Verdict: use ML.NET (or SharpLearning for a sklearn-like API), except text vectorization, which is the gap filled natively by Lodestar.Text (exact CountVectorizer/TfidfVectorizer semantics), and the two sklearn.decomposition estimators that work on a sparse matrix.

dotnet add package Microsoft.ML

using Microsoft.ML;

var ml = new MLContext(seed: 0);
IDataView data = ml.Data.LoadFromTextFile<Row>("data.csv", hasHeader: true, separatorChar: ',');
var pipeline = ml.Transforms.Concatenate("Features", "f1", "f2")
    .Append(ml.Regression.Trainers.Sdca(labelColumnName: "Label"));
var model = pipeline.Fit(data);

• TfidfVectorizer is non-standard. The sklearn formula (smooth_idf, per-row L2 normalization) must be reproduced to the character — ML.NET's FeaturizeText does not reproduce it. That is exactly the reason for Lodestar.Text. See ../equivalence.md.
• min_df / max_df, n-gram bounds: on the Lodestar side, not ML.NET.
• Preprocessing is mostly a coupling gap, not an absence. ML.NET has NormalizeMeanVariance, NormalizeMinMax, OneHotEncoding, ReplaceMissingValues, TrainTestSplit and CrossValidationSplit — read on Microsoft.ML 5.0.0's exported surface, every one of them is reached through an IDataView or through a catalog naming columns. NormalizeMeanVariance(TransformsCatalog, string inputColumn, string outputColumn, …) never sees a value. Lodestar.Preprocessing answers the entry point, not the capability: a caller holding a double[] gets a double[] back. Two exceptions are real: ML.NET cannot stratify a split — samplingKeyColumnName groups — and nothing in .NET does SMOTE. decisions/0004 has the reading and the order the rest is written in; until then, ML.NET remains the answer if you are already inside a pipeline.

This is the pitfall that used to read "check the definitions before comparing to sklearn", which names the trap without getting anyone out of it.

precision_score(y_true, y_pred, average=…) returns a different number, not a different presentation, for each mode. On an imbalanced problem the modes do not disagree slightly — they disagree by a factor of two, and every one of them is arithmetically correct.

A worked example, taken from this repository's own oracle corpus (binary_imbalanced: 190 samples of class 0, 10 of class 1, a classifier with 30 % label noise). Its confusion matrix is [[133, 57], [4, 6]], so the model finds 6 of the 10 positives and calls 57 negatives positive:

Macro F1 says 0.489, weighted F1 says 0.781, for one model on one dataset. Report either without naming the mode and the reader learns nothing. The two are answering different questions: macro asks how the model does on a class picked at random, weighted asks how it does on a sample picked at random.

In C#, the mode is an enum rather than a string, so a typo is a compile error instead of a ValueError at the end of a run. One ConfusionMatrix.Compute pass feeds both F1.Score and ClassificationReport.Compute:

using Lodestar.Metrics;

ConfusionMatrix cm = ConfusionMatrix.Compute(yTrue, yPred);   // one O(samples) pass
double macro    = F1.Score(cm, Averaging.Macro);              // 0.489
double weighted = F1.Score(cm, Averaging.Weighted);           // 0.781
double[] perClass = F1.PerClass(cm);                          // [0.813, 0.164]

Console.WriteLine(ClassificationReport.Compute(cm).ToText()); // what sklearn prints

Two differences from the Python spelling are deliberate. average=None becomes F1.PerClass, a method, because it returns one value per class rather than a scalar — an enum member cannot change its method's return type. And Averaging.Binary throws on a target with more than two classes instead of guessing which class was meant. Both are recorded in ../decisions/0003.

Absent classes. A class with no predictions gives 0/0. sklearn returns 0 and emits an UndefinedMetricWarning; a warning is easy to miss in a log and has no natural .NET equivalent. Lodestar.Metrics makes the choice explicit — ZeroDivision.Zero (sklearn's value), One, NaN, or Throw, which raises UndefinedMetricException rather than letting a silent 0 flow into a report.

Every function, with its sklearn call and its deliberate divergences, is in ../equivalence.md.

dotnet add package Lodestar.Metrics

Guide to be expanded as real needs arise.