为何GridSearchCV在同一数据集上未返回最优得分？

Let's break down what's happening here—you've run into a common pitfall with combining GridSearchCV and Pipeline, plus some confusion about how your models are actually being trained.

The Core Issue: Duplicate Fitting & Data Scaling Mismatch #

Looking at both code blocks, you're making the same critical mistake:

1. First, you fit the GridSearchCV object directly on your raw data (X).
2. Then, you wrap that already-fit GridSearchCV into a Pipeline with RobustScaler and fit the pipeline again.

This does two things that mess up your results:

• The second fit call on the pipeline overwrites all the results (best_score_, best_params_) from the first fit.
• The first fit uses raw, unscaled data, while the second fit uses data transformed by RobustScaler. The scores you're comparing are from two completely different datasets!

Let's Walk Through Your Code #

First Code Block

• When you run grid.fit(X,train_labels), GridSearchCV runs cross-validation on your raw X and stores results.
• Then gridCV.fit(X,train_labels) scales X first, then runs GridSearchCV again on the scaled data. This overwrites the original grid object's results.
• The best_score_ and best_params_ you see are from the scaled data, where the combination {'C': 25, 'epsilon': 0.005, 'gamma': 0.0003} truly has the highest score (0.90146).

Second Code Block

• Same pattern: you first fit GridSearchCV on raw data, then overwrite it by fitting the pipeline on scaled data.
• The score 0.90103 is for the single parameter set on the scaled data, which is lower than the best score from the full grid search. That's why GridSearchCV didn't pick it—it wasn't the best option on the data it was actually evaluated on.

The Fix: Proper Pipeline Usage #

To avoid data leakage and get accurate results, you should never fit GridSearchCV separately before putting it in a pipeline. The pipeline should handle scaling within each cross-validation fold (so validation data never sees the scaling parameters from the training data). Here's the correct approach:

from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import RobustScaler
from sklearn.svm import SVR
from sklearn.model_selection import GridSearchCV

# Define your parameter grid
params = {'C':[15,20,25],'epsilon':[.003,.005,.008],'gamma':[.0001,.0003,.0006]}

# Create pipeline with scaler and GridSearchCV (no pre-fitting!)
gridCV = make_pipeline(RobustScaler(), GridSearchCV(SVR(), param_grid=params))

# Fit once—this handles scaling *per fold* during cross-validation
gridCV.fit(X, train_labels)

# Extract the GridSearchCV results from the pipeline
grid_search_results = gridCV.named_steps['gridsearchcv']
print("Best score:", grid_search_results.best_score_)
print("Best parameters:", grid_search_results.best_params_)

Key Takeaway #

Your original results were skewed because you overwrote the GridSearchCV object with a second fit on scaled data. The full grid search did find the true best parameter combination for the scaled dataset, which outperformed the single parameter set you tested later.

内容的提问来源于stack exchange，提问作者ankur.salunke