Video popularity prediction in data streams based on context-independent features.

Data Stream is the name given to the model under which a considerable portion of data is generated today, showing an unlimited and unpredictable behavior. Data stream mining appears as an alternative to conventional methods of extracting patterns from these data while maintaining comparable performance. As an example of this phenomenon of data volume growth, video sharing services counts every day millions of hours of content watched, generating billions of views. This work applies a straightforward data stream mining technique, namely Hoeffding Tree, to a subset of video data stream. To do so, this work presents the concepts of data stream mining, focusing on classification methods and detailing the Hoeffding Tree algorithm. We propose a problem of predicting video popularity on video streaming services by applying this algorithm. A number of predictive models is generated with the intention of exploring video characteristics that can be used to predict its popularity before the video being sent to the sharing service. Two final predictive models are generated, reaching accuracy values of 34.5% and 43.4%. Also, the algorithm is implemented in Python and made publicly available, aiming to address the lack of this tool in the libraries of this language.