Unveiling Nonlinear Patterns between Gamma-Ray Burst Observational Features and Redshift via Machine Learning
利用机器学习揭示伽马射线暴观测特征与红移之间的非线性模式
The study of multifeature cosmological correlations in gamma-ray bursts (GRBs) holds significant importance for establishing them as effective cosmological probes. As GRB observational datasets continually grow, the limitations of traditional linear models in handling multidimensional, nonlinear, and heavily selection-biased data have become increasingly apparent. And the calculation of intrinsic features heavily depends on specific physical models, which introduces prior biases. Thus, this study avoids strong reliance on physical models by starting directly from raw observational data and employing machine learning to autonomously uncover the nonlinear patterns linking redshift to observational features. Given the complexity of GRBs, traditional network models suffer from issues such as underfitting, weak feature interaction capabilities, and unstable training. We thus propose a hybrid model integrating XGBoost with a deep cross network (DCN), leveraging XGBoost's strength in efficient nonlinear transformation and feature interaction capture alongside DCN's capacity for automatic high-order interaction modeling through a fusion correction strategy and optimization mechanism to enhance training stability. Using public Swift satellite GRB data, we construct an input space of diverse observational features and systematically evaluate model performance via K-fold cross validation. Experimental results show that the proposed hybrid model can capture nonlinear relationships between multiobservational GRB features and redshift, outperforming linear regression and single-network baselines under the current sample. Regression diagnostics, normality tests, and correlation analyses further indicate that the residuals are approximately normal and that the predicted redshifts exhibit a statistically significant positive correlation with true values. Overall, this study provides a data-driven pathway to circumvent the limitations of traditional methods in addressing model dependence.
展开 ▾创新在于混合 XGBoost 与深度交叉网络,通过融合校正策略和优化机制提升训练稳定性,并在当前样本上优于线性回归和单一网络基线。
传统 GRB 宇宙学相关研究多采用线性模型,且内禀特征计算高度依赖具体物理模型,容易引入先验偏差。随着 Swift 等公开 GRB 观测样本不断增长,数据驱动方法为处理高维、非线性和选择偏差严重的数据提供了新路径。本文从原始观测特征出发,构建 XGBoost+DCN 混合模型,经 K 折交叉验证表明预测红移与真值显著正相关,残差近似正态。未来可结合更大样本和多信使观测,进一步提升模型可解释性与稳健性,推动 GRB 成为更可靠的宇宙学探针。
接收 2026-06-22 · 刊出 2026-07-27 · 收录 2026-08-20