如何使用ML.NET基于多特征（含情感文本）实现二分类任务

Hey there! Let's break down how to tackle this binary classification task with ML.NET step by step—since you already have multiclass experience, this should feel familiar with a few key targeted tweaks.

First, pivot your focus to ML.NET's Binary Classification module. Unlike multiclass (which handles 3+ categories), binary classification is built specifically for true/false (0/1) label tasks, with dedicated trainers and evaluation metrics. Your core workflow will mirror what you did for multiclass, but you'll swap out the trainer and adjust how you handle the boolean label.

Start small: nail down the feature processing pipeline first, then plug in a lightweight binary trainer to validate the end-to-end flow before optimizing.

1. Define Your Data Models #

First, create strongly-typed classes to represent your input data and model output—this keeps things clean and aligns with ML.NET's conventions:

public class InputData
{
    public int FeatureInt1 { get; set; }
    public string FeatureStr1 { get; set; }
    public int FeatureInt2 { get; set; }
    public string FeatureStr2 { get; set; }
    public int FeatureInt3 { get; set; }
    public string FeatureStr3 { get; set; }
    public int FeatureInt4 { get; set; }
    public string FeatureText { get; set; } // 你的随机文本特征
    public bool Label { get; set; }
}

public class PredictionResult
{
    public bool Label { get; set; }
    public bool PredictedLabel { get; set; }
    public float Score { get; set; } // 模型输出的置信度分数
    public float Probability { get; set; } // 转换后的概率值
}

2. Build the Feature Processing Pipeline #

This is where you'll handle each feature type separately, then combine everything into a single feature vector for the model:

• First 3 String Features (Lookup Conversion)
  Use MapValueToKey to convert discrete string values into numerical IDs (this is ML.NET's implementation of a lookup table). It automatically assigns unique keys to each distinct string in your dataset:

  var pipeline = mlContext.Transforms.Categorical.MapValueToKey("FeatureStr1_Encoded", nameof(InputData.FeatureStr1))
      .Append(mlContext.Transforms.Categorical.MapValueToKey("FeatureStr2_Encoded", nameof(InputData.FeatureStr2)))
      .Append(mlContext.Transforms.Categorical.MapValueToKey("FeatureStr3_Encoded", nameof(InputData.FeatureStr3)));
  

• Random Text Feature (Sentiment-style Processing)
  For unstructured text, start with ML.NET's FeaturizeText—it handles tokenization, stopword removal, and converts text into a numerical feature vector (think bag-of-words or TF-IDF under the hood). If you later want to add dedicated sentiment analysis, you can swap in a pre-trained sentiment model, but this is perfect for getting started:

  pipeline = pipeline.Append(mlContext.Transforms.Text.FeaturizeText("FeatureText_Features", nameof(InputData.FeatureText)));
  

• Combine All Features
  Merge your raw integer features, encoded string features, and text features into one Features column—this is what the trainer will use to learn patterns:

  pipeline = pipeline.Append(mlContext.Transforms.Concatenate("Features",
      nameof(InputData.FeatureInt1),
      "FeatureStr1_Encoded",
      nameof(InputData.FeatureInt2),
      "FeatureStr2_Encoded",
      nameof(InputData.FeatureInt3),
      "FeatureStr3_Encoded",
      nameof(InputData.FeatureInt4),
      "FeatureText_Features"));
  

3. Add a Binary Classification Trainer #

Swap out your multiclass trainer for a binary-specific one. A great starting point is SdcaLogisticRegressionBinaryTrainer—it's fast, interpretable, and works well for most binary tasks:

pipeline = pipeline.Append(mlContext.BinaryClassification.Trainers.SdcaLogisticRegression(
    labelColumnName: nameof(InputData.Label),
    featureColumnName: "Features"))
// Convert the trainer's numeric predicted label back to a boolean
.Append(mlContext.Transforms.Conversion.MapKeyToValue("PredictedLabel", "PredictedLabel"));

4. Train, Evaluate, and Predict #

The rest of the workflow is similar to multiclass, but with binary-focused evaluation metrics:

1. Split your data into training and test sets with mlContext.Data.TrainTestSplit()
2. Train the model with var model = pipeline.Fit(trainingData);
3. Evaluate performance using binary-specific metrics like AUC-ROC, precision, and recall:

   var predictions = model.Transform(testData);
   var metrics = mlContext.BinaryClassification.Evaluate(predictions);
   Console.WriteLine($"AUC-ROC: {metrics.AreaUnderRocCurve:P2}");
   Console.WriteLine($"Accuracy: {metrics.Accuracy:P2}");
   

4. Make predictions by creating a prediction engine:

   var predictor = mlContext.Model.CreatePredictionEngine<InputData, PredictionResult>(model);
   var sampleInput = new InputData
   {
       FeatureInt1 = 123,
       FeatureStr1 = "CategoryA",
       // Fill in other features...
       FeatureText = "This is a sample random text input"
   };
   var result = predictor.Predict(sampleInput);
   Console.WriteLine($"Predicted Label: {result.PredictedLabel}");
   

• Start with the logistic regression trainer first—don't jump to complex models like neural networks until you have a working baseline.
• If you want more control over text processing, break FeaturizeText into discrete steps: TokenizeWords → RemoveDefaultStopWords → CountVectorize.
• Use mlContext.Data.CreateEnumerable<InputData>(processedData, reuseRowObject: false) to inspect your transformed features and confirm everything looks right.

内容的提问来源于stack exchange，提问作者Ethan DeLong