Learning individual preferences from aggregate data: A genetic algorithm for discovering baskets of television shows with affinities to political and social interests

• Published in: Expert systems with applications 2021-04, Vol.168, p.114184, Article 114184
• Main Authors: Padmanabhan, Balaji, Barfar, Arash
• Format: Article
• Language: English
• Subjects: Affinity; Aggregate data; Audience analytics; Baskets; Elections; Evolutionary algorithms; Genetic algorithms; Machine learning; Micro-targeting; Multi-objective evolutionary algorithms; Multiple objective analysis; Politics; Television; TV panel data
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2020.114184

•A framework based on genetic and multi-objective evolutionary algorithms is proposed.•The framework estimates user political preferences from unlabeled TV viewership data.•The framework discovers sets of shows whose viewership determines social preferences.•We apply and test the framework to Nielsen National Database on TV viewership in 2016.•We examine politics, global warming, same-sex marriage, and abortion in the US. This paper presents a flexible general-purpose framework using genetic and multi-objective evolutionary algorithms that can leverage “unlabeled” (and anonymized) panel data on television viewership along with aggregate-level vote or public opinions statistics to (i) identify sets of programs that have affinities with politics and social issues, and (ii) estimate individual preferences from unlabeled data. The applications of this framework are significant given the wide interest in using big data for political advertising and building election forecasting models with non-polling data. Analyzing viewership spanning over seven billion minutes from Nielsen’s TV panel for an entire year (2016), we illustrate how this framework can learn interesting baskets of programs whose viewership can help estimate individual attitudes toward politics, global warming, same-sex marriage, and abortion.