Bug fix
- corrected two notes (on documentation and vignettes) from CRAN checks
- replaced aes_string by aes
Misc
- changed CITATION format according to new standards
- removed cache from vignettes
Fixed bugs:
- fixed a bug related to issue #4 on Github
New features:
- numeric SOM can now handle missing data: training is done without accounting
- for missing entries and resulting SOM can be used to impute missing entries
Fixed bugs:
- proto0 was incorrectly used for ‘numeric’ (see bug report #4 on Github)
- dist.type was improperly handled for types other than ‘euclidean’ and ‘letremy’ when radius.type was letremy
- minor bug in computation of tolerance for numerical approximations with Gaussian radius
- removed ‘tolerance’ from the prediction (not used any more)
- fixed minor issues in OO for trainSOM
Improvements:
- Heskes’ affectation now account for total neighbord weight in order to not favor too much the borders of the map
Misc:
- refactorization of most of the code
Improvements:
- make trainSOM a method for different classes
- create a trainSOM method for outputs of mixKernel package
Fixed bugs:
eps0 parameter taken into account in shiny app correct use of “what” parameter for superclasses plot fixed a minor problem in Heskes’ assignment when distance is Euclidean
Improvements:
- hexagonal topography is now available
- plots are now ggplot2 based and former plot arguments are no longer valid since ggplot2 syntax is now used
- the shiny app has been entirely remade
- data from the environment can be loaded directly in the shiny application and 3 examples datasets are loaded automatically in the shiny app
- titles and legends in plots have been improved
- code used in the app is printed to ease reproducibility
- protoDist (for neighbors) has been improved by adding, for the case ‘neighbors’ a radius argument. When this radius is passed the Euclidean distance between unit coordinates
- print.titles is deprecated and has been replaced by show.names
Misc:
- radar plots are no longer available
Fixed bugs:
- adapted script to the new stringsAsFactor policy on devel
- fixed a bug for non significant variables in summary
Misc:
- roxygenized documentation
- changed maintainer name and address, added new contributors
- fixed URI in package vignette
Fixed bugs:
- changed unitary tests because of changes in RGN for devel
Misc:
- changed maintainer name and address
Fixed bugs:
- removed unecessary scaling checks for ‘relational’ in predict
Improvements:
- used testthat to implement unitary tests
- remove redundant code in trainSOM
- fixed seed in unitary tests for reproducibility reasons
Misc:
- changed message at start (that might be construed as offensive by colorblinds)
Fixed bugs:
- fix a bug in plot.somSC for customized colors (reported by Duncan Murdoch)
- fix a bug in shiny app (downloading CSV files with rownames)
- fix a bug in ‘plot.legend’ for SC plot of type ‘color’
Improvements:
- improved rendering of data in shiny app
- added ‘legend’ to all SC plots for shiny app
- added features to display titles (cluster number) in package and shiny app
- added CSV download for result of the clustering in shiny app
Misc:
- updated (and improved references)
- changed maintener’s email
Fixed bugs:
- fixed heskes predictions for relation SOM
- updated scripts for latest igraph version
- removed predict.somRes for cosine preprocessing (relational)
Code improvements:
- faster version of dissimilarity SOM as in Mariette et al., 2016
Fixed bugs:
- Changed observation weights in KORRESP: they now correspond to the 2005 KDISJ version and not to the 1993 version
- Fixed bug when passing proto0 (reported by Renaud Dufour)
Code improvements:
- changed the function predict.somRes to speed it up
SOMbrero 1.0 (CRAN first release)
New features:
- projectIGraph now handles weighted graphs
- added a type in plot.somSC that displays a projected graph based on super-graph (see Olteanu & Villa-Vialaneix, Journal de la SFdS, 2015)
Fixed bugs:
- Fixed a bug related to proto0 when type was not set in trainSOM
- Fixed a bug in summary.somSC when the number of clusters is not chosen yet
- Removed parameter ‘view’ that was not used in internal function ‘plotAdd’
Misc:
- Changed default neighbourhood to Gaussian. The default parameters for this neighbourhood are those already implemented in the R package ‘yasomi’
- Changed the internal function projGraph to handle projection for the SOM grid and for the super-clustering in a single function
- Added the neighborhood type of the SOM in method print for the class somRes
- New features:
- Added an option for affectation type (“standard”/hard or “heskes”/soft affectation)
- Updated shiny interface accordingly
- Fixed bugs:
- Fixed a bug related to dist.type in selectNei. This bug did not affect the learning in previous versions but could provide false values for protoDist after the training.
- New features:
- Added PCA initialization
- Added Gaussian neighborhood (default parameters are still tentative)
- Added cosine normalization for dissimilarity
- Added ANOVA like analyses for relational SOM interpretation
- Included a function to project an ‘igraph’ object on the grid
- Fixed bugs:
- Fixed a bug in the computation of projected graph. Only edges with non zero counts for the number of links are now listed in the edge list.
- Misc:
- Added tests: to check the equivalence between automatic and manual scaling and the equivalence between predicted clustering and the output of predict
- Updated vignettes
- New features:
- Added a parameter that tunes the gradient descent step
- Added a web user interface, implemented with shiny
- Added the topographic error to somRes.summary
- Improved the summary of somSC (table of neuron numbers and superclass numbers)
- Code improvements:
- Changed default value of init.proto to “obs” for “relational” algorithm
- Fixed bugs:
- Changed scree plot title and ylabel to “proportion of unexplained variables”
- Misc:
- Added test: equivalence of numeric SOM and relational SOM on squared euclidian distance
- New features:
- Implementation of the “dissimilarity” algorithm to handle dissimilarity data
- Plots, quality functions and super-clustering have been updated to handle tjos case
- Code improvements:
- New way to handle distance calculations
- New way to handle the ‘korresp’ preprocessing (full matrix is now included)
- Slightly changed default values for Letremy’s radius and neighborhood: they are now exactly the same than in the original programs
- Fixed bugs:
- Fixed bug with default options in korresp case
- Fixed bug with affectation step in korresp case
- Fixed bug for what=“add” and type=“graph” (node sizes were wrong)
- Misc:
- Added markdown vignettes: one general and one for each case; hence, the package now depends on R version 3.0 or higher
- Added citation file
- Added demos
- Removed unecessary tests and examples
- New features:
- Implementation of the “korresp” algorithm to handle contingency table
- Plots, quality functions and super-clustering have been updated to handle this case
- Fixed bugs:
- Bug fixed in the dendrogram of super-clustering (percentage of explained variance was incorrectly displayed)
- Bug fixed in smooth.dist and umatrix (distances were reversed)
- New options:
- The argument ‘the.titles’ has been introduced in plot.somRes
- New features:
- Implementation of a quality function for somRes object
- Implementation of a super-clustering for the resulting prototypes
- Implementation of a plot method for the results of the super-clustering
- Name changes:
- The argument ‘scale’ of ‘somRes’ object is now called ‘scaling’ (conflict with a R base function)
Argument changes: The argument ‘dist.type’ has been transfered from ‘somRes’ to ‘myGrid’ objects.
- New options:
- Distance type on the grid : maximum, euclidean, manhattan, canberra, binary or minkowski
- New plots:
- Observations : color, lines, barplot, radar, boxplot, and names
- Prototypes : lines, barplot, radar, 3d, poly.dist, umatrix, and smooth.disy
- Additional variable : pie, color, lines, barplot, radar, boxplot, words, names, and graph