As linear TV remains a major source for media consumption, multiple stakeholders such as broadcasters and advertisers are interested in the prediction of the next programs to be watched by TV viewers. Such predictions are quite challenging given the nature of the domain, where viewing TV is not just an individual activity but is also influenced by various contextual factors. We aim to address this challenge by integrating temporal aspects of linear TV – in the form of sequences – into the next program prediction process. A user profile is built from sequences of 24-hour TV programs’ views, at intervals of 15 min. Such profiles allow us to capture viewing preferences and sequential patterns, for predicting the next program/genre/category to be viewed at any time. We conducted several experiments using naive approaches, hidden Markov models and deep learning, juggling between accuracy and interpretability of the model. The precision@1 results were extremely promising (0.836 for next category prediction with LSTM, comparing to 0.57 an 0.34 with Naive approaches, and 0.366 with hidden Markov models).
Bibliographical noteFunding Information:
This research was supported by THE ISRAEL SCIENCE FOUNDATION (grant No. 262/2017 ) and was partially supported by the “ InfoMedia ” project, a project of the Israeli Innovation Authority, the Ministry of Economy and Industry.
© 2021 Elsevier Ltd
- Recommender systems
- TV domain
ASJC Scopus subject areas
- Engineering (all)
- Computer Science Applications
- Artificial Intelligence