Interpretability & Simplicity

While interpretability and simplicity are non-mathematical concepts, they have become increasingly central in statistical modeling. This shift in focus has been driven by the rise of large––possibly overparameterized––machine learning models that can be challenging to understand.

Interpretability: The growing emphasis on human-interpretable models stems from several key benefits.

1. Interpretable models allow researchers and practitioners to grasp the reasoning behind predictions and decisions.

2. When models are transparent, it is easier to identify and address potential biases or errors.

3. Interpretable models can reveal valuable patterns and relationships in data that might otherwise remain hidden in a "black box."

4. In fields like healthcare and public health, where decisions can have life-altering consequences, stakeholders need to understand the models informing those decisions.

Simplicity: The notion that "simplicity is a sign of truth" has deep roots in philosophy and science. This concept, closely related to Occam's razor principle, suggests that when faced with competing explanations, we should prefer the simplest one that adequately accounts for the observed phenomena. In the context of artificial intelligence, simplicity takes on a nuanced role. While complex models can capture intricate patterns in data, they risk overfitting––essentially "memorizing" noise rather than learning meaningful relationships. Simpler models, on the other hand, often generalize better to unseen data, making them more robust and reliable in real-world applications.

Starting from this, my research goal is to develop models that not only perform well but also remain simple, interpretable, and trustworthy. Below is a selection of these works:

1. Interpretable statistical modeling

• Bargagli Stoffi, F.J., Cadei, R., Lee, K., Dominici, F. Causal Rule Ensemble: Interpretable Discovery and Inference of Heterogeneous Treatment Effects. arXiv preprint arXiv:2009.09036.

  [preprint]

• Bargagli Stoffi, F.J., Tortu, C., Forastiere, L. Heterogeneous Treatment and Spillover Effects under Clustered Network Interference. Annals of Applied Statistics, 2024.

  [preprint]

• Bargagli Stoffi, F.J., De-Witte, K., Gnecco, G. Heterogeneous Causal Effects with Imperfect Compliance: A Bayesian Machine Learning Approach. Annals of Applied Statistics, 16 (3), 1986-2009, 2022.

  [paper]

• Bargagli Stoffi, F.J., Gnecco, G. Causal Tree with Instrumental Variable: An Extension of the Causal Tree Framework to Irregular Assignment Mechanisms. International Journal of Data Science and Analytics, 9, 315–337, 2020.

  [paper]

2. Theory of simplicity in statistical machine learning

• Bargagli Stoffi, F.J., Cevolani, G., Gnecco, G. Simple Models in Complex Worlds: Occam’s Razor and Statistical Learning Theory. Minds and Machines, 32(1), 13-42, 2022.

  [paper]

  
  

• Bargagli Stoffi, F.J., Cevolani, G., Gnecco, G. Should Simplicity Be Always Preferred to Complexity in Supervised Machine Learning? In the International Conference on Machine Learning, Optimization, and Data Science (LOD) (pp. 55-59), 2021. Springer, Cham.

  [paper]

Image: “Leander's Tower on the Bosphorus", Sanford Robinson Gifford, 1876. Harvard Art Museum, Cambridge, USA.