One 、Scikit-learn brief introduction

Scikit-learn Is an open source machine learning library that supports supervised and unsupervised learning . It also fits the model 、 Data preprocessing 、 Model selection and evaluation and many other utilities provide a variety of tools .

Two 、 Fitting and Forecasting ： Estimator Foundation Fitting and predicting: estimator basics

（ One ）estimator

estimators：sklearn Provide built-in machine learning algorithms and models , Collectively referred to as estimators . Each estimator can use its fitting method fit To fit some data .

（ Two ）fit Method

fit Methods usually take two inputs ： Sample matrix / Design matrix X ,X Of size by (n_samples, n_features), That is, the number of samples is rows , Characteristic is column . The target y Is the real number of the regression task , Or categorical integer （ Or any other discrete value set ）. For unsupervised learning tasks , You don't have to specify y . y It's usually a one-dimensional array , Among them the first i Entries correspond to X Of the i Samples （ That's ok ） The goal of .X and y Is usually numpy Array or equivalent data type .
Once the estimator is fitted , It can be used to predict the target value of new data . And there is no need to retrain the estimator .

3、 ... and 、 Converters and preprocessors Transformers and pre-processors

Machine learning workflow usually consists of different parts . A typical pipe （pipeline） It consists of a preprocessing step of converting or interpolating data and a final predictor for predicting the target value .
pipeline = a pre-processing step（transform or impute the data）+ a final predictor/estimator（predicts target value）
stay scikit-learn in , The preprocessor and converter follow estimator The objects are the same API（ They all actually inherit from the same BaseEstimator class ）. Converter object has no prediction method , Instead, output the sample matrix of the new transformation X The transformation method of .
ColumnTransformer Column converter ： Apply different transformations to different features

Four 、 The Conduit ： Link preprocessor and estimator Pipelines: chaining pre-processors and estimators

The pipeline provides the same performance as a conventional estimator API： It can go through fit and predict To be consolidated for forecasting . Using pipes also prevents data leakage .

5、 ... and 、 Model to evaluate Model evaluation

many-ways , Especially for cross-validation

6、 ... and 、 Automatic parameter search Automatic parameter searches

All estimators have adjustable parameters （ Hyperparameters hyper-parameters）. The generalization ability of estimators usually depends mainly on several parameters . The determination of super parameters mainly depends on the data .
Scikit-learn It provides a tool to automatically find the best parameter combination （ Through cross validation ）.

Be careful ： In practice , You almost always want to search the pipeline , Instead of a single estimator . One of the main reasons is , If you apply preprocessing steps to the entire dataset without using pipes , Then perform any type of cross validation , You will break the basic assumption of independence between training and test data . actually , Because you use the entire dataset to preprocess the data , Therefore, some information about the test set can be used by the training set . This will lead to overestimating the generalization ability of the estimator （ You can go to Kaggle Read more in the article ）.
Using pipelines for cross validation and search will largely avoid this common trap .

Reference resources ：
scikit-learn Official documents
scikit-learn Chinese document
Of machine learning framework sklearn brief introduction
scikit-learn Using examples