Personalized Channel Recommendation Deep Learning From a Switch Sequence

Internet protocol TV (IPTV) services could enhance personalized viewing experience in a more interactive way than traditional broadcast TV systems, but it is still difficult for subscribers to quickly find interesting channels to watch from a huge selection. This paper focuses on a framework for personalized live channel recommending via deep learning from a historical switching sequence with a long short-term memory (LSTM) neural network. Using real-world IPTV watching logs, we first obtained insights into user behaviors when watching live channels, and then proposed a learning scheme on how to dynamically generate a recommended channel list for each user with an independent LSTM net trained using the channel watching history during a slide window. For designing a good data architecture and representation scheme for a dynamically learning framework, we then studied the performance of the proposed recommendation method by varying the width of the slide window for training, the length of input sequence for prediction, and the mode to process input and label space. We finally developed a separate learning method to fairly recommend for popular (hot) or unpopular (cold) channels, respectively, based on channel popularity in the training set with an extra price of a possible hit lag after recommendation, in order to alleviate the Matthew effect arising from the conventional recommendation based on historical information. The experimental results show LSTM succeeds in learning from a historical channel switching sequence, outperforms several baseline recommendation methods, especially for hot channels, and the classified recommendation by separate learning brings an overall performance gain.