Predictive self-learning content recommendation system for multimedia contents

Millions of users use the Internet for entertainment, education, shopping and many other purposes. For instance; one billion hours of YouTube videos are watched every day. One of the key features of such platforms such as the entertainment and shopping platforms is the recommendation system based on past activities of users and the contents of the visited sites to provide related contents to decrease search time and increase the data availability. However, the content suggestions have several challenges based on the platforms. Some challenges are: collected data from users can be noisy, the media contents are not well-defined, and users do not want to cooperate. To solve those issues, YouTube recommendation, Netflix, AWS re:invent, and similar commercial sites proposed the context-aware personalized recommendation systems. Most of the recommendation systems use implicit and explicit user past activities with mapping relation between contents. However, the recommendation systems are general and cannot be changed according to user characteristics after visiting the suggested data. For example, a user mostly visits the higher-ranking content which is related to the visited content, and another user can visit the recent content which has high feedback during the different days or even in an hour according to mood. Therefore, in this paper, we propose a predictive self-learning recommendation system. The algorithm predicts what a user searches next by using prior collected information and using machine learning to analyze the user behaviors for the future activity. The results show that our proposed recommendation system is efficient in terms of CPU usage and response time while characterizing users' behaviors in short and long terms. In this paper, we only analyze the short term characterizations of the proposed method. The proposed method and related analysis can assist the shopping, entertainment and similar recommendation systems to increase their efficiency by well-characterizing users' behaviors.