BinaryClassification_SpamDetection

Spam Detection for Text Messages

ML.NET version	API type	Status	App Type	Data type	Scenario	ML Task	Algorithms
v0.9	Dynamic API	Up-to-date	Console app	.tsv files	Spam detection	Two-class classification	SDCA (linear learner)

In this sample, you'll see how to use ML.NET to predict whether a text message is spam. In the world of machine learning, this type of prediction is known as binary classification.

Problem

Our goal here is to predict whether a text message is spam (an irrelevant/unwanted message). We will use the SMS Spam Collection Data Set from UCI, which contains close to 6000 messages that have been classified as being "spam" or "ham" (not spam). We will use this dataset to train a model that can take in new message and predict whether they are spam or not.

This is an example of binary classification, as we are classifying the text messages into one of two categories.

Solution

To solve this problem, first we will build an estimator to define the ML pipeline we want to use. Then we will train this estimator on existing data, evaluate how good it is, and lastly we'll consume the model to predict whether a few examples messages are spam.

1. Build estimator

To build the estimator we will:

• Define how to read the spam dataset that will be downloaded from https://archive.ics.uci.edu/ml/datasets/SMS+Spam+Collection.

• Apply several data transformations:

  • Convert the label ("spam" or "ham") to a boolean ("true" represents spam) so we can use it with a binary classifier.
  • Featurize the text message into a numeric vector so a machine learning trainer can use it

• Add a trainer (such as StochasticDualCoordinateAscent).

The initial code is similar to the following:

// Set up the MLContext, which is a catalog of components in ML.NET.
let mlContext = MLContext(seed = Nullable 1)
let reader = 
    mlContext.Data.CreateTextReader(
        columns = 
            [|
                TextLoader.Column("LabelText" , Nullable DataKind.Text, 0)
                TextLoader.Column("Message" , Nullable DataKind.Text, 1)
            |],
            hasHeader = false,
            separatorChar = '\t')

let data = reader.Read(trainDataPath)

// Create the estimator which converts the text label to a key sorted by value (with ham < spam we get ham -> false and spam -> true)
// then featurizes the text, and add a linear trainer.
let estimator = 
    EstimatorChain()
        .Append(mlContext.Transforms.Conversion.MapValueToKey("LabelText", "Label", sort = ValueToKeyMappingTransformer.SortOrder.Value))
        .Append(mlContext.Transforms.Text.FeaturizeText("Message","Features"))
        .AppendCacheCheckpoint(mlContext)
        .Append(mlContext.BinaryClassification.Trainers.StochasticDualCoordinateAscent("Label", "Features"))

2. Evaluate model

For this dataset, we will use cross-validation to evaluate our model. This will partition the data into 5 'folds', train 5 models (on each combination of 4 folds), and test them on the fold that wasn't used in training.

let cvResults = mlContext.BinaryClassification.CrossValidate(data, downcastPipeline estimator, numFolds = 5);
let avgAuc = cvResults |> Seq.map (fun struct (metrics,_,_) -> metrics.Auc) |> Seq.average
printfn "The AUC is %f" avgAuc

Note that usually we evaluate a model after training it. However, cross-validation includes the model training part so we don't need to do Fit() first. However, we will later train the model on the full dataset to take advantage of the additional data.

3. Train model

To train the model we will call the estimator's Fit() method while providing the full training data.

let model = estimator.Fit(data)

4. Consume model

After the model is trained, you can use the Predict() API to predict whether new text is spam. In this case, we change the threshold of the model to get better predictions. We do this because our data is skewed with most messages not being spam.

// The dataset we have is skewed, as there are many more non-spam messages than spam messages.
// While our model is relatively good at detecting the difference, this skewness leads it to always
// say the message is not spam. We deal with this by lowering the threshold of the predictor. In reality,
// it is useful to look at the precision-recall curve to identify the best possible threshold.
let newModel = 
    let lastTransformer = 
        BinaryPredictionTransformer<IPredictorProducing<float32>>(
            mlContext, 
            model.LastTransformer.Model, 
            model.GetOutputSchema(data.Schema), 
            model.LastTransformer.FeatureColumn, 
            threshold = 0.15f, 
            thresholdColumn = DefaultColumnNames.Probability);
    let parts = model |> Seq.toArray
    parts.[parts.Length - 1] <- lastTransformer :> _
    TransformerChain<ITransformer>(parts)


// Create a PredictionFunction from our model 
let predictor = newModel.CreatePredictionEngine<SpamInput, SpamPrediction>(mlContext);

// Make a prediction
let prediction = p.Predict({LabelText = ""; Message = "That's a great idea. It should work."})
printfn "The message '%s' is %b" x prediction.PredictedLabel