pyGWRx 模型手册

本手册详细说明 19 个正式公开模型，包括模型要解决的问题、数学形式、算法流程、pyGWRx 当前实现、适用场景、限制、推荐图件和完整可运行示例。

模型能力表

模型	类型	输入	新位置能力
GWR	Classic local regression	X, y, coordinates	Validated local re-calibration at new coordinates.
MGWR	Multiscale local regression	X, y, coordinates	Independent-target prediction is intentionally unavailable in the current validated API.
RGWR	Robust local regression	X, y, coordinates	Validated local prediction using the fitted robust calibration state.
STWR	Stage-based spatiotemporal regression	Lists of X, y, and coordinates by stage, plus time intervals	Prediction for the current/latest stage using the fitted historical-stage weighting structure.
GTWR	Row-wise spatiotemporal regression	X, y, coordinates, and row-wise times	Validated at new space-time targets; causal filtering is available when configured.
GWGLM	Generalized local regression	X, response, coordinates; optional exposure for Poisson	Validated for Gaussian means, binomial probabilities, and Poisson means.
GWLasso	Locally regularized regression	X, y, coordinates	Validated local prediction with the learned local penalties and scaling state.
MixedGWR	Semiparametric global-local regression	X, y, coordinates, and global/local variable assignments	Validated using global coefficients and re-estimated local components.
GWPCA	Local multivariate transformation	Multivariate X and coordinates	Not a response predictor; transform() returns local component scores.
GWDA	Local spatial classification	X, class labels, coordinates	Validated class labels and local class probabilities.
GWSS	Local descriptive statistics	Multivariate X and coordinates	Not applicable; this is a local-statistics estimator.
ScalableGWR	Approximate scalable local regression	X, y, coordinates	Validated using the fitted scalable kernel approximation.
LCRGWR	Collinearity-compensated local regression	X, y, coordinates	Validated local prediction with fitted or locally adjusted ridge terms.
BootstrapGWR	Spatial inference	X, y, coordinates	Not applicable; the estimator performs coefficient-variability inference.
SGWR	Geography-plus-similarity regression	X, y, coordinates, and similarity-variable specification	Validated by recomputing geographic and attribute-similarity weights for targets.
SGTWR	Geography-time-similarity regression	X, y, coordinates, times, and similarity variables	Validated at target space-time points with optional causal filtering.
MGTWR	Multiscale spatiotemporal regression	X, y, coordinates, times; optional per-column bandwidths and taus	当前已验证 API 不提供独立目标位置预测；模型拟合与推断由 pyGWRx 内部实现。
LGGWR	Original research model	X, y, coordinates, and contextual attributes	Validated using the learned geometry transform and target attributes.
GRGWR	Original research model	X, y, coordinates, regime count, and connectivity settings	Validated using learned regime structure and target assignment logic.

使用原则

1. 先建立全局模型和标准 GWR 基线，再使用更复杂模型。
2. 固定带宽是距离，自适应带宽是近邻数，不能直接比较数值大小。
3. 局部系数不是自动的因果效应，必须结合不确定性、共线性和残差诊断。
4. 时空模型必须使用防止未来信息泄漏的验证方式。
5. LGGWR 和 GRGWR 是原创研究模型，应报告初始化、敏感性和当前验证边界。

返回中文首页 · 英文模型手册