A monad for classification workflows


In this document we describe the design and implementation of a (software programming) monad for classification workflows specification and execution. The design and implementation are done with Mathematica / Wolfram Language (WL).

The goal of the monad design is to make the specification of classification workflows (relatively) easy, straightforward, by following a certain main scenario and specifying variations over that scenario.

The monad is named ClCon and it is based on the State monad package "StateMonadCodeGenerator.m", [AAp1, AA1], the classifier ensembles package "ClassifierEnsembles.m", [AAp4, AA2], and the package for Receiver Operating Characteristic (ROC) functions calculation and plotting "ROCFunctions.m", [AAp5, AA2, Wk2].

The data for this document is read from WL’s repository using the package "GetMachineLearningDataset.m", [AAp10].

The monadic programming design is used as a Software Design Pattern. The ClCon monad can be also seen as a Domain Specific Language (DSL) for the specification and programming of machine learning classification workflows.

Here is an example of using the ClCon monad over the Titanic data:



The table above is produced with the package "MonadicTracing.m", [AAp2, AA1], and some of the explanations below also utilize that package.

As it was mentioned above the monad ClCon can be seen as a DSL. Because of this the monad pipelines made with ClCon are sometimes called "specifications".

Contents description

The document has the following structure.

  • The sections "Package load" and "Data load" obtain the needed code and data.
    (Needed and put upfront from the "Reproducible research" point of view.)

  • The sections "Design consideration" and "Monad design" provide motivation and design decisions rationale.

  • The sections "ClCon overview" and "Monad elements" provide technical description of the ClCon monad needed to utilize it.
    (Using a fair amount of examples.)

  • The section "Example use cases" gives several more elaborated examples of ClCon that have "real life" flavor.
    (But still didactic and concise enough.)

  • The section "Unit test" describes the tests used in the development of the ClCon monad.
    (The random pipelines unit tests are especially interesting.)

  • The section "Future plans" outlines future directions of development.
    (The most interesting and important one is the "conversational agent" direction.)

  • The section "Implementation notes" has (i) a diagram outlining the ClCon development process, and (ii) a list of observations and morals.
    (Some fairly obvious, but deemed fairly significant and hence stated explicitly.)

Remark: One can read only the sections "Introduction", "Design consideration", "Monad design", and "ClCon overview". That set of sections provide a fairly good, programming language agnostic exposition of the substance and novel ideas of this document.

Package load

The following commands load the packages [AAp1–AAp10, AAp12]:


Importing from GitHub: MathematicaForPredictionUtilities.m
Importing from GitHub: MosaicPlot.m
Importing from GitHub: CrossTabulate.m
Importing from GitHub: StateMonadCodeGenerator.m
Importing from GitHub: ClassifierEnsembles.m
Importing from GitHub: ROCFunctions.m
Importing from GitHub: VariableImportanceByClassifiers.m
Importing from GitHub: SSparseMatrix.m
Importing from GitHub: OutlierIdentifiers.m

Data load

In this section we load data that is used in the rest of the document. The "quick" data is created in order to specify quick, illustrative computations.

Remark: In all datasets the classification labels are in the last column.

The summarization of the data is done through ClCon, which in turn uses the function RecordsSummary from the package "MathematicaForPredictionUtilities.m", [AAp7].

WL resources data

The following commands produce datasets using the package [AAp10] (that utilizes ExampleData):

dsTitanic = GetMachineLearningDataset["Titanic"];
dsMushroom = GetMachineLearningDataset["Mushroom"];
dsWineQuality = GetMachineLearningDataset["WineQuality"];

Here is are the dimensions of the datasets:

Dataset[Dataset[Map[Prepend[Dimensions[ToExpression[#]], #] &, {"dsTitanic", "dsMushroom", "dsWineQuality"}]][All, AssociationThread[{"name", "rows", "columns"}, #] &]]


Here is the summary of dsTitanic:

ClConUnit[dsTitanic]⟹ClConSummarizeData["MaxTallies" -> 12];


Here is the summary of dsMushroom in long form:

ClConUnit[dsMushroom]⟹ClConSummarizeDataLongForm["MaxTallies" -> 12];


Here is the summary of dsWineQuality in long form:

ClConUnit[dsWineQuality]⟹ClConSummarizeDataLongForm["MaxTallies" -> 12];


"Quick" data

In this subsection we make up some data that is used for illustrative purposes.

dsData = RandomInteger[{0, 1000}, {100}];
dsData = Dataset[
   Transpose[{dsData, Mod[dsData, 3], Last@*IntegerDigits /@ dsData, ToString[Mod[#, 3]] & /@ dsData}]];
dsData = Dataset[dsData[All, AssociationThread[{"number", "feature1", "feature2", "label"}, #] &]];

(* {100, 4} *)

Here is a sample of the data:

RandomSample[dsData, 6]


Here is a summary of the data:



Here we convert the data into a list of record-label rules (and show the summary):

mlrData = ClConToNormalClassifierData[dsData];


Finally, we make the array version of the dataset:

arrData = Normal[dsData[All, Values]];

Design considerations

The steps of the main classification workflow addressed in this document follow.

  1. Retrieving data from a data repository.

  2. Optionally, transform the data.

  3. Split data into training and test parts.

    • Optionally, split training data into training and validation parts.
  4. Make a classifier with the training data.

  5. Test the classifier over the test data.

    • Computation of different measures including ROC.

The following diagram shows the steps.


Very often the workflow above is too simple in real situations. Often when making "real world" classifiers we have to experiment with different transformations, different classifier algorithms, and parameters for both transformations and classifiers. Examine the following mind-map that outlines the activities in making competition classifiers.


In view of the mind-map above we can come up with the following flow-chart that is an elaboration on the main, simple workflow flow-chart.


In order to address:

  • the introduction of new elements in classification workflows,

  • workflows elements variability, and

  • workflows iterative changes and refining,

it is beneficial to have a DSL for classification workflows. We choose to make such a DSL through a functional programming monad, [Wk1, AA1].

Here is a quote from [Wk1] that fairly well describes why we choose to make a classification workflow monad and hints on the desired properties of such a monad.

[…] The monad represents computations with a sequential structure: a monad defines what it means to chain operations together. This enables the programmer to build pipelines that process data in a series of steps (i.e. a series of actions applied to the data), in which each action is decorated with the additional processing rules provided by the monad. […]

Monads allow a programming style where programs are written by putting together highly composable parts, combining in flexible ways the possible actions that can work on a particular type of data. […]

Remark: Note that quote from [Wk1] refers to chained monadic operations as "pipelines". We use the terms "monad pipeline" and "pipeline" below.

Monad design

The monad we consider is designed to speed-up the programming of classification workflows outlined in the previous section. The monad is named ClCon for "Classification with Context".

We want to be able to construct monad pipelines of the general form:



ClCon is based on the State monad, [Wk1, AA1], so the monad pipeline form (1) has the following more specific form:



This means that some monad operations will not just change the pipeline value but they will also change the pipeline context.

In the monad pipelines of ClCon we store different objects in the contexts for at least one of the following two reasons.

  1. The object will be needed later on in the pipeline.

  2. The object is hard to compute.

Such objects are training data, ROC data, and classifiers.

Let us list the desired properties of the monad.

  • Rapid specification of non-trivial classification workflows.

  • The monad works with different data types: Dataset, lists of machine learning rules, full arrays.

  • The pipeline values can be of different types. Most monad functions modify the pipeline value; some modify the context; some just echo results.

  • The monad works with single classifier objects and with classifier ensembles.

    • This means support of different classifier measures and ROC plots for both single classifiers and classifier ensembles.
  • The monad allows of cursory examination and summarization of the data.
    • For insight and in order to verify assumptions.
  • The monad has operations to compute importance of variables.

  • We can easily obtain the pipeline value, context, and different context objects for manipulation outside of the monad.

  • We can calculate classification measures using a specified ROC parameter and a class label.

  • We can easily plot different combinations of ROC functions.

The ClCon components and their interaction are given in the following diagram. (The components correspond to the main workflow given in the previous section.)


In the diagram above the operations are given in rectangles. Data objects are given in round corner rectangles and classifier objects are given in round corner squares.

The main ClCon operations implicitly put in the context or utilize from the context the following objects:

  • training data,

  • test data,

  • validation data,

  • classifier (a classifier function or an association of classifier functions),

  • ROC data,

  • variable names list.

Note the that the monadic set of types of ClCon pipeline values is fairly heterogenous and certain awareness of "the current pipeline value" is assumed when composing ClCon pipelines.

