James Davey

Guiding feature subset selection with an interactive visualization

We propose a method for the semi-automated refinement of the results of feature subset selection algorithms. Feature subset selection is a preliminary step in data analysis which identifies the most useful subset of features (columns) in a data table. So-called filter techniques use statistical ranking measures for the correlation of features. Usually a measure is applied to all entities (rows) of a data table. However, the differing contributions of subsets of data entities are masked by statistical aggregation. Feature and entity subset selection are, thus, highly interdependent. Due to the difficulty in visualizing a high-dimensional data table, most feature subset selection algorithms are applied as a black box at the outset of an analysis. Our visualization technique, SmartStripes, allows users to step into the feature subset selection process. It enables the investigation of dependencies and interdependencies between different feature and entity subsets. A user may even choose to control the iterations manually, taking into account the ranking measures, the contributions of different entity subsets, as well as the semantics of the features.

Using Signposts for Navigation in Large Graphs

In this paper we present a new Focus & Context technique for the exploration of large, abstract graphs. Most Focus & Context techniques present context in a visual way. In contrast, our technique uses a symbolic representation: while the focus is a set of visible nodes, labelled signposts provide cues for the context — off‐screen regions of the graph — and indicate the direction of the shortest path linking the visible nodes to these regions. We show how the regions are defined and how they are selected dynamically, depending on the visible nodes. To define the set of visible nodes we use an approach developed by van Ham and Perer that dynamically extracts a subgraph based on an initial focal node and a degree‐of‐interest function. This approach is extended to support multiple focal nodes. With the symbolic visualization, potentially interesting regions of a graph may be represented with a very small visual footprint. We conclude the paper with an initial user study to evaluate the effectiveness of the signposts for navigation tasks.

Visual Analysis of Sets of Heterogeneous Matrices Using Projection-Based Distance Functions and Semantic Zoom

Matrix visualization is an established technique in the analysis of relational data. It is applicable to large, dense networks, where node‐link representations may not be effective. Recently, domains have emerged in which the comparative analysis of sets of matrices of potentially varying size is relevant. For example, to monitor computer network traffic a dynamic set of hosts and their peer‐to‐peer connections on different ports must be analysed. A matrix visualization focused on the display of one matrix at a time cannot cope with this task.

Smart Grid monitoring through visual analysis

In this paper we present a visualization system for the real-time monitoring of Smart Grids. In particular, it supports the control room operators of electric grids with large amounts of distributed power generation. As measurements are continuously read from connected Smart Meters, an expert system performs a classification of these events to relieve the operator from the manual inspection of irrelevant and trivial information items. A coordinated, multiple view environment displays the electric grid model which is annotated with status indicators. This gives the operator both a reasonable overview on the entire system, but also enables him to inspect specific parts in detail. Other sources of information such as ICT coverage and weather conditions are included to provide additional context. The visualization system thus provides monitoring and control support for operators to keep the grid in a stable condition.

A Visual Analytics Field Experiment to Evaluate Alternative Visualizations for Cyber Security Applications

The analysis and exploration of emerging threats in the Internet is important to better understand the behaviour of attackers and develop new methods to enhance cyber security. Fully automated algorithms alone are often not capable of providing actionable insights about the threat landscape. We therefore combine a multi-criteria clustering algorithm, tailor-made for the identification of such attack campaigns with three interactive visualizations, namely treemap representations, interactive node-link diagrams, and chord diagrams, to allow the analysts to visually explore and make sense of the resulting multi-dimensional clusters. To demonstrate the potential of the system, we share our lessons learned in conducting a field experiment with experts in a security response team and show how it helped them to gain new insights into various threat landscapes.

Visual analytics: towards intelligent interactive internet and security solutions

In the Future Internet, Big Data can not only be found in the amount of traffic, logs or alerts of the network infrastructure, but also on the content side. While the term Big Data refers to the increase in available data, this implicitly means that we must deal with problems at a larger scale and thus hints at scalability issues in the analysis of such data sets. Visual Analytics is an enabling technology, that offers new ways of extracting information from Big Data through intelligent, interactive internet and security solutions. It derives its effectiveness both from scalable analysis algorithms, that allow processing of large data sets, and from scalable visualizations. These visualizations take advantage of human background knowledge and pattern detection capabilities to find yet unknown patterns, to detect trends and to relate these findings to a holistic view on the problems. Besides discussing the origins of Visual Analytics, this paper presents concrete examples of how the two facets, content and infrastructure, of the Future Internet can benefit from Visual Analytics. In conclusion, it is the confluence of both technologies that will open up new opportunities for businesses, e-governance and the public.

Matrix-based visual correlation analysis on large timeseries data

In recent years, the quantity of time series data generated in a wide variety of domains grown consistently. Thus, it is difficult for analysts to process and understand this overwhelming amount of data. In the specific case of time series data another problem arises: time series can be highly interrelated. This problem becomes even more challenging when a set of parameters influences the progression of a time series. However, while most visual analysis techniques support the analysis of short time periods, e.g. one day or one week, they fail to visualize large-scale time series, ranging over one year or more. In our approach we present a time series matrix visualization that tackles this problem. Its primary advantages are that it scales to a large number of time series with different start and end points and allows for the visual comparison / correlation analysis of a set of influencing factors. To evaluate our approach, we applied our technique to a real-world data set, showing the impact of local weather conditions on the efficiency of photovoltaic power plants.

Combining statistical independence testing, visual attribute selection and automated analysis to find relevant attributes for classification

We present an iterative strategy for finding a relevant subset of attributes for the purpose of classification in high-dimensional, heterogeneous data sets. The attribute subset is used for the construction of a classifier function. In order to cope with the challenge of scalability, the analysis is split into an overview of all attributes and a detailed analysis of small groups of attributes. The overview provides generic information on statistical dependencies between attributes. With this information the user can select groups of attributes and an analytical method for their detailed analysis. The detailed analysis involves the identification of redundant attributes (via classification or regression) and the creation of summarizing attributes (via clustering or dimension reduction). Our strategy does not prescribe specific analytical methods. Instead, we recursively combine the results of different methods to find or generate a subset of attributes to use for classification.