MLConf 2013 notes (4)

mlconf_twitter.md

Personalization and Recommenders with Content-Based Approaches

Jake Mannix, Applied Machine Learning Engineer, Twitter

• Does collaborative filtering == RecSys?
• cautions that recommenders are not just user->products
• you might actually recommend groups to a user, user to a groups, candidates to a posting (example: Linkedin)

math techniques:

• singular value decomposition or SVD
• [ALS and other matrix factorizations] (http://www.slideshare.net/srowen/matrix-factorization)
• stacked RBM

"At he end of the day the mathematical approaches reduce the user/item matrix to GUIDs but often you need to bring domain knowledge to bear."

• each user has a feature vector
• each item has a feature vector
• dimensionality may (WILL) vary

Doing content-based recommendation

• helps with cold-start problem (you don't need to wait for ratings)
• it scales really well
• flexible, can learn to rank using logistic regression, SVM, GBDT ("Gradient Boosted Decision Tree")

but

• it's not very general, you are using domain-specific model
• features can be too coarse, feature selection may be difficult
• low-latency from large item set may be difficult

Sometimes you can just keep it simple:

• combine with "popularity": find the popular things and then take them and re-rank those in the context of content-based recommendations
• Twitter's own recommendations combine content social and collaborative filtering features
• And remember: collaborative filtering is "unreasonably effective"

Further

• [Content-based filtering] (http://en.wikipedia.org/wiki/Recommender_system#Content-based_filtering) on Wikipedia
• "Combining Predictions for Accurate Recommender Systems"