View Count of Online Videos Prediction Using Clustering View Count Patterns with Multivariate Linear Model

Abstract

Social media of online video content should have some interesting applications, which provide convenience to users. For example, online advertising should be shown to the targeted users and the recommender system suggests appropriate videos to a user. Video view count prediction can help improve these recommendations because the video will be more interesting if it gets more view count. On YouTube, video view count and revenue are positively correlated. However, to correctly predict the view count of a video on YouTube, we divide problems into two issues. First, the view count has several patterns, which cause poor prediction. They should be grouped and trained specifically. Second, some patterns cause more errors in prediction because they are rare. Therefore, they should be separated to prevent overfitting. In this paper, we aim to design a model, which accurately predicts the short-term view count of videos on YouTube. We present F7NML (The First 7-day Normalization for clustering with Multivariate Linear model), a predictive model that can group the view count patterns and remove outliers. First, we grouped the patterns into many groups using the clustering model, which is presented in the paper. Then, we removed the groups of rare patterns, which are called outliers. Next, the video view count in the test dataset was matched to the groups using 1-nearest neighbor method. Finally, the Multivariate Linear model was trained for each group specifically. The experimental results show that F7NML with an appropriate clustering model could reduce error when it was compared to the best baseline model from the literature by about 27% on the 30th day view count prediction.