Automated Machine Learning for Humans
MLJAR Automated Machine Learning for Humans
Source Code: https://github.com/mljar/mljar-supervised
Table of Contents
- Automated Machine Learning
- What's good in it?
- Automatic Documentation
- Available Modes
Automated Machine Learning :rocket:
mljar-supervised is an Automated Machine Learning Python package that works with tabular data. It is designed to save time for a data scientist. It abstracts the common way to preprocess the data, construct the machine learning models, and perform hyper-parameters tuning to find the best model :trophy:. It is no black-box as you can see exactly how the ML pipeline is constructed (with a detailed Markdown report for each ML model).
mljar-supervised will help you with:
- explaining and understanding your data (Automatic Exploratory Data Analysis),
- trying many different machine learning models (Algorithm Selection and Hyper-Parameters tuning),
- creating Markdown reports from analysis with details about all models (Atomatic-Documentation),
- saving, re-running and loading the analysis and ML models.
It has four built-in modes of work:
Explainmode, which is ideal for explaining and understanding the data, with many data explanations, like decision trees visualization, linear models coefficients display, permutation importances and SHAP explanations of data,
Performfor building ML pipelines to use in production,
Competemode that trains highly-tuned ML models with ensembling and stacking, with a purpose to use in ML competitions.
Optunamode that can be used to search for highly-tuned ML models, should be used when the performance is the most important, and computation time is not limited (it is available from version
Of course, you can further customize the details of each
mode to meet the requirements.
What's good in it? :boom:
- It is using many algorithms:
Neural Networks, and
- It can compute Ensemble based on greedy algorithm from Caruana paper.
- It can stack models to build level 2 ensemble (available in
Competemode or after setting
- It can do features preprocessing, like: missing values imputation and converting categoricals. What is more, it can also handle target values preprocessing.
- It can do advanced features engineering, like: Golden Features, Features Selection, Text and Time Transformations.
- It can tune hyper-parameters with
not-so-random-searchalgorithm (random-search over defined set of values) and hill climbing to fine-tune final models.
- It can compute the
Baselinefor your data. That you will know if you need Machine Learning or not!
- It has extensive explanations. This package is training simple
max_depth <= 5, so you can easily visualize them with amazing dtreeviz to better understand your data.
mljar-supervisedis using simple linear regression and include its coefficients in the summary report, so you can check which features are used the most in the linear model.
- It cares about explainability of models: for every algorithm, the feature importance is computed based on permutation. Additionally, for every algorithm the SHAP explanations are computed: feature importance, dependence plots, and decision plots (explanations can be switched off with
- There is automatic documnetation for every ML experiment run with AutoML. The
mljar-supervisedcreates markdown reports from AutoML training full of ML details, metrics and charts.
The AutoML Report
The report from running AutoML will contain the table with infomation about each model score and time needed to train the model. For each model there is a link, which you can click to see model's details. The performance of all ML models is presented as scatter and box plots so you can visually inspect which algorithms perform the best :trophy:.
Decision Tree Report
The example for
Decision Tree summary with trees visualization. For classification tasks additional metrics are provided:
- confusion matrix
- threshold (optimized in the case of binary classification task)
- F1 score
- Precision, Recall, MCC
The example for
Available Modes :books:
In the docs you can find details about AutoML modes are presented in the table .
automl = AutoML(mode="Explain")
It is aimed to be used when the user wants to explain and understand the data.
- It is using 75%/25% train/test split.
- It is using:
Neural Networkalgorithms and ensemble.
- It has full explanations: learning curves, importance plots, and SHAP plots.
automl = AutoML(mode="Perform")
It should be used when the user wants to train a model that will be used in real-life use cases.
- It is using 5-fold CV.
- It is using:
Neural Network. It uses ensembling.
- It has learning curves and importance plots in reports.
automl = AutoML(mode="Compete")
It should be used for machine learning competitions.
- It adapts the validation strategy depending on dataset size and
total_time_limit. It can be: train/test split (80/20), 5-fold CV or 10-fold CV.
- It is using:
Nearest Neighbors. It uses ensemble and stacking.
- It has only learning curves in the reports.
automl = AutoML(mode="Optuna", optuna_time_budget=3600)
It should be used when the performance is the most important and time is not limited.
- It is using 10-fold CV
- It is using:
CatBoost. Those algorithms are tuned by
optuna_time_budgetseconds, each. Algorithms are tuned with original data, without advanced feature engineering.
- It is using advanced feature engineering, stacking and ensembling. The hyperparameters found for original data are reused with those steps.
- It produces learning curves in the reports.
