Author Topic: learning  (Read 7542 times)

Offline pibar

  • Newbie
  • *
  • Posts: 5
    • View Profile
learning
« on: August 16, 2010, 21:23:13 »
I have historical data that produces a very interconnected network that uses terabytes of memory. My data corresponds to 15 continuous value variables, discretized in 10 intervals (minimum).

How can I get only the arcs more representatives, keeping the number of parents of a node limited?

Offline Anders L Madsen

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 2283
    • View Profile
Re: learning
« Reply #1 on: August 18, 2010, 15:13:14 »
Quote
How can I get only the arcs more representatives, keeping the number of parents of a node limited?

This is one of the weaknesses of the constraint-based approach to structure learning implemented in HUGIN.

More or less the only option is to change the significance level of the statistical tests performed. You should reduce the number to obtain fewer edges in the graph.

Alternatively, you can try one of the restricted models (e.g., Naive Bayes) or use constraints. Using constraints can have some side-effects though.
HUGIN EXPERT A/S