Approximated and User Steerable tSNE for Progressive Visual Analytics (1512.01655v3)

Abstract: Progressive Visual Analytics aims at improving the interactivity in existing analytics techniques by means of visualization as well as interaction with intermediate results. One key method for data analysis is dimensionality reduction, for example, to produce 2D embeddings that can be visualized and analyzed efficiently. t-Distributed Stochastic Neighbor Embedding (tSNE) is a well-suited technique for the visualization of several high-dimensional data. tSNE can create meaningful intermediate results but suffers from a slow initialization that constrains its application in Progressive Visual Analytics. We introduce a controllable tSNE approximation (A-tSNE), which trades off speed and accuracy, to enable interactive data exploration. We offer real-time visualization techniques, including a density-based solution and a Magic Lens to inspect the degree of approximation. With this feedback, the user can decide on local refinements and steer the approximation level during the analysis. We demonstrate our technique with several datasets, in a real-world research scenario and for the real-time analysis of high-dimensional streams to illustrate its effectiveness for interactive data analysis.

• The paper presents A-tSNE as a novel method that approximates traditional tSNE using AKNN queries to drastically reduce initialization time.
• It enables dynamic user steerable refinement, allowing interactive focus and real-time adjustments in high-dimensional datasets.
• Experiments on datasets such as MNIST demonstrate a two orders of magnitude speed improvement while preserving embedding quality.

Approximated and User Steerable tSNE for Progressive Visual Analytics

In the context of Progressive Visual Analytics (PVA), this paper addresses a significant challenge in interactive high-dimensional data exploration by introducing A-tSNE, an approximation of the t-Distributed Stochastic Neighbor Embedding (tSNE) algorithm. Traditional tSNE, despite its effectiveness in dimensionality reduction and visualization, suffers from high computational demands, especially during its initialization phase. This limitation hinders its applicability in PVA, where immediate feedback is essential for user-driven exploration.

The authors propose A-tSNE as a method to overcome the computational bottlenecks of tSNE by implementing a controllable approximation mechanism. The key innovation lies in utilizing Approximated K-Nearest Neighbor (AKNN) queries, enhancing tSNE's ability to provide near-instantaneous intermediate results, crucial for real-time data analysis. This approach allows users to interactively modify, refine, and manipulate high-dimensional data sets without significant delays.

Key Contributions and Methodology

A-tSNE introduces the following enhancements over the traditional tSNE framework:

1. Performance Improvement: A-tSNE demonstrates a marked reduction in initialization time, facilitating immediate data inspection. This is achieved through the use of Forests of Randomized KD-Trees for approximating KNN queries, allowing for a trade-off between the accuracy of the neighbors and computational efficiency.
2. User-Steerable Refinement: The approximation can be adjusted dynamically by the user, enabling a focus on areas of interest within the data. This is particularly beneficial in scenarios where iterative refinements are necessary to uncover subtle structures in the data.
3. Visualization of Approximation Levels: Users are kept informed about the degree of approximation used in their analysis through visual feedback mechanisms. Real-time density-based visualizations complement point-based visualizations to highlight data clusters and support exploration.
4. Data Manipulation: The proposed method allows for real-time modifications, including the addition and removal of data points and dimensions. This is achieved with minimal disruption to the ongoing analysis, paving the way for applications in dynamic and streaming data environments.

Experimental Results and Performance Analysis

The authors provide a robust performance evaluation of A-tSNE using several high-dimensional datasets, including the MNIST and NORB datasets. A significant result is the reduction in computation time by over two orders of magnitude compared to Barnes-Hut SNE (BH-SNE), while still maintaining comparable embedding quality. The paper validates these results using standard benchmark datasets, demonstrating that A-tSNE can achieve a precise balance between rapid computation and maintaining the integrity of the data structures within the embeddings.

Practical Implications and Future Work

A-tSNE opens new avenues in PVA, allowing for efficient handling of large datasets and paving the way for its integration into interactive systems for visual data analysis. The method's ability to manage high-dimensional data streams in real-time also highlights its potential application in monitoring and telecommunications sectors, where quick insights from vast quantities of continuously generated data are necessary.

Looking forward, one promising direction for further research could involve combining A-tSNE with advanced visualization techniques and machine learning models to enhance user interpretability and engagement. Exploration of extended applications, such as climate science and finance, could leverage A-tSNE's capabilities to handle extremely large datasets in domains where dimensionality reduction is paramount.

In conclusion, the introduction of A-tSNE represents a significant contribution to the field of dimensionality reduction and data visualization. By focusing on user interaction and real-time data modification, the authors offer a practical solution for addressing the challenges of high-dimensional data analysis in dynamic and progressive analytical environments.