CoLiDE: Concomitant Linear DAG Estimation
Authors: Seyed Saman Saboksayr, Gonzalo Mateos, Mariano Tepper
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate Co Li DE s effectiveness through comprehensive experiments with both simulated (linear SEM) and real-world datasets. When benchmarked against state-of-the-art DAG learning algorithms, Co Li DE consistently attains better recovery performance across graph ensembles and exogenous noise distributions, especially when the DAGs are larger and the noise level profile is heterogeneous. We also find Co Li DE exhibits enhanced stability manifested via reduced standard deviations in various domain-specific metrics, underscoring the robustness of our novel estimator. |
| Researcher Affiliation | Collaboration | Seyed Saman Saboksayr, Gonzalo Mateos University of Rochester {ssaboksa,gmateosb}@ur.rochester.edu Mariano Tepper Intel Labs mariano.tepper@intel.com |
| Pseudocode | Yes | Algorithm 1: Co Li DE optimization |
| Open Source Code | Yes | We have made our code publicly available at https://github.com/SAMiatto/colide. |
| Open Datasets | No | The paper mentions using |
| Dataset Splits | No | The paper mentions data generation and number of samples (e.g., n = 1000 samples) but does not specify how the data is split into training, validation, and test sets, nor does it refer to standard splits with citations. |
| Hardware Specification | Yes | All experiments were executed on a 2-core Intel Xeon processor E5-2695v2 with a clock speed of 2.40 GHz and 32GB of RAM. For models like GOLEM and DAGuerreotype that necessitate GPU processing, we utilized either the NVIDIA A100, Tesla V100, or Tesla T4 GPUs. |
| Software Dependencies | No | The paper mentions using "Python" and "Tensorflow" but does not provide specific version numbers for these or any other libraries/solvers. |
| Experiment Setup | Yes | Hence, our algorithm uses λ = 0.05 and employs K = 4 decreasing values of µk [1.0, 0.1, 0.01, 0.001], and the maximum number of BCD iterations is specified as Tk = [2 104, 2 104, 2 104, 7 104]. Furthermore, early stopping is incorporated, activated when the relative error between consecutive values of the objective function falls below 10 6. The learning rate for ADAM is 3 10 4. We adopt sk [1, 0.9, 0.8, 0.7]. |