Show Posts

This section allows you to view all posts made by this member. Note that you can only see posts made in areas you currently have access to.

Messages - Anders L Madsen

Pages: 1 2 [3] 4 5 ... 20
General Discussion / Re: Analysing AIC, BIC and LL scores
« on: February 24, 2015, 14:34:54  »
My doubts concern the analysis of the results reached after the training of the model: when I test the network with a test set (of examples that don't belong to the training set), how should I interpret the AIC, BIC and log-likelihood scores? When are they consistent or good enough? Have they significance on their own or only compared with the scores of another model? And in the comparison with another model, how to choose the best one?

The AIC, BIC and LogLikelihood (LL) scores are criteria for selecting between a set of candidate models representing a data set. The LL specifies how well a model represents a data set and the LL can be increased by making the model more complex. So, this score should only be used to compare models with the same complexity. Both the BIC and AIC scores are based on the LL with a penalty score for complexity. The penalty score is different for BIC and AIC.

You should use these scores to select between candidate models as a representation of a data set. The higher the score, the better the model represents the data.

can I deduce something (good or bad) about my network?
No, not as far as I know. You can use it as a relative score to compare models and select a model with highest score.

Another parameter of the analysis is the ROC curve: which role does it play in the valuation of the goodness of the model?
The ROC can be used for assessing the performance of a classification model, i.e., your model should be used to assign a class label to a set of instances. The ROC (and the area under the ROC) is a measure of classification performance showing the True Positive rate as a function of the False Positive rate.

You can find some introductory material on these concepts on Wikipedia.

FAQ / Re: Analysis Wizard
« on: September 08, 2014, 09:09:37  »

In the Analysis Wizard, the Error rate for Test Data Accuracy pane of how much is acceptable?

This depends on the domain and the application. It may also be relevant to take a look at AUC of the ROC (which should be at least 0.5). By searching the internet you will be able to find rules of thumb on how to interpret/classify the performance of a classifier based on the AUC of the ROC.

In section Case table, The probability values 0.5 or less, Large difference between the actual data and test data are available (By multiplying the probabilities of each category for discrete data.), the cause is?what to do to fix it?

I do not understand this comment. If the performance of the model is not sufficient, then the model should be improved. When building a Bayesian network classifier from data a number of design choices have to be made, e.g., which variables to include, how do discretize numerical variables and which edges should be present in the model. It is impossible to say how you model could be improved without a detailed description of the model and data.

AMIDST / Project Description
« on: August 11, 2014, 10:35:15  »
The web-site is dedicated to hosting information on the AMIDST project with the main focus on the use of Bayesian networks.

ActiveX / Re: Is it possible to unselect a state ?
« on: August 02, 2014, 08:21:54  »

Use RetractFindings for this

FAQ / Re: Model validation
« on: August 02, 2014, 08:18:51  »
You cannot compute the AIC and BIS scores from this information. See section 12.2 of the HUGIN API Reference Manual on how graphical models are scored.

FAQ / Re: Model validation
« on: June 14, 2014, 13:59:26  »
The number of free parameters in a discrete CPT is (n-1) * m, where n is the number of states in the child and m is the number of parent configurations.

You can find a lot of information on AIC and BIC by performing a few Google searches.

General Discussion / Re: Decision Trees
« on: May 01, 2014, 11:13:58  »
Dear Marco,

It is correct that HUGIN does not directly support decision trees.

Instead HUGIN supports the use of influence diagrams and limited memory influence diagrams (LIMIDs). LIMIDs were introduced as part of version 7.0. With the introduction of LIMIDs, the solution algorithm was changed from being based on Jensen, Jensen & Dittmer (1994) to being based on Lauritzen & Nilsson  (2001). The solution algorithm is referred to as Single Policy Updating (SPU). SPU requires that all informational links are specified in the model meaning a change in the interpretation of the structure of the influence diagram compared to previous versions of HUGIN.

Hope this helps

Lauritzen, S. L. and Nilsson, D., (2001), Representing and solving decision problems with limited information. Management Science, 47, 1238 - 1251.

Jensen, F., Jensen, F. V., Dittmer, S. L., (1994), From influence diagrams to junction trees, Proceedings of the Tenth Conference on Uncertainty in Artificial Intelligence, pages 367-373.

