How AutoGluon Dominated Kaggle Competitions and How You Can Beat It. The algorithm that beats 99% of Data Scientists with 4 lines of code.

Cristian Leo

·

Follow

Published in

Towards Data Science

·

19 min read

·

4 days ago

--

3

Listen

Share

In two popular Kaggle competitions, AutoGluon beat 99% of the participating data scientists after merely 4h of training on the raw data (AutoGluon Team. “AutoGluon: AutoML for Text, Image, and Tabular Data.” 2020)

This statement, taken from the AutoGluon research paper, perfectly captures what we will explore today: a machine-learning framework that delivers impressive performance with minimal coding. You only need four lines of code to set up a complete ML pipeline, a task that could otherwise take hours. Yes, just four lines of code! See for yourself:

from autogluon.tabular import TabularDataset, TabularPredictor

train_data = TabularDataset('train.csv')
predictor = TabularPredictor(label='Target').fit(train_data, presets='best_quality')
predictions = predictor.predict(train_data)

These four lines handle data preprocessing by automatically recognizing the data type of each column, feature engineering by finding useful column combinations, and model training through ensembling to identify the best-performing model within a given time. Notice that I didn’t even specify the type of machine learning task (regression/classification). AutoGluon examines the label and determines the task on its own.

Am I advocating for this algorithm? Not necessarily. While I appreciate the power of AutoGluon, I prefer solutions that don’t reduce data science to mere accuracy scores in a Kaggle competition. However, as these models become increasingly popular and widely adopted, it’s important to understand how they work, the math and code behind them, and how you can leverage or outperform them.

1: AutoGluon Overview

AutoGluon is an open-source machine-learning library created by Amazon Web Services (AWS). It's designed to handle the entire ML process for you, from preparing your data to selecting the best model and tuning its settings.

AutoGluon combines simplicity with top-notch performance. It employs advanced techniques like ensemble learning and automatic hyperparameter tuning to ensure that the models you create are highly accurate. This means you can develop powerful machine-learning solutions without getting bogged down in the technical details.

The library takes care of data preprocessing, feature selection, model training, and evaluation, which significantly reduces the time and effort required to build robust machine-learning models. Additionally, AutoGluon scales well, making it suitable for both small projects and large, complex datasets.

For tabular data, AutoGluon can handle both classification tasks, where you categorize data into different groups, and regression tasks, where you predict continuous outcomes. It also supports text data, making it useful for tasks like sentiment analysis or topic categorization. Moreover, it can manage image data, assisting with image recognition and object detection. Although several variations of AutoGluon were built to better handle time-series data, text, and image, here we will focus on the variation to handle tabular data. Let me know if you liked this article and would like future deep dives into its variations. (AutoGluon Team. “AutoGluon: AutoML for Text, Image, and Tabular Data.” 2020)

2: The Space of AutoML

2.1: What is AutoML?

AutoML, short for Automated Machine Learning, is a technology that automates the entire process of applying machine learning to real-world problems. The main goal of AutoML is to make machine learning more accessible and efficient, allowing people to develop models without needing deep expertise. As we’ve already seen, it handles tasks like data preprocessing, feature engineering, model selection, and hyperparameter tuning, which are usually complex and time-consuming (He et al., “AutoML: A Survey of the State-of-the-Art,” 2019).

The concept of AutoML has evolved significantly over the years. Initially, machine learning required a lot of manual effort from experts who had to carefully select features, tune hyperparameters, and choose the right algorithms. As the field grew, so did the need for automation to handle increasingly large and complex datasets. Early efforts to automate parts of the process paved the way for modern AutoML systems. Today, AutoML uses advanced techniques like ensemble learning and Bayesian optimization to create high-quality models with minimal human intervention (Feurer et al., “Efficient and Robust Automated Machine Learning,” 2015).

Several players have emerged in the AutoML space, each offering unique features and capabilities. AutoGluon, developed by Amazon Web Services, is known for its ease of use and strong performance across various data types (AutoGluon Team, “AutoGluon: AutoML for Text, Image, and Tabular Data,” 2020). Google Cloud AutoML provides a suite of machine-learning products that allow developers to train high-quality models with minimal effort. H2O.ai offers H2O AutoML, which provides automatic machine-learning capabilities for both supervised and unsupervised learning tasks (H2O.ai, “H2O AutoML: Scalable Automatic Machine Learning,” 2020). DataRobot focuses on enterprise-level AutoML solutions, offering robust tools for model deployment and management. Microsoft’s Azure Machine Learning also features AutoML capabilities, integrating seamlessly with other Azure services for a comprehensive machine learning solution.

