Supervised vs. Unsupervised ML

In a world where data has become more valuable and sought after than oil, the mechanisms we use to process and understand data have become increasingly complex. Central to data interpretation in 2024 (almost 2025!) are machine learning algorithms - the products of modern-day artificial intelligence. 

While I have many a blog post on machine learning (especially for medicine and biology related applications), I haven’t talked much about the two main subsets of machine learning. Supervised and unsupervised learning represent two of the most pivotal paradigms, shaping everything from our social media page to predicting stock prices. Understanding the differences between these approaches is key to being better equipped in utilizing machine learning for your own applications. 

Supervised machine learning is my more commonly employed type of algorithms. It’s complex, and is supported by too many calculus rules (chain rule being the most evident) to count, but the logic of supervised algorithms is undeniably intuitive. Its defining feature is its reliance on labeled data.

Earlier this year, I created a machine learning algorithm that diagnosed an eye illness; diabetic retinopathy through being trained on a sample set of 5,000 retinal images. This algorithm was supervised, because every retinal image that I fed into it had a label. I told the algorithm which image represented a healthy retina and which image represented a retina nearing blindness. Using these examples, the supervised learning algorithm learned to generalize patterns so it could diagnose this illness in unseen pictures with remarkable accuracy.

The main advantage of supervised learning lies in its precision. Since the model has access to labeled examples, it learns to make predictions with clear guidance. Additionally, supervised learning models can often provide interpretable results, meaning, they provide some justification as to why it gave the output it did. This is critical and has much untapped potential, especially in fields like medicine and law, where understanding the “why” behind a prediction is just as important as the prediction itself.

Now unsupervised learning is a field I was initially more hesitant to explore. Unsupervised learning works without labeled data, and the code structure for unsupervised algorithms is predictably more complex. The algorithm is tasked with finding hidden patterns, structures, or relationships within the data. It operates somewhat ambiguously, uncovering insights that even the programmer might not have known.

Unsupervised learning does however, shine in exploratory tasks. Clustering algorithms, such as k-means and hierarchical clustering, group data points into clusters based on similarity. This has many applications; the first one I learned about being customer segmentation, where businesses group customers by purchasing behavior to tailor marketing strategies. Another popular application is dimensionality reduction, which reduces the complexity of large datasets while preserving essential features. Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE) are common techniques used in this space.

In fields that I’ve spent more time studying, like bioinformatics, unsupervised learning identifies gene expressions and mutations without prior labels. Similarly, in anomaly detection, it highlights unusual patterns in network traffic or financial transactions or even patient diagnostics that could indicate something abnormal or even dangerous.

Perhaps what is even more important than the different characteristics of these algorithms is when to employ which one. The choice between supervised and unsupervised learning is not a binary one; rather, it depends on the problem at hand. Supervised learning is ideal for predictive tasks where labeled data is available and accuracy is crucial. Unsupervised learning, on the other hand, is better suited for exploratory analyses, where the goal is to uncover hidden structures or relationships within the data.

One cool new intersection I found (only when researching things for this article; and while we’re on the topic of research, https://www.ibm.com/think/topics/supervised-vs-unsupervised-learning and https://cloud.google.com/discover/supervised-vs-unsupervised-learning are great resources to check out) is semi-supervised learning, which combined small amounts of labeled data with larger volumes of unlabeled data. This hybrid approach is gaining traction in fields like natural language processing, where labeling datasets can be prohibitively expensive but large volumes of text are readily available.

As machine learning continues to evolve and grow, the lines between supervised and unsupervised learning are becoming increasingly blurred. Advances in transfer learning, reinforcement learning, and self-supervised learning (all of which I will cover in future blog posts!) are pushing the boundaries of what these algorithms can achieve. For example, self-supervised learning, a subset of unsupervised learning, has been instrumental in training large language models like GPT by generating labels from the data itself.

Supervised and unsupervised machine learning are both important, and work together to impact our digital landscape and the way we view our world. Comparing one to the other will inherently limit your ability to harness both in important fields. However, employing the right algorithm for the right application can transform industries. Together, supervised and unsupervised machine learning algorithms work together to create the foundation of intelligent systems that are reshaping the future.