Obviously, we can put in the context any object through the generic operations of the State monad of the package "StateMonadGenerator.m", [AAp1].

ClCon overview

When using a monad we lift certain data into the "monad space", using monad’s operations we navigate computations in that space, and at some point we take results from it.

With the approach taken in this document the "lifting" into the ClCon monad is done with the function ClConUnit. Results from the monad can be obtained with the functions ClConTakeValue, ClConContext, or with the other ClCon functions with the prefix "ClConTake" (see below.)

Here is a corresponding diagram of a generic computation with the ClCon monad:


Remark: It is a good idea to compare the diagram with formulas (1) and (2).

Let us examine a concrete ClCon pipeline that corresponds to the diagram above. In the following table each pipeline operation is combined together with a short explanation and the context keys after its execution.



Here is the output of the pipeline:



In the specified pipeline computation the last column of the dataset is assumed to be the one with the class labels.

The ClCon functions are separated into four groups:

  • operations,

  • setters,

  • takers,

  • State Monad generic functions.

An overview of the those functions is given in the tables in next two sub-sections. The next section, "Monad elements", gives details and examples for the usage of the ClCon operations.

Monad functions interaction with the pipeline value and context

The following table gives an overview the interaction of the ClCon monad functions with the pipeline value and context.



Several functions that use ROC data have two rows in the table because they calculate the needed ROC data if it is not available in the monad context.

State monad functions

Here are the ClCon State Monad functions (generated using the prefix "ClCon", [AAp1, AA1]):



Monad elements

In this section we show that ClCon has all of the properties listed in the previous section.

The monad head

The monad head is ClCon. Anything wrapped in ClCon can serve as monad’s pipeline value. It is better though to use the constructor ClConUnit. (Which adheres to the definition in [Wk1].)

ClCon[{{1, "a"}, {2, "b"}}, <||>]⟹ClConSummarizeData;


Lifting data to the monad

The function lifting the data into the monad ClCon is ClConUnit.

The lifting to the monad marks the beginning of the monadic pipeline. It can be done with data or without data. Examples follow.





(See the sub-section "Setters and takers" for more details of setting and taking values in ClCon contexts.)

Currently the monad can deal with data in the following forms:

  • datasets,

  • matrices,

  • lists of example->label rules.

The ClCon monad also has the non-monadic function ClConToNormalClassifierData which can be used to convert datasets and matrices to lists of example->label rules. Here is an example:

Short[ClConToNormalClassifierData[dsData], 3]

 {{639, 0, 9} -> "0", {121, 1, 1} -> "1", {309, 0, 9} ->  "0", {648, 0, 8} -> "0", {995, 2, 5} -> "2", {127, 1, 7} -> "1", {908, 2, 8} -> "2", {564, 0, 4} -> "0", {380, 2, 0} -> "2", {860, 2, 0} -> "2",
 {464, 2, 4} -> "2", {449, 2, 9} -> "2", {522, 0, 2} -> "0", {288, 0, 8} -> "0", {51, 0, 1} -> "0", {108, 0, 8} -> "0", {76, 1, 6} -> "1", {706, 1, 6} -> "1", {765, 0, 5} -> "0", {195, 0, 5} -> "0"}

When the data lifted to the monad is a dataset or a matrix it is assumed that the last column has the class labels. WL makes it easy to rearrange columns in such a way the any column of dataset or a matrix to be the last.

Data splitting

The splitting is made with ClConSplitData, which takes up to two arguments and options. The first argument specifies the fraction of training data. The second argument — if given — specifies the fraction of the validation part of the training data. If the value of option Method is "LabelsProportional", then the splitting is done in correspondence of the class labels tallies. ("LabelsProportional" is the default value.) Data splitting demonstration examples follow.

Here are the dimensions of the dataset dsData:


(* {100, 4} *)

Here we split the data into 70% for training and 30% for testing and then we verify that the corresponding number of rows add to the number of rows of dsData:

val = ClConUnit[dsData]⟹ClConSplitData[0.7]⟹ClConTakeValue;
Map[Dimensions, val]
Total[First /@ %]

 <|"trainingData" -> {69, 4}, "testData" -> {31, 4}|>

Note that if Method is not "LabelsProportional" we get slightly different results.

val = ClConUnit[dsData]⟹ClConSplitData[0.7, Method -> "Random"]⟹ClConTakeValue;
Map[Dimensions, val]
Total[First /@ %]

  <|"trainingData" -> {70, 4}, "testData" -> {30, 4}|>

In the following code we split the data into 70% for training and 30% for testing, then the training data is further split into 90% for training and 10% for classifier training validation; then we verify that the number of rows add up.

val = ClConUnit[dsData]⟹ClConSplitData[0.7, 0.1]⟹ClConTakeValue;
Map[Dimensions, val]
Total[First /@ %]

 <|"trainingData" -> {61, 4}, "testData" -> {31, 4}, "validationData" -> {8, 4}|>

Classifier training

The monad ClCon supports both single classifiers obtained with Classify and classifier ensembles obtained with Classify and managed with the package "ClassifierEnsembles.m", [AAp4].

Single classifier training

With the following pipeline we take the Titanic data, split it into 75/25 % parts, train a Logistic Regression classifier, and finally take that classifier from the monad.

cf =

Here is information about the obtained classifier:

ClassifierInformation[cf, "TrainingTime"]

(* Quantity[3.84008, "Seconds"] *)

If we want to pass parameters to the classifier training we can use the Method option. Here we train a Random Forest classifier with 400 trees:

cf =
   ClConMakeClassifier[Method -> {"RandomForest", "TreeNumber" -> 400}]⟹

ClassifierInformation[cf, "TreeNumber"]

(* 400 *)

Classifier ensemble training

With the following pipeline we take the Titanic data, split it into 75/25 % parts, train a classifier ensemble of three Logistic Regression classifiers and two Nearest Neighbors classifiers using random sampling of 90% of the training data, and finally take that classifier ensemble from the monad.

ensemble =
   ClConMakeClassifier[{{"LogisticRegression", 0.9, 3}, {"NearestNeighbors", 0.9, 2}}]⟹

The classifier ensemble is simply an association with keys that are automatically assigned names and corresponding values that are classifiers.



Here are the training times of the classifiers in the obtained ensemble:

ClassifierInformation[#, "TrainingTime"] & /@ ensemble

 <|"LogisticRegression[1,0.9]" -> Quantity[3.47836, "Seconds"], 
   "LogisticRegression[2,0.9]" -> Quantity[3.47681, "Seconds"], 
   "LogisticRegression[3,0.9]" -> Quantity[3.4808, "Seconds"], 
   "NearestNeighbors[1,0.9]" -> Quantity[1.82454, "Seconds"], 
   "NearestNeighbors[2,0.9]" -> Quantity[1.83804, "Seconds"]|>

A more precise specification can be given using associations. The specification

<|"method" -> "LogisticRegression", "sampleFraction" -> 0.9, "numberOfClassifiers" -> 3, "samplingFunction" -> RandomChoice|>

says "make three Logistic Regression classifiers, for each taking 90% of the training data using the function RandomChoice."

Here is a pipeline specification equivalent to the pipeline specification above:

ensemble2 =
       <|"method" -> "LogisticRegression", 
         "sampleFraction" -> 0.9, 
         "numberOfClassifiers" -> 3, 
         "samplingFunction" -> RandomSample|>, 
       <|"method" -> "NearestNeighbors", 
         "sampleFraction" -> 0.9, 
         "numberOfClassifiers" -> 2, 
         "samplingFunction" -> RandomSample|>}]⟹



Classifier testing

Classifier testing is done with the testing data in the context.

Here is a pipeline that takes the Titanic data, splits it, and trains a classifier:

p =

Here is how we compute selected classifier measures:

 ClConClassifierMeasurements[{"Accuracy", "Precision", "Recall", "FalsePositiveRate"}]⟹

 <|"Accuracy" -> 0.792683, 
   "Precision" -> <|"died" -> 0.802691, "survived" -> 0.771429|>, 
   "Recall" -> <|"died" -> 0.881773, "survived" -> 0.648|>, 
   "FalsePositiveRate" -> <|"died" -> 0.352, "survived" -> 0.118227|>|>

(The measures are listed in the function page of ClassifierMeasurements.)

Here we show the confusion matrix plot:



Here is how we plot ROC curves by specifying the ROC parameter range and the image size:

p⟹ClConROCPlot["FPR", "TPR", "ROCRange" -> Range[0, 1, 0.1], ImageSize -> 200];


Remark: ClCon uses the package ROCFunctions.m, [AAp5], which implements all functions defined in [Wk2].

