated collaborative filtering, a ‘content less’ approach as our recommendation strategy. We describe how we improve the ACF technique by leveraging a light content-based technique that attempts to capture the user’s current listening ‘context’. This involves a two stage retrieval process where ACF recommendations are ranked according to the user’s current interests. Finally, we demonstrate a novel on-line evaluation strategy that pits the ACF strategy against the context-boosted strategy in a real time competition. 1. Introduction This paper describes a personalised web-based music service called Smart Radio, which has been in operation in the computer science department at Trinity College Dublin for the past three years. The service was set up to examine how a personalised service of radio programming could be achieved over the web. The advent of on-line music services poses similar problems of information overload often described for textual material. However, the filtering/recommendation of audio resources has its own difficulties. Chief amongst these is the absence of good content description required by content-based or knowledge-based systems. This drawback is conventionally overcome using collaborative filtering, a technique that leverages similarity between users to make recommendations. As such it is often termed a ‘contentless’ approach to recommendation because description of the items being recommended is not required. Apart from the obvious advantage of a ‘knowledge–light’ approach to recommendation, ACF is often credited with being able find recommendations that would otherwise escape content-based recommender strategies. This is because it relies upon user preferences that may capture subtle qualities such as aesthetic merit that may escape current content-based systems. However, ACF does have well documented drawbacks such as the problem of bootstrapping new users and new content into the system. In this paper we examine a less documented weakness, that of context insensitivity, and provide a solution using a lightweight, case-based approach. Since our technique imposes a ranking based on what the user is currently listening to in the system we do not consider off-line approaches to evaluation such as cross validation or measures of recall/precision appropriate for this situation. Instead we measure whether a user was inclined to make use of the recommendations