Clustering data using k-means in ML.NET

2018-08-07

Microsoft recently released a preview of a machine learning framework for .NET developers---ML.NET.

I needed to perform a clustering analysis from existing data in one of my applications. This is a pretty common machine learning task, so I decided to document the basic approach in this article.

We'll use the well-worn iris data set from the UCI Machine Learning Repository to demonstrate how to perform a cluster analysis using ML.NET. The iris data set contains three fairly distinct clusters of different types of iris---with a few difficult to classify specimens. This will make it easy to check the output of the analysis.

If you wish to review/run the completed application, you can clone/download it from GitHub.

Data preparation

The ML.NET framework allows you to import your data directly as part of the analysis pipeline. However, I have a number of libraries I like to use to load and transform my data, so I used CsvHelper to load the data. This resulted in an IEnumerable of Iris objects.

public class Iris
{
    public double PetalLength { get; set; }
    public double PetalWidth { get; set; }
    public double SepalLength { get; set; }
    public double SepalWidth { get; set; }
    public string Type { get; set; }
}

public class Observation
{
    [VectorType(4)]
    public float[] Features;

    public static Observation Create(Iris iris)
    {
        return new Observation
        {
            Features = new[]
            {
                (float)iris.SepalLength,
                (float)iris.SepalWidth,
                (float)iris.PetalLength,
                (float)iris.SepalWidth
            }
        };
    }
}

IEnumerable observations = data.Select(Observation.Create).ToList();

public class ClusterPrediction
{
    public unit PredictedLabel;
    public float[] Score;
}

var pipeline = new LearningPipeline
{
    CollectionDataSource.Create(observations),
    new KMeansPlusPlusClusterer
    {
        K = 3,
        NormalizeFeatures = NormalizeOption.Yes,
        MaxIterations = 100
    }
};

PredictionModel model = pipeline.Train();

data.ToList().ForEach(x =>
{
    ClusterPrediction prediction = model.Predict(Observation.Create(x));

    Console.WriteLine($"Type {x.Type} was assigned to cluster {prediction.PredictedLabel}");
});