Tutorial

The Tutorial section aims to provide working code samples demonstrating how to use the abess library to solve real world issues. In the following pages, the abess Python package is used for illustration. The counterpart for R package is available at here.

Generalized Linear Model

Linear Regression

Linear Regression

Linear Regression
Classification: Logistic Regression and Beyond

Classification: Logistic Regression and Beyond

Classification: Logistic Regression and Beyond
Multi-Response Linear Regression

Multi-Response Linear Regression

Multi-Response Linear Regression
Survival Analysis: Cox Regression

Survival Analysis: Cox Regression

Survival Analysis: Cox Regression
Positive Responses: Poisson & Gamma Regressions

Positive Responses: Poisson & Gamma Regressions

Positive Responses: Poisson & Gamma Regressions
Power of **abess** Library: Empirical Comparison

Power of abess Library: Empirical Comparison

Power of **abess** Library: Empirical Comparison
ABESS Algorithm: Details

ABESS Algorithm: Details

ABESS Algorithm: Details

Principal Component Analysis

Principal Component Analysis

Principal Component Analysis

Principal Component Analysis
Robust Principal Component Analysis

Robust Principal Component Analysis

Robust Principal Component Analysis

Advanced Generic Features

When analyzing the real world datasets, we may have the following targets:

  1. identifying predictors when group structure are provided (a.k.a., best group subset selection);

  2. certain variables must be selected when some prior information is given (a.k.a., nuisance regression);

  3. selecting the weak signal variables when the prediction performance is mainly interested (a.k.a., regularized best-subset selection).

These targets are frequently encountered in real world data analysis. Actually, in our methods, the targets can be properly handled by simply change some default arguments in the functions. In the following content, we will illustrate the statistic methods to reach these targets in a one-by-one manner, and give quick examples to show how to perform the statistic methods in LinearRegression and the same steps can be implemented in all methods.

Besides, abess library is very flexible, i.e., users can flexibly control many internal computational components. Specifically, users can specify: (i) the division of samples in cross validation (a.k.a., cross validation division), (ii) specify the initial active set before splicing (a.k.a., initial active set), and so on. We will also describe these in the following.

Best Subset of Group Selection

Best Subset of Group Selection

Best Subset of Group Selection
Nuisance Regression

Nuisance Regression

Nuisance Regression
Regularized Best Subset Selection

Regularized Best Subset Selection

Regularized Best Subset Selection
Cross-Validation Division

Cross-Validation Division

Cross-Validation Division
Initial Active Set

Initial Active Set

Initial Active Set

Computational Tips

The generic splicing technique certifiably guarantees the best subset can be selected in a polynomial time. In practice, the computational efficiency can be improved to handle large scale datasets. The tips for computational improvement are applicable for:

  1. ultra-high dimensional data via

    • feature screening;

    • focus on important variables;

  2. large-sample data via

    • golden-section searching;

    • early-stop scheme;

  3. sparse inputs via

    • sparse matrix computation;

  4. specific models via

    • covariance update for LinearRegression and MultiTaskRegression;

    • quasi Newton iteration for LogisticRegression, PoissonRegression, CoxRegression, etc.

More importantly, the technique in these tips can be use simultaneously. For example, abess allow algorithms to use both feature screening and golden-section searching such that algorithms can handle datasets with large-sample and ultra-high dimension. The following contents illustrate the above tips.

Besides, abess efficiently implements warm-start initialization and parallel computing, which are very useful for fast computing. To help use leverage them, we will also describe their implementation details in the following.

Ultra-High dimensional data

Ultra-High dimensional data

Ultra-High dimensional data
Large-Sample Data

Large-Sample Data

Large-Sample Data
Sparse Inputs

Sparse Inputs

Sparse Inputs
Specific Models

Specific Models

Specific Models