fdm2id 1.0.2

Residuals

• resplot (index = "fitted") plots the studentized residuals against the fitted values, the plot that was missing: no value of index gave it. index = 0, which plots them against the response, now warns – a least-squares residual is correlated with the response (sqrt (1 - R^2)) whatever the model, so that plot shows a trend that says nothing about the fit. The examples used it; they now use "fitted". New xlab argument, and a vector given as index names the axis.
• residuals(), fitted() and rstudent() apply to a model built by the package, without going through model$model.

Bugs

• LR() learnt from a one-column data frame can now predict: the variable kept its own name when learning, and was renamed X when predicting.
• treeplot(): the rectangles take the colour of their cluster in plotclus() (its cutree() number + 1); they were coloured from left to right of the tree, so the dendrogram and the scatter plot of a clustering did not match.
• evaluation.kappa() returned NA beyond 46 340 predictions (n * n overflowed as an integer), as repeated bootstraps easily reach.
• plotclus() and scatterplot() placed their legend by the first two columns of the data rather than by the plotted coordinates – the PCA plane, beyond two variables – and so often on the points.
• DBSCAN (epsilonDist), EM (clusters) and params in SVR(), SVRl(), SVRr() and MLPREG(), renamed in 1.0.0, were swallowed by ... without a word: an old script ran with the defaults and gave another result. They are now an error that names the new argument.

Graphics

• Legends are placed where they hide the least of the plot: of the eight positions legend() offers along the edges, the one covering the fewest points – or the fewest stretches of curve, bar or box – and, at equal count, the one furthest from them. legendpos = "auto" is the new default of plotdata(), plot() on a factorial analysis, a CDA or a feature selection, and boxclus(); scatterplot (legend = "auto") replaces "auto1" and "auto2", still accepted. An explicit position is honoured. The legends that had a fixed place – ROC and cost curves, kmeans.getk (graph = TRUE), the PCR/PLS and ridge/lasso tuning curves, plotzipf(), the scree plot – are placed the same way.
• New plot() method for HCA(): the dendrogram, the branches of each cluster in the colour plotclus() gives it, those above the cut dashed.
• treeplot (highlight = "branches") colours the branches of the dendrogram below the cut with the colour of their cluster – the colours of plotclus(), so that the dendrogram and the scatter plot of the same clustering match – and dashes the ones above it, all in the current line width (par (lwd = 2) for thick lines). ylab is now an argument, and the other arguments are passed to plot() (axes = FALSE).
• plot() on a SOM(), type = "mapping": the cells are coloured with a quarter of the colour of their cluster instead of two thirds, so that the points, in the full colour of the same cluster, stand out.
• The legends drawn by the package follow par ("cex.lab"), as the axis labels do: unchanged by default, and enlarged with them (par (cex.lab = 1.5)).

Changes

• selectfeatures() and FEATURESELECTION(): the default multivariate criterion is CFS rather than mRMR. With algorithm = "ranking", the criterion scores the nested subsets of the best-ranked features; mRMR, fstat and inertiaratio are averages, which only decrease as less relevant features join, so the default always kept a single feature.
• kmeans.getk() re-seeds before each number of clusters, so the partition behind each value is the one KMEANS (d, k, nstart, seed) returns, and the retained k the one the pseudo-F of those partitions gives.
• plotdata(): new argument scale (FALSE by default, as before), to project scaled variables with type = "pca" and "scatter".

Documentation

• LINREG (quali): the four values described.
• performance (methodparameters): the tuning done beforehand when it is missing, its cost, and how to pass pre-tuned parameters. BAGGING() and ADABOOST(): how to give the base learner through performance().

Packaging

• datasets is no longer imported in the NAMESPACE. The package code never used it – only the examples, the tests and one vignette do, and they attach it themselves – and R-devel now notes a base package that is imported while listed under Suggests.

fdm2id 1.0.1

• The vignettes are pre-computed: their code is run when the sources are prepared, by make-vignettes.R, rather than every time the package is built. Rebuilding them took between six and eleven minutes on the CRAN check machines – most of the check time, and more than its budget. It now costs a pandoc pass. The results they print are unchanged.
• Shorter introductions in the vignettes.

fdm2id 1.0.0

The package was reviewed with Claude Code. Some forty bugs were fixed as a result, and a test suite was added.

Breaking changes

Results that change: evaluation.precision(), evaluation.recall(), evaluation.kappa(), evaluation.adjr2(), compare.jaccard(), intern.dunn(), CDA(), MLPREG(), LINREG(), TSNE(), correlated(), predict() on SVM, performance (type = "roc" / "cost"), feature selection (CFS, Relief, mRMR), and every split or cross-validation, now stratified (stratify). performance() also re-seeds before each method it is given, so a method that searches a grid no longer scores differently depending on its position in the list.

Arguments:

• the final arguments of every learning method are now nfolds, tune, methodparameters, graph, seed, in that order
• HCA(): second argument
• DBSCAN (epsilonDist) is now eps; EM (clusters) is now k; params is now methodparameters in SVR(), SVRl(), SVRr() and MLPREG()
• new defaults: graph = FALSE everywhere, STUMP (randomvar), FEATURESELECTION (multieval), LINREG (regeval, nrep, validation), CART (xval), loadtext (dir), performance (fuzzy), CA (ncp), MCA (ncp)

Return values: compare (comp = "cluster"), LINREG (reg = "subset"), MLPREG (tune = TRUE), FEATURESELECTION (tune = TRUE).

Errors raised where a value was returned: misspelt comp, type, quali and reg, CDA(), LR(), intern.dunn(), performance().

Removed: HDBSCAN(), ISOFOREST(), UMAP(), the setClass() declarations.

Datasets: ozone has 13 variables instead of 10; movies is now simulated and rated from 1 to 10.

New features

• new methods: GBREG(), PAM(), penalized logistic regression (LR (reg = ))
• predict() on factorial analyses and on every clustering method
• multiclass precision, recall and F-measure (average)
• performance() on a test set given directly, without resampling
• number of clusters: silhouette, gap statistic, elbow
• selectfeatures (unieval = "randomforest") and plot() on a selection
• plotdata (type = "correlation") and plotdata (target = )
• HCA (engine = ), automatic eps in DBSCAN() and number of clusters in HCA()
• seed in every learning method, k in every clustering method
• package objects print a summary
• legendpos in plot() on a CDA object
• faster hyperparameter search, SPECTRAL(), HCA() and feature selection
• eight vignettes, one per case study of the courses, listed by vignette (package = "fdm2id"); they print the same figures as the course handouts, and flag in comments every call whose result moves from one run to the next without a seed