OpenNESS / Re: zero intervals and benefit cost analysis
« on: April 14, 2014, 19:36:44  »
Can you help?

We can try.

1) Explanation.

The zero width interval -feature was introduced to extend the Table generator functionality. We did not consider including this as part of the Learning Wizards. So, the discretization operator in the Learning Wizard simply ignores the zero width interval.

2) Workaround.

Here is a workaround (I assume that you are using the Learning Wizard and not only the EM Learning Wizard):

  • do the discretization  and structure learning in Learning Wizard
  • leave the Learning Wizard prior to the EM part (parameter estimation)
  • manually add the "zero width interval" to the appropriate node
  • use the EM learning wizard to perform the parameter estimation on the adjusted model

If your are using the EM learning wizard, then the steps should be adjusted accordingly.

As you have no data on zero, then you will probably not learning anything about the relation between the parents for this value. The result will probably be a uniform likelihood.

OpenNESS / Re: OpenNESS forum
« on: April 14, 2014, 18:55:41  »
Hi David,

This is not a restricted access area. Everyone can read and post messages here.

Only members of the OpenNESS group can read this part of the forum:,29.0.html

best regards

FAQ / Re: Credible Interval
« on: February 11, 2014, 15:26:21  »
Maybe you can use the variance, which can be displayed in the monitor window of the node in the  Graphical User Interface (GUI)?

Display of the variance in the GUI can be enabled in "Network->Network Properties" under the "Monitors"-tab.

This functionality is not available in the API.

Hope this helps

Probabilistic Graphical Models (PGM2014) / Call for papers
« on: February 06, 2014, 15:36:21  »
By request of Silja Renooij, Universiteit Utrecht we post this CfP:

Call for Papers
 PGM 2014: The Seventh European Workshop on Probabilistic Graphical Models
 September 17-19, 2014
 Utrecht, The Netherlands

 Important Dates:

 Deadline for abstract submissions: May 12, 2014
 Deadline for paper submissions: May 16, 2014
 Notification of acceptance: June 23, 2014
 Final versions due: July 18, 2014
 Workshop dates September 17-19, 2014


 Call for papers:

The aim of the workshop is to bring together people interested in
probabilistic graphical models and to provide a forum for discussion
of the latest research developments in this field. To promote interactions
among the participants, parallel sessions will be avoided and the workshop
will be organised around a single thread of plenary and poster sessions.
We welcome theoretical and applied contributions related to various aspects
of probabilistic graphical models, such as:

 - Principles of Bayesian (belief) networks, chain graphs, decision networks, influence diagrams,
   probabilistic relational models, and other probabilistic graphical models (PGMs).
 - Information processing in PGMs (exact and approximate inference).
 - Learning and data mining in the context of PGMs: machine learning approaches,
   statistical testing and search methods, MCMC simulation.
 - Exploitation for the construction of PGMs of results from related disciplines
   such as statistics, information theory, optimization, and decision making under uncertainty.
 - Software systems based on PGMs.
 - Application of PGMs to real-world problems.

Papers submitted for review should report on original, previously unpublished work.
Each submitted paper will be reviewed by at least three reviewers.

All accepted papers will be included in the workshop proceedings.
At this stage, we are negotiating publication by a renowned publisher.

Submission procedure:

PGM 2014 requires electronic submission of papers according to the
instructions found on the web site of the workshop:

The authors must prepare their papers according to the Springer LNCS
format, and submit the full version of the papers by the deadline stated above.
The page limit is 16 pages, including figures and bibliography.
The web page for author instructions contains further format information
and provides access to style files and templates.

Paper presentation:

Papers will be accepted for plenary or poster presentation.
In the proceedings, no distinction is made between the two.
At least one author of accepted papers is required to register for the workshop.

After the workshop, authors of selected papers will be invited to submit
an extended version of their paper for a special issue of a journal, which is yet to be decided.
The extended papers will undergo a full reviewing process.


The workshop is hosted by the Department of Information and Computing Sciences,
Utrecht University, The Netherlands.

OpenNESS / Project Description
« on: January 22, 2014, 19:44:25  »
The web-site is dedicated to hosting information on the OpenNESS project with the main focus on the use of Bayesian networks.

Pages: 1 2 [3] 4 5 ... 20