SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsOther AI toolsVisualizing Data using t-SNE

Visualizing Data using t-SNE

t-SNE is a technique for visualizing high-dimensional data, proposed by Laurens van der Maaten and Geoffrey Hinton in 2008 in the paper "Visualizing Data using t-SNE" published in the Journal of Machine Learning Research. This method assigns a position to each data point in a two-dimensional or three-dimensional map, thereby presenting the structural relationships of high-dimensional data in a visual way. It is a variant of Stochastic Neighbor Embedding; compared with the original method, it is easier to optimize and significantly reduces the tendency of data points to cluster in the center of the visualization result, thus producing images of better quality. t-SNE is superior to existing non-parametric visualization techniques at the time, such as Sammon mapping, Isomap, and Locally Linear Embedding, in constructing a single visualization map to reveal data structure at multiple different scales. The paper points out that this property is especially important for high-dimensional data distributed on multiple different but related low-dimensional manifolds, such as images of multi-class objects observed from multiple viewpoints. For very large datasets, the paper shows how t-SNE can use random walks on a neighborhood graph so that the implicit structure of all the data influences the way a subset is displayed. The paper demonstrates the performance of t-SNE on a variety of datasets and compares it with a variety of non-parametric visualization techniques; on almost all datasets, the visualizations generated by t-SNE are clearly better.

Open website

About this tool

Purpose

t-SNE is a data visualization technique whose goal is to map high-dimensional data into two-dimensional or three-dimensional space, assigning a position to each data point, thereby presenting the intrinsic structure of high-dimensional data on a low-dimensional map. It is used to observe the proximity relationships and clustering patterns among data points.

Main Capabilities

According to the paper, t-SNE has the following characteristics:

  • It is a variant of Stochastic Neighbor Embedding (Hinton and Roweis, 2002), and is easier to optimize than the original method.
  • By reducing the tendency of data points to cluster in the center of the visualization map, it produces significantly better visualizations.
  • Compared with existing techniques, it is better at creating a single map to reveal structure at multiple different scales.
  • The paper points out that this characteristic is especially important for high-dimensional data distributed on multiple different but related low-dimensional manifolds, such as images of multi-class objects observed from multiple viewpoints.
  • For very large datasets, t-SNE can use random walks on a neighborhood graph so that the implicit structure of all the data influences the way a subset is displayed.

Usage

The paper presents the technique of t-SNE itself as well as its principles and experimental comparisons. Information about specific installation methods, code invocation interfaces, or usage steps is not provided in the supplied material.

Applicable Scenarios

The applicable scenarios discussed in the paper include:

  • Visualization of high-dimensional data.
  • High-dimensional data lying on multiple different but related low-dimensional manifolds, such as images of multi-class objects seen from multiple viewpoints.
  • Structure visualization of very large datasets (with the help of random walks on a neighborhood graph, so that the implicit structure of all the data influences the display of a subset).

Limitations and Comparison

The paper compares t-SNE with a variety of non-parametric visualization techniques, including Sammon mapping, Isomap, and Locally Linear Embedding. The paper states that on almost all tested datasets, the visualizations generated by t-SNE are significantly better than these other techniques.

Publication Information

  • Authors: Laurens van der Maaten, Geoffrey Hinton
  • Journal volume, issue, and page numbers: 9(86):2579−2605, 2008
  • Paper page:https://jmlr.org/papers/v9/vandermaaten08a.html

Information about fees, pricing, platform support, runtime environment, and similar details is not provided in the supplied material.

Related tools

Browse category