Unsupervised Learning
Unsupervised learning methods aim to support insights from data by revealing relevant structure in function of a goal. To do this efficiently when many variables present themselves, we will focus on methods that cluster the data into subsets that share common features or have observations that are `close’ in terms of some well-chosen distance measure. Once the instructions are formalized, this can happen semi-automatically without much prior knowledge of the detailed meaning of individual variables that make up the data set. The term Unsupervised learning collects clustering methods, dimensionality reduction algorithms and associations that are not aiming for the prediction or estimation of a jointly observed outcome variable.
Practical information:
Want to register?
- Prerequisites: The course is open to all interested persons, but basic Python or R knowledge is expected.
- Price: Industry, private sector, profession: € 455 ; Non profit, government, higher education staff, (Doctoral) students, unemployed: € 210
Schedule: February 24th and March 3th, 2026, from 5.30 pm to 9 pm
Georganiseerd door:
In this course we will focus on clustering and dimensionality reduction methods. We start by reviewing different types of applications where unsupervised learning methods are called upon. We show how dimensionality reduction methods are used to simplify data sets to enhance data visualization, reduce noise and speed up performance and we illustrate how the use of clustering methods can reduce complexity of large data sets into interpretable categories and segments.
The different approaches are then situated in the wide field of AI, machine learning and data analysis. We continue with the introduction of Principle Component Analysis, or PCA. PCA is often used as a pre-processing method before application of other methods, including a clustering method. It is a linear dimensionality reduction technique which transforms existing, possibly highly correlated variables, into fewer but independent variables retaining most information contained in the dataset. Afterwards we move on to the clustering methods. In this course, we will discuss three methods: K-Means, agglomerative hierarchical clustering and density-based clustering, more specifically DBSCAN. In addition to these methods, we present different metrics to evaluate the performance of the applied method. Finally, we discuss the applicability of these algorithms and their differences. Depending on the distributions of the variables and the research question or goal at hand, choices have to be made. Is the data `well-behaved’ (e.g. Gaussian distributions) or do the data exhibit complex patterns or noise? Is the data strongly correlated? Do we want to explore a single level of grouping, or do we want to keep granularity in our findings? We will aim to properly address these questions and formulate a suitable framework based on the presented algorithms.
We alternate theoretical sessions with practical sessions to allow participants to apply and practice the new algorithms. This will further enhance insight in the method as well as programming skills in interaction with the teacher. The student can choose to work with either R or Python.
Course material
The course material includes slides, datasets and notebooks of examples. Solutions to additional exercises will also be given.
Objective of the course
The objective of the course is that you can perform the new algorithms on simulated data sets and in context.