/// <summary>
/// Iterative random selection of ensemble models.
/// </summary>
/// <param name="metric">Metric to minimize</param>
/// <param name="ensembleStrategy">Strategy for ensembling models</param>
/// <param name="numberOfModelsToSelect">Number of models to select</param>
/// <param name="iterations">Number of iterations to try random selection</param>
/// <param name="selectWithReplacement">If true the same model can be selected multiple times.
/// This will correspond to weighting the models. If false each model can only be selected once</param>
/// <param name="seed"></param>
public RandomClassificationEnsembleSelection(IMetric <double, ProbabilityPrediction> metric, IClassificationEnsembleStrategy ensembleStrategy,
int numberOfModelsToSelect, int iterations, bool selectWithReplacement, int seed = 42)
{
if (metric == null)
{
throw new ArgumentNullException("metric");
}
if (ensembleStrategy == null)
{
throw new ArgumentNullException("ensembleStrategy");
}
if (numberOfModelsToSelect < 1)
{
throw new ArgumentException("numberOfModelsToSelect must be at least 1");
}
if (iterations < 1)
{
throw new ArgumentException("Number of iterations");
}
m_metric = metric;
m_ensembleStrategy = ensembleStrategy;
m_numberOfModelsToSelect = numberOfModelsToSelect;
m_selectWithReplacement = selectWithReplacement;
m_iterations = iterations;
m_random = new Random(seed);
}
/// <summary>
/// Greedy forward selection of ensemble models.
/// </summary>
/// <param name="metric">Metric to minimize</param>
/// <param name="ensembleStrategy">Strategy for ensembling models</param>
/// <param name="numberOfModelsToSelect">Number of models to select</param>
/// <param name="numberOfModelsFromStart">Number of models from start of the search.
/// The top n models will be selected based in their solo performance</param>
/// <param name="selectWithReplacement">If true the same model can be selected multiple times.
/// This will correspond to weighting the models. If false each model can only be selected once</param>
public ForwardSearchClassificationEnsembleSelection(IMetric <double, ProbabilityPrediction> metric, IClassificationEnsembleStrategy ensembleStrategy,
int numberOfModelsToSelect, int numberOfModelsFromStart, bool selectWithReplacement)
{
if (metric == null)
{
throw new ArgumentNullException("metric");
}
if (ensembleStrategy == null)
{
throw new ArgumentNullException("ensembleStrategy");
}
if (numberOfModelsToSelect < 1)
{
throw new ArgumentException("numberOfModelsToSelect must be at least 1");
}
if (numberOfModelsFromStart < 1)
{
throw new ArgumentException("numberOfModelsFromStart must be at least 1");
}
if (numberOfModelsFromStart > numberOfModelsToSelect)
{
throw new ArgumentException("numberOfModelsFromStart must be smaller than numberOfModelsToSelect");
}
m_metric = metric;
m_ensembleStrategy = ensembleStrategy;
m_numberOfModelsToSelect = numberOfModelsToSelect;
m_numberOfModelsFromStart = numberOfModelsFromStart;
m_selectWithReplacement = selectWithReplacement;
}
/// <summary>
/// Classification model selecting EnsembleLearner.
/// Trains several models and selects the best subset of models for the ensemble.
/// The selection of the best set of models is based on cross validation.
/// http://www.cs.cornell.edu/~alexn/papers/shotgun.icml04.revised.rev2.pdf
/// </summary>
/// <param name="learners">Learners in the ensemble</param>
/// <param name="crossValidation">Cross validation method</param>
/// <param name="ensembleStrategy">Strategy on how to combine the models</param>
/// <param name="ensembleSelection">Ensemble selection method used to find the beset subset of models</param>
public ClassificationModelSelectingEnsembleLearner(
IIndexedLearner <ProbabilityPrediction>[] learners,
ICrossValidation <ProbabilityPrediction> crossValidation,
IClassificationEnsembleStrategy ensembleStrategy,
IClassificationEnsembleSelection ensembleSelection)
: this(learners, crossValidation, () => ensembleStrategy, ensembleSelection)
{
}
/// <summary>
/// Classification ensemble learner. Combines several models into a single ensemble model.
/// </summary>
/// <param name="learners">Learners in the ensemble</param>
/// <param name="ensembleStrategy">Strategy on how to combine the models</param>
/// <param name="subSampleRatio">Default is 1.0. All models are trained on all data.
/// If different from 1.0 models are trained using bagging with the chosen sub sample ratio</param>
/// <param name="seed">Seed for the bagging when used</param>
public ClassificationEnsembleLearner(
IIndexedLearner <ProbabilityPrediction>[] learners,
IClassificationEnsembleStrategy ensembleStrategy,
double subSampleRatio = 1.0,
int seed = 24)
: this(learners.Select(l => new Func <F64Matrix, double[], int[], IPredictorModel <ProbabilityPrediction> >((o, t, i) => l.Learn(o, t, i))).ToArray(),
() => ensembleStrategy, subSampleRatio, seed)
{
}
/// <summary>
/// Classification model selecting EnsembleLearner.
/// Trains several models and selects the best subset of models for the ensemble using greedy backward elimination.
/// The selection of the best set of models is based on cross validation.