Here we plot ROC functions values (y-axis) over the ROC parameter (x-axis):

p⟹ClConROCListLinePlot[{"ACC", "TPR", "FPR", "SPC"}];


Note of the "ClConROC*Plot" functions automatically echo the plots. The plots are also made to be the pipeline value. Using the option specification "Echo"->False the automatic echoing of plots can be suppressed. With the option "ClassLabels" we can focus on specific class labels.

  ClConROCListLinePlot[{"ACC", "TPR", "FPR", "SPC"}, "Echo" -> False, "ClassLabels" -> "survived", ImageSize -> Medium]⟹


Variable importance finding

Using the pipeline constructed above let us find the most decisive variables using systematic random shuffling (as explained in [AA3]):


 <|None -> 0.792683, "id" -> 0.664634, "passengerClass" -> 0.75, "passengerAge" -> 0.777439, "passengerSex" -> 0.612805|>

We deduce that "passengerSex" is the most decisive variable because its corresponding classification success rate is the smallest. (See [AA3] for more details.)

Using the option "ClassLabels" we can focus on specific class labels:

p⟹ClConAccuracyByVariableShuffling["ClassLabels" -> "survived"]⟹ClConTakeValue

 <|None -> {0.771429}, "id" -> {0.595506}, "passengerClass" -> {0.731959}, "passengerAge" -> {0.71028}, "passengerSex" -> {0.414414}|>

Setters and takers

The values from the monad context can be set or obtained with the corresponding "setters" and "takers" functions as summarized in previous section.

For example:


(* ClassifierFunction[__] *) 


  {<|"id" -> 858, "passengerClass" -> "3rd", "passengerAge" -> 30, "passengerSex" -> "male", "passengerSurvival" -> "survived"|>, <<979>> }


(* {<|"id" -> 285, "passengerClass" -> "1st", "passengerAge" -> 60, "passengerSex" -> "female", "passengerSurvival" -> "survived"|> , <<327>> } 


(* {"id", "passengerClass", "passengerAge", "passengerSex", "passengerSurvival"} *)

If other values are put in the context they can be obtained through the (generic) function ClConTakeContext, [AAp1]:

p = ClConUnit[RandomReal[1, {2, 2}]]⟹ClConAddToContext["data"];


(* {{0.815836, 0.191562}, {0.396868, 0.284587}} *)

Another generic function from [AAp1] is ClConTakeValue (used many times above.)

Example use cases

Classification with MNIST data

Here we show an example of using ClCon with the reasonably large dataset of images MNIST, [YL1].

mnistData = ExampleData[{"MachineLearning", "MNIST"}, "Data"];

p =
  ClConUnit[RandomSample[mnistData, 20000]]⟹
   ClConClassifierMeasurements[{"Accuracy", "ConfusionMatrixPlot"}]⟹


Here we plot the ROC curve for a specified digit:

p⟹ClConROCPlot["ClassLabels" -> 5];

Conditional continuation

In this sub-section we show how the computations in a ClCon pipeline can be stopped or continued based on a certain condition.

The pipeline below makes a simple classifier ("LogisticRegression") for the WineQuality data, and if the recall for the important label ("high") is not large enough makes a more complicated classifier ("RandomForest"). The pipeline marks intermediate steps by echoing outcomes and messages.

res =
  ClConUnit[dsWineQuality[All, Join[#, <|"wineQuality" -> If[#wineQuality >= 7, "high", "low"]|>] &]]⟹
   ClConSplitData[0.75, 0.2]⟹
   ClConSummarizeData(* summarize the data *)⟹
   ClConMakeClassifier[Method -> "LogisticRegression"](* training a simple classifier *)⟹
   ClConROCPlot["FPR", "TPR", "ROCPointCallouts" -> False]⟹
   ClConClassifierMeasurements[{"Accuracy", "Precision", "Recall", "FalsePositiveRate"}]⟹
   ClConIfElse[#["Recall", "high"] > 0.70 & (* criteria based on the recall for "high" *),
    ClConEcho["Good recall for \"high\"!", "Success:"],
      ClConEcho[Style["Recall for \"high\" not good enough... making a large random forest.", Darker[Red]], "Info:"]⟹
      ClConMakeClassifier[Method -> {"RandomForest", "TreeNumber" -> 400}](* training a complicated classifier *)⟹
      ClConROCPlot["FPR", "TPR", "ROCPointCallouts" -> False]⟹
      ClConClassifierMeasurements[{"Accuracy", "Precision", "Recall", "FalsePositiveRate"}]⟹
      ClConEchoValue &];


We can see that the recall with the more complicated is classifier is higher. Also the ROC plots of the second classifier are visibly closer to the ideal one. Still, the recall is not good enough, we have to find a threshold that is better that the default one. (See the next sub-section.)

Classification with custom thresholds

(In this sub-section we use the monad from the previous sub-section.)

Here we compute classification measures using the threshold 0.3 for the important class label ("high"):

 ClConClassifierMeasurementsByThreshold[{"Accuracy", "Precision", "Recall", "FalsePositiveRate"}, "high" -> 0.3]⟹

(* <|"Accuracy" -> 0.782857,  "Precision" -> <|"high" -> 0.498871, "low" -> 0.943734|>, 
     "Recall" -> <|"high" -> 0.833962, "low" -> 0.76875|>, 
     "FalsePositiveRate" -> <|"high" -> 0.23125, "low" -> 0.166038|>|> *)

We can see that the recall for "high" is fairly large and the rest of the measures have satisfactory values. (The accuracy did not drop that much, and the false positive rate is not that large.)

Here we compute suggestions for the best thresholds:

res (* start with a previous monad *)⟹
  ClConROCPlot[ImageSize -> 300] (* make ROC plots *)⟹
  ClConSuggestROCThresholds[3] (* find the best 3 thresholds per class label *)⟹
  ClConEchoValue (* echo the result *);


The suggestions are the ROC points that closest to the point {0, 1} (which corresponds to the ideal classifier.)

Here is a way to use threshold suggestions within the monad pipeline:

    ClConClassifierMeasurementsByThreshold[{"Accuracy", "Precision", "Recall"}, "high" -> First[#1["high"]]] &)⟹

value: <|high->{0.35},low->{0.65}|>
value: <|Accuracy->0.825306,Precision-><|high->0.571831,low->0.928736|>,Recall-><|high->0.766038,low->0.841667|>|> 

Unit tests

The development of ClCon was done with two types of unit tests: (1) directly specified tests, [AAp11], and (2) tests based on randomly generated pipelines, [AAp12].

Both unit test packages should be further extended in order to provide better coverage of the functionalities and illustrate — and postulate — pipeline behavior.

Directly specified tests

Here we run the unit tests file "MonadicContextualClassification-Unit-Tests.wlt", [AAp11]:

 testObject = TestReport["~/MathematicaForPrediction/UnitTests/MonadicContextualClassification-Unit-Tests.wlt"]


The natural language derived test ID’s should give a fairly good idea of the functionalities covered in [AAp11].

Values[Map[#["TestID"] &, testObject["TestResults"]]]

(* {"LoadPackage", "EvenOddDataset", "EvenOddDataMLRules", \
"DataToContext-no-[]", "DataToContext-with-[]", \
"ClassifierMaking-with-Dataset-1", "ClassifierMaking-with-MLRules-1", \
"AccuracyByVariableShuffling-1", "ROCData-1", \
"ClassifierEnsemble-different-methods-1", \
"ClassifierEnsemble-different-methods-2-cont", \
"ClassifierEnsemble-different-methods-3-cont", \
"ClassifierEnsemble-one-method-1", "ClassifierEnsemble-one-method-2", \
"ClassifierEnsemble-one-method-3-cont", \
"ClassifierEnsemble-one-method-4-cont", "AssignVariableNames-1", \
"AssignVariableNames-2", "AssignVariableNames-3", "SplitData-1", \
"Set-and-take-training-data", "Set-and-take-test-data", \
"Set-and-take-validation-data", "Partial-data-summaries-1", \
"Assign-variable-names-1", "Split-data-100-pct", \
"MakeClassifier-with-empty-unit-1", \
"No-rocData-after-second-MakeClassifier-1"} *)

Random pipelines tests

Since the monad ClCon is a DSL it is natural to test it with a large number of randomly generated "sentences" of that DSL. For the ClCon DSL the sentences are ClCon pipelines. The package "MonadicContextualClassificationRandomPipelinesUnitTests.m", [AAp12], has functions for generation of ClCon random pipelines and running them as verification tests. A short example follows.

