Ground-breaking paper on Integrative Causal Analysis published in JMLR

Is it possible to predict a disease is correlated with an exposure factor without ever performing a study to measure them? The answer is shown to be yes, in some cases, by analyzing existing datasets that measure these two quantities, but not jointly. The idea is to induce all causal models that are simultaneously consistent with several existing datasets and make novel statistical inferences, sometimes regarding quantities never jointly measured together. We have named the approach Integrative Causal Analysis (INCA). A recent important publication in the Journal of Machine Learning Research shows that INCA often makes new and very accurate inferences. It works ubiquitously on all domains we have tried it. It is strong evidence that theories of Computational Causal Discovery can go further than classical statistics: modeling and inducing causal relations gives more inferences capabilities.

STATegra project is accepted for funding

The STATegra collaborative FP7 proposal has been accepted for funding by the EC! Our group is a member of the STATegra consortium and will be leading the Work Package related to the development of machine learning, statistical, and data mining methods to integratively analyze heterogeneous biological datasets. STATegra funding will allow us to grow, expand, and perform leading edge research on quite novel machine learning methods, particularly in the direction of Integrative Causal Analysis. This is an ambitious project, excellent consortium and we are very excited to be a part of it. Hiring should start soon.

Website: http://stategra.eu/
Facebook page: http://www.facebook.com/pages/STATegra/372832349410163
Twitter page: https://twitter.com/#!/STATegra

ECML Tutorial on causal discovery

Can we learn causal relationships from observations? In this 3-hour tutorial we attempt to give the intuition behind causal discovery from observational data, discuss the main assumptions, present the theory of Bayesian networks and some basic causal discovery algorithms, demonstrate some of their most successful applications to feature selection and biomedical data, and also share what the typical causal discovery enthusiast won’t tell you: what can go wrong.

Watch the tutorial, which was presented in ECML- PKDD 2011, here: http://videolectures.net/ecmlpkdd2011_tsamardinos_discovery/

Bioinformatics Laboratory is founded in ICS-FORTH

The management of the Institute of Computer Science, Foundation of Research and Technology, Hellas has recently (Oct. 2010) decided to found the Bioinformatics Laboratory, with the Mens X Machina group members as its founding and core members and Prof. Tsamardinos as its first Head. The decision reflects the commitment of the management to progressing and advancing Bioinformatics in the Institute, to provide our group with administrative autonomy, and recognizes the contributions of our group to the field. The Mens X Machina group will continue its mission as Bioinformatics Laboratory. The Mens X Machina website will still be maintained but a new updated Bioinformatics Laboratory website will be available within a couple of months.

New MATLAB toolboxes versions 0.9.2

New versions (0.9.2) of our MATLAB® toolboxes have been released.
New features in version 0.9.2 of Mens X Machina Probabilistic Graphical Model Toolbox (MxM PGM) include:

• Skeleton identification error estimation/calculation utilities
• Templates for custom functions, handles to which can be used as values of skeleton identification function parameters.

New features in version 0.9.2 of Mens X Machina Commons Toolbox include multiple testing error estimation/calculation utilities.

Also, a supplement to MxM PGM called Mens X Machina Probabilistic Graphical Model Samples (MxM PGM Samples) is available to download. MxM PGM Samples is a collection of samples from several Bayesian networks.

New MATLAB toolboxes versions 0.9.1

New versions (0.9.1) of our MATLAB® toolboxes have been released. The Mens X Machina Bayesian Network Toolbox has been renamed to Probabilistic Graphical Model Toolbox, because we intent to include functionality that covers other types of graphical models such as Maximal Ancestral Graphs (MAGs) and Pairwise Casual Graphs (PCGs) in the future.
New features in version 0.9.1 of Mens X Machina Probabilistic Graphical Model Toolbox include:

• Creating and sampling Gaussian Bayesian Networks
• More customizable MMPC via new parameters
• MMPC support for continuous data by employing the Fisher’s Z test of conditional independence

The most prominent addition in version 0.9.1 of Mens X Machina Commons Toolbox is the Fisher’s Z test of conditional independence.
Unfortunately, because the previous versions were the first ones and many things had not yet been settled, the new versions are not backward-compatible in general, but the Compatibility Considerations sections in the accompanying release notes should be helpful in migration. Hopefully, from now on the packages will be backward-compatible.

The MMPC algorithm adapted for survival data

A new Bioinformatics paper adapts and compares MMPC for molecular signature discovery in high-dimensional, micro-array gene-expression survival data outperforming the state-of-the-art in the field. The source code of the algorithm and experimental scripts are available in the Software section.

New toolboxes available for MATLAB

We have just released the first versions of our toolboxes for MATLAB®:

Mens X Machina Bayesian Network Toolbox aims to provide a comprehensive set of tools for bayesian network inference and structure learning. Currently only bayesian network skeleton learning using the MMPC algorithm is implemented. Support for full structure learning and inference will be added in future versions.

Mens X Machina Commons Toolbox is a MATLAB toolbox with common functions (to be) used in other toolboxes provided by Mens X Machina, including the Bayesian Network Toolbox.

New paper in JMLR, Special Topic on Causality

Our latest journal paper entitled Local Causal and Markov Blanket Induction for Causal Discovery and Feature Selection for Classification has appeared in two parts in the Journal of Machine Learning Research, Special Topic on Causality, vol. 11, 2010 (see Publications).

Upcoming AI & Statistics 2010 paper

A paper based on Ms. Sofia Triantafilou’s master thesis, Learning Causal Structure From Overlapping Variable Sets, will be presented at the Thirteenth International Conference on Artificial Intelligence and Statistics which will be held in Chia Laguna Resort, Sardinia, Italy on May 13-15, 2010.