Milos Krstajic

CloudLines: Compact Display of Event Episodes in Multiple Time-Series

We propose incremental logarithmic time-series technique as a way to deal with time-based representations of large and dynamic event data sets in limited space. Modern data visualization problems in the domains of news analysis, network security and financial applications, require visual analysis of incremental data, which poses specific challenges that are normally not solved by static visualizations. The incremental nature of the data implies that visualizations have to necessarily change their content and still provide comprehensible representations. In particular, in this paper we deal with the need to keep an eye on recent events together with providing a context on the past and to make relevant patterns accessible at any scale. Our technique adapts to the incoming data by taking care of the rate at which data items occur and by using a decay function to let the items fade away according to their relevance. Since access to details is also important, we also provide a novel distortion magnifying lens technique which takes into account the distortions introduced by the logarithmic time scale to augment readability in selected areas of interest. We demonstrate the validity of our techniques by applying them on incremental data coming from online news streams in different time frames.

EventRiver: Visually Exploring Text Collections with Temporal References

Many text collections with temporal references, such as news corpora and weblogs, are generated to report and discuss real life events. Thus, event-related tasks, such as detecting real life events that drive the generation of the text documents, tracking event evolutions, and investigating reports and commentaries about events of interest, are important when exploring such text collections. To incorporate and leverage human efforts in conducting such tasks, we propose a novel visual analytics approach named EventRiver. EventRiver integrates event-based automated text analysis and visualization to reveal the events motivating the text generation and the long term stories they construct. On the visualization, users can interactively conduct tasks such as event browsing, tracking, association, and investigation. A working prototype of EventRiver has been implemented for exploring news corpora. A set of case studies, experiments, and a preliminary user test have been conducted to evaluate its effectiveness and efficiency.

Event-Based Analysis of People's Activities and Behavior Using Flickr and Panoramio Geotagged Photo Collections

Photo-sharing websites such as Flickr and Panoramio contain millions of geotagged images contributed by people from all over the world. Characteristics of these data pose new challenges in the domain of spatio-temporal analysis. In this paper, we define several different tasks related to analysis of attractive places, points of interest and comparison of behavioral patterns of different user communities on geotagged photo data. We perform analysis and comparison of temporal events, rankings of sightseeing places in a city, and study mobility of people using geotagged photos. We take a systematic approach to accomplish these tasks by applying scalable computational techniques, using statistical and data mining algorithms, combined with interactive geo-visualization. We provide exploratory visual analysis environment, which allows the analyst to detect spatial and temporal patterns and extract additional knowledge from large geotagged photo collections. We demonstrate our approach by applying the methods to several regions in the world.

Processing online news streams for large-scale semantic analysis

While Internet has enabled us to access a vast amount of online news articles originating from thousands of different sources, the human capability to read all these articles has stayed rather constant. Usually, the publishing industry takes over the role of filtering this enormous amount of information and presenting it in an appropriate way to the group of their subscribers. In this paper, the semantic analysis of such news streams is discussed by introducing a system that streams online news collected by the Europe Media Monitor to our proposed semantic news analysis system. Thereby, we describe in detail the emerging challenges and the corresponding engineering solutions to process incoming articles close to real-time. To demonstrate the use of our system, the case studies show a) temporal analysis of entities, such as institutions or persons, and b) their co-occurence in news articles.

Story Tracker: Incremental visual text analytics of news story development

Online news sources produce thousands of news articles every day, reporting on local and global real-world events. New information quickly replaces the old, making it difficult for readers to put current events in the context of the past. The stories about these events have complex relationships and characteristics that are difficult to model: they can be weakly or strongly related or they can merge or split over time. In this article, we present a visual analytics system for temporal analysis of news stories in dynamic information streams, which combines interactive visualization and text mining techniques to facilitate the analysis of similar topics that split and merge over time. Text clustering algorithms extract stories from online news streams in consecutive time windows and identify similar stories from the past. The stories are displayed in a visualization, which (1) sorts the stories by minimizing clutter and overlap from edge crossings, (2) shows their temporal characteristics in different time frames with different levels of detail, and (3) allows incremental updates of the display without recalculating the past data. Stories can be interactively filtered by their duration and connectivity in order to be explored in full detail. To demonstrate the system’s capabilities for detailed dynamic text stream exploration, we present a use case with real news data about the Arabic Uprising in 2011.

