TopicLens: Efficient Multi-Level Visual Topic Exploration of Large-Scale Document Collections.

Area of Science:

• Visual Analytics
• Natural Language Processing
• Information Visualization

Background:

• Topic modeling is widely used for analyzing large document collections.
• Current visual analytics systems lack interactive topic modeling due to processing time and non-interactive nature.
• Existing systems are limited to using a fixed, initial set of topics.

Purpose of the Study:

• To propose TopicLens, a novel interaction technique for dynamic data exploration.
• To enable efficient, on-the-fly computation of topic models and 2D embeddings within a visual analytics workflow.
• To address the gap in integrating interactive topic modeling into visual analytics.

Main Methods:

• Developed a novel interaction technique, TopicLens, featuring a lens interface.
• Proposed an efficient topic modeling method and a semi-supervised 2D embedding algorithm for real-time computation.
• Improved upon state-of-the-art methods like nonnegative matrix factorization and t-distributed stochastic neighbor embedding.

Main Results:

• Demonstrated real-time computation of topic models and 2D embeddings.
• Achieved view consistency during dynamic exploration.
• Evaluated performance and visualization quality using a web-based visual analytics system with real-world datasets.

Conclusions:

• TopicLens facilitates dynamic and efficient exploration of large document collections.
• The proposed methods enable tight integration of topic modeling into visual analytics workflows.
• The system showcases practical capabilities for interactive data analysis.