Multifactor Dimensionality Reduction / News: Recent posts

Added SURF*nTuRF
Stopped disabling CV spinner when forced search type since still relevant
Sped up SURF and SURF* by only evaluating neighbor combinations once. This will change the
scores generated by these algorithms but will preserve the rank order.
Convert scores to z-scores during all versions of TuRF variants. C implementations on which published papers
report always were converting to z-scores. Didn't affect ReliefF much but important for SURF. Still not ideal -- scores chould be ranked in order they were removed from dataset.

added explicit test of interaction which permutes in such a way that main affects are not changed. This has the affect of raising the balanced accuracy needed for an interaction to be considered significant.

fixed a display problem inadvertently caused by change to SystemLookAndFeel

added workaround for known java bug 6429812 http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6429812. On Windows machines we got lots of NULL
POINTER EXCEPTIONS involving sun.swing.table.DefaultTableCellHeaderRenderer.getColumnSortOrder.
Fix was to slightly change way we made custom cell renderers.

Added SURF
Added SURFnTURF
Added support for 'minimal_output' in console when filtering. This will output only the
attributes and their scores to standard out. When minimal_output=true the output filename
for the filtered dataset becomes optional and the filtered dataset is written only when
the name is supplied
Fixed bugs in EDA Linear/Rank
disable cross validation controls when searchMethod set to Forced

2.0 beta 2
Improved reliefF and TURF filtering. Now faster and floating point values better (e.g. .025 instead of .024999999)
Reverted the first change from 2.0 beta 1 because it changed results without adding value
Warn user if java has less than 100mb of RAM
Fixed bug reloading analysis containing filtered data
Slight changed text output for batch mdr console and gui mdr Best Model and CV Results for readability
WARNING -- change to text output may affect parsing of batch mdr results!
Fixed bug in calculating Cross validation consistency for Random and EDA searches. The code to
determine if the same attribute combination was found assumed the attribute
Made parsing of forced attributes looser to allow various common delimiters and whitespace
If forced attributes incorrect, show warning dialog when analysis run is attempted
Improved error messages that occur when reading data files.
Added new feature: Top Models. For each attribute level, this will track the n best
models where n is a user supplied number from a spinner on the configuration tab. This
was added because 'compute landscape' can involve thousands or even millions of models
and can cause mdr to run out of memory.
Tried to prevent 'compute landscape' from killing mdr with out of memory. If you check
'compute landscape' but nothing is shown then we probably ran out of memory and threw
away the landscape.

2.0 beta 1
Fixed bug in calculating testing accuracy. Previously, if testing dataset contained an attribute combination (genotype)
that was not in the training dataset, those instances would be ignored. It is more correct to treat those instances as having a
'tie status' and therefore be assigned whatever the tiePriority is set to. This is a very minor issue and affects only
small datasets where the counts might be zero in the testing set.... read more

This new version will load and analyze a genome-wide association study (GWAS) using stochastic search algorithms. The key to a successfull GWAS using MDR is expert or domain-specific knowledge. This can be in the form of LOD scores from a linkage analysis, ReliefF score on the SNPs or any other number you can assign to each SNP such as that derived from biochemical pathways or Gene Ontology, for example. I am a firm believer that expert knowledge of this type is the only thing that will make a GWAS of epistasis technically feasible.... read more

1.2.5
Fix attribute construction bug. The bug also occurred when running with many levels. If the number of combinations of the attributes exceeded the number of rows then the code would switch from a dense matrix to a sparse matrix. The bug was in the filling up of the sparse matrix and the result was we would get a null pointer exception and the attribute would silently not be created.

Fixed a similar bug when reading back a saved analysis file: need to use same logic as buildCounts to decide whether to use DenseMatrix or SparseMatrix.

The objective of this project is to make available an open-source version of our Multifactor Dimensionality Reduction (MDR) software. MDR is a nonparametric alternative to logistic regression for detecting and characterizing nonlinear interactions.

1.2.4
Fix saved analysis output bug. In the '@results' section for each model, where the training and
testing confusion matrix are output for each cross validation test, the testing was incorrect --
it was the training confusion matrix repeated rather than the testing matrix. The averages in
the line with 'AvgTrain' were correct so the overall training and testing accuracies were still
available.... read more

0.4.7
re-generated MDR_PT_Overview.pdf since somehow lost in 0.4.6
0.4.6
Enhanced dataset loading to support space delimited files (since mdr does this)

Does a better job opening files with thousands of attributes.

Removed the Sign test column from the summary table results.

Added new chart types showing attribute interactions. Renamed the 'Dendrogram' results tab to 'Entropy' so all chart types are available under 'Entropy'

This version uses a new MDR engine, so it runs much faster than previous versions. It has also been updated visually, so it should look nicer.

After several months of testing the release candidate by our user base, we have decided that MDR is ready for version 1.0. Only minor changes have been made from rc1, and they are listed in the change log.

This is a bugfix release.

We've added some neat new features to MDR. We feel that this is the feature set we want in version 1.0.0, so we've made this the first release candidate. Try it out!

Changed to be consistent with MDR 0.6.2.

This version uses balanced accuracy as a fitness measure, and always uses the ratio from the dataset as the ratio threshold. This combination works well with both balanced and unbalanced data.

A bugfix release.

This is an unsupported library for use by people who need to run MDR on a supercomputing cluster, or generally need more performance from MDR.

Added a new wrapper framework. The only new wrapper is random search, but there will be more to come. Also added an odds ratio filter, and fixed some bugs.

This is the first major update to MDR in quite a while. It includes lots of new features, like pre-processing filters, new statistics, and new options. Check it out!

Another bugfix release... will they never end?

This one fixes the -saveanalysis problem again, as well as adjusting the algorithm slightly to cope with data sets that don't generate any predictions at all in cross-validation.

Bugfix releases. Most notable is that now, null attribute values work correctly in all three applications.