ML.NET多分类训练报错：Label列类型不匹配求助

错误信息 #

System.ArgumentOutOfRangeException: 'Schema mismatch for label column 'Label': expected Key, got Single Arg_ParamName_Name'

初始代码 #

using Microsoft.ML.Data;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace MLPredictor1
{
    internal class TicketInputDataModel
    {
        [LoadColumn(8)]
        public float state { get; set; }
        [LoadColumn(1)]
        public string? short_description { get; set; }
        [LoadColumn(2)]
        public string? description { get; set; }
        [LoadColumn(3)]
        public string? email { get; set; }
        [LoadColumn(4)]
        public float priority { get; set; }
        [LoadColumn(5)]
        public bool active { get; set; }
        [LoadColumn(6)]
        public DateTime opened_at { get; set; }
        [LoadColumn(7)]
        public float child_incidents { get; set; }
        [LoadColumn(0), ColumnName("Label")]
        public float num_of_days_com { get; set; }
    }
}

using Microsoft.ML.Data;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace MLPredictor1
{
    internal class TicketOutputDataModel
    {
        [ColumnName("PredictedLabel")]
        public float num_of_days_com { get; set; }
    }
}

// See https://aka.ms/new-console-template for more information
using Microsoft.ML;
using MLPredictor1;

Console.WriteLine("Hello, World!");

List<TicketInputDataModel> ticketInputModel = new List<TicketInputDataModel>();
ticketInputModel.Add(new TicketInputDataModel() { 
active = true,
child_incidents = 4,
description = "Great to hear that this will be resolved soon",
email = "john.doe@telefonicatech.uk",
num_of_days_com = 5,
opened_at = DateTime.Now,
priority = 5,
short_description = "will be resolved soon",
state = 2
});

ticketInputModel.Add(new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    num_of_days_com = 3,
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
});

ticketInputModel.Add(new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    num_of_days_com = 3,
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
});
ticketInputModel.Add(new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    num_of_days_com = 3,
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
});
List<TicketInputDataModel> ticketInputModel2 = new List<TicketInputDataModel>();
ticketInputModel2.Add(new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    num_of_days_com = 3,
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
});
ticketInputModel2.Add(new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    num_of_days_com = 3,
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
});
ticketInputModel2.Add(new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    num_of_days_com = 3,
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
});

MLContext mlContext = new MLContext();

IDataView? trainingData = mlContext.Data.LoadFromEnumerable<TicketInputDataModel>(ticketInputModel);

var pipeline = mlContext.Transforms.Concatenate("Features", "state", "child_incidents", "priority");

var multiclassTrainer = pipeline.Append(mlContext.MulticlassClassification.Trainers
    .NaiveBayes(labelColumnName: "Label", featureColumnName: "Features")).Append(mlContext.Transforms.Categorical.OneHotEncoding(outputColumnName: "short_descriptionEncoded", inputColumnName: nameof(TicketInputDataModel.short_description)));


ITransformer model = multiclassTrainer.Fit(trainingData); // --- ERROR

var predictor = mlContext.Model.CreatePredictionEngine<TicketInputDataModel, TicketOutputDataModel>(model);

var batchData = mlContext.Data.LoadFromEnumerable<TicketInputDataModel>(ticketInputModel2);
IDataView predictions = model.Transform(batchData);


ITransformer model2 = multiclassTrainer.Fit(predictions);

var predictor2 = mlContext.Model.CreatePredictionEngine<TicketInputDataModel, TicketOutputDataModel>(model2);


TicketInputDataModel tt = new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
};

TicketOutputDataModel ttt = predictor2.Predict(tt);

更新后的代码 #

MLContext mlContext = new MLContext();

IDataView? trainingData = mlContext.Data.LoadFromEnumerable<TicketInputDataModel>(ticketInputModel);

var multiclassTrainer = mlContext.Transforms.Concatenate("FeaturesText", new[]
{
    "description",
    "short_description",
    "email",
})
.Append(mlContext.Transforms.Text.FeaturizeText("Features", "FeaturesText")).Append(mlContext.MulticlassClassification.Trainers
        .NaiveBayes(labelColumnName: "Label", featureColumnName: "Features"));


ITransformer model = multiclassTrainer.Fit(trainingData); // -- ERROR

var predictor = mlContext.Model.CreatePredictionEngine<TicketInputDataModel, TicketOutputDataModel>(model);

var batchData = mlContext.Data.LoadFromEnumerable<TicketInputDataModel>(ticketInputModel2);
IDataView predictions = model.Transform(batchData);


ITransformer model2 = multiclassTrainer.Fit(predictions);

var predictor2 = mlContext.Model.CreatePredictionEngine<TicketInputDataModel, TicketOutputDataModel>(model2);


TicketInputDataModel tt = new TicketInputDataModel()
{
    active = true,
    child_incidents = 6,
    description = "This is taking a bit of time but making progress",
    email = "john.doe@telefonicatech.uk",
    opened_at = DateTime.Now,
    priority = 1,
    short_description = "progress being made",
    state = 3
};

TicketOutputDataModel ttt = predictor2.Predict(tt);

问题原因与解决方法 #

原因 #

ML.NET的多分类训练器要求标签列必须是Key类型（代表离散的类别标识），但当前代码中作为Label的num_of_days_com是float（Single）类型，属于连续数值类型，不符合多分类任务的标签格式要求，因此触发Schema不匹配错误。

解决方法 #

根据实际业务需求，分两种情况处理：

情况1：确实是多分类任务（预测离散的天数类别）

如果num_of_days_com是离散的类别（比如天数区间：1-3天、4-6天等），需要做以下修改：

1. 修改输入模型的标签类型：将num_of_days_com的类型改为uint或int，并添加[KeyType]属性标注类别数量（根据实际类别数调整参数）：

internal class TicketInputDataModel
{
    // 其他属性保持不变
    [LoadColumn(0), ColumnName("Label"), KeyType(5)] // 假设共有5个类别，按需修改
    public uint num_of_days_com { get; set; }
}

2. 修改输出模型：PredictedLabel的类型要与输入标签一致，还可以添加Score属性获取各类别的概率：

internal class TicketOutputDataModel
{
    [ColumnName("PredictedLabel")]
    public uint num_of_days_com { get; set; }
    [ColumnName("Score")]
    public float[]? CategoryProbabilities { get; set; } // 可选，获取每个类别的概率值
}

情况2：实际是回归任务（预测连续的天数数值）

如果num_of_days_com是连续的天数，你应该使用回归训练器而非多分类训练器，修改训练器部分代码：

// 替换多分类训练器为回归训练器
var regressionTrainer = mlContext.Transforms.Concatenate("FeaturesText", new[]
{
    "description",
    "short_description",
    "email",
})
.Append(mlContext.Transforms.Text.FeaturizeText("Features", "FeaturesText"))
.Append(mlContext.Regression.Trainers.Sdca(labelColumnName: "Label", featureColumnName: "Features"));

// 后续训练和预测逻辑保持一致，输出模型的PredictedLabel类型仍为float即可

内容的提问来源于stack exchange，提问作者redoc01