Contact: Daichi Kuroda
Dimensionality reduction aims to embed high-dimensional data into a lower-dimensional space while preserving as much of the original structure as possible. It is widely used for data visualization, exploratory analysis, and preprocessing for downstream machine learning tasks, including vector retrieval and Retrieval-Augmented Generation (RAG) [1].
Many popular nonlinear dimensionality reduction methods preserve pairwise distances or local neighborhood relationships in the original data. However, high-dimensional datasets often contain multiple independent latent structures that are not fully captured by these relationships. For example, a dataset of coloured geometric shapes may naturally admit two valid clusterings: one based on color and another based on shape. Since conventional dimensionality reduction methods typically optimize a single geometric representation of the data, these distinct views may become entangled in the embedding, making it difficult for downstream algorithms to recover each latent structure separately.
Although dimensionality reduction has been extensively studied and surveyed [2], comparatively little attention has been paid to evaluating how well existing methods preserve multiview structure.
In this project, we will evaluate several widely used dimensionality reduction methods on benchmark multiview datasets to assess how effectively they preserve multiple latent views. If time permits, we also aim to propose a new dimensionality-reduction method that explicitly preserves multiview structure.
[1] Sarthi, Parth, et al. "Raptor: Recursive abstractive processing for tree-organized retrieval." International Conference on Learning Representations. Vol. 2024. 2024.
[2] Espadoto, Mateus, et al. "Toward a quantitative survey of dimension reduction techniques." IEEE transactions on visualization and computer graphics 27.3 (2019): 2153-2173.
Send your CV and transcript to Daichi at daichi.kuroda@epfl.ch.