Generate pipelines:

pipelines = MakeClConRandomPipelines[300];

(* 300 *)

Here is sample of the generated pipelines:

Block[{DoubleLongRightArrow, pipelines = RandomSample[pipelines, 6]}, 
 pipelines = pipelines /. {_Dataset -> "ds", _?DataRulesForClassifyQ -> "mlrData"};
  Map[List@ToString[DoubleLongRightArrow @@ #, FormatType -> StandardForm] &, pipelines], 
  TableHeadings -> {"pipeline"}]


Here we run the pipelines as unit tests:

 res = TestRunClConPipelines[pipelines, "Echo" -> True];

(* {350.083, Null} *)

From the test report results we see that a dozen tests failed with messages, all of the rest passed.

rpTRObj = TestReport[res]


(The message failures, of course, have to be examined — some bugs were found in that way. Currently the actual test messages are expected.)

Future plans

Workflow operations


Better outliers finding and manipulation incorporation in ClCon. Currently only outlier finding is surfaced in [AAp3]. (The package internally has other related functions.)

ClConUnit[dsTitanic[Select[#passengerSex == "female" &]]]⟹

(* {4, 17, 21, 22, 25, 29, 38, 39, 41, 59} *)

Dimension reduction

Support of dimension reduction application — quick construction of pipelines that allow the applying different dimension reduction methods.

Currently with ClCon dimension reduction is applied only to data the non-label parts of which can be easily converted into numerical matrices.

  ClConReduceDimension[2, "Echo" -> True]⟹
  ClConEchoFunctionValue["SVD dimensions:", Dimensions /@ # &]⟹


Conversational agent

Using the packages [AAp13, AAp15] we can generate ClCon pipelines with natural commands. The plan is to develop and document those functionalities further.

Implementation notes

The ClCon package, MonadicContextualClassification.m, [AAp3], is based on the packages [AAp1, AAp4-AAp9]. It was developed using Mathematica and the Mathematica plug-in for IntelliJ IDEA, by Patrick Scheibe , [PS1]. The following diagram shows the development workflow.


Some observations and morals follow.

  • Making the unit tests [AAp11] made the final implementation stage much more comfortable.
    • Of course, in retrospect that is obvious.
  • Initially "MonadicContextualClassification.m" was not real a package, just a collection of global context functions with the prefix "ClCon". This made some programming design decisions harder, slower, and more cumbersome. By making a proper package the development became much easier because of the "peace of mind" brought by the context feature encapsulation.
  • The making of random pipeline tests, [AAp12], helped catch a fair amount of inconvenient "features" and bugs.
    • (Both tests sets [AAp11, AAp12] can be made to be more comprehensive.)
  • The design of a conversational agent for producing ClCon pipelines with natural language commands brought a very fruitful viewpoint on the overall functionalities and the determination and limits of the ClCon development goals. See [AAp13, AAp14, AAp15].

  • "Eat your own dog food", or in this case: "use ClCon functionalities to implement ClCon functionalities."

    • Since we are developing a DSL it is natural to use that DSL for its own advancement.

    • Again, in retrospect that is obvious. Also probably should be seen as a consequence of practicing a certain code refactoring discipline.

    • The reason to list that moral is that often it is somewhat "easier" to implement functionalities thinking locally, ad-hoc, forgetting or not reviewing other, already implemented functions.

  • In order come be better design and find inconsistencies: write many pipelines and discuss with co-workers.

    • This is obvious. I would like to mention that a somewhat good alternative to discussions is (i) writing this document and related ones and (ii) making, running, and examining of the random pipelines tests.



[AAp1] Anton Antonov, State monad code generator Mathematica package, (2017), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/MonadicProgramming/StateMonadCodeGenerator.m .

[AAp2] Anton Antonov, Monadic tracing Mathematica package, (2017), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/MonadicProgramming/MonadicTracing.m .

[AAp3] Anton Antonov, Monadic contextual classification Mathematica package, (2017), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/MonadicProgramming/MonadicContextualClassification.m .

[AAp4] Anton Antonov, Classifier ensembles functions Mathematica package, (2016), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/ClassifierEnsembles.m .

[AAp5] Anton Antonov, Receiver operating characteristic functions Mathematica package, (2016), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/ROCFunctions.m .

[AAp6] Anton Antonov, Variable importance determination by classifiers implementation in Mathematica,(2015), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/VariableImportanceByClassifiers.m .

[AAp7] Anton Antonov, MathematicaForPrediction utilities, (2014), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/MathematicaForPredictionUtilities.m .

[AAp8] Anton Antonov, Cross tabulation implementation in Mathematica, (2017), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/CrossTabulate.m .

[AAp9] Anton Antonov, SSparseMatrix Mathematica package, (2018), MathematicaForPrediction at GitHub.

[AAp10] Anton Antonov, Obtain and transform Mathematica machine learning data-sets, (2018), MathematicaVsR at GitHub.

[AAp11] Anton Antonov, Monadic contextual classification Mathematica unit tests, (2018), MathematicaVsR at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/UnitTests/MonadicContextualClassification-Unit-Tests.wlt .

[AAp12] Anton Antonov, Monadic contextual classification random pipelines Mathematica unit tests, (2018), MathematicaVsR at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/UnitTests/MonadicContextualClassificationRandomPipelinesUnitTests.m .

ConverationalAgents Packages

[AAp13] Anton Antonov, Classifier workflows grammar in EBNF, (2018), ConversationalAgents at GitHub, https://github.com/antononcube/ConversationalAgents.

[AAp14] Anton Antonov, Classifier workflows grammar Mathematica unit tests, (2018), ConversationalAgents at GitHub, https://github.com/antononcube/ConversationalAgents.

[AAp15] Anton Antonov, ClCon translator Mathematica package, (2018), ConversationalAgents at GitHub, https://github.com/antononcube/ConversationalAgents.

MathematicaForPrediction articles

[AA1] Anton Antonov, Monad code generation and extension, (2017), MathematicaForPrediction at GitHub, https://github.com/antononcube/MathematicaForPrediction.

[AA2] Anton Antonov, "ROC for classifier ensembles, bootstrapping, damaging, and interpolation", (2016), MathematicaForPrediction at WordPress. URL: https://mathematicaforprediction.wordpress.com/2016/10/15/roc-for-classifier-ensembles-bootstrapping-damaging-and-interpolation/ .

[AA3] Anton Antonov, "Importance of variables investigation guide", (2016), MathematicaForPrediction at GitHub. URL: https://github.com/antononcube/MathematicaForPrediction/blob/master/MarkdownDocuments/Importance-of-variables-investigation-guide.md .


[Wk1] Wikipedia entry, Monad, URL: https://en.wikipedia.org/wiki/Monad_(functional_programming) .

[Wk2] Wikipedia entry, Receiver operating characteristic, URL: https://en.wikipedia.org/wiki/Receiver_operating_characteristic .

[YL1] Yann LeCun et al., MNIST database site. URL: http://yann.lecun.com/exdb/mnist/ .

[PS1] Patrick Scheibe, Mathematica (Wolfram Language) support for IntelliJ IDEA, (2013-2018), Mathematica-IntelliJ-Plugin at GitHub. URL: https://github.com/halirutan/Mathematica-IntelliJ-Plugin .


Progressive Machine Learning Examples


In this MathematicaVsR at GitHub project we show how to do Progressive machine learning using two types of classifiers based on:

  • Tries with Frequencies, [AAp2, AAp3, AA1],

  • Sparse Matrix Recommender framework [AAp4, AA2].

Progressive learning is a type of Online machine learning. For more details see [Wk1]. The Progressive learning problem is defined as follows.

Problem definition:

  • Assume that the data is sequentially available.
    • Meaning, at a given time only part of the data is available, and after a certain time interval new data can be obtained.

    • In view of classification, it is assumed that at a given time not all class labels are presented in the data already obtained.

    • Let us call this a data stream.

  • Make a machine learning algorithm that updates its model continuously or sequentially in time over a given data stream.

    • Let us call such an algorithm a Progressive Learning Algorithm (PLA).

In comparison, the typical (classical) machine learning algorithms assume that representative training data is available and after training that data is no longer needed to make predictions. Progressive machine learning has more general assumptions about the data and its problem formulation is closer to how humans learn to classify objects.

Below we are shown the applications of two types of classifiers as PLA’s. One is based on Tries with Frequencies (TF), [AAp2, AAp3, AA1], the other on an Item-item Recommender (IIR) framework [AAp4, AA2].

Remark: Note that both TF and IIR come from tackling Unsupervised machine learning tasks, but here they are applied in the context of Supervised machine learning.

