MatrixIRLS

Nicholas Cassarino

​

Christian Kuemmerle, Tyler Allen

Recommender Systems and Matrix Factorization

• Matrix Factorization is an algorithm most commonly used for recommender systems, but also image processing, etc
• Recommender systems solve for unknown user/product pairs
• “In its basic form, matrix factorization characterizes both items and users by vectors of factors inferred from item rating patterns” (Koren, Yehuda, et al, 2009).

Koren, Yehuda, et al. “MATRIX FACTORIZATION TECHNIQUES FOR RECOMMENDER SYSTEMS.” IEEE Intelligent Systems, 2009.

Defining MatrixIRLS

• Iteratively Reweighted Least Squares (IRLS) is used as a method for matrix factorization.
• IRLS is faster and similarly accurate compared to competing algorithms.
• In this project, we will use a GPU-accelerated implementation.
• To handle large-scale data
• Test effectiveness on large-scale problems w/ limited data

Reproducing Google Research

• Successfully reproduced figures Google Research entitled: ‘iALS++: Speeding up Matrix Factorization with Subspace Optimization’
• We can accurately compare our approach to these:

​

Future Plans

• Understand the algorithmic steps of MatrixIRLS in detail.
• We have just begun working on a linear C implementation of MatrixIRLS.
• Convert into CUDA code to use GPU parallelization.

Thank You

Q & A