16/ One registry of classifiers and feature selection methods (no change to results)

DESCRIPTION

@@ -69,6 +69,7 @@ Collate:
     'interfaceRandomForest.R'
     'interfaceRandomForestSurvival.R'
     'interfaceSVM.R'
+    'interfaceTemplate.R'
     'interfaceXGB.R'
     'performancePlot.R'
     'plotFeatureClasses.R'
@@ -90,12 +91,12 @@ Collate:
     'rankingPairsDifferences.R'
     'rankingPlot.R'
     'reexports.R'
+    'registry.R'
     'runTest.R'
     'runTests.R'
     'samplesMetricMap.R'
     'selectMulti.R'
     'selectionPlot.R'
-    'simpleParams.R'
     'subtractFromLocation.R'
     'utilities.R'
 URL: https://sydneybiox.github.io/ClassifyR/

R/available.R

@@ -6,6 +6,8 @@
 #' @aliases available
 #' @param what Default: \code{"classifier"}. Either \code{"classifier"}, \code{"selectionMethod"}
 #' or \code{"multiViewMethod"}.
+#' @return A data frame of keywords and their descriptions. For classifiers and feature selection methods, the column
+#' \code{Outcomes} says whether each one accepts classes, survival outcomes or both.
 
 #' @author Dario Strbenac
 #' @examples

R/constants.R

@@ -25,80 +25,7 @@ NULL
 
 .ClassifyRenvir[["performanceTypes"]] <- .ClassifyRenvir[["performanceInfoTable"]][, "type"]
 
-# Nice-looking names for feature selection and classification functions, to automatically use
-# in a variety of performance plots.
-.ClassifyRenvir[["functionsTable"]] <- matrix(
-  c("subtractFromLocation", "Subtraction From Training Set Location",
-    "bartlettRanking", "Bartlett Test",
-    "classifyInterface", "Poisson LDA",
-    "differentMeansRanking", "Difference in Means",
-    "DLDAtrainInterface", "Diagonal LDA",
-    "DMDranking", "Differences of Medians and Deviations",
-    "edgeRranking", "edgeR LRT",
-    "GLMtrainInterface", "Logistic Regression",
-    "penalisedGLMtrainInterface", "Penalised GLMs (Ridge, Elastic net, LASSO)",
-    "fisherDiscriminant", "Fisher's LDA",
-    "kNNinterface", "k Nearest Neighbours",
-    "KolmogorovSmirnovRanking", "Kolmogorov-Smirnov Test",
-    "kTSPclassifier", "k Top-Scoring Pairs",
-    "KullbackLeiblerRanking", "Kullback-Leibler Divergence",
-    "leveneRanking", "Levene Test",
-    "likelihoodRatioRanking", "Likelihood Ratio Test (Normal)",
-    "limmaRanking", "Moderated t-test",
-    "mixModelsTrain", "Mixtures of Normals",
-    "naiveBayesKernel", "Naive Bayes Kernel",
-    "NSCtrainInterface",  "Nearest Shrunken Centroids",
-    "pairsDifferencesRanking", "Pairs Differences",
-    "previousSelection", "Previous Selection", 
-    "previousTrained", "Previous Trained",
-    "randomSelection", "Random Selection",
-    "randomForestTrainInterface", "Random Forest",
-    "SVMtrainInterface", "Support Vector Machine",
-    "coxphTrainInterface", "Cox Proportional Hazards",
-    "coxphRanking", "Cox Proportional Hazards",
-    "coxnetTrainInterface", "Penalised Cox Proportional Hazards",
-    #"NEMOEtrainInterface", "Nutrition-Ecotype Mixture of Experts",
-    "rfsrcTrainInterface", "Random Survival Forest",
-    "extremeGradientBoostingTrainInterface", "Extreme Gradient Boosting"),
-  ncol = 2, byrow = TRUE, dimnames = list(NULL, c("character", "name"))
-) |> as.data.frame()
-
-.ClassifyRenvir[["selectKeywords"]] <- matrix(
-  c("none", "Skip selection procedure and use all input features.",
-    "t-test", "T-test.",
-    "limma", "Moderated t-test.",
-    "edgeR", "edgeR likelihood ratio test.",
-    "Bartlett", "Bartlett's test for different variance.",
-    "Levene", "Levene's test for different variance.",
-    "DMD", "Differences in means/medians and/or deviations.",
-    "likelihoodRatio", "Likelihood ratio test (normal distribution).",
-    "KS", "Kolmogorov-Smirnov test for differences in distributions.",
-    "KL", "Kullback-Leibler divergence between distributions.",
-    "CoxPH", "Cox proportional hazards Wald test per-feature.",
-    "randomSelection", "Randomly selects a specified number of features."
-    ),
-  ncol = 2, byrow = TRUE, dimnames = list(NULL, c("selectionMethod Keyword", "Description"))
-) |> as.data.frame()
-
-.ClassifyRenvir[["classifyKeywords"]] <- matrix(
-  c("randomForest", "Random forest.",
-    "DLDA", "Diagonal Linear Discriminant Analysis.",
-    "kNN", "k Nearest Neighbours.",
-    "GLM", "Logistic regression.",
-    "ridgeGLM", "Ridge GLM multinomial regression (alpha = 0).",
-    "elasticNetGLM", "Elastic net GLM multinomial regression (alpha = 0.5).",
-    "LASSOGLM", "LASSO GLM multinomial regression (alpha = 1).",  
-    "SVM", "Support Vector Machine.",
-    "NSC", "Nearest Shrunken Centroids.",
-    "naiveBayes", "Naive Bayes kernel feature voting classifier.",
-    "mixturesNormals", "Mixture of normals feature voting classifier.",
-    "CoxPH", "Cox proportional hazards.",
-    "CoxNet", "Penalised Cox proportional hazards.",
-    "randomSurvivalForest", "Random survival forest.",
-    "XGB", "Extreme gradient booster."
-    ),
-  ncol = 2, byrow = TRUE, dimnames = list(NULL, c("classifier Keyword", "Description"))
-) |> as.data.frame()
+# The classifiers and feature selection methods, their keywords and display names are in the registry (registry.R).
 
