理解问题

本教程需要解决的问题是根据网站内评论的意见采取合适的行动。

可用的训练数据集中，网站评论可能是有毒(toxic)(1)或者无毒(not toxic)(0)两种类型。这种场景下，机器学习中的分类任务最为适合。

分类任务用于区分数据内的类别(category)，类型(type)或种类(class)。常见的例子有：

识别情感是正面或是负面
将邮件按照是否为垃圾邮件归类
判定病人的实验室样本是否为癌症
按照客户的偏好进行分类以响应销售活动

分类任务可以是二元又或是多元的。这里面临的是二元分类的问题。

准备数据

首先建立一个控制台应用程序，基于.NET Core。完成搭建后，添加Microsoft.ML类库包。接着在工程下新建名为Data的文件夹。

之后，下载WikiPedia-detox-250-line-data.tsv与wikipedia-detox-250-line-test.tsv文件，并将它们放入Data文件夹，值得注意的是，这两个文件的Copy to Output Directory属性需要修改成Copy if newer。

加载数据

在Program.cs文件的Main方法里加入以下代码：

MLContext mlContext = new MLContext(seed: 0);
_textLoader = mlContext.Data.TextReader(new TextLoader.Arguments()
{
    Separator = "tab",
    HasHeader = true,
    Column = new[]
                {
                    new TextLoader.Column("Label", DataKind.Bool, 0),
                    new TextLoader.Column("SentimentText", DataKind.Text, 1)
                }
});

其目的是通过使用TextLoader类为数据的加载作好准备。

Column属性中构建了两个对象，即对应数据集中的两列数据。不过第一列这里必须使用Label而不是Sentiment。

提取特征

新建一个SentimentData.cs文件，其中加入SentimentData类与SentimentPrediction。

public class SentimentData
{
    [Column(ordinal: "0", name: "Label")]
    public float Sentiment;
    [Column(ordinal: "1")]
    public string SentimentText;
}
public class SentimentPrediction
{
    [ColumnName("PredictedLabel")]
    public bool Prediction { get; set; }
    [ColumnName("Probability")]
    public float Probability { get; set; }
    [ColumnName("Score")]
    public float Score { get; set; }
}

SentimentData类中的SentimentText为输入数据集的特征，Sentiment则是数据集的标记(label)。

SentimentPrediction类用于模型被训练后的预测。

训练模型

在Program类中加入Train方法。首先它会读取训练数据集，接着将特征列中的文本型数据转换为浮点型数组并设定了训练时所使用的决策树二元分类模型。之后，即是实际训练模型。

public static ITransformer Train(MLContext mlContext, string dataPath)
{
    IDataView dataView = _textLoader.Read(dataPath);
    var pipeline = mlContext.Transforms.Text.FeaturizeText("SentimentText", "Features")
        .Append(mlContext.BinaryClassification.Trainers.FastTree(numLeaves: 50, numTrees: 50, minDatapointsInLeaves: 20));
    Console.WriteLine("=============== Create and Train the Model ===============");
    var model = pipeline.Fit(dataView);
    Console.WriteLine("=============== End of training ===============");
    Console.WriteLine();
    return model;
}

评估模型

加入Evaluate方法。到了这一步，需要读取的是用于测试的数据集，且读取后的数据仍然需要转换成合适的数据类型。

public static void Evaluate(MLContext mlContext, ITransformer model)
{
    IDataView dataView = _textLoader.Read(_testDataPath);
    Console.WriteLine("=============== Evaluating Model accuracy with Test data===============");
    var predictions = model.Transform(dataView);
    var metrics = mlContext.BinaryClassification.Evaluate(predictions, "Label");
    Console.WriteLine();
    Console.WriteLine("Model quality metrics evaluation");
    Console.WriteLine("--------------------------------");
    Console.WriteLine($"Accuracy: {metrics.Accuracy:P2}");
    Console.WriteLine($"Auc: {metrics.Auc:P2}");
    Console.WriteLine($"F1Score: {metrics.F1Score:P2}");
    Console.WriteLine("=============== End of model evaluation ===============");
}

使用模型

训练及评估模型完成后，就可以正式使用它了。这里需要建立一个用于预测的对象(PredictionFunction)，其预测方法的输入参数是SentimentData类型，返回结果为SentimentPrediction类型。

private static void Predict(MLContext mlContext, ITransformer model)
{
    var predictionFunction = model.MakePredictionFunction<SentimentData, SentimentPrediction>(mlContext);
    SentimentData sampleStatement = new SentimentData
    {
        SentimentText = "This is a very rude movie"
    };
    var resultprediction = predictionFunction.Predict(sampleStatement);
    Console.WriteLine();
    Console.WriteLine("=============== Prediction Test of model with a single sample and test dataset ===============");
    Console.WriteLine();
    Console.WriteLine($"Sentiment: {sampleStatement.SentimentText} | Prediction: {(Convert.ToBoolean(resultprediction.Prediction) ? "Toxic" : "Not Toxic")} | Probability: {resultprediction.Probability} ");
    Console.WriteLine("=============== End of Predictions ===============");
    Console.WriteLine();
}

