kapelner
commented
11 months ago Cool project! Is there a reason you can't use the x, y coordinates (or r, rho coordinates) as regressors?
On Sat, Sep 30, 2023, 10:12 Nikolaos Tziokas @.***> wrote:
First of all, thank you for creating this package, the documentation is very clear and easy to follow. I apologize in advance for the long post, but I'd like to make my problem as clear as possible.
I have a response variable which is the nighttime light luminosity (called ntl in the data.frame I am using). As you can see in the image below, it has some bright spots (highlighted in red) which are areas with high brightness (outliers in the data.frame).
[image: ntl] https://user-images.githubusercontent.com/45749751/271774959-7cfc88c9-d753-4fd3-862c-b93f7f184ac5.png
Also, here is the histogram of the response. It's clear that the distribution is right-skewed (and possibly there is bimodality?). [image: histogram] https://user-images.githubusercontent.com/45749751/271775040-dbcc26e7-23c6-496b-8e38-6738a62fc56c.png
Purpose of my analysis My goal is to predict the ntl from the coarse spatial resolution to a higher. This means that, I need to maintain these bright spots (the outliers) in the predicted image. Because at a later stage, I will downscale the residuals of the regression using area-to-point Kriging, I need them to be random (no spatial structure).
Analysis Following the approach found in your paper https://www.jstatsoft.org/article/view/v070i04, I created this code:
options(java.parameters = "-Xmx5g") library("bartMachine") set_bart_machine_num_cores(3)
set working directory
wd <- "path/"
Projected reference system (in order to convert the residuals into a raster image)
provoliko <- "EPSG:24313"
original df
df <- read.csv(paste0(wd, 'block.data.csv'))
extract the x and y columns (coordinates) from the df
crds <- df[, 1:2]
here I keep only the necessary columns for my analysis
keep <- c("ntl", "pop", "agbh", "nir", "ebbi", "ndbi", "road", "pan", "nbai", "tirs") df <- df[keep]
x <- df[, 2:10] y <- df[, 1]
bart_machine <- bartMachine(x, y) bart_machine
The output of the default bartMachine model is:
bart_machine bartMachine v1.3.4.1 for regression
training data size: n = 5658 and p = 9 built in 11.7 secs on 1 core, 50 trees, 250 burn-in and 1000 post. samples
sigsq est for y beforehand: 76.35 avg sigsq estimate after burn-in: 40.8314
in-sample statistics: L1 = 21551.53 L2 = 208688.47 rmse = 6.07 Pseudo-Rsq = 0.83 p-val for shapiro-wilk test of normality of residuals: 0 p-val for zero-mean noise: 0.98127
Using the function plot_convergence_diagnostics(bart_machine), the result is: [image: residuals_diagnostics] https://user-images.githubusercontent.com/45749751/271776302-cf80f431-e820-48e2-9827-8912641d970f.png
Again, for a "good" model I would like to see a symmetric scatter of points around the horizontal like at zero, indicating random deviations of predictions from the observed values, but from the plot above I see that this isn't the case.
Moreover, a map of the residuals is shown below. In red, I highlighted the areas were I believe the model couldn't model well (you can compare these areas to the areas in the first image above). [image: bart_rsds] https://user-images.githubusercontent.com/45749751/271777072-2b645d5a-6b19-4a9a-90a2-1d8f80c6e610.png As you can see, there is clearly a spatial structure (i.e., the residuals do not show a random pattern).
My question is, is there a way tell BART to consider the outliers (i.e., bright spots in the study area) as "more important" when modelling the NTL? What are your recommendations?
Because the csv I'm using has several thousands of rows, I can share it via a link, from here https://drive.google.com/drive/folders/1V115zpdU2-5fXssI6iWv_F6aNu4E5qA7?usp=drive_link. Just so you know, running a model with default parameters takes less than 30 seconds on my laptop (8 gigs of RAM, 4 cores processor (CPU: Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz)).
Lastly, I tried to use a model computed by the bartMachineCV function (as well as variable selection) but the results are not better.
— Reply to this email directly, view it on GitHub https://github.com/kapelner/bartMachine/issues/52, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAFAV6G5RQTQQZL2S6HKUPDX5ASEVANCNFSM6AAAAAA5NSV7GA . You are receiving this because you are subscribed to this thread.Message ID: @.***>
nikosGeography
First of all, thank you for creating this package, the documentation is very clear and easy to follow. I apologize in advance for the long post, but I'd like to make my problem as clear as possible.
I have a response variable which is the nighttime light luminosity (called ntl in the
data.frameI am using). As you can see in the image below, it has some bright spots (highlighted in red) which are areas with high brightness (outliers in thedata.frame).Also, here is the histogram of the response. It's clear that the distribution is right-skewed (and possibly there is bimodality?).
Purpose of my analysis My goal is to predict the ntl from the coarse spatial resolution to a higher. This means that, I need to maintain these bright spots (the outliers) in the predicted image. Because at a later stage, I will downscale the residuals of the regression using area-to-point Kriging, I need them to be random (no spatial structure).
Analysis Following the approach found in your paper, I created this code:
The output of the default
bartMachinemodel is:Using the function
plot_convergence_diagnostics(bart_machine), the result is:Again, for a "good" model I would like to see a symmetric scatter of points around the horizontal like at zero, indicating random deviations of predictions from the observed values, but from the plot above I see that this isn't the case.
Moreover, a map of the residuals is shown below. In red, I highlighted the areas were I believe the model couldn't model well (you can compare these areas to the areas in the first image above).
As you can see, there is clearly a spatial structure (i.e., the residuals do not show a random pattern).
My question is, is there a way tell
BARTto consider the outliers (i.e., bright spots in the study area) as "more important" when modelling the NTL? What are your recommendations?Because the
csvI'm using has several thousands of rows, I can share it via a link, from here. Just so you know, running a model with default parameters takes less than 30 seconds on my laptop (8 gigs of RAM, 4 cores processor (CPU: Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz)).Lastly, I tried to use a model computed by the
bartMachineCVfunction (as well as variable selection) but the results are not better.