SubmodularBatchStrategy<T, TInput, TOutput>

Models & Types · AiDotNet.ActiveLearning.Batch

Submodular batch selection strategy using facility location objectives.

Implements: ISubmodularBatchStrategy<T, TInput, TOutput>

For Beginners

Submodular functions have a special property called “diminishing returns” - adding more similar samples provides less and less benefit. This naturally encourages selecting diverse, representative samples.

How It Works

How It Works:

Facility Location Objective:

F(S) = Σ max_{s∈S} sim(x, s) for all x in the data

This objective ensures selected samples cover the entire data space well.

Advantages:

Reference: Wei et al. “Submodularity in Data Subset Selection and Active Learning” (ICML 2015)

Constructors

Constructor Summary
SubmodularBatchStrategy() Initializes a new SubmodularBatchStrategy with default settings.
SubmodularBatchStrategy(SubmodularObjective objective, double lambda) Initializes a new SubmodularBatchStrategy with specified parameters.

Properties

Property Summary
T DiversityTradeoff { get; set; }
string Name { get; }

Methods

Method Summary
T ComputeDiversity(TInput sample1, TInput sample2)
T ComputeMarginalGain(IReadOnlyList<int> currentSelection, int candidateIndex, IDataset<T, TInput, TOutput> unlabeledPool)
int[] GreedyMaximization(int[] candidateIndices, IDataset<T, TInput, TOutput> unlabeledPool, int batchSize)
int[] SelectBatch(int[] candidateIndices, Vector<T> scores, IDataset<T, TInput, TOutput> unlabeledPool, int batchSize)