Best Feature Selection for Forecasting
Linear regression was invented at the beginning of the 19th century and today, after more than 200 years, it is still used extensively in practical applications for description and prediction purposes:
- In econometrics, it is useful to estimate the price elasticity of a particular product by regressing sales revenue on price and possibly other features such as demographics and competitor and retail information.
- In health sciences, it can be applied to predict how long a patient will remain, (i.e. length of stay) in the ER of a hospital-based on patient information, triage assessment, medical test results and the date/time of arrival.
- In social sciences, it may shed light on the future academic performance of students, so proactive measures can be taken to improve their learning outcomes.
In general, linear regression is used to model the relationship between a continuous variable and other explanatory variables, which can be either continuous or categorical. When applying this technique, finding the subset of features that maximizes its performance is often of interest.
This modeling example is at the intermediate level, where we assume that you know Python and are familiar with the Gurobi Python API. In addition, you should have some knowledge about building mathematical optimization models.
A Mixed Integer Quadratic Programming (MIQP) formulation that finds the weight estimates for a linear regression problem, where exactly $s$ of those weights must be nonzero.
Key Features of the Solution
- Minimization of the training Residual Sum of Squares (RSS).
- Cross-validation and grid search to select how many features to include in the model.
- Benchmark against scikit-learn’s implementations of Ordinary Least Squares (OLS) Regression and the Lasso.
It is shown that, unlike the Lasso, L0-Regression is scale invariant and does not add bias to the weight estimates. Furthermore, this approach is amenable to the specification of additional linear constraints, such as:
- Enforcing group sparsity among features.
- Limiting pairwise multicollinearity.
- Limiting global multicollinearity.
- Considering a fixed set of nonlinear transformations.
Bertsimas et. al. (2016) successfully applied mathematical programming to solve the best-subset-selection problem for linear regression.
Request a Gurobi Evaluation License or Free Academic License
Modeling examples are coded using the Gurobi Python API in Jupyter Notebook. In order to use the Jupyter Notebooks, you must have a Gurobi License. If you do not have a license, you can request an Evaluation License as a Commercial User or download a free license as an Academic User.
Access the Jupyter Notebook Modeling Example
Click on the button below to be directed to the GitHub HTML page, where you can download the repository for the Best Feature Selection for Forecasting Jupyter Notebook modeling example.