AutoML BinaryClassificationExperiment ignores ColumnInformation.ExampleWeightColumnName — trainers are fitted unweighted

Aberta
  • Algo errado?
#7,823 0 comentários 0 reações 0 responsáveis Ver no GitHub

Mantenedores costumam responder em até 1 dia

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
3/5
Tempo estimado
1-2 dias
Facilidade para iniciantes
70/100
Tipo de issue
Bug
Clareza
Claramente especificada
Status de atividade
Ativa
Stack de tecnologia
csharp
Domínio
machine-learning

Direção de pesquisa

Start at BinaryClassificationExperiment.Execute and trace how ColumnInformation.ExampleWeightColumnName is passed to the trainers and validation metrics; the issue names no source files or tests. Compare the AutoML result with the direct LbfgsLogisticRegression example using the supplied weighted fixture. Done when the experiment either applies weights to fitting and validation or clearly warns or refuses a weighted request.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

area-AutoML untriaged
System information
  • Microsoft.ML 5.0.0, Microsoft.ML.AutoML 0.23.0
  • .NET 10 (net10.0), macOS arm64
Describe the bug

ColumnInformation.ExampleWeightColumnName is accepted by BinaryClassificationExperiment.Execute, but the weight column does not reach the trainers the experiment builds, nor its validation metrics. No warning is emitted, so the caller gets an unweighted model while believing it is weighted.

To reproduce

Fixture: 1,000 X values in [0,1] (plus a noise column Y), each row written twice — once Label=true, once Label=false — with weight W = 10 / 1 for true / false when X > 0.5, and 1 / 10 otherwise. Unweighted, the label is independent of X (P = 0.5 everywhere); weighted, P(true | X=0.9) ≈ 0.9.

  1. AutoML (15 s budget):

    ml.Auto().CreateBinaryClassificationExperiment(15).Execute(train, new ColumnInformation
{
    LabelColumnName = "Label",
    ExampleWeightColumnName = "W",
    NumericColumnNames = { "X", "Y" },
});

    Best run LbfgsLogisticRegressionBinary: P(true | X=0.9) = 0.49, P(true | X=0.1) = 0.49 — identical to the same experiment with W placed in IgnoredColumnNames. The per-trainer best validation AUCs were also identical in both runs (FastTree 0.41, FastForest 0.40, SdcaLogisticRegression 0.50, LbfgsLogisticRegression 0.53).

  2. The same trainer fitted directly:

    ml.Transforms.Concatenate("Features", "X", "Y")
  .Append(ml.BinaryClassification.Trainers.LbfgsLogisticRegression(exampleWeightColumnName: "W"));

    gives 0.96 / 0.04.

Expected behavior

Either the trainers AutoML constructs receive exampleWeightColumnName (the binary, multiclass and regression trainers it sweeps all accept it), with validation metrics weighted the same way — or Execute refuses / warns when ExampleWeightColumnName is set, so a weighted request cannot silently produce an unweighted fit.

Why it matters

Labels collected by importance sampling, or from sources of differing reliability, are trained with per-row weights. A silently unweighted fit learns the sampling bias, and nothing in the result tells the caller.

Related: #6708 (documentation of ExampleWeightColumnName).

Linguagem predominante
C#
Estrelas
9.4k
Forks
2k
Merge médio
2d 18h
PRs com merge (30d)
64

Preparar o ambiente

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de dotnet/machinelearning

Todas as issues de dotnet/machinelearning

Issues semelhantes

Mais issues de C#

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.