For recommendations engine what is the advantage and disadvantage of those technique (matrix factorization:ALS, pearson or cossin correlation) and how we deside which technique to use.
1
votes
1 Answers
3
votes
Collaborative filtering using a pearson correlation suffers from a few issues. I'll list a few of the big ones:
Scalability
When your database grows, it will take increasingly longer to calculate a pearson correlation between users. Since you have to calculate a correlation between each user, this will grow exponentially.Data sparsity
This is a big problem for most recommendation engines. If you have a lot of user and a lot of items with a few reviews, it will become difficult to generate recommendations, since you have too little data to calculate a correlation between users.Cold start problem
Technically, each method suffers from this problem, but matrix factorization handles this better than a collaborative filter. A cold start problem basically means that you have no or barely any data about a user. You can't solve this when you are using a collaborative filter. Period.
You can use alternate techniques to surpass this problem, i.e. content-based filtering.
Determining which method you should use can be difficult. Matrix factorization outperforms traditional user-based and item-based collaborative filtering, but you have to decide if it would suit your model best.
If you don't have a sparse database, a collaborative filter would work well, but so would a matrix factorization method.
Here are some interesting websites containing data about these methods. In the end, it's up to you or your team to decide which method works best.
Matrix Factorization Example: http://www.quuxlabs.com/blog/2010/09/matrix-factorization-a-simple-tutorial-and-implementation-in-python/
Content and collaborative filtering: http://recommender-systems.org/
Feel free to ask more if this wasn't clear enough!
