GraphLab ML library's underlying Numerical Computing Library

User 649 | 8/30/2014, 10:50:05 AM

I am new to Graphlab. I was wondering if the math libraries on which the ML libraries are written are built by the graphlab team or do you use some other well-known math libraries that can be distributed easily and are also thread-safe? For example, computing matrix inverses, eigenvectors of very large matrices?

Comments

User 6 | 8/31/2014, 3:58:54 PM

Unlike scikit-learn which builds upon numpy and scipy python libraries, GraphLab Create has its own computation engine and this is the reason it can scale to much larger problems. We plan to open source this engine in a few months time. At this point we do not support eigen decomposition or have plans to export such primitives.

We do have an efficient SVD implementation as part of our older open source GraphChi/ PowerGraph. We would love to get feedback which algorithms do you had in mind and for what problem domain. We are quickly adding new functionality based on user requests.