In my opinion, both articles use artificial neural networks to find geographical weights.
So, can you help me a little bit? Thank you ~~

The article address is as follows:
https://www.tandfonline.com/doi/full/10.1080/13658816.2019.1707834?src=recsys

Hello, jhagenauer,

I would like to know if it would be possible to export only the trained model for later prediction of values.

In my case, I have a dataset with 6253 observations and 14 independent variables. I want to test the impact on prediction by changing some independent variables. It has been computationally expensive to run the entire model to perform this test. Would it be possible to separate the code into training and prediction?

I thank you in advance.

Hello, Mr. Julian.
I tried to install gwann with this command in Windows 10 :
Sys.setenv("R_REMOTES_NO_ERRORS_FROM_WARNINGS" = "true")
if (!require("devtools"))
install.packages("devtools")
devtools::install_github("jhagenauer/gwann",INSTALL_opts=c("--no-multiarch"))

And i got this error:
Downloading GitHub repo jhagenauer/gwann@HEAD
Error in utils::download.file(url, path, method = method, quiet = quiet, :
download from 'https://api.github.com/repos/jhagenauer/gwann/tarball/HEAD' failed
Warning message:
In devtools::install_github("jhagenauer/gwann", INSTALL_opts = c("--no-multiarch")) :
Arguments in ... must be used.
✖ Problematic argument:
• INSTALL_opts = c("--no-multiarch")
ℹ Did you misspell an argument name?

Do you have any suggestion how to solve the problem?
Thank you so much in advance. May you have a blessed day.

Hi there,

I have installed and loaded rJava, but when I run the following code, I get the error message:

SCRIPT

**library(rJava)

Sys.setenv("R_REMOTES_NO_ERRORS_FROM_WARNINGS" = "true")
if (!require("devtools"))
install.packages("devtools",INSTALL_opts="--no-multiarch")
devtools::install_github("jhagenauer/gwann")**

CONSOLE

Error: package or namespace load failed for 'rJava':
.onLoad failed in loadNamespace() for 'rJava', details:
call: fun(libname, pkgname)
error: JAVA_HOME cannot be determined from the Registry
Error : package 'rJava' could not be loaded
Error: loading failed
Execution halted
*** arch - x64
ERROR: loading failed for 'i386'

• removing 'C:/Users/patri/Documents/R/win-library/4.1/gwann'
  Warning message:

Hello, jhagenauer,

Recently I have done some analysis using this wonderful job. A problem I came across is that, when I organize data in test set in different ways, the prediction results may change. This really confused me. It can be illustrated with the example code in README.md. Below I give three cases to describle the problem in detail.

Case One

library(viridis)
library(gwann)
library(ggplot2)

set.seed(1)
data(toy4)

x<-as.matrix(toy4[,c("x1","x2")])
y<-as.numeric(toy4[,c("y")] )
dm<-as.matrix(dist(toy4[,c("lon","lat")])  )
s_test<-sample(nrow(x),0.3*nrow(x)) # indices of test samples

x_pred1 <- x[s_test,]
w_pred1 <- dm[-s_test, s_test]
r<-gwann(x_train=x[-s_test,],y_train=y[-s_test],w_train=dm[-s_test,-s_test],
         x_pred=x_pred1,w_pred=w_pred1,
         nrHidden=30,batchSize=50,lr=0.1,
         adaptive=F,
         bwSearch="goldenSection",bwMin=min(dm)/4, bwMax=max(dm)/4,
         threads=8
)
p<-diag(r$predictions)
print(p[1:5])

This is exactly the example code in the readme file, except that I have added a seed to make the result comparable. The predition results, i.e., the varaible p in the code is of my interest.

The output is

Golden section search...
Cross-validation results for hyperparameter search (folds: 10, repeats: 1):
* Bandwidth: 2.0720938004226177
* Iterations: 1606
* RMSE: 0.3637624585267835
Building final model with bandwidth 2.0720938004226177 and 1606 iterations...
[1] 4.6746140 8.2461332 0.6771165 0.2575546 2.5992905

Case Two

library(viridis)
library(gwann)
library(ggplot2)

set.seed(1)
data(toy4)

x<-as.matrix(toy4[,c("x1","x2")])
y<-as.numeric(toy4[,c("y")] )
dm<-as.matrix(dist(toy4[,c("lon","lat")])  )
s_test<-sample(nrow(x),0.3*nrow(x)) # indices of test samples

x_pred2 <- x[s_test[1:5],]
w_pred2 <- dm[-s_test, s_test[1:5]]
r<-gwann(x_train=x[-s_test,],y_train=y[-s_test],w_train=dm[-s_test,-s_test],
         x_pred=x_pred2,w_pred=w_pred2,
         nrHidden=30,batchSize=50,lr=0.1,
         adaptive=F,
         bwSearch="goldenSection",
         bwMin=min(dm)/4, bwMax=max(dm)/4,
         threads=8
)
p<-diag(r$predictions)
print(p[1:5])

