About the Book
See the flipbook version at https://betaandbit.github.io/RML/
All right, but how do you build predictive models in a responsible way?
This is a question I am often asked by data scientists at different levels of experience. Seemingly simple but at the same time challenging because there are several orthogonal threads and perspectives of different stakeholders that should be addressed.
Model developers focus on automation of model training, monitoring of performance, debugging, and other MLOps-related matters. Users of predictive models are more interested in explainability, transparency and security, while fairness, bias, ethics are issues of interest to society. Regulators are interested in the consequences of model deployments, especially those with large-scale impacts.
Taking these perspectives into account we focus on three essential elements related to Responsible Machine Learning (RML).
Algorithms - Often, to capture complex relationships in data, you need to use advanced and elastic machine learning algorithms. These, however, should not be used without understanding how they work. So a discussion about responsible modelling must touch on the topic of how complex models work.
Software - Training of advanced models is a computationally demanding process. The libraries that allow for efficient training are low-level engineering masterpieces. Professionals use good tools, so a story about responsible modelling must include a section related to good software.
Process - Predictive modelling is not only about tools but also about planning, logistics, communication, deadlines and objectives. The process of data and model exploration is iterative, as in each iteration, we head towards better and better models. Knowing the tools does not help much if you do not know when and how to use them. Therefore, to talk about responsible modelling, we need to talk about the processes behind modelling.
This book is a unique entanglement of all these aspects together at the same time. You will find here selected modern machine learning techniques and the intuition behind them. Methods are supplemented by code snippets with examples in R language. The process is shown through a comic book describing the adventures of two characters, Beta and Bit.
The interaction of these two shows the decisions that analysts often face, whether to try a different model, try another technique for exploration or look for other data --- questions like: how to compare models or how to validate them.
Model development is responsible and challenging work but also an exciting adventure. Sometimes textbooks focus only on the technical side, losing all the fun.
Here we are going to have it all.
About the Author
18+ years of experience in teaching, training and using machine learning models.
What have I learned from building predictive models in business and academia over the last 18 years? A simple truth - there is no free lunch. Machine learning and artificial intelligence are atomic energy. They can support radiologists, improve credit risk models or streamline business operations. But if not implemented responsibly then discrimination, model drift hidden artefacts can kill any initiative. In 2017 I set up MI2-DataLab a group that works on new methods, tools and initiatives to deliver AI/ML solutions responsibly.