c# - 获取用于 PFI 分析的 BinaryClassification FastTree FeatureNames

Question

我使用 trainingDataView 中的列子集在 ML.net 1.0.0 中构建了一个简单的 BinaryClassification FastTree 模型。现在，我想执行 PFI 分析，但似乎无法仅隔离模型中使用的列/功能与 IDataView 中的所有列。

我一直在引用 this link 的示例二进制分类的 PFI。

var trainingDataView = mlContext.Data.LoadFromTextFile<FPPCNTKData>(TrainDataPath, hasHeader: false, separatorChar: ' ');

Var pipeline = mlContext.Transforms.Concatenate("Features",
                                                "mCalc_FPP_Legs_Range",
                                                "mCalc_FPP_Legs_Ticks",
                                                "mCalc_FPP_Legs_Bars",
                                                "mCalc_FPP_Legs_TMins",
                                                "mCalc_FPP_Diag_RangeBars",
                                                "mCalc_FPP_Diag_RangeTMins",
                                                "mCalc_FPP_Diag_TicksBars",
                                                "mCalc_FPP_Diag_TicksTMins",
                                                "mCalc_XD_XA_Mult_Ticks",
                                                "mCalc_AB_XA_Mult_Ticks",
                                                "mCalc_AD_XA_Mult_Ticks",
                                                "mCalc_BC_XA_Mult_Ticks",
                                                "mCalc_BC_AB_Mult_Ticks",
                                                "mCalc_CD_AB_Mult_Ticks",
                                                "mCalc_CD_BC_Mult_Ticks",
                                                "mCalc_CD_BD_Mult_Ticks")
     .Append(mlContext.BinaryClassification.Trainers.FastTree(labelColumnName: "mHiProfitOneHot", featureColumnName: "Features"));

var trainedModel = pipeline.Fit(trainingDataView);

正如您在下面看到的，由于我是从原始 trainingDataView 中收集特征名称，而不是模型中使用的特征名称，因此 PFI 项目被错误标记。

//// Compute the permutation metrics using the properly normalized data.
var linearPredictor = trainedModel.LastTransformer;
var transformedData = trainedModel.Transform(trainingDataView);
var permutationMetrics = mlContext.BinaryClassification.PermutationFeatureImportance(
                linearPredictor, transformedData, labelColumnName: "mHiProfitOneHot", permutationCount: 3);

// Now let's look at which features are most important to the model overall.
// Get the feature indices sorted by their impact on AUC.
var sortedIndices = permutationMetrics.Select((MetricStatistics, index) => new { index, metrics.AreaUnderRocCurve })
                .OrderByDescending(feature => Math.Abs(feature.AreaUnderRocCurve))
                .Select(feature => feature.index);

// Get the feature names from the training set
var featureNames =
    trainingDataView.Schema.AsEnumerable()
    .Select(column => column.Name) // Get the column names
    .Where(name => name != "mHiProfitOneHot") // Drop the Label
    .ToArray();


Console.WriteLine("Feature\tModel Weight\tChange in AUC\t95% Confidence in the Mean Change in AUC");
var auc = permutationMetrics.Select(x => x.AreaUnderRocCurve).ToArray();
foreach (int i in sortedIndices)
{
    Console.WriteLine("{0}\t{1:0.00}\t{2:G4}\t{3:G4}",
         featureNames[i],
         linearPredictor.Model.SubModel.TrainedTreeEnsemble.TreeWeights[i],
         auc[i].Mean,
         1.96 * auc[i].StandardError);
}

是否可以直接从模型中提取特征名称的子集？谢谢。

Answer 1

您可以搜索您的模型(假设它是一个 TransformerChain，就像您的情况一样)寻找 ColumnConcatenatingTransformer 并获取输入列名称。

string[] columnNames = (model
                    .FirstOrDefault(t => t is ColumnConcatenatingTransformer) as ColumnConcatenatingTransformer)
                    ?.Columns
                    ?.FirstOrDefault(c => c.outputColumnName == "Features")
                    .inputColumnNames;
Console.WriteLine(String.Join(", ", columnNames));