Binary classification: Spam Detection for Text Messages

#load "../../FSharpML.fsx"


open System;
open Microsoft.ML
open Microsoft.ML.Data;
open FSharpML
open FSharpML.EstimatorModel
open FSharpML.TransformerModel


/// Type representing the Message to run analysis on.
[<CLIMutable>] 
type SpamInput = 
    { 
        [<LoadColumn(0)>] LabelText : string
        [<LoadColumn(1)>] Message : string 
    }




//Create the MLContext to share across components for deterministic results
let mlContext = MLContext(seed = Nullable 1) // Seed set to any number so you
                                             // have a deterministic environment

// STEP 1: Common data loading configuration   
let fullData = 
    __SOURCE_DIRECTORY__  + "./data/SMSSpamCollection.txt"
    |> DataModel.fromTextFileWith<SpamInput> mlContext '\t' false 

let trainingData, testingData = 
    fullData
    |> DataModel.trainTestSplit 0.1 


//STEP 2: Process data, create and train the model 
let model = 
    EstimatorModel.create mlContext
    // Process data transformations in pipeline
    |> EstimatorModel.appendBy (fun mlc -> mlc.Transforms.Conversion.MapValue(DefaultColumnNames.Label, dict ["ham", false; "spam", true], "LabelText") )
    |> EstimatorModel.appendBy (fun mlc -> mlc.Transforms.Text.FeaturizeText(DefaultColumnNames.Features, "Message"))
    |> EstimatorModel.appendCacheCheckpoint
    // Create the model
    |> EstimatorModel.appendBy (fun mlc -> mlc.BinaryClassification.Trainers.SdcaLogisticRegression(DefaultColumnNames.Label, DefaultColumnNames.Features))
    // Train the model
    |> EstimatorModel.fit trainingData.Dataview

// STEP3: Run the prediciton on the test data
let predictions =
    model
    |> TransformerModel.transform testingData.Dataview

// STEP4: Evaluate accuracy of the model
let metrics = 
    model
    |> Evaluation.BinaryClassification.evaluate testingData.Dataview

metrics.Accuracy





//// STEP5: Create prediction engine function related to the loaded trained model
//let predict = 
//    TransformerModel.createPredictionEngine<_,SpamInput,SpamInput> model

//// Score
//let prediction = predict sampleStatement

//// Test a few examples
//[
//    "That's a great idea. It should work."
//    "free medicine winner! congratulations"
//    "Yes we should meet over the weekend!"
//    "you win pills and free entry vouchers"
//] 
//|> List.iter (classify predictor)