In the past couple of weeks, I started to use sklearn pipelines more intensively. It seemed like a good project to find out more about them and share my experiences in a blog post. So here it is: a sklearn pipeline tutorial.
See the full code!
In most machine learning projects, you won’t start modeling before you went to a whole load of preprocessing steps such as handling missing data and engineering new features to feed to the model. Because you’ll want to work with a train, (validation) and test set, you’ll go through those preprocessing steps multiple times. If you’re not an efficient coder the result will be a huge notebook with many repetitions. Enter: pipelines.
Pipelines are a great way to apply sequential transformations on your data and to feed the result to a classifier. It is an end-to-end procedure that forces you to structure your code and thought process in a specific way.
I wouldn’t recommend it as a tool in an exploratory phase of your project. However, I tend to use it in parallel. Everytime I finished an exploratory step, I add it to the pipeline. Structuring your machine learning project in a pipeline has some benefits:
- The possibility to hypertune not only the classifier, but also the feature engineering process
- A standardized workflow — i.e. less chance you’ll have data leakage
- Readability, interpretability and reproducibility
If you want to use scikit-learn like a pro, you should think like scikit-learn. Because that is what is expected of you. Every step in your thought process should be structured in a fit and transform. The easiest way is by creating new classes that inherit from the BaseEstimator class in scikit-learn.
“All estimators in the main scikit-learn codebase should inherit from sklearn.base.BaseEstimator.”— Developing scikit-learn estimators
In the donors dataset, there are some missing values. Since it’s only a few, a quick imputation method is simply using the mode as the imputation value. But to prevent data leakage, you need to calculate the mode on the train set, but apply the transformation on both train and validation set. That’s why both steps are done in two different class methods.
In the following chunk of code, the PreProcessor class inherits from the BaseEstimator. In the fit method, the mode is calculated. And in the transform method, it is imputed in the missing values.
class PreProcessor(BaseEstimator): def __init__(self): print('PreProcessor initiated.') pass def fit(self, x, y=None): self.teacher_prefix_mode = statistics.mode(x.loc[~x['teacher_prefix'].isna(),'teacher_prefix']) return self def transform(self, x): x_dataset = x.copy() x_dataset['teacher_prefix'] = x_dataset['teacher_prefix'].fillna(self.teacher_prefix_mode) return x_dataset
All the steps in my machine learning project come together in the pipeline. The syntax is as follows: (1) each step is named, (2) each step is done within a sklearn object. To get an overview of all the steps I took, please take a look at the notebook. The cool thing about this chunk of code is that it only takes you a couple of seconds to understand all the steps in your Machine Learning project.
The final object in the pipeline is a voting classifier. It is an ensemble model of a random forest, an adaboost and a K-nearest-neighbour model. Why? For demonstrational purposes. I want to show that pipelines are very flexible: you can have one classifier, stacked classifiers, feature classifiers, voting classifier, etc…
See the full code!
model_pipeline = Pipeline( steps = [ ('pre_process', PreProcessor()), ('datetime_features', DateTimeFeatureCreator()), ('essay_features', EssayFeatureCreator()), ('essay_ngrams', EssayNGramCreator()), ('resources_ngrams', ResourcesNGramCreator()), ('OHE_categoricals', OneHotTransformer()), ('post_process', PostProcessor()), ('scale', StandardScaler()), ('voting_classifier', VotingClassifier( estimators=[('rf', RandomForestClassifier()), ('ab', AdaBoostClassifier()), ('knn', KNeighborsClassifier())], voting='hard') ) ] )
One thing is missing from this blog post: hypertuning. In a next blog post, I want to show you how you can hypertune both your learning algorithm and the feature engineering steps.
I hope you enjoyed this sklearn pipeline tutorial. Great success!