Recommendation overview
Recommendation systems are one of the most successful and widespread applications of machine learning for businesses. You can use a recommendation system to help your users find compelling content in a large body of content. For example, Google Play Store offers millions of apps, while YouTube offers billions of videos, with more apps and videos added every day. Users can use search to find new content, but that is limited by the search terms they use. A recommendation system can suggest content that users might not have thought to search for on their own. For more information, see Recommendation Systems Overview.
Machine learning algorithms in recommendation systems are typically classified into the following categories:
- Content-based filtering: uses similarity between items to provide recommendations. For example, if a user watches two cute cat videos, then the recommendation system can recommend more cute animal videos to that user.
- Collaborative filtering: uses similarities between users (based on user queries) to provide recommendations. For example, if user A searches for similar things to user B, and user B likes video 1, then the recommendation system can recommend video 1 to user A, even if user A hasn't watched any videos similar to video 1.
Matrix factorization models
Matrix factorization models are widely used as a collaborative filtering method for recommendation systems.
In a matrix factorization model, user-item pairs are mapped to a two-dimensional matrix, with the unique users on one axis and the unique items on the other axis. Ratings that a user has given items reside in the cells of the matrix. This matrix doesn't need to be completely full; the majority of the time, users won't have a value for each item. The goal of the matrix factorization model is to create two smaller, dense matrixes of weights that when multiplied together, approximate the original matrix cell values and provide predicted ratings for the empty matrix cells.
One of the smaller matrixes contains the unique users on one axis and the number
of latent factors on the other axis, as specified by the
NUM_FACTORS option
of the CREATE MODEL statement. The other smaller matrix contains the unique
items on one axis and the number of latent factors on the other axis. In this
matrix, the latent factor weights are generated by the algorithm used to train
the model, based on the user-item combinations from the input matrix.
For more information, see Matrix Factorization.
You can use a matrix factorization model with the
ML.RECOMMEND function
to make recommendations.
Other models for recommendation
To extend a collaborative filtering-based recommendation system beyond what is
possible with a matrix factorization model, you can use
deep neural network (DNN)
and
Wide-and-Deep
models with the
ML.PREDICT function
to make recommendations. These models can incorporate query and item features to improve the
relevance of recommendations. For more information, see the following resources:
Disable comment: * Recommendation using Deep Neural Network Models * Deep Neural Networks for YouTube Recommendations * Wide & Deep Learning for Recommender Systems