A classifier ensemble approach to the TV-viewer profile adaptation problem

作者：Christou Ioannis T^*; Gekas George; Kyrikou Anna

来源：International Journal of Machine Learning and Cybernetics, 2012, 3(4): 313-326.

DOI：10.1007/s13042-011-0066-4

• Recommender systems
• Pattern recognition
• Information search and retrieval
• Classifier ensembles
• Autonomous systems
• Adaptive systems
• Continuous learning

For the past few years the digital video broadcasting (DVB) and Internet TV (IP-TV) market has experienced a significant growth throughout the world. As a result of this trend, many broadcasters offer such services, and many homes now feature a set-top box in their living rooms providing them with a myriad of digital TV channels. This plethora of choices dictates the need for a system that can autonomously monitor a user's choices on TV and be able to predict what content the user is likely to appreciate. We describe a system called CPACE (continuous profile adaptation via classifier ensembles) that continuously monitors user choices, and using machine learning techniques, it classifies not-seen-before content from an electronic program guide (EPG) and suggests to the user content that he/she is likely to be interested in viewing. Since we use no collaborative filtering at all, our architecture is fundamentally different from other commercially successful approaches such as Ti-Vo or the winners of the Netflix Prize that utilize collaborative filtering techniques that necessarily suffer from privacy related issues. We use time-windows combined with online learning to adapt to changing user behavior and classifier ensemble techniques to enhance classification accuracy. As a result, the system after a short training period is capable of classifying new broadcasted content with very high accuracy reaching up to 86%. The system has been shown to exhibit very high precision in its recommendations in a real-world installation where it usually achieves rates of more than 90%, and for a significant fraction of the user population it achieves 100% correct recommendations.

• 出版日期2012-12