General workflow

The Mathematica and R notebooks follow the steps in the following flow chart.


For detailed explanations see any of the notebooks.

Project organization

Mathematica files

R files

Example runs

(For details see Progressive-machine-learning-examples.md.)

Using Tries with Frequencies

Here is an example run with Tries with Frequencies, [AAp2, AA1]:


Here are the obtained ROC curves:


We can see that with the Progressive learning process does improve its success rates in time.

Using an Item-item recommender system

Here is an example run with an Item-item recommender system, [AAp4, AA2]:


Here are the obtained ROC curves:




[AAp1] Anton Antonov, Obtain and transform Mathematica machine learning data-sets, GetMachineLearningDataset.m, (2018), MathematicaVsR at GitHub.

[AAp2] Anton Antonov, Java tries with frequencies Mathematica package, JavaTriesWithFrequencies.m, (2017), MathematicaForPrediction at GitHub.

[AAp3] Anton Antonov, Tries with frequencies R package, TriesWithFrequencies.R, (2014), MathematicaForPrediction at GitHub.

[AAp4] Anton Antonov, Sparse matrix recommender framework in Mathematica, SparseMatrixRecommenderFramework.m, (2014), MathematicaForPrediction at GitHub.


[Wk1] Wikipedia entry, Online machine learning.

[AA1] Anton Antonov, "Tries with frequencies in Java", (2017), MathematicaForPrediction at WordPress.

[AA2] Anton Antonov, "A Fast and Agile Item-Item Recommender: Design and Implementation", (2011), Wolfram Technology Conference 2011.

ROC for classifier ensembles, bootstrapping, damaging, and interpolation


The main goals of this document are:

i) to demonstrate how to create versions and combinations of classifiers utilizing different perspectives,

ii) to apply the Receiver Operating Characteristic (ROC) technique into evaluating the created classifiers (see [2,3]) and

iii) to illustrate the use of the Mathematica packages [5,6].

The concrete steps taken are the following:

  1. Obtain data: Mathematica built-in or external. Do some rudimentary analysis.

  2. Create an ensemble of classifiers and compare its performance to the individual classifiers in the ensemble.

  3. Produce classifier versions with from changed data in order to explore the effect of records outliers.

  4. Make a bootstrapping classifier ensemble and evaluate and compare its performance.

  5. Systematically diminish the training data and evaluate the results with ROC.

  6. Show how to do classifier interpolation utilizing ROC.

In the steps above we skip the necessary preliminary data analysis. For the datasets we use in this document that analysis has been done elsewhere. (See [,,,].) Nevertheless, since ROC is mostly used for binary classifiers we want to analyze the class labels distributions in the datasets in order to designate which class labels are "positive" and which are "negative."

ROC plots evaluation (in brief)

Assume we are given a binary classifier with the class labels P and N (for "positive" and "negative" respectively).

Consider the following measures True Positive Rate (TPR):

 TPR:= \frac {correctly \:  classified \:  positives}{total \:  positives}.

and False Positive Rate (FPR):

 FPR:= \frac {incorrectly \:  classified \:  negatives}{total \:  negatives}.

Assume that we can change the classifier results with a parameter \theta and produce a plot like this one:


For each parameter value \theta _{i} the point {TPR(\theta _{i}), FPR(\theta _{i})} is plotted; points corresponding to consecutive \theta _{i}‘s are connected with a line. We call the obtained curve the ROC curve for the classifier in consideration. The ROC curve resides in the ROC space as defined by the functions FPR and TPR corresponding respectively to the x-axis and the y-axis.

The ideal classifier would have its ROC curve comprised of a line connecting {0,0} to {0,1} and a line connecting {0,1} to {1,1}.

Given a classifier the ROC point closest to {0,1}, generally, would be considered to be the best point.

The wider perspective

This document started as being a part of a conference presentation about illustrating the cultural differences between Statistics and Machine learning (for Wolfram Technology Conference 2016). Its exposition become both deeper and wider than expected. Here are the alternative, original goals of the document:

i) to demonstrate how using ROC a researcher can explore classifiers performance without intimate knowledge of the classifiers` mechanisms, and

ii) to provide concrete examples of the typical investigation approaches employed by machine learning researchers.

To make those points clearer and more memorable we are going to assume that exposition is a result of the research actions of a certain protagonist with a suitably selected character.

A by-product of the exposition is that it illustrates the following lessons from machine learning practices. (See [1].)

  1. For a given classification task there often are multiple competing models.

  2. The outcomes of the good machine learning algorithms might be fairly complex. I.e. there are no simple interpretations when really good results are obtained.

  3. Having high dimensional data can be very useful.

In [1] these three points and discussed under the names "Rashomon", "Occam", and "Bellman". To quote:

Rashomon: the multiplicity of good models;
Occam: the conflict between simplicity and accuracy;
Bellman: dimensionality — curse or blessing."

The protagonist

Our protagonist is a "Simple Nuclear Physicist" (SNP) — someone who is accustomed to obtaining a lot of data that has to be analyzed and mined sometimes very deeply, rigorously, and from a lot of angles, for different hypotheses. SNP is fairly adept in programming and critical thinking, but he does not have or care about deep knowledge of statistics methods or machine learning algorithms. SNP is willing and capable to use software libraries that provide algorithms for statistical methods and machine learning.

SNP is capable of coming up with ROC if he is not aware of it already. ROC is very similar to the so called phase space diagrams physicists do.

Used packages

These commands load the used Mathematica packages [4,5,6]:


Data used

The Titanic dataset

These commands load the Titanic data (that is shipped with Mathematica).

data = ExampleData[{"MachineLearning", "Titanic"}, "TrainingData"];
columnNames = (Flatten@*List) @@ ExampleData[{"MachineLearning", "Titanic"}, "VariableDescriptions"];
data = ((Flatten@*List) @@@ data)[[All, {1, 2, 3, -1}]];
trainingData = DeleteCases[data, {___, _Missing, ___}];

(* {732, 4} *)

RecordsSummary[trainingData, columnNames]


data = ExampleData[{"MachineLearning", "Titanic"}, "TestData"];
data = ((Flatten@*List) @@@ data)[[All, {1, 2, 3, -1}]];
testData = DeleteCases[data, {___, _Missing, ___}];

(* {314, 4} *)

RecordsSummary[testData, columnNames]


nTrainingData = trainingData /. {"survived" -> 1, "died" -> 0, "1st" -> 0, "2nd" -> 1, "3rd" -> 2, "male" -> 0, "female" -> 1};

Classifier ensembles

This command makes a classifier ensemble of two built-in classifiers "NearestNeighbors" and "NeuralNetwork":

aCLs = EnsembleClassifier[{"NearestNeighbors", "NeuralNetwork"}, trainingData[[All, 1 ;; -2]] -> trainingData[[All, -1]]]


A classifier ensemble of the package [6] is simply an association mapping classifier IDs to classifier functions.

The first argument given to EnsembleClassifier can be Automatic:

aCLs = EnsembleClassifier[Automatic, trainingData[[All, 1 ;; -2]] -> trainingData[[All, -1]]];

With Automatic the following built-in classifiers are used:


(* {"NearestNeighbors", "NeuralNetwork", "LogisticRegression", "RandomForest", "SupportVectorMachine", "NaiveBayes"} *)

Classification with ensemble votes

Classification with the classifier ensemble can be done using the function EnsembleClassify. If the third argument of EnsembleClassify is "Votes" the result is the class label that appears the most in the ensemble results.

EnsembleClassify[aCLs, testData[[20, 1 ;; -2]], "Votes"]

(* "died" *)

The following commands clarify the voting done in the command above.

Map[#[testData[[20, 1 ;; -2]]] &, aCLs]

(* <|"NearestNeighbors" -> "died", "NeuralNetwork" -> "survived", "LogisticRegression" -> "survived", "RandomForest" -> "died", "SupportVectorMachine" -> "died", "NaiveBayes" -> "died"|> *)

(* {{"died", 4}, {"survived", 2}} *)

Classification with ensemble averaged probabilities

If the third argument of EnsembleClassify is "ProbabilitiesMean" the result is the class label that has the highest mean probability in the ensemble results.

EnsembleClassify[aCLs, testData[[20, 1 ;; -2]], "ProbabilitiesMean"]

(* "died" *)

The following commands clarify the probability averaging utilized in the command above.

