Predict future values of a PRNN

My data set consists of 4 input variables, and 1 target variable with 5 categories. My input and target variables are the same length (~2000 values each). I followed the tutorials with the Pattern Recognition Neural Network, and then used my data in the network with okay results. I would like to use the network to predict the next 5-20 values, but having trouble. When I use the 'sim' function, I get a matrix the size of my target (5x2200) instead of the desired size of my prediction (5x20). How do I correct the prediction error?

% Load INPUT and TARGET
Load data.mat    % Contains 'inputs' & 'targets' variables

% Create a Pattern Recognition Network
hiddenLayerSize = 17;
net = patternnet(hiddenLayerSize);

% Set up Division of Data for Training, Validation, Testing
net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

% Train the Network
[net,tr] = train(net,inputs,targets);

% Test the Network
outputs = net(inputs);
errors = gsubtract(targets,outputs);
performance = perform(net,targets,outputs);

% Predict next 20 values
y = sim(net,inputs);

Any help is greatly appreciated!

Thanks,

Dan

Answer 1

[Asvin Kumar]

Explanation for the output you’re getting:

Your patternnet is learning a mapping/relationship between the inputs and the target. Using the sim command, you’re predicting the values for every data point in inputs. Since inputs is of, I presume, size 4x2200, your predictions are of size 5x2200. Essentially, the network is predicting for every 4-length data point.

The requirement:

What you want is for the network to predict the future values. You haven’t mentioned whether there is temporal information in your dataset. That is, you haven’t mentioned whether these data points are sequential, in some sense, and have enough information among them to predict future target values. I am going to assume they do. There are, at least, two ways to achieve this.

The first way is for you train your network to predict the k-th future target which can be done with the following statement for some constant ‘k’:

[net, tr] = train(net, inputs(1:end-k), targets(1+k:end));

In this approach, you’re not learning/using any temporal relation among the datapoints. The network only learns a mapping/relation between the input for a given time instant and the target k steps into the future. This leads to the shortcoming that you’ll have to construct different networks for different values of ‘k’ if you’re constructing your networks using patternnet. You could make this better by constructing an equivalent network using fullyConnectedLayers of appropriate sizes. This has the advantage that you can drop the necessity for one-hot vectors. You can then replace the targets with a 5-20 length vectors containing the categorical target information for the next 5-20 time steps. Here’s an example to help you understand how fullyConnectedLayers are used: https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.fullyconnectedlayer.html#mw_bad3166a-93e7-42c0-8bd2-740cd2a842ad

The second approach is to use an LSTM. This is a type of network which can learn relations between consecutive samples. This approach can be implemented like the previous approach except it also uses temporal information. You could train the network to learn a temporal relation between the current input and a 5-20 length target vector of future categorical information. Here’s a simple example for you to model your network: https://www.mathworks.com/help/deeplearning/examples/time-series-forecasting-using-deep-learning.html

[Dan Griffiths]

Thank you very much for your reply!

If I understand it correctly, patternnet is good at finding relationships between an input and target, but not designed to make time series predictions. It sounds like LSTM might be the best solution... but how do I structure my input data? The example you gave predicts future values only from past values. We would like to use our input to predict patient outcomes. Currently, I have four input parameters to determine 1 of 5 categories for the patient. Each row of the input is a different parameter, and each column is a different time point. Each row of the target is a patient category, and each column is the corresponding timepoint... the target is populated with 0s and 1s, with 1s indicatng the patient status. This structure seemed to work well for patternnet, but doesn't seem to work for LSTM.

[Asvin Kumar]

You can modify the targets from a one-hot representation to a single category number. You can then also create another vector of the next ‘k’ categories to be predicted for the current time point. You can do this with the following code:

[category, max_categories] = vec2ind(targets); 
k = 3; % or 5 or 20 or as many future steps as you want to predict. 
newtargets = fliplr(toeplitz(category(k:end),category(k:-1:1))).'; 

This code gives you the sequence of new targets of length k x (N-k+1) where N is the sequence length. This means you will accordingly train on a shorter sequence.

Your testing will also be different from the code given in the example. You can try using something like:

net = predictAndUpdateState(net,XTrain); 
 
numTimeStepsTest = numel(XTest); 
for i = 1:numTimeStepsTest 
    [net,YPred(:,i)] = predictAndUpdateState(net,XTest(:,i),'ExecutionEnvironment','cpu'); 
end 

This way, when you predict with the patient’s data at any time point, you will get an estimate of the patient’s category in the next k time points.