 .ClassifyRenvir[["multiViewKeywords"]] <- matrix(
   c("none", "Keep assays separate.",

R/crossValidate.R

@@ -159,6 +159,8 @@ setMethod("crossValidate", "DataFrame",
                 if(isCategorical) selectionMethod <- "t-test" else selectionMethod <- "CoxPH"
               if(length(classifier) == 1 && classifier == "auto")
                 if(isCategorical) classifier <- "randomForest" else classifier <- "CoxPH"
+              .checkOutcomeType(classifier, outcome, "classifiers")
+              .checkOutcomeType(selectionMethod, outcome, "selections")
 
 
               # Which data-types or data-views are present?
@@ -413,14 +415,16 @@ cleanClassifier <- function(classifier, measurements, nFeatures){
     if(!is.null(names(obsFeatures)) && all(names(obsFeatures) %in% names(classifier)) & is(classifier, "character")) classifier <- as.list(classifier[names(obsFeatures)])
 
     nFeatures <- nFeatures[names(classifier)]
-    checkENs <- which(classifier %in% c("ridgeGLM", "elasticNetGLM", "LASSOGLM"))
-    if(length(checkENs) > 0)
+    # A classifier needing more features than an assay has is replaced by its fallback classifier.
+    for(index in seq_along(classifier))
     {
-      replacements <- sapply(checkENs, function(checkEN) ifelse(any(nFeatures[[checkEN]] == 1), "GLM", classifier[[checkEN]]))
-      classifier[checkENs] <- replacements
-      if(any(replacements == "GLM"))
-        warning("Penalised GLM requires two or more features as input but there is only one.
-Using an ordinary GLM instead.")
+      entry <- if(length(classifier[[index]]) == 1) .ClassifyRenvir[["classifiers"]][[classifier[[index]]]]
+      if(!is.null(entry[["fewFeaturesClassifier"]]) && any(nFeatures[[index]] < entry[["minFeatures"]]))
+      {
+        warning(classifier[[index]], " requires ", entry[["minFeatures"]], " or more features as input but there are fewer.\n",
+                "Using ", entry[["fewFeaturesClassifier"]], " instead.")
+        classifier[[index]] <- entry[["fewFeaturesClassifier"]]
+      }
     }
     classifier
 }

R/interfaceTemplate.R

@@ -0,0 +1,33 @@
+################################################################################
+#
+# The contract of a classifier or feature selection interface. Copy this outline for a new method, then register it
+# in registry.R.
+#
+# Training: newTrainInterface(measurementsTrain, outcomeTrain, <parameters>, ..., verbose = 3)
+# - measurementsTrain is a DataFrame of the training samples, with features in columns. Categorical features are
+#   factors; methods needing a numeric matrix encode them with .encodeTrain and, at prediction, .encodeTest.
+# - outcomeTrain (named classesTrain by the classifiers of classes only) is a factor of classes or a Surv object, as
+#   the registry entry's outcomes allow. It is passed by position.
+# - <parameters> are the method's settings. Those in the entry's tunePresets are tuned when tuneParams is "auto".
+# - It returns the trained model.
+#
+# Prediction: newPredictInterface(model, measurementsTest, ..., returnType = c("both", "class", "score"), verbose = 3)
+# - Classes: "class" gives a factor with the training levels, "score" a samples x classes numeric matrix of class
+#   scores or probabilities with the class names as column names, and "both" a data frame of a column "class"
+#   followed by the score columns.
+# - Survival: a numeric vector of risk scores, where higher means a higher risk of the event.
+# - A classifier that trains and predicts in one function takes measurementsTest as its third argument and has no
+#   prediction interface (predict = NULL in its registry entry).
+#
+# Features used by a model (optional), for classifiers that rank or select features themselves: newFeatures(model)
+# returns a list of two vectors of column indices of measurementsTrain, the features ranked from best to worst and
+# the features the model uses. Other arguments, if any, are named after variables of runTest (for example,
+# measurementsTrain), which supplies them. It is the entry's getFeatures.
+#
+# Feature ranking: newRanking(measurementsTrain, outcomeTrain, ..., verbose = 3) returns the indices of the columns
+# of measurementsTrain, from the best feature to the worst.
+#
+# Each function carries its name as an attribute, which results and plots use to show its display name:
+#   attr(newTrainInterface, "name") <- "newTrainInterface"
+#
+################################################################################