Collaborative-Filtering versus Content-based RS:
Considering OTT platform’s data as an example, which is basically a table with user profiles are listed in rows, movies in columns and the ratings are the values for each pair (user,movie) respectively. Collaborative Filering(CF) Recommender Systems tends to use the rating values for future predictions. The core idea on which CF works is that -“users who have agreed in the past will also agree in the future.”
Whereas, the Content-based Recommender Systems doesn’t use ratings. Content based (CB) Recommender Systems typically tend to use other features inorder to predict future ratings. For each movie, a vector (Ij)is created based on few keypoints such as “movie_genre”, “year_released”, “actor_roles”, “director”, etc. And for each user data point, it will create a vector(Ui) based on other data such as “gender” , “ age group” , “location”, etc. And then proceeds to apply Classification/Regression model on the vectors Ui, Ij.
Collaborative Filtering RS can be thought of as User-User similarity algorithm in which cosine similarity is typically used to obtain the similarity matrix. Content Based RS can be thought of as Item-Item similarity algorithm.
Which one to opt?
Content based Recommender System is preferred over the other as it would take into account the fact that user likes/dislikes changes over time. As users preferences play a cruical role in vectorization, content based model will be less prone to errors leading to better recommendations. Item-Item RS tends to not change over time and thats the reason why major OTT platforms choose this for their business.
References:
