GLAD: Learning Sparse Graph Recovery

Authors: Harsh Shrivastava, Xinshi Chen, Binghong Chen, Guanghui Lan, Srinivas Aluru, Han Liu, Le Song

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we show that the AM architecture provides very good inductive bias, allowing the model to learn very effective sparse graph recovery algorithm with a small amount of training data.In this section, we report several experiments to compare GLAD with traditional algorithms and other data-driven algorithms.
Researcher Affiliation Collaboration {1School of Computational Science & Engineering, 2School of Mathematics, 3School of Industrial and Systems Engineering }at Georgia Institute of Technology, 4Computer Science Department at Northwestern University, 5Ant Financial Services Group
Pseudocode Yes Algorithm 1: GLAD and Algorithm 2: ADMMu
Open Source Code Yes Python implementation (tested on P100 GPU) is available1. 1code: https://drive.google.com/open?id=16POE4TMp7UUie_Lc_Lq_Rz_STqzk_VHm2stl_M
Open Datasets Yes For sections 5.1 and 5.2, the synthetic data was generated based on the procedure described in Rolfs et al. (2012).The Syn TRe N (Van den Bulcke et al., 2006) is a synthetic gene expression data generator specifically designed for analyzing the structure learning algorithms.We use the real data from the DREAM 5 Network Inference challenge (Marbach et al., 2012).
Dataset Splits Yes For finetuning the traditional algorithms, a validation dataset of 10 graphs was used. For the GLAD algorithm, 10 training graphs were randomly chosen and the same validation set was used.
Hardware Specification Yes Python implementation (tested on P100 GPU) is available1.Convergence results and average runtime of different algorithms on Nvidia s P100 GPUs are shown in Figure 4 and Table 2 respectively.
Software Dependencies No The paper mentions 'Python implementation' and 'sklearn package Pedregosa et al. (2011)' but does not provide specific version numbers for Python, scikit-learn, or any other software libraries.
Experiment Setup Yes GLAD parameter settings: ρnn was a 4 layer neural network and Λnn was a 2 layer neural network. Both used 3 hidden units in each layer. The non-linearity used for hidden layers was tanh, while the final layer had sigmoid (σ) as the non-linearity for both, ρnn and Λnn (refer Figure 3). The learnable offset parameter of initial Θ0 was set to t = 1. It was unrolled for L = 30 iterations. The learning rates were chosen to be around [0.01, 0.1] and multi-step LR scheduler was used. The optimizer used was adam.