Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Local Causal Discovery Without Causal Sufficiency

Authors: Zhaolong Ling, Jiale Yu, Yiwen Zhang, Debo Cheng, Peng Zhou, Xingyu Wu, Bingbing Jiang, Kui Yu

AAAI 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on synthetic datasets validate that the proposed Latent LCD algorithm significantly outperforms the state-of-the-art methods.
Researcher Affiliation Academia 1School of Computer Science and Technology, Anhui University 2Uni SA STEM, University of South Australia 3Department of Computing, Hong Kong Polytechnic University 4School of Information Science and Engineering, Hangzhou Normal University 5School of Computer Science and Information Technology, Hefei University of Technology
Pseudocode Yes Algorithm 1: Latent LCD Algorithm 2: Find BI (Find Bidirectional Edges)
Open Source Code No The paper does not contain any explicit statements about making the source code available, nor does it provide a link to a code repository.
Open Datasets No In all generated datasets, we follow a process similar to RFCI (Colombo et al. 2012) to generate random DAG with latent common causes. [...] We select a discrete Bayesian network, BARLEY, which consists of 48 nodes and 84 arcs. For this network, we treat SAATID as a latent variable, with all other variables considered as observed variables.
Dataset Splits No For each DAG, we further generate datasets with sample sizes of {500, 1000, 1500, 2000, 2500, 3000}. [...] Using the BARLEY network, we randomly generate six sets of data with sample sizes of {500, 1000, 1500, 2000, 2500, 3000}, each containing ten datasets.
Hardware Specification Yes All experiments are conducted on a computer equipped with an Intel Core i7-12700 2.10GHz CPU and 16GB RAM.
Software Dependencies No Since RFCI is implemented in R and Latent LCD is implemented in Matlab, we use the number of CI tests to measure the efficiency of the algorithms.
Experiment Setup Yes The conditional independence test used is Fisher s Z-test, with significant level α set at 0.01. [...] For each DAG, we randomly generate 100 DAGs for each node count p {15, 20, 25, 30, 35}, each with E(N) = 2. [...] We employ the G2 test, setting the significant level α at 0.01.