:point_right: Binary Classification Example
There is a simple interface available with
import pandas as pd from sklearn.model_selection import train_test_split from supervised.automl import AutoML df = pd.read_csv( "https://raw.githubusercontent.com/pplonski/datasets-for-start/master/adult/data.csv", skipinitialspace=True, ) X_train, X_test, y_train, y_test = train_test_split( df[df.columns[:-1]], df["income"], test_size=0.25 ) automl = AutoML() automl.fit(X_train, y_train) predictions = automl.predict(X_test)
fit will print:
Create directory AutoML_1 AutoML task to be solved: binary_classification AutoML will use algorithms: ['Baseline', 'Linear', 'Decision Tree', 'Random Forest', 'Xgboost', 'Neural Network'] AutoML will optimize for metric: logloss 1_Baseline final logloss 0.5519845471086654 time 0.08 seconds 2_DecisionTree final logloss 0.3655910192804364 time 10.28 seconds 3_Linear final logloss 0.38139916864708445 time 3.19 seconds 4_Default_RandomForest final logloss 0.2975204390214936 time 79.19 seconds 5_Default_Xgboost final logloss 0.2731086827200411 time 5.17 seconds 6_Default_NeuralNetwork final logloss 0.319812276905242 time 21.19 seconds Ensemble final logloss 0.2731086821194617 time 1.43 seconds
- the AutoML results in Markdown report
- the Xgboost Markdown report, please take a look at amazing dependence plots produced by SHAP package :sparkling_heart:
- the Decision Tree Markdown report, please take a look at beautiful tree visualization :sparkles:
- the Logistic Regression Markdown report, please take a look at coefficients table, and you can compare the SHAP plots between (Xgboost, Decision Tree and Logistic Regression) :coffee:
:point_right: Multi-Class Classification Example
The example code for classification of the optical recognition of handwritten digits dataset. Running this code in less than 30 minutes will result in test accuracy ~98%.
import pandas as pd # scikit learn utilites from sklearn.datasets import load_digits from sklearn.metrics import accuracy_score from sklearn.model_selection import train_test_split # mljar-supervised package from supervised.automl import AutoML # load the data digits = load_digits() X_train, X_test, y_train, y_test = train_test_split( pd.DataFrame(digits.data), digits.target, stratify=digits.target, test_size=0.25, random_state=123 ) # train models with AutoML automl = AutoML(mode="Perform") automl.fit(X_train, y_train) # compute the accuracy on test data predictions = automl.predict_all(X_test) print(predictions.head()) print("Test accuracy:", accuracy_score(y_test, predictions["label"].astype(int)))
:point_right: Regression Example
Regression example on Boston house prices data. On test data it scores ~ 10.85 mean squared error (MSE).
import numpy as np import pandas as pd from sklearn.datasets import load_boston from sklearn.model_selection import train_test_split from sklearn.metrics import mean_squared_error from supervised.automl import AutoML # mljar-supervised # Load the data housing = load_boston() X_train, X_test, y_train, y_test = train_test_split( pd.DataFrame(housing.data, columns=housing.feature_names), housing.target, test_size=0.25, random_state=123, ) # train models with AutoML automl = AutoML(mode="Explain") automl.fit(X_train, y_train) # compute the MSE on test data predictions = automl.predict(X_test) print("Test MSE:", mean_squared_error(y_test, predictions))
:point_right: More Examples
- Income classification - it is a binary classification task on census data
- Iris classification - it is a multiclass classification on Iris flowers data
- House price regression - it is a regression task on Boston houses data
For details please check mljar-supervised docs.
From PyPi repository:
pip install mljar-supervised
From source code:
git clone https://github.com/mljar/mljar-supervised.git cd mljar-supervised python setup.py install
Installation for development
git clone https://github.com/mljar/mljar-supervised.git virtualenv venv --python=python3.6 source venv/bin/activate pip install -r requirements.txt pip install -r requirements_dev.txt
Running in the docker:
FROM python:3.7-slim-buster RUN apt-get update && apt-get -y update RUN apt-get install -y build-essential python3-pip python3-dev RUN pip3 -q install pip --upgrade RUN pip3 install mljar-supervised jupyter CMD ["jupyter", "notebook", "--port=8888", "--no-browser", "--ip=0.0.0.0", "--allow-root"]
To get started take a look at our Contribution Guide for information about our process and where you can fit in!
mljar-supervised is provided with MIT license.
mljar-supervised is an open-source project created by MLJAR. We care about ease of use in the Machine Learning.
The mljar.com provides a beautiful and simple user interface for building machine learning models.
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size mljar-supervised-0.10.3.tar.gz (101.0 kB)||File type Source||Python version None||Upload date||Hashes View|