User 1586 | 5/5/2015, 2:11:28 PM
I read the Six Degrees of K. Bacon notebook (http://dato.com/learn/gallery/notebooks/graphanalyticsmovies.html) about analyzing a film network. In the Computing the actor network, they calculate the co-occurrence matrix by means of A*A'. Specifically, by extracting a subset of the vertices, creating a new Panda's dataframe and finally calculating the dot product (with numpy) between the dataframe values (a matrix).
My question is: Would it be possible to access directly the matrix multiplication procedures you are presumably using in other functions, as Random Walk, instead of converting the data into a numpy matrix and using its operations ? I imagine that buried in the C++ code there must be a matrix multiplication implementation tailored for the out-of-core setting we have in Graphlab Create. Or maybe these routines in C++ call BLAS/LAPACK as the numpy library does anyway ?
I find Graphlab Create really nice to work with, but I fear that eventually I will have to transform my SFrame structures into sparse matrices and deal with certain operations (as the co-occurrence matrix calculation) as usual, losing the out-of-core parallel cool stuff.