Collaborative Filtering can be divided into following two categories:
1. Memory-based collaborative filtering
The memory based approach can be further divided into user-based similarity method and item-based method. In memory-based methods, we make recommendations by considering similarity between users or items. In user-based methods, for predicting rating given by the user to an item, cosine-similarity is used to find similar users and then their ratings given to the item are averaged out. Likewise for item-based approaces, items closer in space to the correct item are found out and recommended.
2. Model-based collaborative filtering
The model based approach are used to find out latent factors involved in the interaction between users and items. To find out these latent factors, we use matrix-factorization methods such as Singular Value Decomposition and Principal Component Analysi. This handles sparsity of the rating matrix well. Deep-Learning models like Variational AutoEncoders can also be used to obtain thse latent features.
Most of the modern-day recommendations sytems use both memory and model based; a popular example would be Google News.
- Scalability with new users and new items
- Data Sparsity
- Promoting diversity for items with limitted history.