Reliable and Efficient Anytime Skeleton Learning

Authors: Rui Ding, Yanzhi Liu, Jingjing Tian, Zhouyu Fu, Shi Han, Dongmei Zhang10101-10109

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

Reproducibility Variable Result LLM Response
Research Type Experimental Thorough experiments were conducted on both benchmark and realworld datasets demonstrate that REAL significantly outperforms the other state-of-the-art algorithms.
Researcher Affiliation Collaboration 1Microsoft Research 2Beijing University of Posts and Telecommunications 3Peking University 4Alibaba Group
Pseudocode Yes Algorithm 1 Best First Growing Input: priority queue ℚ, variable set 𝒱; Algorithm 2 REAL Input: data 𝐷, variable set 𝒱, cardinalities 𝐶 Output: skeleton 𝐺:; Procedure Prune Input: 𝑇, 𝑅𝑎𝑤𝑃𝐶[𝑇]; Procedure ANDRule Input: 𝑋, 𝑌; Sub-Procedure Robust Corr Input: 𝑇, and its current 𝑅𝑎𝑤𝑃𝐶, 𝑉; Procedure Robust PCFinder Input: 𝑇, and its 𝑅𝑎𝑤𝑃𝐶 Output: A new variable most likely to be a 𝑃𝐶 of 𝑇
Open Source Code No The paper states "ALL proofs for lemmas and theorems are available at our website1: 1 https://www.microsoft.com/en-us/research/project/real/", but this does not explicitly state that the source code for the described methodology is available. The paper mentions their implementation in C# but does not provide access.
Open Datasets Yes We use available benchmark datasets from Bayesian Network Repository (Scutari 2012) for evaluation.
Dataset Splits No The paper states "We sample 20,000 records from each network." but does not specify explicit training, validation, or test dataset splits or a cross-validation setup.
Hardware Specification Yes All experiments are conducted on a machine with 3.2GHz Intel i7-8700 processor and 16 GB RAM.
Software Dependencies No The paper mentions using "C++ implementation of MINOBS", "Java implementation of BLIP", and their own "C# implementation" but does not provide specific version numbers for these languages or any other required libraries.
Experiment Setup Yes Configurations of MINOBS and BLIP are optimized on these datasets: because all networks are with max-indegree 6 except win95pts is with max-in-degree=7. We set max-in-degree threshold for BLIP and MINOBS to 6 (change it to 7 would significantly degrade their efficiency).