Clustering High-dimensional Data: Balancing Abstraction and Representation Tutorial at AAAI 2026

Claudia Plant, Lena G. M. Bauer, Christian Böhm · 2026

Read Clustering High-dimensional Data: Balancing Abstraction and Representation Tutorial at AAAI 2026 by Claudia Plant, Lena G. M. Bauer, Christian Böhm (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-11160.txt.

Introduction

Claudia Plant, Lena G. M. Bauer, and Christian Böhm present an AAAI 2026 tutorial on clustering high-dimensional data. Clustering asks how to find a natural grouping in a large dataset. It demands a balance between abstraction and representation. To find clusters, one must drop needless detail from single objects. Yet the representation must keep the key features that make groups distinct.

High dimensions strain that balance: distances blur and noise piles up. The tutorial gathers methods that keep groups visible in such data. It is aimed at researchers and practitioners who cluster genomes, documents, or sensor streams. A clear map of the trade-offs can save teams from groupings that look neat and mean nothing.

Catalog record

CREATORClaudia Plant, Lena G. M. Bauer, Christian Böhm
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-11160

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Clustering High-dimensional Data: Balancing Abstraction and Representation Tutorial at AAAI 2026 Claudia Plant, Lena G. M. Bauer, Christian Böhm Abstract How to find a natural grouping of a large real data set? Clustering requires a balance between abstraction and representation. To identify clusters, we need to abstract from superfluous details of individual objects. But we also need a rich representation that emphasizes the key features shared by groups of objects that distinguish them from other groups of objects. Each clustering algorithm implements a different trade-off between abstraction and representation. Classical K-means implements a high level of abstraction - details are simply averaged out - combined with a very simple representation - all clusters are Gaussians in the original data space. We will see how approaches to subspace and deep clustering support high-dimensional and complex data by allowing richer representations. However, with increasing representational expressiveness comes the need to explicitly enforce abstraction in the objective function to ensure that the resulting method performs clustering and not just representation learning. We will see how current deep clustering methods define and enforce abstraction through centroid-based and density-based clustering losses. Balancing the conflicting goals of abstraction and representation is challenging. Ideas from subspace clustering help by learning one latent space for the information that is relevant to clustering and another latent space to capture all other information in the data. The tutorial ends with an outlook on future research in clustering. Future methods will more adaptively balance abstraction and representation to improve performance, energy efficiency and interpretability. By automatically finding the sweet spot between abstraction and representation, the human brain is very good at clustering and other related tasks such as single-shot learning. So, there is still much room for improvement. Figure 1: Conflicting Goals Abstraction and Representation. Introduction Clustering requires a balance between abstraction and representation . Abstraction is a central goal in natural and artificial intelligence. When confronted with massive, high-dimensional, sparse and noisy data, our brain is remarkably effective in performing abstraction, i.e. in identifying the relevant information in the flood of sensory input data. Clustering or identifying a natural grouping of massive amounts of data objects is one central approach to abstraction. To identify clusters, we need to abstract from superfluous details of individual objects, such as background or lighting in images as shown in Figure 1 . Abstraction creates attractive forces that move objects within the same cluster together (red arrows). However, to determine which variations are superfluous details and which are distinguishing features, we also need a sufficiently rich representation. Representation learning requires mapping different objects to different locations in the latent space, creating repelling forces indicated by the blue arrows. Figure 2: Different Representations of our Running Example Data (Subset of the German Traffic Sign Benchmark). A representation containing too much information, also minor details, makes it difficult to identify the grouping. If, on the other hand, the representation is too coarse, clustering is also difficult as distinguishing features between clusters are missing. Therefore, clustering requires to find a balance between abstraction of unnecessary details and representation of key distinguishing features. As shown in Figure 1 by red arrows, we need a representation that tightly groups the traffic signs belonging to a common cluster but still preserves distinguishing differences between different objects (black arrows). Let us note that general representation learning methods and also supervised learning methods face related challenges of balancing under- and over-fitting. However, the key distinguishing feature is that we consider the challenge of balancing the level of abstraction with the accuracy of representation for the goal of clustering. As our goal is grouping of objects, the sweet spot is at a different level of abstraction/representation as for general representation learning. The goals of abstraction and representation tend to be more conflicting in the context of clustering, as abstraction is more urgently needed to find a grouping. In contrast to supervised approaches, we do not have any labels that help us identify a good trade-off. Figure 3: Exploring the Quality of the Learned Representation. Comparison of Image Reconstructions of Different Methods. Goal of the Tutorial This tutorial is intended for machine learning researchers and professionals interested in learning more about clustering high-dimensional and complex data. Clustering is a well-established research area with numerous published books and articles. Because of this abundance of literature, choosing a suitable algorithm, parametrizing it, and applying it to real data is challenging. This tutorial is intended for practitioners and researchers. Practitioners will receive an overview of different approaches to clustering high-dimensional data, along with insights into their benefits and limitations. This knowledge will enable them to select an appropriate method for their particular problem. Researchers will find starting points for contributing to the topic, as we will illustrate foundational and current approaches with Python code examples. We will summarize the evaluation methodology and provide pointers to benchmark data. We will also highlight open problems that require further research. This tutorial can therefore serve as a starting point for actively contributing to this active and fascinating research topic. Motivation: K-Means The tutorial starts by introducing the task of clustering, which is the task of automatically detecting a natural grouping of a set of data objects. Clustering is an unsupervised machine learning task, we do not have any labeled data. Data objects are often high-dimensional feature vectors, images or texts. Clustering of high-dimensional data is challenging due to the effects of the curse of dimensionality. Furthermore, high-dimensional data can often be clustered in multiple meaningful ways. As a running example, we will use throughout the tutorial a subset of the German Traffic Sign Benchmark (GTSB) data set, which contains images of different traffic signs, see Figure 1 for different instances of the ”keep right” sign. Each image has approximately 3,000 pixels. To review the fundamentals of clustering and demonstrate the curse of dimensionality, we will first introduce the classic K-means algorithm and apply it to this data set. Participants will be able to scan a QR code pointing to a Jupyter notebook containing the running example, https://tinyurl.com/bdha8zca . This allows them to follow the demonstration and run the algorithms in the Google Colab app, if desired. As expected, the clustering quality of K-Means is low, with an NMI of 0.28. For labeled benchmark data, the Normalized Mutual Information (NMI) of given class labels and the cluster labels assigned by the algorithm is an established quality measure for clustering results. The NMI scales between 0 and 1, where numbers close to 0 characterize poor clustering results and 1 represents a perfect clustering. For attendees who are not familiar with evaluating clustering methods on labeled benchmark data, we will briefly introduce the corresponding methodology ( 31 ) . Participants will experience that successfully clustering high-dimensional data requires balancing abstraction and representation. Each clustering method has a unique trade-off between abstraction and representation. K-means clustering in the 3,000-dimensional original space involves significant abstraction because clusters are modeled by their mean vectors. However, it is unsuccessful because the original feature representation does not allow for cluster detection. Therefore, we need to learn a better representation for clustering. One obvious option is to perform a Principle Component Analysis (PCA). Figure 2 ( 1^{\text{st}} subfigure) shows the

Read the complete text