Neural Network toolbox. Training and testing the model

Hi everyone. I have my input train and output train dataset which are 5184x3 and 5184x8. I am a new for all staff about neural network and etc. My output dataset colums values varies hugely i.e first coloums max value is 0.8 wherease seconds is 150.should I scale down the values to between 0 and 1? Actually, I want to construct network to predict the other values and I have additonally test data. However, I have a big error.

Here is my code:

T= xlsread("1data.xlsx");

intrain=T(1:5184,1:3);

outtrain=T(1:5184,4:11);

intest=T(5185:8640,1:3);

outtest=T(5185:8640,4:11);

net=feedforwadnet(10);

intrans=intrain';

outtrans=outtrain';

n=train(net,intrans,outtrans);

y=n(intrans);

yy=y';

error=outtrain(1,1)-yy(1,1);

Answer 1

[Aditya]

Hi Medet,

Absolutely, scaling your data is a crucial step when working with neural networks, especially when your output variables have significantly different ranges. If you do not scale your outputs, the neural network’s loss function will be dominated by the outputs with larger values, making it difficult for the network to learn the smaller-valued outputs effectively. This imbalance often leads to poor prediction accuracy and large errors, as you have observed.

To address this, it is recommended to normalize both your input and output data to a common range, typically [0, 1]. MATLAB provides the mapminmax function for this purpose. This function scales the data and also returns settings that can be used to reverse the scaling (i.e., to get predictions back to their original scale after the network has made its predictions).

Below is an example of how you can properly scale your data, train your neural network, and then reverse the scaling on the predictions:

% Read data
T = xlsread("1data.xlsx");
intrain = T(1:5184,1:3);
outtrain = T(1:5184,4:11);
intest = T(5185:8640,1:3);
outtest = T(5185:8640,4:11);

% Transpose for neural network toolbox
intrans = intrain';
outtrans = outtrain';

% Scale inputs and outputs to [0, 1]
[intransNorm, inSettings] = mapminmax(intrans, 0, 1);
[outtransNorm, outSettings] = mapminmax(outtrans, 0, 1);

% Create and train the neural network
net = feedforwardnet(10);
net = train(net, intransNorm, outtransNorm);

% Predict on training data (normalized)
yNorm = net(intransNorm);

% Reverse the scaling to get predictions in original scale
y = mapminmax('reverse', yNorm, outSettings);
yy = y';

% Calculate error on original scale
error = outtrain(1,1) - yy(1,1);