This is an implementation of the following paper:

DAGs with NO TEARS: Continuous Optimization for Structure Learning (NeurIPS 2018, Spotlight)

Xun Zheng, Bryon Aragam, Pradeep Ravikumar, Eric Xing.

If you find it useful, please consider citing:

@inproceedings{zheng2018dags,
    author = {Zheng, Xun and Aragam, Bryon and Ravikumar, Pradeep and Xing, Eric P.},
    booktitle = {Advances in Neural Information Processing Systems},
    title = {{DAGs with NO TEARS: Continuous Optimization for Structure Learning}},
    year = {2018}
}

Check out notears.py for a complete, end-to-end implementation of the NOTEARS algorithm in fewer than 60 lines.

This includes L2, Logistic, and Poisson loss functions with L1 penalty.

A directed acyclic graphical model (aka Bayesian network) with d nodes defines a distribution of random vector of size d. We are interested in the Bayesian Network Structure Learning (BNSL) problem: given n samples from such distribution, how to estimate the graph G?

A major challenge of BNSL is enforcing the directed acyclic graph (DAG) constraint, which is combinatorial. While existing approaches rely on local heuristics, we introduce a fundamentally different strategy: we formulate it as a purely continuous optimization problem over real matrices that avoids this combinatorial constraint entirely. In other words,

where h is a smooth function whose level set exactly characterizes the space of DAGs.

• Python 3.5+
• numpy
• scipy
• python-igraph: Install igraph C core and pkg-config first.

• notears.py - the 60-line implementation of NOTEARS with l1 regularization for various losses
• utils.py - graph simulation, data simulation, and accuracy evaluation

The simplest way to try out NOTEARS is to run a simple example:

$ git clone https://github.com/xunzheng/notears.git
$ cd notears/
$ python notears.py

This runs the l1-regularized NOTEARS on a randomly generated 20-node Erdos-Renyi graph with 100 samples. Within a few seconds, you should see output like this:

{'fdr': 0.0, 'tpr': 1.0, 'fpr': 0.0, 'shd': 0, 'nnz': 20}

The data, ground truth graph, and the estimate will be stored in X.csv, W_true.csv, and W_est.csv.

• Ground truth: d = 20 nodes, 2d = 40 expected edges.

  

• Estimate with n = 1000 samples: lambda = 0, lambda = 0.1, and FGS (baseline).

  

  Both lambda = 0 and lambda = 0.1 are close to the ground truth graph when n is large.

• Estimate with n = 20 samples: lambda = 0, lambda = 0.1, and FGS (baseline).

  

  When n is small, lambda = 0 perform worse while lambda = 0.1 remains accurate, showing the advantage of L1-regularization.

• Ground truth: d = 20 nodes, 4d = 80 expected edges.

  

  The degree distribution is significantly different from the Erdos-Renyi graph. One nice property of our method is that it is agnostic about the graph structure.

• Estimate with n = 1000 samples: lambda = 0, lambda = 0.1, and FGS (baseline).

  

  The observation is similar to Erdos-Renyi graph: both lambda = 0 and lambda = 0.1 accurately estimates the ground truth when n is large.

• Estimate with n = 20 samples: lambda = 0, lambda = 0.1, and FGS (baseline).

  

  Similarly, lambda = 0 suffers from small n while lambda = 0.1 remains accurate, showing the advantage of L1-regularization.

• Python: https://github.com/jmoss20/notears
• Tensorflow with Python: https://github.com/ignavier/notears-tensorflow