Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LangLasso: Interactive Cluster Descriptions through LLM Explanation

Raphael Buchmüller, Dennis Collaris, Linhao Meng, Angelos Chatzimparmpas

TL;DR

The paper tackles the interpretability gap in dimensionality-reduction-driven clustering by introducing LangLasso, which uses LLMs to generate natural-language explanations of clusters. It integrates with existing visual analytics workflows and evaluates three prompting strategies—S1 statistics, S2 subsampling, S3 full data—against ground-truth statistics across four tabular datasets. Findings show that the statistics-backed approach (S1) provides the most precise and reliable descriptions, while S2 is more ambiguous and S3 is often constrained by token limits, highlighting the need to ground LLM outputs with conventional statistics. The method lowers the barrier to cluster interpretation for non-experts and enables incorporation of external context, but it should be used as a complementary component rather than a stand-alone explanation. The authors provide a public demo and discuss reliability considerations, suggesting directions for hybrid strategies and future user studies.

Abstract

Dimensionality reduction is a powerful technique for revealing structure and potential clusters in data. However, as the axes are complex, non-linear combinations of features, they often lack semantic interpretability. Existing visual analytics (VA) methods support cluster interpretation through feature comparison and interactive exploration, but they require technical expertise and intense human effort. We present \textit{LangLasso}, a novel method that complements VA approaches through interactive, natural language descriptions of clusters using large language models (LLMs). It produces human-readable descriptions that make cluster interpretation accessible to non-experts and allow integration of external contextual knowledge beyond the dataset. We systematically evaluate the reliability of these explanations and demonstrate that \langlasso provides an effective first step for engaging broader audiences in cluster interpretation. The tool is available at https://langlasso.vercel.app

LangLasso: Interactive Cluster Descriptions through LLM Explanation

TL;DR

The paper tackles the interpretability gap in dimensionality-reduction-driven clustering by introducing LangLasso, which uses LLMs to generate natural-language explanations of clusters. It integrates with existing visual analytics workflows and evaluates three prompting strategies—S1 statistics, S2 subsampling, S3 full data—against ground-truth statistics across four tabular datasets. Findings show that the statistics-backed approach (S1) provides the most precise and reliable descriptions, while S2 is more ambiguous and S3 is often constrained by token limits, highlighting the need to ground LLM outputs with conventional statistics. The method lowers the barrier to cluster interpretation for non-experts and enables incorporation of external context, but it should be used as a complementary component rather than a stand-alone explanation. The authors provide a public demo and discuss reliability considerations, suggesting directions for hybrid strategies and future user studies.

Abstract

Dimensionality reduction is a powerful technique for revealing structure and potential clusters in data. However, as the axes are complex, non-linear combinations of features, they often lack semantic interpretability. Existing visual analytics (VA) methods support cluster interpretation through feature comparison and interactive exploration, but they require technical expertise and intense human effort. We present \textit{LangLasso}, a novel method that complements VA approaches through interactive, natural language descriptions of clusters using large language models (LLMs). It produces human-readable descriptions that make cluster interpretation accessible to non-experts and allow integration of external contextual knowledge beyond the dataset. We systematically evaluate the reliability of these explanations and demonstrate that \langlasso provides an effective first step for engaging broader audiences in cluster interpretation. The tool is available at https://langlasso.vercel.app
Paper Structure (6 sections, 3 figures, 2 tables)

This paper contains 6 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Results
  5. Discussion
  6. Conclusion

Figures (3)

  • Figure 1: Pipeline of LangLasso. (a) We upload a dataset and a file with per-feature descriptions. (b) We select a cluster in the projection and test three strategies (S1--S3). (c) LangLasso generates a natural-language explanation of differences between selected and non-selected points. (d) We evaluate the explanations using feature distributions, prior human findings, and/or statistics computed from the data.
  • Figure 2: Penguins' projection with three classes but many clusters.
  • Figure 3: Further analysis of the bank marketing dataset. (A) Projection of a distinct, manually selected cluster containing mixed class labels. (B.1) The LLM’s response using the statistics strategy is highly accurate. (B.2) The LLM’s response using the subsampling strategy is vague, limiting human understanding. (C) Data values are used as a validation method to assess the correctness of the LLM’s responses.