How to reshape data for LSTM training in multivariate sequence prediction

Question

I want to build an LSTM model for customer behaviour. It's the first time for me working on a timeseries, so some concepts are not clear to me at all.

My prediction problem is multidimensional, meaning that I also want to predict many informations associated to an action for each customer.

The dataset is currently shaped as a list of 2d padded arrays of one-hot encoded features (customer actions + other informations), for example:

   customer_id               encoded_features
0   25464205      [[0,1,0],..,[1,1,1],[1,0,1],..,[1,0,1]]
1   56456574      [[0,1,1],..,[1,0,1],[1,0,1],..,[1,1,1]]

where each element in the encoded_features entries represents a specific timestep.

My idea here is to use keras input shape

(n. customers, n. timesteps, length of features encoding)

In the example above it would be (2,#timesteps,3).

I have two main questions:

1. Is this whole setting rigth for the prediction of next single customer action? I would like to simply give a new sequence of features for a certain customer and predict all features in the next timestep.

2. I am thinking about splitting the data (according to a certain ratio) into sequential training and test sets, in order to test the trained model on unseen feature vectors. In the example above it would be:

  customer_id       X_train          y_train   
0   25464205       [[0,1,0],..]      [1,1,1]  
1   56456574       [[0,1,1],..]      [1,0,1] 

  customer_id       X_test           y_test    
0   25464205       [[1,0,1],..]      [1,0,1]
1   56456574       [[1,0,1],..]      [1,1,1]

Notice that X_train and X_test will generally contain all Train/Test events, except for the last one which has to be predicted. Is this a correct interpretation?

Answer 1

This makes sense. It should work for input and first couple of layers. For output layers, you can have a softmax if you need to generate only next record in sequence.

Following Keras code has an example that :

1. Accepts multi-dimensional inputs (Each sample is a Sequence of video frames)
2. Predicts next few frames of video ( Multi dimensional since each pixel is a feature)

https://github.com/keras-team/keras/blob/master/examples/conv_lstm.py