ANN training using two time series as input

Hello everyone. A) Please and I need help on how I can train a Neural network using two sets of time series (same size) with each being row vectors ranging up to 1500 data points (1x1500) as input variable and to return a single class of output.

Input 1= [-20 -20.30 -20.61 -20.91 -21.20 -21.49 -21.77 -22.05....]

Input 2= [-15.81 -15.44 -15.05 -14.67 -14.28 -13.88 -13.48 -13.08....]

Output=[H1]

I have about four classes of output that each case of the double time series can define and I have about 700 cases of the above data type that I want to use for training, validation and testing.

B) In another case I also want repeat the above procedure but my output will be two vectors i.e something like:

Input 1= [-20 -20.30 -20.61 -20.91 -21.20 -21.49 -21.77 -22.05....]

Input 2= [-15.81 -15.44 -15.05 -14.67 -14.28 -13.88 -13.48 -13.08....]

Output= [0.2331 -3.221]

A typical sample of the code for training and testing using MATLAB functions will be appreciated

Please I am new to this area of study, simple and easily understandable terms will be preferred

Answer 1

[Greg Heath]

You have a mischaracterization of the standard classification and timeseries models.

1. Classification input data is divided into class-labeled areas of I-dimensional input space. However, the areas of a single class are not necessarily connected.

   The input matrix for c classes contains N   
I-dimensional samples and has dimension [ I N ].

   The output matrix for c classes contains N 
c-dimensional 0-1 unit vectors where the row of 
the 1 indicates the class index of the  
corresponding class.

   In general there is no correlation between 
the physical location of inputs and their matrix 
location.

2. Typically, time-series are not used for classification. Instead they consist of a series of correlated inputs that so that , for example, a string of m samples are used to estimate the following sample of the same series and/or an associated output series.

Hope this helps

Thank you for formally accepting my answer

Greg

[Kenneth Afebu]

Thanks and really appreciate your explicit explanation. Please maybe I should let you know what I am trying to do, perhaps there can be a way around it using Neural Network. A) I want to use ANN to predict the type of plot two sets of time series will give rise to. I am restricting myself to just four types of plots which can be H11, H12, H21 or H22. Normally, you have to plot these two time series against each other to know the type of plot, but right now I want to explore the ability of ANN to learn the sequential relationship between the values of these two time series and tell me the type of plot without plotting them. Each of the times series will have a dimension of at least (1x1500) to be able to make the complete plot. The ability of the ANN to use lesser dimension to predict the plot type is also to be explored.

B) Another thing is that, there are two values (each belonging to each of the time series) which controls the sequential progression of the two time series and thus the type of plot that comes out. Some plots are more desirable than others. I want to explore if ANN can learn the relationship between the two sequential time series and the controlling values such that if it (ANN) is supplied with an undesirable type of plot (defined by the two time series), it can predict the controlling values that can transform it to a desirable one.

You contribution will be highly appreciated. Thanks