Surfing the Digital WaveGenerating Personalised TV Listingsusing Collaborative, Case-Based RecommendationBarry Smyth & Paul CotterDepartment of Computer ScienceUniversity College DublinBelfield, Dublin 4, Ireland{Barry.Smyth, Paul.Cotter}@ucd.ieAbstract. In the future digital TV will offer an unprecedented level ofprogramme choice. We are told that this will lead to dramatic increases inviewer satisfaction as all viewing tastes are catered for all of the time. However,the reality may be somewhat different. We have not yet developed the tools todeal with this increased level of choice (for example, conventional TV guideswill be virtually useless), and viewers will face a significant and frustratinginformation overload problem. This paper describes a solution in the form ofthe PTV system. PTV employs user profiling and information filteringtechniques to generate web-based TV viewing guides that are personalised forthe viewing preferences of individual users. The paper explains how PTVconstructs graded user profiles to drive a hybrid recommendation technique,combining case-based and collaborative information filtering methods. Theresults of an extensive empirical study to evaluate the quality of PTV’s case-based and collaborative filtering strategies are also described.1 IntroductionWith the advent of new cable and satellite services, and the next generation of digitalTV systems, we will soon be faced with an unprecedented level of programme choice.Where we have tens of TV channels today, tomorrow we will have hundreds, andsoon after that it will be thousands. If we believe the hype, we are entering a new ageof television viewing, an age of incredible choice and unprecedented viewingsatisfaction. However, while increased programme choice does offer the potential forimproved viewing satisfaction, the reality may be somewhat different. We have notyet developed the tools to deal with this new level of choice, and it will becomeincreasingly difficult to find out what programmes are on in a given week, never mindlocating a small set of relevant programmes for a quiet evening’s viewing.Consider for example the traditional TV guide, listing programming informationon local channels for up to a week in advance. The days of a slim, easy to digest 30page volume are essentially gone. Looking to the US for a sign of things to come wenotice, with some consternation, that the current issue of the TV Guide (that weeklybible for American channel surfers) runs to nearly 400 pages of indigestible schedulecharts. Moreover, the way that we interact with our TV sets will also have to change.Those rapid “remote-controlled surfs”, that prove so effective (and so irritating toyour partner) for 10 or 20 channels, will no longer be a viable means of finding outwhat is on at a given time. A 10 second per channel surf over even a modest 200channel service will take about 35 minutes to complete! The digital TV vendors dorecognise this as a serious information overload problem, and in response they arenow offering electronic programme guides to help users to navigate this digital maze.However, these guides are relatively crude and offer little more than a static categorybased view of the evenings programming; the burden of search remains with the user.This paper describes the PTV system (http://ptv.ucd.ie), which offers a workingsolution to the problem of locating relevant programme information quickly andeasily. PTV combines user profiling and case-based reasoning (CBR) techniques togenerate electronic TV viewing guides that are carefully personalised for the viewingpreferences of individual users ([2, 6, 7]). At the present time, these electronic guidesare Web based and delivered over the Internet to desktop PCs, but of course theadvent of WebTV and cable-internet services will allow PTV to deliver personalisedprogramme information directly to the TV set.The remainder of this paper is organised in the following way. The next sectionprovides an overview of the PTV system, describing its various sources of knowledgeand main functional components. Section 3 focuses on PTV’s user profiling and case-based recommendation strategy. Before concluding, section 4 reports the results of anextensive empirical study to evaluate the quality of the personalised programmeguides that are produced, and the effectiveness of the case-based and collaborativerecommendation strategies. Finally, a new appendix has been added to indicate thecurrent state of the PTV system including the results of a recent online survey that addfurther support to the PTV concept and, we believe, paves the way for the use ofcollaborative, case-based recommendation methods in a wide range of personalisedmedia service in the future.2 The PTV SystemPTV is a client-server system operating over the Web, allowing users to register,login, and view their personalised TV guides as specially customised Web pages. Thearchitecture of PTV is shown in Figure 1. A standard Web browser provides therequired client functionality, and all user interaction is handled via the HTML Formsinterface. The heart of PTV lies with its server-side components, which handle all themain information processing functions such as user registration and authentication,user profiling, guide compilation, and the all-important programme recommendationand grading.In the following sections we will concentrate on user profiling in PTV, focusing onhow these profiles are used to deliver personalised content and, in particular, howPTV can make intelligent recommendations to PTV subscribers. However, in thissection we will provide a suitable backdrop for these future discussions by taking abroad look at the form and function of PTV’s main components.Profile Database & Profiler: The key to PTV’s personalisation facility is an accuratedatabase of user profiles. Each user profile encodes the TV preferences of a givenuser, listing channel information, preferred viewing time, liked and dislikedprogrammes, subject preferences, etc (see Figure 1). Preliminary profile informationis collected from the user at registration time in order to bootstrap the personalisationprocess. However, the majority of information is learned from grading feedbackprovided by the user; each recommended programme is accompanied with