Inês Areosa Rodrigues

Inês Areosa Rodrigues

Data Analyst

GVC Lebanon

Inês graduated in Aerospace Engineering, with MSc thesis in Data Science. She is currently working as a Data Analyst at GVC Lebanon.

MSc Thesis

Title: Visual Tools for Understanding Regression Performance

Supervisor: Luis Torgo; co-supervisor: Luis Custodio (IST, Lisboa, Portugal)

MSc in in Aerospace Engineering

Finished: Nov/2019


Lack of transparency has become a significant barrier to the widespread adoption of Machine Learning techniques in many areas of human society, despite the outstanding performance of recent algorithms in terms of accuracy. When accounting for important and costly decisions, end users need to understand the model to be able to rely on the predictions. In that regard, explaining black-box models has become a hot topic in Machine Learning. There are plenty of methods one can utilize for better understanding the behaviour of a model. Here we focus on explaining regression prediction problems through the usage of visual methods, since these are more adequate for conveying information to end users with reduced technical background.

While most existing work analyses the output of the model, we claim that explaining the performance of the model is also of high relevance. The contributions of our work are then more focused on this latter aspect. We develop a novel approach to inspect the estimated risks of using a black-box regression model for a concrete test case. We describe, evaluate and propose tools that visually convey the relationship between the expected error and the values of a predictor variable.

Moreover, we address a real-world problem, in which we our tools and other state-of-the-art methods to understand factors that drive fishing effort around Large Scale Marine Protected Areas. Additionally, we compare some predictive algorithms to select the most suitable for the problem and then provide an overview analysis of the performance of the chosen one.


  • Interpretability in Machine Learning
  • Regression models