An implementation of the learning algorithm from https://github.com/iivek/dag-constrained-nmf in an older version of Graphlab, by team from Carnegie Mellon University. Hyperparameter optimization has not been implemented. For more information, please visit the aforementioned link.
Decomposition results on a small synthetic dataset are shown below.
The generated (oracle) matrix is low-rank, with known correlations in the latent factors. The algorithm has been presented with only 10% of the oracle data with the goal of reconstructing the rest. In case correlations are embedded into the algorithm, reconstruction is closer to the oracle as compared to the correlation-ignorant decomposition.
Currently, both the input and output data are in form of .mat files and MATLAB API is used for I/O. An example of how to format the data can be bound under /resources. How the data and parameters are fed to the algorithm and how the algorithm is run can be seen in main.cpp.
No CMake is available for the time being. The package depends on CMU Graphlab, available here https://github.com/iivek/graphlab-cmu-mirror. MATLAB API is required to read from and write to .mat files. Also, Atlas, Blas, Boost.
CMake Command line interface. Provide other means of input/output (currenlty it goes through .mat files exclusively)
If you have used this code, please cite the following repo: I.Ivek, "DAG-constrained Variational Bayesian NMF." [Online]. Available: https://github.com/iivek/dag-constrained-nmf
Published under GPL-3.0 License.