Map[#[testData[[20, 1 ;; -2]], "Probabilities"] &, aCLs]

(* <|"NearestNeighbors" -> <|"died" -> 0.598464, "survived" -> 0.401536|>, "NeuralNetwork" -> <|"died" -> 0.469274, "survived" -> 0.530726|>, "LogisticRegression" -> <|"died" -> 0.445915, "survived" -> 0.554085|>, 
"RandomForest" -> <|"died" -> 0.652414, "survived" -> 0.347586|>, "SupportVectorMachine" -> <|"died" -> 0.929831, "survived" -> 0.0701691|>, "NaiveBayes" -> <|"died" -> 0.622061, "survived" -> 0.377939|>|> *)

(* <|"died" -> 0.61966, "survived" -> 0.38034|> *)

ROC for ensemble votes

The third argument of EnsembleClassifyByThreshold takes a rule of the form label->threshold; the fourth argument is eighter "Votes" or "ProbabiltiesMean".

The following code computes the ROC curve for a range of votes.

rocRange = Range[0, Length[aCLs] - 1, 1];
aROCs = Table[(
    cres = EnsembleClassifyByThreshold[aCLs, testData[[All, 1 ;; -2]], "survived" -> i, "Votes"]; ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres]), {i, rocRange}];
ROCPlot[rocRange, aROCs, "PlotJoined" -> Automatic, GridLines -> Automatic]


ROC for ensemble probabilities mean

If we want to compute ROC of a range of probability thresholds we EnsembleClassifyByThreshold with the fourth argument being "ProbabilitiesMean".

EnsembleClassifyByThreshold[aCLs, testData[[1 ;; 6, 1 ;; -2]], "survived" -> 0.2, "ProbabilitiesMean"]

(* {"survived", "survived", "survived", "survived", "survived", "survived"} *)

EnsembleClassifyByThreshold[aCLs, testData[[1 ;; 6, 1 ;; -2]], "survived" -> 0.6, "ProbabilitiesMean"]

(* {"survived", "died", "survived", "died", "died", "survived"} *)

The implementation of EnsembleClassifyByThreshold with "ProbabilitiesMean" relies on the ClassifierFunction signature:

ClassifierFunction[__][record_, "Probabilities"]

Here is the corresponding ROC plot:

rocRange = Range[0, 1, 0.025];
aROCs = Table[(
    cres = EnsembleClassifyByThreshold[aCLs, testData[[All, 1 ;; -2]], "survived" -> i, "ProbabilitiesMean"]; ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres]), {i, rocRange}];
rocEnGr = ROCPlot[rocRange, aROCs, "PlotJoined" -> Automatic, PlotLabel -> "Classifier ensemble", GridLines -> Automatic]


Comparison of the ensemble classifier with the standard classifiers

This plot compares the ROC curve of the ensemble classifier with the ROC curves of the classifiers that comprise the ensemble.

rocGRs = Table[
   aROCs1 = Table[(
      cres = ClassifyByThreshold[aCLs[[i]], testData[[All, 1 ;; -2]], "survived" -> th];
      ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres]), {th, rocRange}]; 
   ROCPlot[rocRange, aROCs1, PlotLabel -> Keys[aCLs][[i]], PlotRange -> {{0, 1.05}, {0.6, 1.01}}, "PlotJoined" -> Automatic, GridLines -> Automatic],
   {i, 1, Length[aCLs]}];

GraphicsGrid[ArrayReshape[Append[Prepend[rocGRs, rocEnGr], rocEnGr], {2, 4}, ""], Dividers -> All, FrameStyle -> GrayLevel[0.8], ImageSize -> 1200]


Let us plot all ROC curves from the graphics grid above into one plot. For that the single classifier ROC curves are made gray, and their threshold callouts removed. We can see that the classifier ensemble brings very good results for \theta = 0.175 and none of the single classifiers has a better point.

Show[Append[rocGRs /. {RGBColor[___] -> GrayLevel[0.8]} /. {Text[p_, ___] :> Null} /. ((PlotLabel -> _) :> (PlotLabel -> Null)), rocEnGr]]


Classifier ensembles by bootstrapping

There are several ways to produce ensemble classifiers using bootstrapping or jackknife resampling procedures.

First, we are going to make a bootstrapping classifier ensemble using one of the Classify methods. Then we are going to make a more complicated bootstrapping classifier with six methods of Classify.

Bootstrapping ensemble with a single classification method

First we select a classification method and make a classifier with it.

clMethod = "NearestNeighbors";
sCL = Classify[trainingData[[All, 1 ;; -2]] -> trainingData[[All, -1]], Method -> clMethod];

The following code makes a classifier ensemble of 12 classifier functions using resampled, slightly smaller (10%) versions of the original training data (with RandomChoice).

aBootStrapCLs = Association@Table[(
     inds = RandomChoice[Range[Length[trainingData]], Floor[0.9*Length[trainingData]]];
     ToString[i] -> Classify[trainingData[[inds, 1 ;; -2]] -> trainingData[[inds, -1]], Method -> clMethod]), {i, 12}];

Let us compare the ROC curves of the single classifier with the bootstrapping derived ensemble.

rocRange = Range[0.1, 0.9, 0.025];
 aSingleROCs = Table[(
    cres = ClassifyByThreshold[sCL, testData[[All, 1 ;; -2]], "survived" -> i]; ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres]), {i, rocRange}];
 aBootStrapROCs = Table[(
    cres = EnsembleClassifyByThreshold[aBootStrapCLs, testData[[All, 1 ;; -2]], "survived" -> i]; ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres]), {i, rocRange}];

(* {6.81521, Null} *)

   ROCPlot[rocRange, aSingleROCs, "ROCColor" -> Blue, "PlotJoined" -> Automatic, GridLines -> Automatic],
   ROCPlot[rocRange, aBootStrapROCs, "ROCColor" -> Red, "PlotJoined" -> Automatic]}],
 SwatchLegend @@ Transpose@{{Blue, Row[{"Single ", clMethod, " classifier"}]}, {Red, Row[{"Boostrapping ensemble of\n", Length[aBootStrapCLs], " ", clMethod, " classifiers"}]}}]


We can see that we get much better results with the bootstrapped ensemble.

Bootstrapping ensemble with multiple classifier methods

This code creates an classifier ensemble using the classifier methods corresponding to Automatic given as a first argument to EnsembleClassifier.

 aBootStrapLargeCLs = Association@Table[(
      inds = RandomChoice[Range[Length[trainingData]], Floor[0.9*Length[trainingData]]];
      ecls = EnsembleClassifier[Automatic, trainingData[[inds, 1 ;; -2]] -> trainingData[[inds, -1]]];
      AssociationThread[Map[# <> "-" <> ToString[i] &, Keys[ecls]] -> Values[ecls]]
     ), {i, 12}];

(* {27.7975, Null} *)

This code computes the ROC statistics with the obtained bootstrapping classifier ensemble:

 aBootStrapLargeROCs = Table[(
     cres = EnsembleClassifyByThreshold[aBootStrapLargeCLs, testData[[All, 1 ;; -2]], "survived" -> i]; ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres]), {i, rocRange}];

(* {45.1995, Null} *)

Let us plot the ROC curve of the bootstrapping classifier ensemble (in blue) and the single classifier ROC curves (in gray):

aBootStrapLargeGr = ROCPlot[rocRange, aBootStrapLargeROCs, "PlotJoined" -> Automatic];
Show[Append[rocGRs /. {RGBColor[___] -> GrayLevel[0.8]} /. {Text[p_, ___] :> Null} /. ((PlotLabel -> _) :> (PlotLabel -> Null)), aBootStrapLargeGr]]


Again we can see that the bootstrapping ensemble produced better ROC points than the single classifiers.

Damaging data

This section tries to explain why the bootstrapping with resampling to smaller sizes produces good results.

In short, the training data has outliers; if we remove small fractions of the training data we might get better results.

The procedure described in this section can be used in conjunction with the procedures described in the guide for importance of variables investigation [7].

Ordering function

Let us replace the categorical values with numerical in the training data. There are several ways to do it, here is a fairly straightforward one:

nTrainingData = trainingData /. {"survived" -> 1, "died" -> 0, "1st" -> 0, "2nd" -> 1, "3rd" -> 2, "male" -> 0, "female" -> 1};

Decreasing proportions of females

First, let us find all indices corresponding to records about females.

femaleInds = Flatten@Position[trainingData[[All, 3]], "female"];

The following code standardizes the training data corresponding to females, finds the mean record, computes distances from the mean record, and finally orders the female records indices according to their distances from the mean record.

t = Transpose@Map[Rescale@*Standardize, N@Transpose@nTrainingData[[femaleInds, 1 ;; 2]]];
m = Mean[t];
ds = Map[EuclideanDistance[#, m] &, t];
femaleInds = femaleInds[[Reverse@Ordering[ds]]];

The following plot shows the distances calculated above.

