The forensic investigation of communication datasets which contain unstructured text, social network information, and metadata is a complex task that is becoming more important due to the immense amount of data being collected. Currently there are limited approaches that allow an investigator to explore the network, text and metadata in a unified manner. We developed Beagle as a forensic tool for email datasets that allows investigators to flexibly form complex queries in order to discover important information in email data. Beagle was successfully deployed at a security firm which had a large email dataset that was difficult to properly investigate. We discuss our experience developing Beagle as well as the lessons we learned applying visual analytic techniques to a difficult real-world problem.

In this paper, we define the concept of exploratory labeling: the use of computational and interactive methods to help analysts categorize groups of documents into a set of unknown and evolving labels. While many computational methods exist to analyze data and build models once the data is organized around a set of predefined categories or labels, few methods address the problem of reliably discovering and curating such labels in the first place. In order to move first steps towards bridging this gap, we propose an interactive visual data analysis method that integrates human-driven label ideation, specification and refinement with machine-driven recommendations. The proposed method enables the user to progressively discover and ideate labels in an exploratory fashion and specify rules that can be used to automatically match sets of documents to labels. To support this process of ideation, specification, as well as evaluation of the labels, we use unsupervised machine learning methods that provide suggestions and data summaries. We evaluate our method by applying it to a real-world labeling problem as well as through controlled user studies to identify and reflect on patterns of interaction emerging from exploratory labeling activities.

In this paper we present a set of four user studies aimed at exploring the visual design space of what we call keyword summaries: lists of words with associated quantitative values used to help people derive an intuition of what information a given document collection (or part of it) may contain. We seek to systematically study how different visual representations may affect people's performance in extracting information out of keyword summaries. To this purpose, we first create a design space of possible visual representations and compare the possible solutions in this design space through a variety of representative tasks and performance metrics. Other researchers have, in the past, studied some aspects of effectiveness with word clouds, however, the existing literature is somewhat scattered and do not seem to address the problem in a sufficiently systematic and holistic manner. The results of our studies showed a strong dependency on the tasks users are performing. In this paper we present details of our methodology, the results, as well as, guidelines on how to design effective keyword summaries based in our discoveries.

We describe TextTile, a data visualization tool for investigation of datasets and questions that require seamless and flexible analysis of structured data and unstructured text. TextTile is based on real-world data analysis problems gathered through our interaction with a number of domain experts and provides a general purpose solution to such problems. The system integrates a set of operations that can interchangeably be applied to the structured as well as to unstructured text part of the data to generate useful data summaries. Such summaries are then organized in visual tiles in a grid layout to allow their analysis and comparison. We validate TextTile with task analysis, use cases and a user study showing the system can be easily learned and proficiently used to carry out nontrivial tasks.

We present a study aimed at understanding how human observers judge scatter plot similarity when presented with a large set of iconic scatter plot representations. The work we present involves 18 participants with a scientific background in a similarity perception study. The study asks participants to group a carefully selected set of plots according to their subjective perceptual judgement of similarity, and it integrates the results into a consensus similarity grouping. We then use this consensus grouping to generate insights on similarity perception. The main output of this work is a list of concepts we derive to describe major perceptual features, and a description of how these concepts relate and rank. We also evaluate scagnostics (scatter plot diagnostics), a popular and established set of scatter plot descriptors, and show that they do not reliably reproduce our participants judgements. Finally, we discuss the major implications of this study and how these results can be used for future research.