I am trying to analyze a time series in R using convolutional neural network function provided in the mxnet package. Please let me know 1) What should be the value of num.filter in mx.symbol.Convolution? 2) What changes are to be done in the code here, so that it becomes fit for 1D CNN (Time Series)?
Reference: http://mxnet.io/api/r/mxnet-r-reference-manual.pdf
The num.filter parameter is a hyper-parameter which will effect the expressiveness of your model. A larger number of filters will give you a more expressive model, which can find more subtle patterns given enough data, but is also more likely to over-fit. So in general there's no "best" answer, but this is something you'll need to experiment with for your dataset.
As for building a time-series model with a CNN, again, there's no simple answer. It's certainly possible to use a CNN for time-series analysis, but I wouldn't start with an image-processing CNN like the one you link to. This question https://stats.stackexchange.com/questions/127542/convolutional-neural-network-for-time-series gives a bunch of good references on how to build time-series models with neural networks.
You might also consider using an RNN, which are generally more naturally suited for time-series analysis. Here's a good example of running an RNN in R with MXNet: http://mxnet.io/tutorials/r/charRnnModel.html
