Dick-AI-Lab/ExaTune_landscapes

ExaTune景观数据集是一个详尽的超参数搜索结果集合，通过对七个数据集-模型家族对在密集笛卡尔积网格中的每个配置进行运行而生成。每一行对应一个评估的配置，记录了交叉验证的性能指标、每折时间以及（如适用）补充指标。这七个景观共覆盖了五个scikit-learn/XGBoost模型家族和六个公开可用的表格数据集中的301,516个完全评估的配置，使其成为表格机器学习中最大的公开完整超参数适应度景观集合。主要用途是超参数景观拓扑的实证分析：空间自相关（Morans I）、崎岖性、局部最优检测、适应度-距离相关性、参数重要性和成对交互强度。详尽枚举消除了从随机或贝叶斯搜索轨迹中推导出的景观分析中的抽样偏差。ExaTune Landscapes is a collection of exhaustive hyperparameter search results produced by running every configuration in a dense Cartesian-product grid for seven dataset–model-family pairs. Each row corresponds to one evaluated configuration and records the cross-validated performance metric(s), per-fold timing, and (where applicable) supplementary metrics. Together the seven landscapes span 301,516 fully evaluated configurations across five scikit-learn/XGBoost model families and six publicly available tabular datasets, making this the largest openly available collection of complete hyperparameter fitness landscapes for tabular ML. The primary intended use is empirical analysis of hyperparameter landscape topology: spatial autocorrelation (Morans I), ruggedness, local optima detection, fitness-distance correlation, parameter importance, and pairwise interaction strength. The exhaustive enumeration eliminates sampling bias that confounds landscape analyses derived from random or Bayesian search trajectories.