2.2: Key Components of AutoML

The first step in any machine learning pipeline is data preprocessing. This involves cleaning the data by handling missing values, removing duplicates, and correcting errors. Data preprocessing also includes transforming the data into a format suitable for analysis, such as normalizing values, encoding categorical variables, and scaling features. Proper data preprocessing is crucial because the quality of the data directly impacts the performance of the machine learning models.

Once the data is cleaned, the next step is feature engineering. This process involves creating new features or modifying existing ones to improve the model’s performance. Feature engineering can be as simple as creating new columns based on existing data or as complex as using domain knowledge to create meaningful features. The right features can significantly enhance the predictive power of the models.

With the data ready and features engineered, the next step is model selection. There are many algorithms to choose from, each with its strengths and weaknesses depending on the problem at hand. AutoML systems evaluate multiple models to identify the best one for the given task. This might involve comparing models like decision trees, support vector machines, neural networks, and others to see which performs best with the data.

After selecting a model, the next challenge is hyperparameter optimization. Hyperparameters are settings that control the behavior of the machine learning algorithm, such as the learning rate in neural networks or the depth of decision trees. Finding the optimal combination of hyperparameters can greatly improve model performance. AutoML uses techniques like grid search, random search, and more advanced methods like Bayesian optimization to automate this process, ensuring the model is fine-tuned for the best results.

The final step is model evaluation and selection. This involves using techniques like cross-validation to assess how well the model generalizes to new data. Various performance metrics, such as accuracy, precision, recall, and F1-score, are used to measure the model’s effectiveness. AutoML systems automate this evaluation process, ensuring that the model selected is the best fit for the task. Once the evaluation is complete, the best-performing model is chosen for deployment (AutoGluon Team. “AutoGluon: AutoML for Text, Image, and Tabular Data.” 2020).

2.3: Challenges of AutoML

While AutoML saves time and effort, it can be quite demanding in terms of computational resources. Automating tasks like hyperparameter tuning and model selection often requires running many iterations and training multiple models, which can be a challenge for smaller organizations or individuals without access to high-performance computing.

Another challenge is the need for customization. Although AutoML systems are highly effective in many situations, they might not always meet specific requirements right out of the box. Sometimes, the automated processes may not fully capture the unique aspects of a particular dataset or problem. Users may need to tweak parts of the workflow, which can be difficult if the system doesn’t offer enough flexibility or if the user lacks the necessary expertise.

Despite these challenges, the benefits of AutoML often outweigh the drawbacks. It greatly enhances productivity, broadens accessibility, and offers scalable solutions, enabling more people to leverage the power of machine learning (Feurer et al., “Efficient and Robust Automated Machine Learning,” 2015).

3: The Math Behind AutoGluon

3.1: AutoGluon’s Architecture

AutoGluon’s architecture is designed to automate the entire machine learning workflow, from data preprocessing to model deployment. This architecture consists of several interconnected modules, each responsible for a specific stage of the process.

The first step is the Data Module, which handles loading and preprocessing data. This module deals with tasks such as cleaning the data, addressing missing values, and transforming the data into a suitable format for analysis. For example, consider a dataset X with missing values. The Data Module might impute these missing values using the mean or median:

from sklearn.impute import SimpleImputer
imputer = SimpleImputer(strategy='mean')
X_imputed = imputer.fit_transform(X)

Once the data is preprocessed, the Feature Engineering Module takes over. This component generates new features or transforms existing ones to enhance the model’s predictive power. Techniques such as one-hot encoding for categorical variables or creating polynomial features for numeric data are common. For instance, encoding categorical variables might look like this:

from sklearn.preprocessing import OneHotEncoder
encoder = OneHotEncoder()
X_encoded = encoder.fit_transform(X)

At the core of AutoGluon is the Model Module. This module includes a wide array of machine-learning algorithms, such as decision trees, neural networks, and gradient-boosting machines. It trains multiple models on the dataset and evaluates their performance. A decision tree, for example, might be trained as follows:

from sklearn.tree import DecisionTreeClassifier
model = DecisionTreeClassifier()
model.fit(X_train, y_train)

