locked
Infer.net - complex factor (exponential function) RRS feed

  • Question

  • Hi

    I want to use infer.NET for inferring networks from time series data. I have N variables, each can have a continuous value at time point t, which is a function of all node values at time (t-1).

    Vi(t) = alphai * Vi(t-1)      // this is the effect of the node's value in time t-1 on its value in time t

              +  betai * 1 / (exp(sumj ( Wji * Vi(t) )))   // this is the effect of other nodes, stored in a matrix w for all nodes

    alph and beta are parameters for tuning the amount of self/other's effect on current state of a node.

    Sum is over all j values except i itself (j != i in sum).

    w is the weight matrix. Wji is the weight of the effect of node j in time t-1 on node i in time t.

    My first question is that can I implement this model in INFER.NET? the exponential part does not seem to be implementable.

    If I have a simpler function for relating nodes in time t to time t-1, then how should I train this model to find weights (w). Should I just write the program, and then break my data into windows (t, t-1) records, and introduce this (t, t-1) data as the observed data? How can I learn the weights?

    Any help is appreciated.

    Thanks a lot,

    Zahra


    • Edited by RazinR Tuesday, May 20, 2014 1:22 PM the first subject was not good for it.
    Thursday, May 15, 2014 5:01 PM

Answers

  • I think you may be looking for Variable.Softmax with the appropriate transformation.
    • Marked as answer by RazinR Tuesday, May 20, 2014 2:39 PM
    Friday, May 16, 2014 6:26 AM

All replies

  • I think you may be looking for Variable.Softmax with the appropriate transformation.
    • Marked as answer by RazinR Tuesday, May 20, 2014 2:39 PM
    Friday, May 16, 2014 6:26 AM
  • Thank you Andrew. Yes it seems to work in this case. I'm going to check it.

    Bests

    Monday, May 19, 2014 7:32 AM