完整示例代码

using System;
using System.Collections.Generic;
using System.IO;
using System.Linq;
using Microsoft.ML;
using Microsoft.ML.Core.Data;
using Microsoft.ML.Runtime.Data;
using Microsoft.ML.Transforms.Text;
namespace SentimentAnalysis
{
    class Program
    {
        static readonly string _trainDataPath = Path.Combine(Environment.CurrentDirectory, "Data", "wikipedia-detox-250-line-data.tsv");
        static readonly string _testDataPath = Path.Combine(Environment.CurrentDirectory, "Data", "wikipedia-detox-250-line-test.tsv");
        static readonly string _modelPath = Path.Combine(Environment.CurrentDirectory, "Data", "Model.zip");
        static TextLoader _textLoader;
        static void Main(string[] args)
        {
            MLContext mlContext = new MLContext(seed: 0);
            _textLoader = mlContext.Data.TextReader(new TextLoader.Arguments()
            {
                Separator = "tab",
                HasHeader = true,
                Column = new[]
                            {
                                new TextLoader.Column("Label", DataKind.Bool, 0),
                                new TextLoader.Column("SentimentText", DataKind.Text, 1)
                            }
            });
            var model = Train(mlContext, _trainDataPath);
            Evaluate(mlContext, model);
            Predict(mlContext, model);
            Console.Read();
        }
        public static ITransformer Train(MLContext mlContext, string dataPath)
        {
            IDataView dataView = _textLoader.Read(dataPath);
            var pipeline = mlContext.Transforms.Text.FeaturizeText("SentimentText", "Features")
                .Append(mlContext.BinaryClassification.Trainers.FastTree(numLeaves: 50, numTrees: 50, minDatapointsInLeaves: 20));
            Console.WriteLine("=============== Create and Train the Model ===============");
            var model = pipeline.Fit(dataView);
            Console.WriteLine("=============== End of training ===============");
            Console.WriteLine();
            return model;
        }
        public static void Evaluate(MLContext mlContext, ITransformer model)
        {
            IDataView dataView = _textLoader.Read(_testDataPath);
            Console.WriteLine("=============== Evaluating Model accuracy with Test data===============");
            var predictions = model.Transform(dataView);
            var metrics = mlContext.BinaryClassification.Evaluate(predictions, "Label");
            Console.WriteLine();
            Console.WriteLine("Model quality metrics evaluation");
            Console.WriteLine("--------------------------------");
            Console.WriteLine($"Accuracy: {metrics.Accuracy:P2}");
            Console.WriteLine($"Auc: {metrics.Auc:P2}");
            Console.WriteLine($"F1Score: {metrics.F1Score:P2}");
            Console.WriteLine("=============== End of model evaluation ===============");
        }
        private static void Predict(MLContext mlContext, ITransformer model)
        {
            var predictionFunction = model.MakePredictionFunction<SentimentData, SentimentPrediction>(mlContext);
            SentimentData sampleStatement = new SentimentData
            {
                SentimentText = "This is a very rude movie"
            };
            var resultprediction = predictionFunction.Predict(sampleStatement);
            Console.WriteLine();
            Console.WriteLine("=============== Prediction Test of model with a single sample and test dataset ===============");
            Console.WriteLine();
            Console.WriteLine($"Sentiment: {sampleStatement.SentimentText} | Prediction: {(Convert.ToBoolean(resultprediction.Prediction) ? "Toxic" : "Not Toxic")} | Probability: {resultprediction.Probability} ");
            Console.WriteLine("=============== End of Predictions ===============");
            Console.WriteLine();
        }
    }
}

程序运行后显示的结果：

=============== Create and Train the Model ===============
=============== End of training ===============
=============== Evaluating Model accuracy with Test data===============
Model quality metrics evaluation
--------------------------------
Accuracy: 83.33%
Auc: 98.77%
F1Score: 85.71%
=============== End of model evaluation ===============
=============== Prediction Test of model with a single sample and test dataset ===============
Sentiment: This is a very rude movie | Prediction: Toxic | Probability: 0.7387648
=============== End of Predictions ===============