Event Detection in Document Streams

Exploration of Temporal Events in Document Streams

Visual Analytics of News Story Development

Detection and Exploration of Event Episodes in Text Streams

When exploring time-stamped data that arrives in data streams, the analysts are usually looking for event episodes, i.e. interesting sequences of data points that are similar in some way. An event episode can be, for example, a news story consisting of news articles that come in a news text stream at irregular time intervals, and report on the same real-world topic. In such application scenarios, it is very often necessary to be able to access individual data points (i.e. events), while keeping an overview of the dataset within a wider time frame. Common methods for displaying temporal data employ aggregation or sampling of data points to reduce clutter and provide information about temporal trends in the dataset, thus making the analysis of data on atomic level difficult.

We have developed CloudLines, an interactive visualization method, which combines density estimation with truncation functions and lens distortion and magnification techniques to make exploration of interesting event patterns possible at any scale. The density estimators are used together with importance functions to enhance high-density regions and reduce low-density regions, creating fine-textured temporal fingerprints of underlying data. The method can be coupled with time-series algorithms to automatically detect pre-defined event episodes of interest and perform automated similarity comparison across multiple time series.

More information about this and related work can be found in the following publications.