High Fidelity Video Prediction with Large Neural Nets

Improving YouTube personalization using clustering of videos Internship at Google, Bangalore. Mentored by Sumit Sanghai May July 2015 We explored new ways to improve YouTube user profiles by trying to find ways of modelling interests that are not well represented by Knowledge Graph entities (e.g. “70s music”). Towards this goal, we used clusters of videos as users' features. The input data for the clustering was derived from video correlations due to user co-watches. We tried k-means, HAC and LDA to generate video clusters. We built a simple video recommendation system using clusters as features. We had to deal with large amounts of input data, and thus, had to use compute clusters for distributing the tasks. Consequently, the project also involved heavy usage of distributed frameworks, like MapReduce.