The Book Look: The Data Path Less Traveled

When is an answer “good enough”?

There are many areas of data science and AI where we need to be satisfied with an answer that is not perfect and yet still provides business value. The data scientist’s problems are often not solved with straightforward statistics and are instead much more complex. That’s where heuristics excel. In fact, the entire field of heuristics is about producing quick solutions that are not perfect but still good enough. A heuristic is a problem-solving approach to reaching a non-optimal yet still useful solution.

Zacharias Voulgaris, PhD has written many books on data science, machine learning, and artificial intelligence, but this is the first book completely dedicated to heuristics. The book’s goal is to allow the reader to “become proficient in using heuristics within the data science pipeline to produce higher quality results in less time.”

Voulgaris illustrates how heuristics can help you solve challenging problems through simple examples and real-life situations. Apply Jaccard Similarity and a variant, F1 score, Entropy, Ectropy, Area Under Curve, Particle Swarm Optimization, and Genetic Algorithms (along with GA variants). Beyond just exhibiting the various known and lesser-known heuristics available today, he examines how you can go about creating your own through a simple and functional framework. The book contains five sections. Part I provides an overview of heuristics and covers each of the various types of heuristics. Part II focuses on data-oriented heuristics and how they apply specifically to data science problems. Part III explains optimization-oriented heuristics and how they solve challenging optimization problems. Part IV is all about designing and implementing your own heuristics to help with particular problems. Finally, Part V contains additional topics on heuristics, such as transparency and limitations.

Chapter summaries reinforce the key points of each chapter, a glossary explains important terms within our field, and several appendices contain reference material on heuristics and related programming tools. I like that several chapters expand on the material with a code notebook (.jl files) in the Julia language.

Here is an excerpt from Chapter 2 of the book, used with permission from the publisher, Technics Publications:

This chapter will explore heuristics as metrics and algorithms and examine some important considerations. This can lay the foundations for the chapters that follow, where we’ll look at heuristics in more detail and, in particular, applications related to both data science and AI. Feel free to consult the glossary if you need more information on any heuristics terms. Also, consider specific examples in your work or experience where this information applies. This can help cement what’s described here and make it your own.

Heuristics as Metrics

Heuristics as metrics are an ideal option for modern data science work since they are the best of both worlds, combining conventional analytics with data-driven analysis. In fact, it is the cornerstone of the data-driven paradigm, while at the same time, they borrow a lot from the model-driven paradigm underlying conventional analytics.

However, heuristics as metrics are not supported by theory. They are just methods that work for a particular problem and measure what we need to gauge. That is what makes them relatively easy to apply. A heuristic metric is a powerful tool, but only when applied to problems relevant to that metric in terms of scope, which we’ll cover later in this book.

Additionally, heuristics as metrics are the most obvious application of these tools, even if many people take them for granted. There are a few well-known heuristic metrics, such as the F1 score for assessing classifier performance in relation to a given class, or the various similarity metrics that show us the similarity between two data points. These metrics are quite popular, and people don’t really think about them very much because of their frequent use. That’s why they often don’t even use the term heuristics to describe them.

Heuristics as metrics are commonly used in data science work, though they aren’t limited to data science. Still, it makes more sense to use them with actual data than with the outputs of arbitrary functions, as in abstract AI applications such as optimization. In any case, they can be very valuable and, in some cases, a flexible tool. Of course, their flexibility depends on how well you know them and your skills as a problem-solver.

Although heuristic metrics are somewhat similar to conventional statistics, they are a very powerful tool in the data-driven paradigm. So, whenever you’re using machine learning and machine learning-related methods, you will have to deal with heuristics in one way or another.

Since heuristics are underused, there is much untapped potential. Heuristics as metrics, in particular, have a lot of room for growth and evolution.

Heuristics as Algorithms

Heuristics as algorithms are more popular than heuristics as metrics, at least for mainstream data science and AI. This is because heuristics as algorithms have been around longer, and therefore, there are many applications in both of these fields. In fact, it is likely that AI would have never evolved much without the use of algorithm-related heuristics. The same goes for data-driven data science (particularly machine learning), which relies greatly on some of these heuristics-based methods.

What’s more, heuristics as algorithms are ideal for complex problems. Such problems may involve high dimensionality, many restrictions, or anything that would require a lot of computational power to handle analytically. Perhaps that’s why heuristics algorithms are particularly useful for optimization problems and AI in general, even though this kind of heuristic can apply to all sorts of scenarios involving processes. In all cases, a heuristics algorithm needs to be relatively flexible and scalable for the heuristic to be useful.

For example, you can use a heuristics as algorithms approach to find the best variables to use in a data set for feature engineering and other scenarios where you deal with Natural Processing (NP) problems. You can even view the common clustering algorithms used today as this type of heuristic algorithm. Also, because heuristics algorithms are ingrained in our understanding of the data-driven approach, many people incorrectly think of this type of heuristic when they hear this term. However, in academia, where people are more aligned with this way of thinking, the term heuristic has a very particular meaning in this kind of work, usually related to metrics and methods for facilitating the solution of a complex problem.

Heuristics as algorithms are a more challenging area and require a good understanding of the problem. Maybe that’s why most practitioners do not choose them when figuring out the solution to a new problem. After all, developing a new heuristic is not a simple matter, as we’ll discuss shortly. Still, it’s often necessary, especially when the problem is too difficult or too computationally expensive to solve otherwise.

Just like in other kinds of heuristics, heuristics as algorithms has a great deal of untapped potential, especially with research. So, there are lots of theoretical methods in the research sector. For example, when it comes to optimization, new algorithmic heuristics often make their way to papers as potential ways to tackle very challenging problems.

Share this post

Steve Hoberman

Steve Hoberman

Steve Hoberman has trained more than 10,000 people in data modeling since 1992. Steve is known for his entertaining and interactive teaching style (watch out for flying candy!), and organizations around the globe have brought Steve in to teach his Data Modeling Master Class, which is recognized as the most comprehensive data modeling course in the industry. Steve is the author of nine books on data modeling, including the bestseller Data Modeling Made Simple. Steve is also the author of the bestseller, Blockchainopoly. One of Steve’s frequent data modeling consulting assignments is to review data models using his Data Model Scorecard® technique. He is the founder of the Design Challenges group, Conference Chair of the Data Modeling Zone conferences, director of Technics Publications, and recipient of the Data Administration Management Association (DAMA) International Professional Achievement Award. He can be reached at me@stevehoberman.com.

scroll to top