A recommender system for the Neverthink mobile application

Abstract

This thesis covers the algorithms for video channels recommendations in Never- think mobile application. It is the application which offers video content manually selected from the Internet and then aggregated in channels based on the main topic of a video. By introducing recommendations, we want to improve our engagement metrics like retention, length of a session, the number of channels watched during a session. We have a hypothesis that unless our channels are ordered based on users’ preference, users might have high churn and low level of interaction with the application since they do not see all the content we have, which might be interesting for them.

During our research, we have covered several of the most well-known approaches for collaborative filtering with explicit feedbacks, like SVD, SVD++, kNN. For the research purpose, we have also looked at some of the more complex techniques like Bayesian Factorisation Machines (FM) and Variational Autoencoders. Bayesian FM gave us the best results in terms of RMSE. However, this approach was not suitable for our application architecture, so we proceeded with kNN, which gave us the second-best outcome.

After we implemented the kNN based recommendations in the application, we ran an A/B test. We split our users into two groups of the same size and gave them different versions of the application. The version with the recommendation system has shown higher engagement rates than the version without it. Users in the version with recommendations have longer sessions inside each channel, and they spend more time in the application during the day.