ListPlot[Sort@ds, PlotRange -> All, PlotTheme -> "Detailed"]


The following code removes from the training data the records corresponding to females according to the order computed above. The female records farthest from the mean female record are removed first.

 femaleFrRes = Association@
    Table[cl ->
        inds = Complement[Range[Length[trainingData]], Take[femaleInds, Ceiling[fr*Length[femaleInds]]]];
        cf = Classify[trainingData[[inds, 1 ;; -2]] -> trainingData[[inds, -1]], Method -> cl]; cfPredictedLabels = cf /@ testData[[All, 1 ;; -2]];
        {fr, ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cfPredictedLabels]}),
       {fr, 0, 0.8, 0.05}],
     {cl, {"NearestNeighbors", "NeuralNetwork", "LogisticRegression", "RandomForest", "SupportVectorMachine", "NaiveBayes"}}];

(* {203.001, Null} *)

The following graphics grid shows how the classification results are affected by the removing fractions of the female records from the training data. The results for none or small fractions of records removed are more blue.

   femaleAROCs = femaleFrRes[cl][[All, 2]];
   frRange = femaleFrRes[cl][[All, 1]]; ROCPlot[frRange, femaleAROCs, PlotRange -> {{0.0, 0.25}, {0.2, 0.8}}, PlotLabel -> cl, "ROCPointColorFunction" -> (Blend[{Blue, Red}, #3/Length[frRange]] &), ImageSize -> 300],
   {cl, Keys[femaleFrRes]}],
  {2, 3}], Dividers -> All]


We can see that removing the female records outliers has dramatic effect on the results by the classifiers "NearestNeighbors" and "NeuralNetwork". Not so much on "LogisticRegression" and "NaiveBayes".

Decreasing proportions of males

The code in this sub-section repeats the experiment described in the previous one males (instead of females).

maleInds = Flatten@Position[trainingData[[All, 3]], "male"];

t = Transpose@Map[Rescale@*Standardize, N@Transpose@nTrainingData[[maleInds, 1 ;; 2]]];
m = Mean[t];
ds = Map[EuclideanDistance[#, m] &, t];
maleInds = maleInds[[Reverse@Ordering[ds]]];

ListPlot[Sort@ds, PlotRange -> All, PlotTheme -> "Detailed"]


 maleFrRes = Association@
    Table[cl ->
        inds = Complement[Range[Length[trainingData]], Take[maleInds, Ceiling[fr*Length[maleInds]]]];
        cf = Classify[trainingData[[inds, 1 ;; -2]] -> trainingData[[inds, -1]], Method -> cl]; cfPredictedLabels = cf /@ testData[[All, 1 ;; -2]];
        {fr, ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cfPredictedLabels]}),
       {fr, 0, 0.8, 0.05}],
     {cl, {"NearestNeighbors", "NeuralNetwork", "LogisticRegression", "RandomForest", "SupportVectorMachine", "NaiveBayes"}}];

(* {179.219, Null} *)

   maleAROCs = maleFrRes[cl][[All, 2]];
   frRange = maleFrRes[cl][[All, 1]]; ROCPlot[frRange, maleAROCs, PlotRange -> {{0.0, 0.35}, {0.55, 0.82}}, PlotLabel -> cl, "ROCPointColorFunction" -> (Blend[{Blue, Red}, #3/Length[frRange]] &), ImageSize -> 300],
   {cl, Keys[maleFrRes]}],
  {2, 3}], Dividers -> All]


Classifier interpolation

Assume that we want a classifier that for a given representative set of n items (records) assigns the positive label to an exactly n_p of them. (Or very close to that number.)

If we have two classifiers, one returning more positive items than n_p, the other less than n_p, then we can use geometric computations in the ROC space in order to obtain parameters for a classifier interpolation that will bring positive items close to n_p; see [3]. Below is given Mathematica code with explanations of how that classifier interpolation is done.

Assume that by prior observations we know that for a given dataset of n items the positive class consists of \approx 0.09 n items. Assume that for a given unknown dataset of n items we want 0.2 n of the items to be classified as positive. We can write the equation:

 {FPR} * ((1-0.09) * n) + {TPR} * (0.09 * n) = 0.2 * n ,

which can be simplified to

 {FPR} * (1-0.09) + {TPR} * 0.09 = 0.2 .

The two classifiers

Consider the following two classifiers.

cf1 = Classify[trainingData[[All, 1 ;; -2]] -> trainingData[[All, -1]], Method -> "RandomForest"];
cfROC1 = ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cf1[testData[[All, 1 ;; -2]]]]
(* <|"TruePositive" -> 82, "FalsePositive" -> 22, "TrueNegative" -> 170, "FalseNegative" -> 40|> *)

cf2 = Classify[trainingData[[All, 1 ;; -2]] -> trainingData[[All, -1]], Method -> "LogisticRegression"];
cfROC2 = ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cf2[testData[[All, 1 ;; -2]]]]
(* <|"TruePositive" -> 89, "FalsePositive" -> 37, "TrueNegative" -> 155, "FalseNegative" -> 33|> *)

Geometric computations in the ROC space

Here are the ROC space points corresponding to the two classifiers, cf1 and cf2:

p1 = Through[ROCFunctions[{"FPR", "TPR"}][cfROC1]];
p2 = Through[ROCFunctions[{"FPR", "TPR"}][cfROC2]];

Here is the breakdown of frequencies of the class labels:

Tally[trainingData[[All, -1]]]
%[[All, 2]]/Length[trainingData] // N

(* {{"survived", 305}, {"died", 427}}
   {0.416667, 0.583333}) *)

We want to our classifier to produce 38% people to survive. Here we find two points of the corresponding constraint line (on which we ROC points of the desired classifiers should reside):

sol1 = Solve[{{x, y} \[Element] ImplicitRegion[{x (1 - 0.42) + y 0.42 == 0.38}, {x, y}], x == 0.1}, {x, y}][[1]]
sol2 = Solve[{{x, y} \[Element] ImplicitRegion[{x (1 - 0.42) + y 0.42 == 0.38}, {x, y}], x == 0.25}, {x, y}][[1]]

(* {x -> 0.1, y -> 0.766667}
   {x -> 0.25, y -> 0.559524} *)

Here using the points q1 and q2 of the constraint line we find the intersection point with the line connecting the ROC points of the classifiers:

{q1, q2} = {{x, y} /. sol1, {x, y} /. sol2};
sol = Solve[ {{x, y} \[Element] InfiniteLine[{q1, q2}] \[And] {x, y} \[Element] InfiniteLine[{p1, p2}]}, {x, y}];
q = {x, y} /. sol[[1]]

(* {0.149753, 0.69796} *)

Let us plot all geometric objects:

Graphics[{PointSize[0.015], Blue, Tooltip[Point[p1], "cf1"], Black, 
  Text["cf1", p1, {-1.5, 1}], Red, Tooltip[Point[p2], "cf2"], Black, 
  Text["cf2", p2, {1.5, -1}], Black, Point[q], Dashed, 
  InfiniteLine[{q1, q2}], Thin, InfiniteLine[{p1, p2}]}, 
 PlotRange -> {{0., 0.3}, {0.6, 0.8}}, 
 GridLines -> Automatic, Frame -> True]


Classifier interpolation

Next we find the ratio of the distance from the intersection point q to the cf1 ROC point and the distance between the ROC points of cf1 and cf2.

k = Norm[p1 - q]/Norm[p1 - p2]
(* 0.450169 *)

The classifier interpolation is made by a weighted random selection based on that ratio (using RandomChoice):

cres = MapThread[If, {RandomChoice[{1 - k, k} -> {True, False}, Length[testData]], cf1@testData[[All, 1 ;; -2]], cf2@testData[[All, 1 ;; -2]]}];
cfROC3 = ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres];
p3 = Through[ROCFunctions[{"FPR", "TPR"}][cfROC3]];
Graphics[{PointSize[0.015], Blue, Point[p1], Red, Point[p2], Black, Dashed, InfiniteLine[{q1, q2}], Green, Point[p3]}, 
 PlotRange -> {{0., 0.3}, {0.6, 0.8}}, 
 GridLines -> Automatic, Frame -> True]


We can run the process multiple times in order to convince ourselves that the interpolated classifier ROC point is very close to the constraint line most of the time.

p3s =
    cres = 
     MapThread[If, {RandomChoice[{1 - k, k} -> {True, False}, Length[testData]], cf1@testData[[All, 1 ;; -2]], cf2@testData[[All, 1 ;; -2]]}]; 
    cfROC3 = ToROCAssociation[{"survived", "died"}, testData[[All, -1]], cres];
    Through[ROCFunctions[{"FPR", "TPR"}][cfROC3]]), {1000}];

Show[{SmoothDensityHistogram[p3s, ColorFunction -> (Blend[{White, Green}, #] &), Mesh -> 3], 
  Graphics[{PointSize[0.015], Blue, Tooltip[Point[p1], "cf1"], Black, Text["cf1", p1, {-1.5, 1}], 
     Red, Tooltip[Point[p2], "cf2"], Black, Text["cf2", p2, {1.5, -1}], 
     Black, Dashed, InfiniteLine[{q1, q2}]}, GridLines -> Automatic]}, 
 PlotRange -> {{0., 0.3}, {0.6, 0.8}}, 
 GridLines -> Automatic, Axes -> True, 
 AspectRatio -> Automatic]



[1] Leo Breiman, Statistical Modeling: The Two Cultures, (2001), Statistical Science, Vol. 16, No. 3, 199[Dash]231.

