Unsupervised vs Supervised

What is supervised learning?

Supervised learning involves training an algorithm to classify images or make predictions by providing it with a labelled dataset and telling it what conclusions (or outputs) you want it to reach. For example, you might give a supervised algorithm a training dataset of labelled mammography scans and tell it to learn how to identify breast cancer. Each image would be labelled "cancer" or "no cancer," and the algorithm's accuracy can then be tested by comparing its performance against that of a human clinician. These algorithms are used to automate tasks humans already know how to do, such as pattern recognition.

What is unsupervised learning?

Unsupervised learning involves using algorithms to analyse datasets that haven't been labelled. These algorithms are used to discover patterns or associations that humans weren't previously aware of. For example, you might give an unsupervised algorithm an unlabelled dataset of de-identified electronic health records (EHRs) and ask it to group patients by similarity, a task known as clustering. This new knowledge can then support research purposes, such as discovering new drugs or new uses for existing drugs, identifying new potential causes of disease, or uncovering new patient subgroups.

How does supervised learning work in more detail?

Supervised learning is used for classification and prediction tasks when the outcome of interest is known and can be "taught" to an algorithm by providing a labelled dataset during training. The goal is to "learn" a function so that, given a specific input, the algorithm produces a desired output, for example, correctly classifying chest X-ray images that show known examples of cancerous lung tumours, or predicting a person's 10-year cardiovascular disease risk based on their age, height, weight, and lifestyle.

The most common supervised learning algorithms are:

• Support vector machine — a classification algorithm best suited to binary data (e.g., clinical situations with two clearly differentiated outputs), which classifies data by finding the linear hyperplane (decision boundary) that separates one group of data points from another.
• K-nearest neighbour — a classification algorithm that groups data by similarity, assuming that proximity equals similarity, i.e., objects closer together in a dataset are more alike than those further apart.
• Logistic regression (linear and non-linear) — a prediction algorithm used to fit a model that predicts the probability of a binary outcome belonging to one class or another (e.g., one clinical outcome versus another), commonly used in risk stratification.
• Neural network — a more complex predictive algorithm made up of highly connected networks of neurons that try to identify relationships between input and output variables, most useful when those relationships are complex and non-linear.
• Decision tree — a prediction algorithm that predicts responses to data inputs by following decisions from one node to another along branches, much like a flow diagram. The number of branches and the weight values are determined during training.

‍

‍

What are the strengths and limitations of supervised learning?

Supervised learning is relatively simple, and its algorithms are relatively interpretable, or "explainable," compared to unsupervised learning. It's also easier to validate, since the outcome of interest is known and baseline (human) performance is available for comparison, making it straightforward to check whether the algorithm is performing well.

However, supervised learning algorithms can't be used to discover genuinely "new" information, making them less useful for exploratory research. They also depend heavily on large, accurately labelled datasets, which aren't always easy to obtain.

How does unsupervised learning work in more detail?

Unsupervised learning is used for clustering tasks when the outcome of interest isn't necessarily known in advance and can't be taught to the algorithm directly. Instead, the algorithm must learn the outcome independently by inferring it from unlabelled data. The goal is typically to generate new knowledge, finding "hidden" or previously unknown patterns in datasets, such as identifying new genotypes within genome datasets, or spotting anomalies and outliers in EHR datasets.

The most common unsupervised algorithms are:

• K-means clustering — groups data containing continuous variables into a set number of clearly delineated, non-overlapping clusters.
• Hierarchical clustering — another clustering algorithm capable of producing multiple different clustering solutions to the same problem, with the desired measure of dissimilarity between clusters set by the user.
• Principal component analysis — "simplifies" data by transforming high-dimensional data into simpler linear functions that explain the total variance in the data, a process known as dimensionality reduction.

What are the strengths and limitations of unsupervised learning?

The advantages and disadvantages of unsupervised learning largely mirror those of supervised learning, but in reverse. Unsupervised algorithms don't rely on large, accurately labelled datasets, and can be used to identify non-linear, complex relationships that aren't known in advance. However, they're generally harder to evaluate and are often less "explainable" than supervised algorithms.