/// http://www.cs.cornell.edu/~alexn/papers/shotgun.icml04.revised.rev2.pdf
/// </summary>
/// <param name="learners">Learners in the ensemble</param>
/// <param name="numberOfModelsToSelect">Number of models to select</param>
/// <param name="crossValidation">Cross validation method</param>
/// <param name="ensembleStrategy">Strategy for ensembling models</param>
/// <param name="metric">Metric to minimize</param>
public ClassificationBackwardEliminationModelSelectingEnsembleLearner(
IIndexedLearner <ProbabilityPrediction>[] learners,
int numberOfModelsToSelect,
ICrossValidation <ProbabilityPrediction> crossValidation,
IClassificationEnsembleStrategy ensembleStrategy,
IMetric <double, ProbabilityPrediction> metric)
: base(learners, crossValidation, ensembleStrategy,
new BackwardEliminationClassificationEnsembleSelection(
metric, ensembleStrategy, numberOfModelsToSelect))
{
}
/// <summary>
/// Classification ensemble model
/// </summary>
/// <param name="ensembleModels">Models included in the ensemble</param>
/// <param name="ensembleStrategy">Strategy on how to combine the models</param>
public ClassificationEnsembleModel(IPredictorModel <ProbabilityPrediction>[] ensembleModels, IClassificationEnsembleStrategy ensembleStrategy)
{
if (ensembleModels == null)
{
throw new ArgumentNullException("ensembleModels");
}
if (ensembleStrategy == null)
{
throw new ArgumentNullException("ensembleStrategy");
}
m_ensembleModels = ensembleModels;
m_ensembleStrategy = ensembleStrategy;
}
/// <summary>
/// Greedy backwards elimination of ensemble models.
/// </summary>
/// <param name="metric">Metric to minimize</param>
/// <param name="ensembleStrategy">Strategy for ensembling models</param>
/// <param name="numberOfModelsToSelect">Number of models to select</param>
public BackwardEliminationClassificationEnsembleSelection(
IMetric <double, ProbabilityPrediction> metric,
IClassificationEnsembleStrategy ensembleStrategy,
int numberOfModelsToSelect)
{
m_metric = metric ?? throw new ArgumentNullException(nameof(metric));
m_ensembleStrategy = ensembleStrategy ?? throw new ArgumentNullException(nameof(ensembleStrategy));
if (numberOfModelsToSelect < 1)
{
throw new ArgumentException("numberOfModelsToSelect must be at least 1");
}
m_numberOfModelsToSelect = numberOfModelsToSelect;
}
/// <summary>
/// Classification model selecting EnsembleLearner.
/// Trains several models and selects the best subset of models for the ensemble using greedy forward selection.
/// The selection of the best set of models is based on cross validation.
/// http://www.cs.cornell.edu/~alexn/papers/shotgun.icml04.revised.rev2.pdf
/// </summary>
/// <param name="learners">Learners in the ensemble</param>
/// <param name="numberOfModelsToSelect">Number of models to select</param>
/// <param name="crossValidation">Cross validation method</param>
/// <param name="ensembleStrategy">Strategy for ensembling models</param>
/// <param name="metric">Metric to minimize</param>
/// <param name="numberOfModelsFromStart">Number of models from start of the search.
/// The top n models will be selected based in their solo performance</param>
/// <param name="selectWithReplacement">If true the same model can be selected multiple times.
/// This will correspond to weighting the models. If false each model can only be selected once. Default is true</param>
public ClassificationForwardSearchModelSelectingEnsembleLearner(
IIndexedLearner <ProbabilityPrediction>[] learners,
int numberOfModelsToSelect,
ICrossValidation <ProbabilityPrediction> crossValidation,
IClassificationEnsembleStrategy ensembleStrategy,
IMetric <double, ProbabilityPrediction> metric,
int numberOfModelsFromStart = 1,
bool selectWithReplacement = true)
: base(learners, crossValidation, ensembleStrategy,
new ForwardSearchClassificationEnsembleSelection(
metric, ensembleStrategy, numberOfModelsToSelect,
numberOfModelsFromStart, selectWithReplacement))
{
}
/// <summary>
/// Classification model selecting EnsembleLearner.
/// Trains several models and selects the best subset of models for the ensemble using iterative random selection.
/// The selection of the best set of models is based on cross validation.
/// http://www.cs.cornell.edu/~alexn/papers/shotgun.icml04.revised.rev2.pdf
/// </summary>
/// <param name="learners">Learners in the ensemble</param>
/// <param name="numberOfModelsToSelect">Number of models to select</param>
/// <param name="crossValidation">Cross validation method</param>
/// <param name="ensembleStrategy">Strategy for ensembling models</param>
/// <param name="metric">Metric to minimize</param>
/// <param name="iterations">Number of iterations to random select model combinations.</param>
/// <param name="selectWithReplacement">If true the same model can be selected multiple times.</param>
/// <param name="seed"></param>
public ClassificationRandomModelSelectingEnsembleLearner(
IIndexedLearner <ProbabilityPrediction>[] learners,
int numberOfModelsToSelect,
ICrossValidation <ProbabilityPrediction> crossValidation,
IClassificationEnsembleStrategy ensembleStrategy,
IMetric <double, ProbabilityPrediction> metric,
int iterations = 50,
bool selectWithReplacement = true,
int seed = 42)
: base(learners, crossValidation, ensembleStrategy,
new RandomClassificationEnsembleSelection(
metric, ensembleStrategy, numberOfModelsToSelect,
iterations, selectWithReplacement, seed))
{
}
/// <summary>
/// Classification ensemble model
/// </summary>
/// <param name="ensembleModels">Models included in the ensemble</param>
/// <param name="ensembleStrategy">Strategy on how to combine the models</param>
public ClassificationEnsembleModel(IPredictorModel <ProbabilityPrediction>[] ensembleModels,
IClassificationEnsembleStrategy ensembleStrategy)
{
m_ensembleModels = ensembleModels ?? throw new ArgumentNullException(nameof(ensembleModels));
m_ensembleStrategy = ensembleStrategy ?? throw new ArgumentNullException(nameof(ensembleStrategy));
}