Incremental visual text analytics of news story development

Online news sources produce thousands of news articles every day, reporting on local and global real-world events. New information quickly replaces the old, making it difficult for readers to put current events in the context of the past. Additionally, the stories have very complex relationships and characteristics that are difficult to model: they can be weakly or strongly connected, or they can merge or split over time. In this paper, we present a visual analytics system for exploration of news topics in dynamic information streams, which combines interactive visualization and text mining techniques to facilitate the analysis of similar topics that split and merge over time. We employ text clustering techniques to automatically extract stories from online news streams and present a visualization that: 1) shows temporal characteristics of stories in different time frames with different level of detail; 2) allows incremental updates of the display without recalculating the visual features of the past data; 3) sorts the stories by minimizing clutter and overlap from edge crossings. By using interaction, stories can be filtered based on their duration and characteristics in order to be explored in full detail with details on demand. To demonstrate the usefulness of our system, case studies with real news data are presented and show the capabilities for detailed dynamic text stream exploration.

Real-Time Visual Analytics for Text Streams

Combining automated analysis and visual-interactive displays helps analysts rapidly sort through volumes of raw text to detect critical events and identify surrounding issues.

Visualization of streaming data: Observing change and context in information visualization techniques

Visualizing data streams poses numerous challenges in the data, image and user space. In the era of big data, we need incremental visualization methods that will allow the analysts to explore data faster and help them make important decisions on time. In this paper, we have reviewed several well-known information visualization methods that are commonly used to visualize static datasets and analyzed their degrees of freedom. By observing which independent visual variables can change in each method, we described how these changes are related to the attribute and structure changes that can occur in the data stream. Most of the changes in the data stream lead to potential loss of temporal and relational context between the new data and the past data. We present potential directions for measuring the amount of change and loss of context by reviewing related work and identify open issues for future work in this domain.

Applied Visual Exploration on Real-Time News Feeds using Polarity and Geo-Spatial Analysis

This paper presents a visual analytics approach to explore large news article collections in the domains of polarity and spatial analysis. The exploration is performed on the data collected with Europe Media Monitor (EMM), a system which monitors over 2500 online sources and processes 90,000 articles per day. By analyzing the news feeds, we want to find out which topics are important in different countries and what is the general polarity of the articles within these topics. To assess the polarity of a news article, automatic techniques for polarity analysis are employed and the results are represented using Literature Fingerprinting for visualization. In the spatial description of the news feeds, every article can be represented by two geographic attributes, the news origin and the location of the event itself. In order to assess these spatial properties of news articles, we conducted our geo-analysis, which is able to cope with the size and spatial distribution of the data. Within this application framework, we show opportunities how real-time news feed data can be analyzed efficiently.

The News Auditor: Visual Exploration of Clusters of Stories

In recent years, the quantity of content generated by news agencies and blogs is constantly growing, making it difficult for readers to process and understand this overwhelming amount of data. Online news aggregators present clusters of similar stories in a simple, list-based manner, where the most important article is shown first, while all the other similar articles appear below as hyperlinked headlines. This layout makes the user unaware of the content differences between articles, thus making it very difficult to get a comprehensive picture. Understanding what was changed, how, when and by whom, would lead to new insights about the content distribution over the internet and help in dealing with the news overload problem. We present a visual analytics tool that allows the user to compare the articles that belong to the same story and understand the differences at three levels of detail. Story matrix provides an overview of a document cluster, where the user can identify articles of interest based on their overall similarity and reorder them by different criteria. Structural view shows document thumbnails with highlighted paragraphs of the text that were copied, modified or repositioned by different sources. Finally, Document level view presents two articles side by side to provide full-text comparison. To evaluate our tool, we present two user scenarios applied on a real world data set.