May 8-26, 2023, Torino (Italy).
This is a short but intensive course on cutting-edge distribution-free methods for predictive inference using complex machine learning algorithms. This course will cover both the theory and practice of distribution-free predictive inference. Computer-based sessions will focus on hand-on implementations of the described methods using the Python programming language (students will be assumed to have some familiarity with Python), and on applications to real data.
Monday classes: 9:15am - 1:00 pm
Friday classes: 9:15am - 1:00 pm
Monday, May 8
- Introduction and motivation.
- Review of parametric predictive inference via linear regression.
- Data exchangeability and conformal prediction without explanatory variables.
- Split-conformal prediction for regression.
- Conformalized quantile regression.
Thursday, May 11
- Testing for outliers with conformal p-values. (Seminar at Collegio Carlo Alberto)
Friday, May 12
Computer session 1: Hands-on applications with Python and Jupyter notebooks.
Monday, May 15
- Conformal prediction for classification.
- The jackknife+ and CV+.
Friday, May 19
- Computer session 2: Hands-on applications with Python and Jupyter notebooks.
- Label-conditional and group-conditional validity of conformal predictions.
Monday, May 22
- Weighted exchangeability and conformal prediction under distribution shift.
- Extra topics depending on student interest (e.g., connections to causal inference, survival analysis, data sketching, training deep neural networks, etc.).
Friday, May 26
- Computer session 3: Hands-on applications with Python and Jupyter notebooks.
- Extra topics depending on student interest (e.g., connections to causal inference, survival analysis, data sketching, training deep neural networks, etc.).
The computer sessions of this course will utilize Python and Jupyter notebooks. Students are expected to bring their laptops and have pre-installed both Python (version 3.7+) and Jupyter prior to the beginning of the first computer session (Friday, May 12). For new (and experienced) users, it is highly recommended to install Anaconda. Anaconda conveniently installs Python, the Jupyter Notebook, and other data science packages that will be utilized in this course.
Installation instructions: https://docs.jupyter.org/en/latest/install/notebook-classic.html
Anaconda download: https://www.anaconda.com/download
After you install Python and Jupyter through Anaconda, read the following start-up guide on how to use Jupyter notebooks: https://realpython.com/jupyter-notebook-introduction/
This online course material is recommended for students who have some familiarity with programming but want to stregthen their knowledge to Python with the goal of becoming proficient in the scientific computing and data science stack.
https://web.stanford.edu/class/cme193/index.html
- “A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification” https://arxiv.org/abs/2107.07511
- “Distribution-Free Predictive Inference For Regression” https://arxiv.org/abs/1604.04173
- “Predictive inference with the jackknife+” https://arxiv.org/abs/1905.02928
- “Conformalized Quantile Regression” https://arxiv.org/abs/1905.03222
- “Classification with Valid and Adaptive Coverage” https://arxiv.org/abs/2006.02544
- “With Malice Towards None: Assessing Uncertainty via Equalized Coverage” https://arxiv.org/abs/1908.05428
- “Conformal Prediction Under Covariate Shift” https://arxiv.org/abs/1904.06019
- “Conformal prediction beyond exchangeability” https://arxiv.org/abs/2202.13415
- “Testing for Outliers with Conformal p-values” https://arxiv.org/abs/2104.08279 < 5203 li>“Conformal Inference of Counterfactuals and Individual Treatment Effects" https://arxiv.org/abs/2006.06138