Abstract
Recommendation Systems for video channels are highly investigated areas in the machine learning domain because of their importance in retaining viewers. This paper presents a comparison between two recommendation systems developed on the basis of a similarity index between a video and all other videos in Vimeo’s staff picks database released in January of 2018. The first model extracts text, thumbnail, and numerical features individually and used categories as a proxy for accuracy of prediction of top k recommended clips. The second model puts these three features in one feature space through normalization and dimensionality reduction and returns a list of top ten similar video clips based on distance. It was found that Model 1 yielded total category matching of 37.1% with text, which consequently validated the proposal of Model 2 both. In the future, Model 2 would be further optimized by user input to weigh features based on similarity or creating a graph where a more robust algorithm such as Page Rank could more accurately offer recommendations, though not necessarily based on similarity.