Modeling Mindsets: The Many Cultures Of Learning From Data

by Christoph Molnar

We use data to advance science, make businesses more profitable, automate annoying tasks, and develop smart products. But there is a middleman between data and its usefulness: the model. The model represents a simplified aspect of the world; it’s the glue that connects data and world.

Statistics versus machine learning, frequentist versus Bayesian inference, causation or association, … There are many mindsets to consider for building models from data. Each of these modeling mindsets has its own assumptions, strengths, and limitations.

The best modelers, researchers, and data scientists don’t stubbornly stick to just one mindset. The best modelers mix and match the mindsets.

Modeling Mindsets opens the door to all these different ways of thinking. The book is packed with intuitive explanations and illustrations. Quickly get an overview of the strengths and the limitations of each modeling mindset. Expand your mind when it comes to modeling the world using data.

Genres: Artificial Intelligence, Nonfiction, Mathematics, Technology, Programming

100 pages, ebook

First published April 28, 2022

Reviews

[Robert · Rating 5 out of 5]

Modelling Mindsets is a wonderful book that places the spotlight on a seldom-discussed aspect of data science: it is not a book about how to do modelling, rather, it is about meta-modelling, introducing and contrasting the various approaches to data science.

In any data science problem, whether you realise it or not, you are implicitly using a modelling mindset (Frequentism, Bayesianism, Causal inference, Machine Learning). This book explains the philosophical assumptions used by each mindset and contrasts their applications. The ideal audience of this book is a reader with some experience in applied statistics – perhaps a physicist who has had to run some regressions, a product manager who has done some A/B testing, or a trader attempting to quantify the distribution of returns surrounding a trade – who may not even be aware that they are subscribing to a particular mindset.

I first got interested in data science via machine learning – like many other naive kiddos, I was attracted by glitzy random forests and SVMs and uninterested in boring classical statistics. It has taken me many years of trying different approaches to attain a fraction of the wisdom that Molnar lucidly conveys. While Modelling Mindsets is impractical in the sense that it does not teach you how to apply any of the mindsets, it is highly practical in the sense that it helps you reason about which approach to use for a certain problem. It’s pitched at an ideal level of abstraction, condensing entire fields of academic literature into several sentences explaining the purpose and context for a particular approach, if not the how-tos.

Modelling Mindsets is an excellent tour of the toolbox for data science and an important read if you want to be someone who uses the right tool for the job.

[Syed Nouman Hasany · Rating 5 out of 5]

This was an excellent read. Most people who work with data usually miss the forest for the trees when it comes to how they approach a data science problem. Even when multiple techniques are used, there is often an overarching mindset under which all of those lie; and this can potentially lead to some issues. One might consider a problem to be ill-defined or unsolvable, but it is entirely possible that this stems not by virtue of the problem but as a consequence of the modelling mindset itself.

[molly · Rating 5 out of 5]

This fills what I view as such a gap in the teaching and understanding of statistics, which is the big picture thinking around WHY different approaches in modeling came about and how exactly they differ from each other. It is perfectly possible to understand the math behind different statistical approaches without understanding the philosophical approach (and vice versa, of course, too). In a larger sense, when you are building a model, you are building an approximation of how you see the world. To me, at least, too much focus is often on the model itself (the unthinking "Golem" robot you set in motion, as I always think of them since listening to Richard McElreath's Statistical Rethinking lectures on Youtube) instead of on the larger picture of what you are assuming about the world in order for your model to be true.

The idea behind the book is to give an overview of the scientific mindsets that gave rise to the different modelling approaches. He covers statistical modelling, frequentism, Bayesianism, likelihoodism, causal inference, machine learning, supervised learning, unsupervised learning, reinforcement learning, and deep learning. Instead of diving immediately into the math behind any of these approaches, the usual approach, he talks about the "cultural mindset" in each of these fields. A lot can be learned simply from the headings within each chapter. I do think that likely this book is of most benefit to someone who already has at least a ground level understanding of at least one of the modeling mindsets. Otherwise, it is simply a lot of information to understand all at once.

As someone who works with data, I often feel a bit overwhelmed these days at the sheer quantity of approaches there are to digest and apply. Sometimes I feel haunted by the constant feeling that there is probably some better, perfect approach to solve the data problem at hand. I default to the frequentist approach I learned first the vast majority of the time. I do think that this book offers an answer to this question. Christoph calls it being a "T-shaped modeler"- aim to know a couple mindsets in great detail, and then have a surface understanding of the other approaches. Trying to be an expert on them all is impossible, and it is freeing to admit this. But understanding the philosophy behind the other approaches can go a long way towards realizing when your problem just doesn't fit into the mindset that you were trained in.

After sitting with this book for a while, my main feeling is being surprised that there is not more out there on "meta-modelling." We use models literally every day in almost everything we do at this point, but the training on what a model is tends to fall into narrow siloes with barely a glimpse at the other worlds out there. And as ever, the number of books with truly accessible writing for the non-expert is even more vanishingly small. Christoph is an excellent narrative science writer. He accomplishes big picture thinking about the world of modelling while avoiding the trap of stilted academic writing. As in Interpretable Machine Learning, he uses short stories to build real world understanding of the principles he is talking about.

A very worthwhile quick read that I hope is expanded in future editions.

[Casan Scott · Rating 5 out of 5]

Excellent!

I really enjoyed this read. I’ve been involved with modeling (in some capacity) for about 10 years now, and I really enjoyed the approach Chris took with this book. It’s casual and nicely navigates some really complicated ideas. I recommend this book for both complete beginners and experienced data scientists.

[Davidius · Rating 3 out of 5]

Cheeky overview of major modeling mindsets - btw the diagram echoes the octopus ;)
Statistical modeling
- frequentist
- Bayesian
- likelihood
- causal inference
Machine learning
- supervised learning
- unsupervised learning
- reinforcement learning
- deep learning

[Vipul Mann · Rating 4 out of 5]

Pretty good high-level overview of the differences between statistical vs machine learning mindsets. But don’t expect in-depth analyses or knowledge gain if you’re already familiar with the basics. The book is truly about ’mindsets’

[Justyn Rodrigues · Rating 4 out of 5]

Modeling Mindsets provides a good overview of the different approaches with data in less than 100 pages. It covers a broad range of topics, without going too in depth, allowing enough coverage of the vast array of methods.

[Sarthak Vishnoi · Rating 4 out of 5]

The premise of the book is interesting and it is a good book for someone who just wants to learn the basics of different statistical and machine learning models.