MING: An interpretative support method for visual exploration of multidimensional data

Benoît Colange, Laurent Vuillon, Sylvain Lespinats, Denys Dutykh

Research output: Contribution to journalArticlepeer-review

Abstract

Dimensionality reduction enables analysts to perform visual exploration of multidimensional data with a low-dimensional map retaining as much as possible of the original data structure. The interpretation of such a map relies on the hypothesis of preservation of neighborhood relations. Namely, distances in the map are assumed to reflect faithfully dissimilarities in the data space, as measured with a domain-related metric. Yet, in most cases, this hypothesis is undermined by distortions of those relations by the mapping process, which need to be accounted for during map interpretation. In this paper, we describe an interpretative support method called Map Interpretation using Neighborhood Graphs (MING) displaying individual neighborhood relations on the map, as edges of nearest neighbors graphs. The level of distortion of those relations is shown through coloring of the edges. This allows analysts to assess the reliability of similarity relations inferred from the map, while hinting at the original structure of data by showing the missing relations. Moreover, MING provides a local interpretation for classical map quality indicators, since the quantitative measure of distortions is based on those indicators. Overall, the proposed method alleviates the mapping-induced bias in interpretation while constantly reminding users that the map is not the data.

Original languageBritish English
Pages (from-to)246-269
Number of pages24
JournalInformation Visualization
Volume21
Issue number3
DOIs
StatePublished - Jul 2022

Keywords

  • Dimensionality reduction
  • distortion visualization
  • interpretative support
  • neighborhood retrieval
  • quality evaluation
  • visual data exploration

Fingerprint

Dive into the research topics of 'MING: An interpretative support method for visual exploration of multidimensional data'. Together they form a unique fingerprint.

Cite this