Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

S+t-SNE -- Bringing Dimensionality Reduction to Data Streams

Pedro C. Vieira, João P. Montrezol, João T. Vieira, João Gama

TL;DR

The paper tackles the challenge of applying t-SNE to infinite data streams by introducing S+t-SNE, a batchwise, incremental embedding method that updates the low-dimensional representation as new data arrives. It combines partial embedding via openTSNE with PEDRUL-based density anchors and convex-hull pruning to bound the number of points in the projection, and it introduces Exponential Cobweb Slicing (ECS) for blind drift management to maintain stable, interpretable visualizations. Experiments on MNIST and a synthetic drift dataset demonstrate that S+t-SNE achieves slower growth in KLD, controlled memory usage, and competitive runtimes while effectively capturing evolving data patterns. The approach offers a real-time visualization tool for high-dimensional streaming data and points toward future work on alternative PEDRUL strategies and drift-robustness metrics for online projections.

Abstract

We present S+t-SNE, an adaptation of the t-SNE algorithm designed to handle infinite data streams. The core idea behind S+t-SNE is to update the t-SNE embedding incrementally as new data arrives, ensuring scalability and adaptability to handle streaming scenarios. By selecting the most important points at each step, the algorithm ensures scalability while keeping informative visualisations. By employing a blind method for drift management, the algorithm adjusts the embedding space, which facilitates the visualisation of evolving data dynamics. Our experimental evaluations demonstrate the effectiveness and efficiency of S+t-SNE, whilst highlighting its ability to capture patterns in a streaming scenario. We hope our approach offers researchers and practitioners a real-time tool for understanding and interpreting high-dimensional data.

S+t-SNE -- Bringing Dimensionality Reduction to Data Streams

TL;DR

The paper tackles the challenge of applying t-SNE to infinite data streams by introducing S+t-SNE, a batchwise, incremental embedding method that updates the low-dimensional representation as new data arrives. It combines partial embedding via openTSNE with PEDRUL-based density anchors and convex-hull pruning to bound the number of points in the projection, and it introduces Exponential Cobweb Slicing (ECS) for blind drift management to maintain stable, interpretable visualizations. Experiments on MNIST and a synthetic drift dataset demonstrate that S+t-SNE achieves slower growth in KLD, controlled memory usage, and competitive runtimes while effectively capturing evolving data patterns. The approach offers a real-time visualization tool for high-dimensional streaming data and points toward future work on alternative PEDRUL strategies and drift-robustness metrics for online projections.

Abstract

We present S+t-SNE, an adaptation of the t-SNE algorithm designed to handle infinite data streams. The core idea behind S+t-SNE is to update the t-SNE embedding incrementally as new data arrives, ensuring scalability and adaptability to handle streaming scenarios. By selecting the most important points at each step, the algorithm ensures scalability while keeping informative visualisations. By employing a blind method for drift management, the algorithm adjusts the embedding space, which facilitates the visualisation of evolving data dynamics. Our experimental evaluations demonstrate the effectiveness and efficiency of S+t-SNE, whilst highlighting its ability to capture patterns in a streaming scenario. We hope our approach offers researchers and practitioners a real-time tool for understanding and interpreting high-dimensional data.
Paper Structure (18 sections, 5 figures, 1 algorithm)

This paper contains 18 sections, 5 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Streaming t-SNE (S+t-SNE)
  4. Problem One - When to start
  5. Problem Two - Reduction of space fingerprint
  6. Handling Drift
  7. Experiments
  8. Datasets.
  9. Configurations.
  10. Results
  11. MNIST.
  12. KLD.
  13. Memory.
  14. Time.
  15. Synthetic Data.
  16. ...and 3 more sections

Figures (5)

  • Figure 1: In shape A) we see the concentric cuts over a convex hull; In B) the median cuts; In C) all cuts together, forming a structure similar to a cobweb.
  • Figure 2: Break-down of KLD, Peak Memory and Time for the MNIST dataset.
  • Figure 3: Projections by S+t-SNE for MNIST. B:400 D:400 IT:700 Slice:0.2
  • Figure 4: Break-down of KLD, Peak Memory and Time for the dataset with drift.
  • Figure 5: S+t-SNE for synthetic data and total accumulated points (bottom-right).