Machine Learning Functions in Expression Language

This pages describes functions and properties that are related to the machine learning functionality in the QPR ProcessAnalyzer expression language.

MLModel

MLModel properties	Description
Type	Returns the exact type of the MLModel.

DateTime functions	Parameters	Description
Train ()	Input data Expected outcomes Parameters	Trains given MLModel using given input data and expected outcomes. Parameters input data: Two dimensional array of data where: The first dimension (rows) specifies different data points. The second dimension (columns) specifies the feature values. expected outcomes/labels: An array of expected outcomes for each row in the input data. Must be in the same order as the rows in the input data. parameters: Additional parameters for the MLModel. Optional. NumberOfTrees - the number of trees in the random forest, default value is 10 SampleRatio - the proportion of samples used to train each of the trees in the decision forest, default value is 0.632 Returns the trained MLModel object itself.
Transform		Transforms given input data using the machine learning model thus generating predictions/classifications. Parameters: input data: Two dimensional array of data where: The first dimension (rows) specifies different data points. The second dimension (columns) specifies the feature values. Returns an array of predictions/classifications. Transformations for each row in the input data can be found at the same index of the returned array.

Examples

Example #1: Train a model using an event log and test its performance by replaying training data itself.

Def("GetOneHotColumnInformation", (
  Let("el", _),
  ToDictionary([
    "et": OrderByValue(el.EventTypes),
    "at": ToDictionary(ConcatTop(OrderByTop(el.CaseAttributes, Name).[_: Values]))
  ])
));

Def("GenerateOneHot", "cases", (
  Let("columnInformation", _),
  cases.(
    Let("cas", _),
    Flatten(
      [
        columnInformation.Get("et").(Let("et", _), If(Count(cas.EventsByType(et)) > 0, 1, 0)),
        (
          Let("atColumns", columnInformation.Get("at")),
          OrderByValue(atColumns.Keys).(
            Let("key", _),
            Let("values", atColumns.Get(key)),
            Let("caseValue", cas.Attribute(key)),
            values.(If(_ == caseValue, 1, 0))
          )
        )
      ]
    )
  )
));

Let("el", EventLogById(1));
Let("columnInformation", el.GetOneHotColumnInformation());
Let("allCases", el.Cases);
Let("allCasesOH", columnInformation.GenerateOneHot(el.Cases));
Let("trainDataOH", allCasesOH);
Let("outcomes", allCases.(Duration > TimeSpan(24)));
Let("testDataOH", allCasesOH);
Let("predictions", 
  MLModel("randomforest")
    .Train(trainDataOH, outcomes)
    .Transform(trainDataOH));
Sum(Zip(outcomes, predictions).(_[0] == _[1] != 0)) / Count(outcomes)


Example #2: Train a model using an a 75% sample of an event log and test its performance by using the rest 25% of the event log.

Def("GetOneHotColumnInformation", (
  Let("el", _),
  ToDictionary([
    "et": OrderByValue(el.EventTypes),
    "at": ToDictionary(ConcatTop(OrderByTop(el.CaseAttributes, Name).[_: Values]))
  ])
));

Def("GenerateOneHot", "cases", (
  Let("columnInformation", _),
  cases.(
    Let("cas", _),
    Flatten(
      [
        columnInformation.Get("et").(Let("et", _), If(Count(cas.EventsByType(et)) > 0, 1, 0)),
        (
          Let("atColumns", columnInformation.Get("at")),
          OrderByValue(atColumns.Keys).(
            Let("key", _),
            Let("values", atColumns.Get(key)),
            Let("caseValue", cas.Attribute(key)),
            values.(If(_ == caseValue, 1, 0))
          )
        )
      ]
    )
  )
));

Let("el", EventLogById(1));
Let("columnInformation", el.GetOneHotColumnInformation());
Let("allCases", Shuffle(el.Cases));
Let("lastTrainCaseIndex", 0.75 * CountTop(el.Cases));
Let("trainCases", allCases[NumberRange(0, lastTrainCaseIndex)]);
Let("testCases", allCases[NumberRange(lastTrainCaseIndex + 1, CountTop(el.Cases) - 1)]);
Let("trainDataOH", columnInformation.GenerateOneHot(trainCases));
Let("testDataOH", columnInformation.GenerateOneHot(testCases));
Let("trainOutcomes", trainCases.(Duration > TimeSpan(24)));
Let("testOutcomes", testCases.(Duration > TimeSpan(24)));
Let("predictions", 
  MLModel("randomforest")
    .Train(trainDataOH, trainOutcomes)
    .Transform(testDataOH));
Sum(Zip(testOutcomes, predictions).(_[0] == _[1] != 0)) / Count(testOutcomes)

Example #3: Three sets of cases: training cases, target cases (subset of training cases) and test cases (independent set of cases). Try to predict which cases in the test set will eventually end up becoming a case in target cases.

Def("GetOneHotColumnInformation", (
  Let("el", _),
  ToDictionary([
    "et": OrderByValue(el.EventTypes),
    "at": ToDictionary(ConcatTop(OrderByTop(el.CaseAttributes, Name).[_: Values]))
  ])
));

Def("GenerateOneHot", "cases", (
  Let("columnInformation", _),
  cases.(
    Let("cas", _),
    Flatten(
      [
        columnInformation.Get("et").(Let("et", _), If(Count(cas.EventsByType(et)) > 0, 1, 0)),
        (
          Let("atColumns", columnInformation.Get("at")),
          OrderByValue(atColumns.Keys).(
            Let("key", _),
            Let("values", atColumns.Get(key)),
            Let("caseValue", cas.Attribute(key)),
            values.(If(_ == caseValue, 1, 0))
          )
        )
      ]
    )
  )
));

Let("el", <event log to use>);
Let("trainCases", <cases to use for training>);
Let("targetCases", <cases representing the properties we want to try to predict (subset of traincases)>);
Let("testCases", <cases to use for testing>);
Let("targetCasesDict", ToDictionary(targetCases:true));
Let("outcomes", traincases.(Let("c", _), targetCasesDict.ContainsKey(c) ? 1 : 0));
Let("columnInformation", el.GetOneHotColumnInformation());

Let("mlModel", MLModel("randomforest"));
mlModel.Train(columnInformation.GenerateOneHot(trainCases), outcomes);
mlModel.Transform(columnInformation.GenerateOneHot(testCases));