Hybrid machine learning approach for popularity prediction of newly released contents of online video streaming services

In the industry of video content providers such as VOD and IPTV, predicting the popularity of video contents in advance is critical, not only for marketing but also for network usage. By successfully predicting user preferences, contents can be optimally deployed among servers which ultimately leads to network cost reduction. Many previous studies have predicted view-counts for this purpose. However, they normally make predictions based on historical view-count data from users, given the assumption that contents are already published to users. This can be a downside for newly released contents, which inherently does not have historical data. To address the problem, this research proposes a hybrid machine learning approach for the popularity prediction of unpublished video contents. In this paper, we propose a framework which effectively predicts the popularity of video contents, via a combination of various methods. First, we divide the entire dataset into two types, according to the characteristics of the contents. Next, the popularity prediction is performed by either using XGBoost or neural net with category embedding, which helps resolving the sparsity of categorical variables and requiring the system to learn efficiently for the specified deep neural net model. In addition, we use the FTRL model to alleviate the volatility of view-counts. Experiments are carried out with a dataset from one of the top streaming service companies, and results display overall better performance compared to various standalone methods.