The Hyperparameter Optimization Module automates the search for the best hyperparameters for each model. It uses methods like grid search, random search, and Bayesian optimization. Bayesian optimization, as detailed in the paper by Snoek et al. (2012), builds a probabilistic model to guide the search process:

from skopt import BayesSearchCV
search_space = {'max_depth': (1, 32)}
bayes_search = BayesSearchCV(estimator=DecisionTreeClassifier(), search_spaces=search_space)
bayes_search.fit(X_train, y_train)

After training, the Evaluation Module assesses model performance using metrics like accuracy, precision, recall, and F1-score. Cross-validation is commonly used to ensure the model generalizes well to new data:

from sklearn.model_selection import cross_val_score
scores = cross_val_score(model, X, y, cv=5, scoring='accuracy')
mean_score = scores.mean()

AutoGluon excels with its Ensemble Module, which combines the predictions of multiple models to produce a single, more accurate prediction. Techniques like stacking, bagging, and blending are employed. For instance, bagging can be implemented using the BaggingClassifier:

from sklearn.ensemble import BaggingClassifier
bagging = BaggingClassifier(base_estimator=DecisionTreeClassifier(), n_estimators=10)
bagging.fit(X_train, y_train)

Finally, the Deployment Module handles the deployment of the best model or ensemble into production. This includes exporting the model, generating predictions on new data, and integrating the model into existing systems:

import joblib
joblib.dump(bagging, 'model.pkl')

These components work together to automate the machine learning pipeline, allowing users to build and deploy high-quality models quickly and efficiently.

3.2: Ensemble Learning in AutoGluon

Ensemble learning is a key feature of AutoGluon that enhances its ability to deliver high-performing models. By combining multiple models, ensemble methods improve predictive accuracy and robustness. AutoGluon leverages three main ensemble techniques: stacking, bagging, and blending.

Stacking
Stacking involves training multiple base models on the same dataset and using their predictions as input features for a higher-level model, often called a meta-model. This approach leverages the strengths of various algorithms, allowing the ensemble to make more accurate predictions. The stacking process can be mathematically represented as follows:

Here, h_1​ represents the base models, and h_2​ is the meta-model. Each base model h_1​ takes the input features x_i​ and produces a prediction. These predictions are then used as input features for the meta-model h_2​, which makes the final prediction y^​. By combining the outputs of different base models, stacking can capture a broader range of patterns in the data, leading to improved predictive performance.

Bagging
Bagging, short for Bootstrap Aggregating, improves model stability and accuracy by training multiple instances of the same model on different subsets of the data. These subsets are created by randomly sampling the original dataset with replacement. The final prediction is typically made by averaging the predictions of all the models for regression tasks or by taking a majority vote for classification tasks.

Mathematically, bagging can be represented as follows:

For regression:

For classification:

Here, h_i​ represents the i-th model trained on a different subset of the data. For regression, the final prediction y^​ is the average of the predictions made by each model. For classification, the final prediction y^​ is the most frequently predicted class among the models.

The variance reduction effect of bagging can be illustrated by the law of large numbers, which states that the average of the predictions from multiple models will converge to the expected value, reducing the overall variance and improving the stability of the predictions. It can be illustrated as:

By training on different subsets of the data, bagging also helps in reducing overfitting and increasing the generalizability of the model.

Blending
Blending is similar to stacking but with a simpler implementation. In blending, the data is split into two parts: the training set and the validation set. Base models are trained on the training set, and their predictions on the validation set are used to train a final model, also known as the blender or meta-learner. Blending uses a holdout validation set, which can make it faster to implement:

# Example of blending with simple train-validation split
train_meta, val_meta, y_train_meta, y_val_meta = train_test_split(X, y, test_size=0.2)
base_model_1.fit(train_meta, y_train_meta)
base_model_2.fit(train_meta, y_train_meta)
preds_1 = base_model_1.predict(val_meta)
preds_2 = base_model_2.predict(val_meta)
meta_features = np.column_stack((preds_1, preds_2))
meta_model.fit(meta_features, y_val_meta)

These techniques ensure that the final predictions are more accurate and robust, leveraging the diversity and strengths of multiple models to deliver superior results.

3.3: Hyperparameter Optimization

Hyperparameter optimization involves finding the best settings for a model to maximize its performance. AutoGluon automates this process using advanced techniques like Bayesian optimization, early stopping, and smart resource allocation.

