Causal Effect Identification in Uncertain Causal Networks
Authors: Sina Akbari, Fateme Jamshidi, Ehsan Mokhtarian, Matthew Vowels, Jalal Etesami, Negar Kiyavash
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We propose efficient algorithms to approximate this problem and evaluate them against both real-world networks and randomly generated graphs. |
| Researcher Affiliation | Academia | Sina Akbari EPFL, Switzerland sina.akbari@epfl.ch; Fateme Jamshidi EPFL, Switzerland fateme.jamshidi@epfl.ch; Ehsan Mokhtarian EPFL, Switzerland ehsan.mokhtarian@epfl.ch; Matthew J. Vowels University of Lausanne, Switzerland matthew.vowels@unil.ch; Jalal Etesami TUM, Germany j.etesami@tum.de; Negar Kiyavash EPFL, Switzerland negar.kiyavash@epfl.ch |
| Pseudocode | Yes | We first present a recursive approach for solving the edge ID problem in Section 4.1, described in Algorithm 1... Algorithm 2 Heuristic algorithm for edge ID... Algorithm 3 Maximal Hedge... Algorithm 4 Heuristic algorithm 2. |
| Open Source Code | Yes | To replicate our findings, our Python implementations are accessible via https://github.com/SinaAkbarii/Causal-Effect-Identification-in-Uncertain-Causal-Networks and https://github.com/matthewvowels1/NeurIPS_testbed_ADMGs. |
| Open Datasets | Yes | The first Psych (22 nodes & 70 directed edges) concerns the putative structure from a causal discovery algorithm Structural Agnostic Model [18] using data collected as part of the Health and Relationships Project [32]. The other three Barley (48 nodes & 84 directed edges), Water (32 nodes & 66 directed edges), and Alarm (37 nodes & 46 directed edges) come from the bnlearn python package [28]. |
| Dataset Splits | No | The paper does not explicitly provide training/test/validation dataset splits. For simulations, it describes generating ADMG structures and sampling edge probabilities, then running algorithms on these. For real-world graphs, it states applying algorithms to existing graph structures without mentioning data splits for evaluation. |
| Hardware Specification | Yes | All experiments were carried out on an Intel i9-9900K CPU running at 3.6GHz. |
| Software Dependencies | No | The paper mentions 'our Python implementations' and 'bnlearn python package [28]' but does not provide specific version numbers for Python or the bnlearn package, nor for any other key software components. |
| Experiment Setup | Yes | For a given number of vertices, we uniformly sample 50 ADMG structures... Edges for each of these 100 graphs are sampled with probability of log(n)/n... For each graph, we sample directed and bidirected edge probabilities pe uniformly between 0.51 and 1.0... The outcome Y is selected to be the last vertex in the topological ordering... If the runtime exceeds 3 minutes, we abort and log that the algorithm has failed to find a solution... Hyperparameter settings for the Structural Agnostic Model used to generate the putative (directed) structure for the Psych real-world dataset. Parameter Value Learning Rate 0.01 DAG Penalty True DAG Penalty Weight 0.05 Number of Runs 50 Train Epochs 3000 Test Epochs 800 Mixed Data True hlayers 2 dhlayers 2 lambda1 10 lambda2 0.001 dlr 0.001 linear False nh 20 dnh 200 |