Movatterモバイル変換

ciu 0.8

• Changed the default colours if influence/contrastive plots to thebrick red / blue used by LIME in R version.
• Corrected the function “ciu.contrastive” that was implemented in avery strange (and presumably incorrect) way.
• Added possibility to get line segment IO-plots for categoricalfeatures, rather than the earlier histogram visualisation. Addedparameter to ggplot.ciu() “categorical_style=”segment”“, where thedefault is now to use line segments, even though that breaks backwardscompatibility a little. However, for some reason setting”ylim” breaksthe old “geom_col” plot so the bar plots are now un-scaled. But, strangeenough, “+ ylim” applied to the returned ggplot works so it’s possibleto adjust afterwards. Something happened somewhere, no change in the CIUcode and it worked in the past.
• Added new function ciu.contrastive that uses CU values of oneclass/instance/whatever as the baseline “normal.CU” values of theother.
• Added new function ciu.contextual.influence that calculatescontextual influence value(s) from CI and CU value(s) and“baseline”.
• Added new file ContextualInfluence.R for functions related tocontextual influence.
• Fixed x-axis label error for influence plot inciu.ggplot.col.

ciu 0.6.0

• Extensive tests on Intermediate Concepts, which are also integratedinto the new README.Rmd file (and README) as well as in test scriptsTestCases.R and TestCases_NoObject.R on Github.
• Fixed inverted axis bug in plot inplot.ciu.3D.
• Integrated support for mlr3. Only tested with “classif.rpart” modelthough because other tested models (xgboost, nnet, …) don’t like R typesand no need/time to dive into mlr3 for the moment. To fix in the futurein case bugs are discovered.
• Added parameter “plot.mode” to “ciu.ggplot.col” function, whichallows for value “overlap” that plots CI as a bar and then CU as anotherbar over it, scaled into CI bar. Colour of CU bar can either be fixed(dark green by default) or set to NULL, which then uses thegreen-yellow-red color scale (unless those colours are changed by thefunction parameters).
• Removed parameter for min/max values of contextual influence becausethe range of contextual influence is (and should be) always one (1)based on the mathematical constructs.
• Added new method “ggplot.ciu” to ciu.new for plotting input/outputgraphs with ggplot. Identical to old plot.ciu, except that 1) ggplotoffers some advantages, notably what comes to figure scaling, 2) addedpossibility to include CIU visualisation (cmin, cmax, neutral) bysetting “illustrate.CIU=TRUE”.
• Added new method “influence” to ciu.new for getting numericalcontextual influence (rather than just seeing them in plots). TO-DO: addcorresponding function to “ciu.R”.
• Corrected axis inplot.ciu.3D.

ciu 0.5.0

• Textual explanations have been implemented with function“ciu.textual”.
• Implemented “meta.explain” method, which returns a self-contained“ciu.meta.result” object with CIU results for a given instance. Thismainly makes it possible to visualize exactly the same CIU result indifferent ways. Before this (and still if no “ciu.meta” parameter isgiven), every visualization method ran “explain” again, so CIU resultscould differ somewhat between different runs.
• Added parameters “use.influence” and “influence.minmax” to“ggplot.col” and “barplot.ciu” functions/methods, which produces aLIME/SHAP/etc-like plot where “influence = (influence.max -influence.min)*ci(cu-neutral.CU)“, bars go either right or left of zeroand there are only two colours.
• Added support for factor-type inputs to method “$plot.ciu”.
• Created new “class” called “ciu”, which is just a list object withthe “instance variables” of a CIU object. This makes it possible tocreate plot functions that are not methods of CIU but rather take a“ciu” object as their first parameter. The main reason for doing thismodification is that CIU objects seem to use much more memory than asimple data-based “ciu” object. The documentation and package tools in Rare also not too aware about “inner functions” (methods), which is aninconvenient. Finally, the plan was indeed to go for this approach inany case because increasing the code length of CIU was not desirable inthe long run. The way in which it is implemented now allows functionsand methods to be used interchangeably in any case.

ciu 0.1.0

• First version of ciu published at CRAN