Bayesian Optimization
Bayesian optimization aims to find the optimal set of hyperparameters by building a probabilistic model of the objective function. It uses past evaluation results to make informed decisions about which hyperparameters to try next. This is particularly useful for efficiently navigating large and complex hyperparameter spaces, reducing the number of evaluations needed to find the best configuration:

where f(θ) is the objective function want to optimize, such as model accuracy or loss. θ represents the hyperparameters. E[f(θ)] is the expected value of the objective function given the hyperparameters θ.

Bayesian optimization involves two main steps:

1. Surrogate Modeling: A probabilistic model, usually a Gaussian process, is built to approximate the objective function based on past evaluations.
2. Acquisition Function: This function determines the next set of hyperparameters to evaluate by balancing exploration (trying new areas of the hyperparameter space) and exploitation (focusing on areas known to perform well). Common acquisition functions include Expected Improvement (EI) and Upper Confidence Bound (UCB).

The optimization iteratively updates the surrogate model and acquisition function to converge on the optimal set of hyperparameters with fewer evaluations compared to grid or random search methods.

Early Stopping Techniques
Early stopping prevents overfitting and reduces training time by halting the training process once the model’s performance stops improving on a validation set. AutoGluon monitors the performance of the model during training and stops the process when further training is unlikely to yield significant improvements. This technique not only saves computational resources but also ensures that the model generalizes well to new, unseen data:

from sklearn.model_selection import train_test_split
from sklearn.metrics import log_loss

X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.2)
model = DecisionTreeClassifier()
best_loss = np.inf

for epoch in range(100):
    model.fit(X_train, y_train)
    val_preds = model.predict(X_val)
    loss = log_loss(y_val, val_preds)
    if loss < best_loss:
        best_loss = loss
    else:
        break

Resource Allocation Strategies
Effective resource allocation is crucial in hyperparameter optimization, especially when dealing with limited computational resources. AutoGluon employs strategies like multi-fidelity optimization, where the system initially trains models with a subset of the data or fewer epochs to quickly assess their potential. Promising models are then allocated more resources for thorough evaluation. This approach balances exploration and exploitation, ensuring that computational resources are used effectively:

In this formula:

• h_i​ represents the i-th model.
• C_i​ is the cost associated with model h_i​, such as computational time or resources used.
• Resource(h_i​) represents the proportion of total resources allocated to model h_i​.

By initially training models with reduced fidelity (e.g., using fewer data points or epochs), multi-fidelity optimization quickly identifies promising candidates. These candidates are then trained with higher fidelity, ensuring that computational resources are used effectively. This approach balances the exploration of the hyperparameter space with the exploitation of known good configurations, leading to efficient and effective hyperparameter optimization.

3.4: Model Evaluation and Selection

Model evaluation and selection ensure the chosen model performs well on new, unseen data. AutoGluon automates this process using cross-validation techniques, performance metrics, and automated model selection criteria.

Cross-Validation Techniques
Cross-validation involves splitting the data into multiple folds and training the model on different subsets while validating it on the remaining parts. AutoGluon uses techniques like k-fold cross-validation, where the data is divided into k subsets, and the model is trained and validated k times, each time with a different subset as the validation set. This helps in obtaining a reliable estimate of the model’s performance and ensures that the evaluation is not biased by a particular train-test split:

Performance Metrics
To evaluate the quality of a model, AutoGluon relies on various performance metrics, which depend on the specific task at hand. For classification tasks, common metrics include accuracy, precision, recall, F1-score, and area under the ROC curve (AUC-ROC). For regression tasks, metrics like mean absolute error (MAE), mean squared error (MSE), and R-squared are often used. AutoGluon automatically calculates these metrics during the evaluation process, providing a comprehensive view of the model’s strengths and weaknesses:

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
y_pred = model.predict(X_val)
accuracy = accuracy_score(y_val, y_pred)
precision = precision_score(y_val, y_pred)
recall = recall_score(y_val, y_pred)
f1 = f1_score(y_val, y_pred)

Automated Model Selection Criteria
After evaluating the models, AutoGluon uses automated criteria to select the best-performing one. This involves comparing the performance metrics across different models and choosing the model that excels in the most relevant metrics for the task. AutoGluon also considers factors like model complexity, training time, and resource efficiency. The automated model selection process ensures that the chosen model not only performs well but is also practical to deploy and use in real-world scenarios. By automating this selection, AutoGluon eliminates human bias and ensures a consistent and objective approach to choosing the best model:

best_model = max(models, key=lambda model: model['score'])

4: AutoGluon in Python

Before diving into using AutoGluon, you need to set up your environment. This involves installing the necessary libraries and dependencies.

