Predicting video popularity of streaming services with machine learning approaches

Live or online video streaming services have gained immense popularity in recent years due to the expansion of the internet and the impact of Covid-19. However, delivering content to an ever-increasing user base poses challenges to online video streaming companies. To respond to user demands and preferences, a prediction model with high accuracy is needed. In this thesis, a predictive model is developed to anticipate a video's popularity as popular or unpopular by applying machine learning algorithms to metadata and textual features of each video from a prominent online video streaming provider in Iran. The study shows that video popularity can be modeled with high accuracy based on video-related attributes and textual features in Persian language. Four classification models are applied, with the Random Forest model achieving the highest accuracy and F1-Score of 86% and 72%, respectively. The Support Vector Machine obtains the most accurate results when new attributes obtained through NLP are combined with metadata. Moreover, the inclusion of word embeddings of the video description as predictive features improves classifiers' performance significantly. The study finds that the number of program episodes, video type, channel, and year of production are the most influential features in predicting video popularity. Predicting video content popularity in advance has enormous benefits for marketing purposes, network usage, and network cost reduction.