CMF.jl

Convolutive Matrix Factorizations (CMF) in Julia.

NOTE: this package is currently under development. You can expect a stable package in June 2020, and major rehauls soon thereafter. Feel free to raise an issue if you have questions in the meantime.

Simple Example

Fitting CMF models is as simple as calling fit_cnmf:

>> using CMF
>> data = CMF.gen_synthetic(N=500, T=2000)
>> 
>> results = fit_cnmf(data; L=10, K=5, alg=:hals)
>> println(results.loss_hist[end])

Produces:

0.012

Basic Usage

The command results = fit_cnmf(data) fits a convolutive NMF model to the dataset. Several optional parameters are available, including:

• alg: the algorithm used to fit the model. The package currently supports:
  • Multiplicative Updates (:mult)
  • Hierarchical Alternating Least Squares (:hals)
  • Alternating Nonnegative Least Squares (:anls)
• K: the number of components in the model (default K=5).
• L: the width, or lag, of each component (default L=10).
• max_time: the maximum runtime in seconds (default max_time=Inf).
• max_itr: the maximum number of iterations (default max_itr=100).

Regularization

Multiplicative Updates (MU) and Hierarchical Alternating Least Squares (HALS) support L1 and L2 regularization. Using regularization with the Alternating Nonnegative Least Squares solver will raise an error.

Regularization is added with additional keyword arguments to the fit_cnmf function. The following keyword arguments may be used:

• l1_H
• l2_H
• l1_W
• l2_W

These add L1 and L2 regularization to W or H as you would expect. Below is an example:

results = fit_cnmf(data; L=10, K=5, alg=:hals, l1_H=0.1, l2_H=0.2, l1_W=0.1, l2_W=0.5)

Installation

First, install the latest version of Julia. Then use the ] key to enter package mode and type:

(v1.0) pkg> add https://github.com/degleris1/CMF.jl