Golden section search...
Cross-validation results for hyperparameter search (folds: 10, repeats: 1):
* Bandwidth: 2.0720938004226177
* Iterations: 1606
* RMSE: 0.3637624585267835
Building final model with bandwidth 2.0720938004226177 and 1606 iterations...
[1] 4.6890841 8.4763250 0.4011214 0.4280682 2.5759752

Here I only select the first five observations of test data in Case One to form a new test set x_pred2 and corresponding weight w_pred2. From the output we can see that although the cross-validation results are the same, the prediction results are quite different from those in Case One (especially for the third and fourth observations), which is very strange.

Case Three

library(viridis)
library(gwann)
library(ggplot2)

set.seed(1)
data(toy4)

x<-as.matrix(toy4[,c("x1","x2")])
y<-as.numeric(toy4[,c("y")] )
dm<-as.matrix(dist(toy4[,c("lon","lat")])  )
s_test<-sample(nrow(x),0.3*nrow(x)) # indices of test samples

s_test3 <- s_test
s_test3[2] <- s_test3[1]
x_pred3 <- x[s_test3,]
w_pred3 <- dm[-s_test, s_test3]
r<-gwann(x_train=x[-s_test,],y_train=y[-s_test],w_train=dm[-s_test,-s_test],
         x_pred=x_pred3,w_pred=w_pred3,
         nrHidden=30,batchSize=50,lr=0.1,
         adaptive=F,
         bwSearch="goldenSection",
         bwMin=min(dm)/4, bwMax=max(dm)/4,
         threads=8
)
p<-diag(r$predictions)
print(p[1:5])

Golden section search...
Cross-validation results for hyperparameter search (folds: 10, repeats: 1):
* Bandwidth: 2.0720938004226177
* Iterations: 1606
* RMSE: 0.3637624585267835
Building final model with bandwidth 2.0720938004226177 and 1606 iterations...
[1] 4.6732400 4.6189392 0.6733432 0.2431097 2.5998484

In this case, I replace the second observation of test set in Case One with the first observation. So the first two observations are totally the same. And I expect the predicted values of these two observations to be the same (because their data and distance vector are the same). However, the ouput shows that they are different.

In summary, there are at least two strange things happened in the above three cases:

1. Using subset of a test data will give a different prediction (Case One vs. Case Two);
2. Two identical observations in the test set will result in different prediction results (Case Three).

Did I do something wrong? Hope for your help!

Hi Julian, may I know the version of the rJave, since every time I try to import the gawnn (this line -- library(gwann)), the R sessions is aborted.

hello，When I finish reading this article, I want to know how to produce estimated coefficient surfaces of GWR

Hello Hagenauer,
I would like to ask if gwann syntax is currently only used for data testing?
Then couldn't gwann display all the predictions for each location?

I thank you in advance.

Hello Mr. Julian,
I tried your program code and encountered an error in .jcall:
SCRIPT :
r<-.jcall(obj="supervised.nnet.gwann.GWANN_RInterface",method="run",returnSig = "Lsupervised/nnet/gwann/Return_R;",

      .jarray(x_train,dispatch=T),
      y_train,
      .jarray(w_train,dispatch=T),
      
      .jarray(x_pred,dispatch=T),
      y_pred,
      .jarray(w_pred,dispatch=T),
      
      norm,nrHidden,batchSize,optimizer,lr,linOut,
      kernel,bandwidth,adaptive,
      bwSearch,
      
      bwMin,bwMax,steps,iterations,patience,folds,repeats,permutations,threads)

CONSOLE :
Error in .jcall(obj = "supervised.nnet.gwann.GWANN_RInterface", method = "run", :
method run with signature ([[D[D[[D[[D[D[[DZDDLjava/lang/String;DZLjava/lang/String;DZLjava/lang/String;DDDDDDDDD)Lsupervised/nnet/gwann/Return_R; not found

Do you have any suggestions on how to solve the problem?
Thank you so much in advance. May you have a blessed day

Hi, I tried to install gwann with this command in Windows 10:
devtools::install_github("jhagenauer/gwann",INSTALL_opts=c("--no-multiarch"))
I got this error:
Error: (converted from warning) package 'rJava' was built under R version 3.6.3
Execution halted
ERROR: lazy loading failed for package 'gwann'
* removing 'C:/Users/boenz/Documents/R/win-library/3.6/gwann'
Error: Failed to install 'gwann' from GitHub:
(converted from warning) installation of package
'C:/Users/boenz/AppData/Local/Temp/Rtmp6pqJmL/filed243a522a3/gwann_0.0.1.tar.gz’ had non-zero exit status
Do yo have any suggestion how to solve the problem?

Thank you very much!