You can install AutoGluon using pip. Open your terminal or command prompt and run the following command:

pip install autogluon

This command will install AutoGluon along with its required dependencies.

Next, you need to download the data. You’ll need to install Kaggle to download the dataset for this example:

pip install kaggle

After installing, download the dataset by running these commands in your terminal. Make sure you’re in the same directory as your notebook file:

mkdir data
cd data
kaggle competitions download -c playground-series-s4e6
unzip "Academic Succession/playground-series-s4e6.zip"

Alternatively, you can manually download the dataset from the recent Kaggle competition “Classification with an Academic Success Dataset”. The dataset is free for commercial use.

Once your environment is set up, you can use AutoGluon to build and evaluate machine learning models. First, you need to load and prepare your dataset. AutoGluon makes this process straightforward. Suppose you have a CSV file named train.csv containing your training data:

from autogluon.tabular import TabularDataset, TabularPredictor

# Load the dataset
train_df = TabularDataset('data/train.csv')

With the data loaded, you can train a model using AutoGluon. In this example, we will train a model to predict a target variable named ‘Target’ and use accuracy as the evaluation metric. We will also enable hyperparameter tuning and automatic stacking to improve model performance:

# Train the model
predictor = TabularPredictor(
    label='Target',
    eval_metric='accuracy',
    verbosity=1
).fit(
    train_df,
    presets=['best_quality'],
    hyperparameter_tune=True,
    auto_stack=True
)

After training, you can evaluate the model’s performance using the leaderboard, which provides a summary of the model’s performance on the training data:

# Evaluate the model
leaderboard = predictor.leaderboard(train_df, silent=True)
print(leaderboard)

The leaderboard gives you a detailed comparison of all the models trained by AutoGluon.

Let’s break down the key columns and what they mean:

• model: This column lists the names of the models. For example, RandomForestEntr_BAG_L1 refers to a Random Forest model using entropy as the criterion, bagged at level 1.
• score_test: This shows the model’s accuracy on the dataset. A score of 1.00 indicates perfect accuracy for some models. Contrary to its name, score_test is the training dataset used during training.
• score_val: This shows the model’s accuracy on the validation dataset. Keep an eye out for this one, as it shows how well the models perform on unseen data.
• eval_metric: The evaluation metric used, which in this case is accuracy.
• pred_time_test: The time taken to make predictions on the test data.
• pred_time_val: The time taken to make predictions on the validation data.
• fit_time: The time taken to train the model.
• pred_time_test_marginal: The additional prediction time added by the model in the ensemble on the test dataset.
• pred_time_val_marginal: The additional prediction time added by the model in the ensemble on the validation dataset.
• fit_time_marginal: The additional training time added by the model in the ensemble.
• stack_level: Indicates the stacking level of the model. Level 1 models are the base models, while level 2 models are meta-models that use the predictions of level 1 models as features.
• can_infer: Indicates whether the model can be used for inference.
• fit_order: The order in which the models were trained.

Looking at the provided leaderboard, we can see some models like RandomForestEntr_BAG_L1 and RandomForestGini_BAG_L1 have perfect train accuracy (1.000000) but slightly lower validation accuracy, suggesting potential overfitting. WeightedEnsemble_L2, which combines the predictions of level 1 models, generally shows good performance by balancing the strengths of its base models.

Models such as LightGBMLarge_BAG_L1 and XGBoost_BAG_L1 have competitive validation scores and reasonable training and prediction times, making them strong candidates for deployment.

The fit_time and pred_time columns offer insights into the computational efficiency of each model, which is crucial for practical applications.

In addition to the leaderboard, AutoGluon offers several advanced features that allow you to customize the training process, handle imbalanced datasets, and perform hyperparameter tuning.

You can customize various aspects of the training process by adjusting the parameters of the fit method. For example, you can change the number of training iterations, specify different algorithms to use, or set custom hyperparameters for each algorithm.

from autogluon.tabular import TabularPredictor, TabularDataset

# Load the dataset
train_df = TabularDataset('train.csv')

# Define custom hyperparameters
hyperparameters = {
    'GBM': {'num_boost_round': 200},
    'NN': {'epochs': 10},
    'RF': {'n_estimators': 100},
}

# Train the model with custom settings
predictor = TabularPredictor(
    label='Target',
    eval_metric='accuracy',
    verbosity=2
).fit(
    train_data=train_df,
    hyperparameters=hyperparameters
)

