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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Pattern-Guided Adaptive Prior for Structure Learning
Authors: Lyuzhou Chen, Yijia Sun, Yanze Gao, Xiangyu Wang, Derui Lyu, Taiyu Ban, Xin Wang, Xiren Zhou, Huanhuan Chen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This section presents a subset of the primary experimental results and analysis. We evaluate the performance of our proposed method across various synthetic data settings, including different graph structures, numbers of nodes, dataset sizes, and noise types. We also assess its effectiveness when combined with different prior integration techniques and compare it against several baseline structure learning algorithms. |
| Researcher Affiliation | Academia | Lyuzhou Chen, Yijia Sun, Yanze Gao, Xiangyu Wang , Derui Lyu, Taiyu Ban, Xin Wang, Xiren Zhou, Huanhuan Chen School of Computer Science and Technology, University of Science and Technology of China 96 Jinzhai Rd, Hefei, China, 230026 {clz31415@mail., sunyijia2002@mail., gaoyz@mail., sa312@,} {drlv@mail., banty@mail., wz520@mail., zhou0612@, hchen@}ustc.edu.cn |
| Pseudocode | Yes | Algorithm 1 PGAP with Soft Prior Constraints and Algorithm 2 PGAP with Hard Prior Constraints |
| Open Source Code | Yes | Justification: We provide access to the code and datasets used. |
| Open Datasets | Yes | Justification: We provide access to the code and datasets used. Synthetic Data Random DAGs are generated using the Erdös-Rényi (ER) model or the Scale-Free (SF) model... The first is the well-established Sachs dataset, which contains 11 nodes and 17 edges from measurements of protein and phospholipid expression levels [Sachs et al., 2005]. We also include three additional datasets: the small-scale Paid Search dataset (7 nodes, 6 edges) Rutz and Bucklin [2007], and two larger-scale datasets, Magic-niab (44 nodes, 66 edges) Scutari et al. [2014] and Ecoli70 (46 nodes, 70 edges) Schäfer and Strimmer [2005]. |
| Dataset Splits | No | Sample sizes are set to 2 and 20 times the number of nodes (2d, 20d), and data is normalized post-generation. While synthetic data is generated and real-world datasets are used, explicit training/test/validation splits are not detailed in the paper. |
| Hardware Specification | Yes | Experiments are conducted on a system equipped with a 4.5 GHz AMD Ryzen 9 7950X CPU, NVIDIA Ge Force RTX 3090 GPU, and 32GB of memory. |
| Software Dependencies | No | No specific version numbers for software dependencies are listed. The paper mentions utilizing base structure learning algorithms like NOTEARS, GOLEM, DAGMA, but does not provide versioned software dependencies for its own implementation. |
| Experiment Setup | Yes | The initial weight of the prior loss is 1. The adaptive reduction parameter α, i.e., the amount of weight reduction for each graph pattern, is 0.1. Other settings follow the default configurations of the baseline methods. |