Chenghao Wei

Bayesian network-based causal structure learning provides an effective framework for uncovering causal relationships among continuous variables. However, many existing methods for continuous data still rely on strong parametric distribution assumptions, which may introduce information loss and reduce Bayesian network modeling accuracy. Kernel density estimation (KDE), a non-parametric statistical method that is more flexible in density estimation form, offers a versatile framework for conducting conditional independence (CI) tests. This approach enables the estimation of mutual information and conditional mutual information, thereby facilitating the identification of underlying structural relationships. Nevertheless, the high computational cost of KDE-based CI testing restricts its practical application in continuous-variable causal learning. To address this issue, this study introduces a radial symmetric kernel-based acceleration scheme within a Fast Fourier Transform (FFT) framework to improve the efficiency of density estimation. On this basis, an enhanced Bayesian network structure learning method is developed for continuous variables, enabling more efficient estimation of mutual information and conditional mutual information while improving the computational efficiency and empirical stability of variable dependency discovery. With proper bandwidth and grid resolution, the proposed MMHC-FFTKDE framework achieves a reduction in computational runtime and improves efficiency compared to MMHC-KDE in the ablation setting, while maintaining competitive F1-scores and SHD for causal structure discovery.

Tree-Augmented Naive Bayes (TAN) is an interpretable graphical structure model. However, its structure learning for continuous attributes depends on the class-conditional mutual information, which is sensitive to one-dimensional or two-dimensional density estimation. Accurate estimation is challenging under complex distributions such as multi-peak, long-tailed and heteroscedastic cases. To address this issue, we propose a structure learning method for TAN based on Fast Generative Bootstrap Maximum Likelihood Estimation (TAN-FGBMLE). FGBMLE consists of two stages of work. In the first stage, resampling weights and random noise are input into a network generator to rapidly produce candidate parameters, efficiently covering the latent density space without repeated independent optimization. In the second stage, optimal mixture weights are estimated by maximum likelihood estimation, assigning appropriate contributions to each candidate component. This design enables fast and accurate complex density estimation for both single and joint attributes, providing reliable computation of class-conditional mutual information. The TAN structure is then constructed using Prim’s maximum spanning tree algorithm. Experiments show that our estimation method attains higher fitting accuracy and lower runtime compared with traditional nonparametric estimators. By using open-source datasets, the TAN-FGBMLE achieves superior accuracy and recall compared to classic methods, demonstrating good robustness and interpretability. On publicly available real air quality data, it has a high classification result and produces graph structures that more accurately capture dependencies among continuous attributes.

The structure learning of a Bayesian network (BN) is a crucial process that aims to unravel the complex dependencies relationships among variables using a given dataset. This paper proposes a new BN structure learning method for data with continuous attribute values. As a non-parametric distribution-free method, kernel density estimation (KDE) is applied in the conditional independence (CI) test. The skeleton of the BN is constructed utilizing the test based on mutual information and conditional mutual information, delineating potential relational connections between parents and children without imposing any distributional assumptions. In the searching stage of BN structure learning, the causal relationships between variables are achieved by using the conditional entropy scoring function and hill-climbing strategy. To further enhance the computational efficiency of our method, we incorporate a locality sensitive hashing (LSH) function into the KDE process. The method speeds up the calculations of KDE while maintaining the precision of the estimates, leading to a notable decrease in the time required for computing mutual information, conditional mutual information, and conditional entropy. A BN classifier (BNC) is established by using the computationally efficient BN learning method. Our experiments demonstrated that KDE using LSH has greatly improved the speed compared to traditional KDE without losing fitting accuracy. This achievement underscores the effectiveness of our method in balancing speed and accuracy. By giving the benchmark networks, the network structure learning accuracy with the proposed method is superior to other traditional structure learning methods. The BNC also demonstrates better accuracy with stronger interpretability compared to conventional classifiers on public datasets.