Imbalanced datasets can be challenging, but AutoGluon provides tools to handle them effectively. You can use techniques such as oversampling the minority class, undersampling the majority class, or applying cost-sensitive learning algorithms. AutoGluon can automatically detect and handle imbalances in your dataset.

from autogluon.tabular import TabularPredictor, TabularDataset

# Load the dataset
train_df = TabularDataset('train.csv')

# Handle imbalanced datasets by specifying custom parameters
# AutoGluon can handle this internally but specifying here for clarity
hyperparameters = {
    'RF': {'n_estimators': 100, 'class_weight': 'balanced'},
    'GBM': {'num_boost_round': 200, 'scale_pos_weight': 2},
}

# Train the model with settings for handling imbalance
predictor = TabularPredictor(
    label='Target',
    eval_metric='accuracy',
    verbosity=2
).fit(
    train_data=train_df,
    hyperparameters=hyperparameters
)

Hyperparameter tuning is crucial for optimizing model performance. AutoGluon automates this process using advanced techniques like Bayesian optimization. You can enable hyperparameter tuning by setting hyperparameter_tune=True in the fit method.

from autogluon.tabular import TabularPredictor, TabularDataset

# Load the dataset
train_df = TabularDataset('train.csv')

# Train the model with hyperparameter tuning
predictor = TabularPredictor(
    label='Target',
    eval_metric='accuracy',
    verbosity=2
).fit(
    train_data=train_df,
    presets=['best_quality'],
    hyperparameter_tune=True
)

Let’s explore how you could potentially outperform an AutoML model. Let’s assume your main goal is to improve the loss metric, rather than focusing on latency, computational costs, or other metrics.

If you have a large dataset that’s well-suited for deep learning, you might find it easier to experiment with deep learning architectures. AutoML frameworks often struggle in this area because deep learning requires a thorough understanding of the dataset, and blindly applying models can be very time and resource-consuming. Here are some resources to get you started with Deep Learning:

However, the real challenge lies in beating AutoML with traditional machine learning tasks. AutoML systems typically use ensembling, which means you’ll likely end up doing the same thing. A good starting strategy could be to first fit an AutoML model. For instance, using AutoGluon, you can identify which models performed best. You can then take these models and recreate the ensemble architecture that AutoGluon used. By optimizing these models further with a technique like Optuna, you might be able to achieve better performance. Here’s a comprehensive guide to master Optuna:

Additionally, applying domain knowledge to feature engineering can give you an edge. Understanding the specifics of your data can help you create more meaningful features, which can significantly boost your model’s performance. If applicable, augment your dataset to provide more varied training examples, which can help improve the robustness of your models.

By combining these strategies with the insights gained from an initial AutoML model, you can outperform the automated approach and achieve superior results.

Conclusion

AutoGluon revolutionizes the ML process by automating everything from data preprocessing to model deployment. Its cutting-edge architecture, powerful ensemble learning techniques, and sophisticated hyperparameter optimization make it an indispensable tool for newcomers and seasoned data scientists. With AutoGluon, you can transform complex, time-consuming tasks into streamlined workflows, enabling you to build top-tier models with unprecedented speed and efficiency.

However, to truly excel in machine learning, it’s essential not to rely solely on AutoGluon. Use it as a foundation to jumpstart your projects and gain insights into effective model strategies. From there, dive deeper into understanding your data and applying domain knowledge for feature engineering. Experiment with custom models and fine-tune them beyond AutoGluon’s initial offerings.

Bibliography

• Erickson, N., Mueller, J., Charpentier, P., Kornblith, S., Weissenborn, D., Norris, E., … & Smola, A. (2020). AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data. arXiv preprint arXiv:2003.06505.
• Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical Bayesian optimization of machine learning algorithms. Advances in neural information processing systems, 25.
• Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., … & Duchesnay, É. (2011). Scikit-learn: Machine learning in Python. Journal of machine learning research, 12(Oct), 2825–2830.
• AutoGluon Team. “AutoGluon: AutoML for Text, Image, and Tabular Data.” 2020.
• Feurer, Matthias, et al. “Efficient and Robust Automated Machine Learning.” 2015.
• He, Xin, et al. “AutoML: A Survey of the State-of-the-Art.” 2020.
• Hutter, Frank, et al. “Automated Machine Learning: Methods, Systems, Challenges.” 2019.
• H2O.ai. “H2O AutoML: Scalable Automatic Machine Learning.” 2020.