Skip to main content

Unsupervised AutoML

CSharpNumerics includes a full unsupervised clustering framework with the same philosophy as the supervised pipeline: easy API, pluggable algorithms, and transparent results.

All clustering models operate directly on Matrix primitives — no target vector needed.

➡️ Clustering Grid​

All algoritms are implemented directly on top of the library’s Matrix and Vector primitives.

Models can be combined with:

  • Scalers (e.g. StandardScaler)
  • Hyperparameter search grids
var clusteringGrid = new ClusteringGrid()
        .AddModel<KMeans>(g => g
            .Add("K", 2, 3, 4, 5, 6, 7, 8)
            .Add("InitMethod", KMeansInit.Random, KMeansInit.PlusPlus)
            .AddScaler<StandardScaler>(s => { }))
        .AddModel<DBSCAN>(g => g
            .Add("Epsilon", 0.3, 0.5, 0.8, 1.0, 1.5)
            .Add("MinPoints", 3, 5, 10)
            .AddScaler<MinMaxScaler>(s => { }))
        .AddModel<AgglomerativeClustering>(g => g
            .Add("K", 2, 3, 4, 5)
            .Add("Linkage", LinkageType.Ward, LinkageType.Complete)

➡️ ClusteringExperiment (Fluent API)​

Run a full experiment with grid search and evaluators in one call:

var experiment = ClusteringExperiment
    .For(X)
    .WithGrid(clusteringGrid)
    .WithEvaluators(
        new SilhouetteEvaluator(),
        new CalinskiHarabaszEvaluator(),
        new DaviesBouldinEvaluator())
    .Run();

🔀 Multi-Algorithm Comparison​

var experiment = ClusteringExperiment
    .For(X)
    .WithAlgorithms(new KMeans(), new AgglomerativeClustering())
    .TryClusterCounts(2, 8)
    .WithEvaluators(new SilhouetteEvaluator(), new CalinskiHarabaszEvaluator())
    .WithScaler(new StandardScaler())
    .Run();

var bestSilhouette = experiment.BestBy<SilhouetteEvaluator>();
var bestCH = experiment.BestBy<CalinskiHarabaszEvaluator>();

Key points:

  • TryClusterCounts(min, max) auto-expands K for algorithms that accept it
  • DBSCAN discovers K on its own — the range is ignored
  • First evaluator added is the primary used for ranking
  • BestBy<T>() retrieves the best result by a specific evaluator