[2] Wikipedia entry, Receiver operating characteristic. URL: http://en.wikipedia.org/wiki/Receiver_operating_characteristic .

[3] Tom Fawcett, An introduction to ROC analysis, (2006), Pattern Recognition Letters, 27, 861[Dash]874. (Link to PDF.)

[4] Anton Antonov, MathematicaForPrediction utilities, (2014), source code MathematicaForPrediction at GitHub, package MathematicaForPredictionUtilities.m.

[5] Anton Antonov, Receiver operating characteristic functions Mathematica package, (2016), source code MathematicaForPrediction at GitHub, package ROCFunctions.m.

[6] Anton Antonov, Classifier ensembles functions Mathematica package, (2016), source code MathematicaForPrediction at GitHub, package ClassifierEnsembles.m.

[7] Anton Antonov, "Importance of variables investigation guide", (2016), MathematicaForPrediction at GitHub, folder Documentation.

Basic example of using ROC with Linear regression


This post is for using the package [2] that provides Mathematica implementations of Receiver Operating Characteristic (ROC) functions calculation and plotting. The ROC framework is used for analysis and tuning of binary classifiers, [3]. (The classifiers are assumed to classify into a positive/true label or a negative/false label. )

The function ROCFuntions gives access to the individual ROC functions through string arguments. Those ROC functions are applied to special objects, called ROC Association objects.

Each ROC Association object is an Association that has the following four keys: "TruePositive", "FalsePositive", "TrueNegative", and "FalseNegative" .

Given two lists of actual and predicted labels a ROC Association object can be made with the function ToROCAssociation .

For more definitions and example of ROC terminology and functions see [3].

Why Linear regression

I was asked in this discussion why Linear regression and not, say, Logistic regression.

Here is my answer:

1. I am trying to do a minimal and quick to execute example — the code of the post is included in the package ROCFunctions.m.

2. I am aware that there are better alternatives of LinearModelFit, but I plan to discuss those in the MathematicaVsR project: “Regression with ROC”. (As the name hints, it is not just about Linear regression.)

3. Another point is that although the Linear regression is not a good method for this classification, nevertheless using ROC it can be made to give better results than, say, the built-in “NeuralNetwork” method. See the last section of “Linear regression with ROC.md”.

Minimal example

Note that here although we use both of the provided Titanic training and test data, the code is doing only training. The test data is used to find the best tuning parameter (threshold) through ROC analysis.

Get packages

These commands load the packages [1,2]:


Using Titanic data

Here is the summary of the Titanic data used below:

titanicData = (Flatten@*List) @@@ExampleData[{"MachineLearning", "Titanic"}, "Data"];
columnNames = (Flatten@*List) @@ExampleData[{"MachineLearning", "Titanic"}, "VariableDescriptions"];
RecordsSummary[titanicData, columnNames]


This variable dependence grid shows the relationships between the variables.

Magnify[#, 0.7] &@VariableDependenceGrid[titanicData, columnNames]


Get training and testing data

data = ExampleData[{"MachineLearning", "Titanic"}, "TrainingData"];
data = ((Flatten@*List) @@@ data)[[All, {1, 2, 3, -1}]];
trainingData = DeleteCases[data, {___, _Missing, ___}];

(* {732, 4} *)

data = ExampleData[{"MachineLearning", "Titanic"}, "TestData"];
data = ((Flatten@*List) @@@ data)[[All, {1, 2, 3, -1}]];
testData = DeleteCases[data, {___, _Missing, ___}];

(* {314, 4} *)

Replace categorical with numerical values

trainingData = trainingData /. {"survived" -> 1, "died" -> 0, "1st" -> 0, "2nd" -> 1, "3rd" -> 2, "male" -> 0, "female" -> 1};

testData = testData /. {"survived" -> 1, "died" -> 0, "1st" -> 1, "2nd" -> 2, "3rd" -> 3, "male" -> 0, "female" -> 1};

Do linear regression

lfm = LinearModelFit[{trainingData[[All, 1 ;; -2]], trainingData[[All, -1]]}]


Get the predicted values

modelValues = lfm @@@ testData[[All, 1 ;; -2]];

Histogram[modelValues, 20]




Obtain ROC associations over a set of parameter values

testLabels = testData[[All, -1]];

thRange = Range[0.1, 0.9, 0.025];
aROCs = Table[ToROCAssociation[{0, 1}, testLabels, Map[If[# > \[Theta], 1, 0] &, modelValues]], {\[Theta], thRange}];

Evaluate ROC functions for given ROC association

Through[ROCFunctions[{"PPV", "NPV", "TPR", "ACC", "SPC"}][aROCs[[3]]]]

(* {34/43, 19/37, 17/32, 197/314, 95/122} *)

Standard ROC plot

ROCPlot[thRange, aROCs, "PlotJoined" -> Automatic, "ROCPointCallouts" -> True, "ROCPointTooltips" -> True, GridLines -> Automatic]


Plot ROC functions wrt to parameter values

ListLinePlot[Map[Transpose[{thRange, #}] &, Transpose[Map[Through[ROCFunctions[{"PPV", "NPV", "TPR", "ACC", "SPC"}][#]] &, aROCs]]],
 Frame -> True, FrameLabel -> Map[Style[#, Larger] &, {"threshold, \[Theta]", "rate"}], PlotLegends -> Map[# <> ", " <> (ROCFunctions["FunctionInterpretations"][#]) &, {"PPV", "NPV", "TPR", "ACC", "SPC"}], GridLines -> Automatic]


Finding the intersection point of PPV and TPR

We want to find a point that provides balanced positive and negative labels success rates. One way to do this is to find the intersection point of the ROC functions PPV (positive predictive value) and TPR (true positive rate).

Examining the plot above we can come up with the initial condition for \(x\).

ppvFunc = Interpolation[Transpose@{thRange, ROCFunctions["PPV"] /@ aROCs}];
tprFunc = Interpolation[Transpose@{thRange, ROCFunctions["TPR"] /@ aROCs}];
FindRoot[ppvFunc[x] - tprFunc[x] == 0, {x, 0.2}]

(* {x -> 0.3} *)

Area under the ROC curve

The Area Under the ROC curve (AUROC) tells for a given range of the controlling parameter "what is the probability of the classifier to rank a randomly chosen positive instance higher than a randomly chosen negative instance, (assuming ‘positive’ ranks higher than ‘negative’)", [3,4]

Calculating AUROC is easy using the Trapezoidal quadrature formula:

 N@Total[Partition[Sort@Transpose[{ROCFunctions["FPR"] /@ aROCs, ROCFunctions["TPR"] /@ aROCs}], 2, 1] 
   /. {{x1_, y1_}, {x2_, y2_}} :> (x2 - x1) (y1 + (y2 - y1)/2)]

 (* 0.698685 *)

It is also implemented in [2]:


(* 0.698685 *)


[1] Anton Antonov, MathematicaForPrediction utilities, (2014), source code MathematicaForPrediction at GitHub, package MathematicaForPredictionUtilities.m.

[2] Anton Antonov, Receiver operating characteristic functions Mathematica package, (2016), source code MathematicaForPrediction at GitHub, package ROCFunctions.m .

[3] Wikipedia entry, Receiver operating characteristic. URL: http://en.wikipedia.org/wiki/Receiver_operating_characteristic .

[4] Tom Fawcett, An introduction to ROC analysis, (2006), Pattern Recognition Letters, 27, 861-874.