Predicting Online Participation in Public Broadcasting Using Machine Learning

Abstract

As it remains unclear how the success of public broadcasting on social media should be measured, this thesis argues that online participation could serve as an alternative, social media metric that aligns with public broadcasting’s traditional aims and draws support from more general public research and audience research. Consequently, public broadcasting benefits from understanding what online participation can be expected of their new media content. For this reason, and to help characterize online participation as an alternative metric of online participation, this thesis aims to investigate to what extent a predictive model of online participation can be built, and which predictors are important in the process. In addition, this thesis aims to investigate whether topic modeling can be applied on the subtitles of public broadcasting shows to generate useful features for the prediction of online participation. Using data from the Dutch public broadcasting service, 22 potentially predictive features of online participation were collected and created, missing values and outliers were dealt with, and 7 models were individually tuned and compared with the aim of achieving the best prediction performance. The results suggest that although most values of online participation can be predicted with decent accuracy, the model performs poorly on large values of online participation. Furthermore, the results indicate that the inclusion of topics as features did not lead to significant improvements in prediction performance but do generate some useful insights. Scholars and public broadcasting organizations may use the results of this thesis to enhance their understanding of online participation as an alternative metric of broadcasting success.