研发ggplot：如何使用自定义平滑器（高斯过程）
我正在使用 R。我在 15 个时间点上有 25 个变量，每个时间点每个变量有 3 个或更多重复。我已将其melt编辑成data.frame，我可以使用（除其他外）ggplot 的facet_wrap() 命令愉快地绘制它。我的融化数据框名为lis；这是它的头部和尾部，因此您可以了解数据：
> head(lis)
time variable value
1 10 SELL 8.170468
2 10 SELL 8.215892
3 10 SELL 8.214246
4 15 SELL 8.910654
5 15 SELL 7.928537
6 15 SELL 8.777784
> tail(lis)
time variable value
145 1 GAS5 10.92248
146 1 GAS5 11.37983
147 1 GAS5 10.95310
148 1 GAS5 11.60476
149 1 GAS5 11.69092
150 1 GAS5 11.70777
我可以使用以下 ggplot2 命令获得所有时间序列的漂亮图，以及拟合样条线和 95% 置信区间
p <- ggplot(lis, aes(x=time, y=value)) + facet_wrap(~variable)
p <- p + geom_point() + stat_smooth(method = "lm", formula = y ~ ns(x,3))
：不符合我的喜好——95% 的置信区间相差甚远。我想使用高斯过程（GP）来更好地回归和估计时间序列的协方差。
拟合 GP
library(tgp)
out <- bgp(X, Y, XX = seq(0, 200, length = 100))
我可以使用类似需要时间 X、观察 Y 并在 XX 中的每个点进行预测的方法来 。对象 out 包含许多有关这些预测的内容，包括我可以使用的协方差矩阵来代替我从 ns() 得到的 95% 置信区间（我认为？）代码>.
问题是我不知道如何包装这个函数以使其与 ggplot2::stat_smooth() 接口。任何有关如何进行的想法或指示将不胜感激！
I'm using R. I have 25 variables over 15 time points, with 3 or more replicates per variable per time point. I've melted this into a data.frame, which I can plot happily using (amongst other things) ggplot's facet_wrap() command. My melted data frame is called lis; here's its head and tail, so you get an idea of the data:
> head(lis)
time variable value
1 10 SELL 8.170468
2 10 SELL 8.215892
3 10 SELL 8.214246
4 15 SELL 8.910654
5 15 SELL 7.928537
6 15 SELL 8.777784
> tail(lis)
time variable value
145 1 GAS5 10.92248
146 1 GAS5 11.37983
147 1 GAS5 10.95310
148 1 GAS5 11.60476
149 1 GAS5 11.69092
150 1 GAS5 11.70777
I can get a beautiful plot of all the time series, along with a fitted spline and 95% confidence intervals using the following ggplot2 commands:
p <- ggplot(lis, aes(x=time, y=value)) + facet_wrap(~variable)
p <- p + geom_point() + stat_smooth(method = "lm", formula = y ~ ns(x,3))
The trouble is that the smoother is not to my liking - the 95% confidence intervals are way off. I would like to use Gaussian Processes (GP) to get a better regression and estimate of covariance for my time series.
I can fit a GP using something like
library(tgp)
out <- bgp(X, Y, XX = seq(0, 200, length = 100))
which takes time X, observations Y and makes predictions at each point in XX. The object out contains a bunch of things about those predictions, including a covariance matrix I can use in place of the 95% confidence interval I get (I think?) from ns().
The trouble is I'm not how to wrap this function to make it interface with ggplot2::stat_smooth(). Any ideas or pointers as to how to proceed would be greatly appreciated!
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（2）
看起来
bgp不遵循标准的 R 函数建模风格。这意味着您无法在 geom_smooth 内部使用它，并且需要在 ggplot2 调用之外调整模型。您可能还想向tgp包作者发送电子邮件，鼓励他们遵循 R 标准。It looks like
bgpdoesn't follow the standard R style for modelling functions. This means that you can't use it insidegeom_smoothand you'll need to fit the model outside of the ggplot2 call. You might also want to email thetgppackage author and encourage them to follow R standards.Stat_smooth 具有
y、ymin和ymax美学，您可以将它们与自定义平滑器一起使用，如下所述：http://had.co.nz/ggplot2/stat_smooth.html。您可以使用自定义平滑器中的预测和 CI 创建一个数据框，并直接在 stat_smooth 中使用它（指定新的数据参数）。您也许可以使用 stat_smooth(method="tgp::bgp",XX=seq(0,200,length=100)) 但我还没有尝试过。Stat_smooth has
y,ymin, andymaxaesthetics that you can use with a custom smoother, as documented here: http://had.co.nz/ggplot2/stat_smooth.html. You create a data frame with the predictions and CI from your custom smoother and use that directly instat_smooth(specifying a new data argument). You might be able to usestat_smooth(method="tgp::bgp",XX=seq(0,200,length=100))but I haven't tried it.