Screenshot of the temporal distribtion chart. 

Context in source publication

Context 1

... this implementation the semantic social network analysis [13] module of Condor, the successor of TecFlow [28], is extended. Condor analyzes and visualizes communication and interaction networks. The data representing the network structure needs to be available in electronic format. Condor natively supports different data sources. Email archives can be imported from Eudora, Microsoft Outlook, or directly from an IMAP server. Weblinks and blogs can be accessed via Google’s blogsearch and Microsoft’s live search API. Data gathered with Social Badges can also be loaded into Condor [29]. Not natively supported data sources can be loaded into Condor via flat files or by parsing the data directly into a MySQL database. In extension to a static analysis of interaction networks, Condor and its predecessors can be used to study dynamic networks and their evolution over time. Monitoring social networks over time helps understanding the evolution of relationships in the networks [15]. Condor visualizes the social networks with a spring-embedder model developed by Fruchterman and Reingold [30]. This algorithm enhances Eades method [31] to places the nodes and the edges of a network on a two-dimensional plane. The screenshot in figure 3 shows the resulting structure of the force-directed algorithm. An important feature of Condor is its ability to process and visualize temporal information on social networks. Especially when analyzing the content of the communication the temporal distribution of terms used by actors in the network is of interest. An increased use of terms pooled in certain clusters at one moment in time could point to the topics discussed in the social network during that time. To support users in assessing the prominence of single nodes in the taxonomies and the concepts these nodes represent this implementation allows to view the temporal distribution of each node in a chart. Figure 1 shows an example of such a distribution chart. In order to contrast the cluster scrutinized by the user to all the other clusters in the taxonomy the distribution of all terms is shown in each chart too. Values used in the temporal distribution chart are weighted with the terms’ importance scores as described in section 4.3. By using those importance scores instead of the bare numbers of term occurrences, the betweenness centrality of the actors using the terms is factored in. Utilizing the betweenness based importance of each term punishes those terms and concepts used by less important actors. An additional view on the data in each cluster is the list of the most important documents shown in figure 2. For each document a summary as well as the sender’s name and the submission date is given. The documents are ordered by their importance for the terms in the selected cluster. On the left side of figure 2 an additional window with the content of the selected document can be seen. The evaluation of clustering algorithms can be conducted with statistical measures as introduced in [32]. Those methods assess the quality of the clustering algorithms. However, the quality of the resulting clustering needs to be judged by human users. A reduced Enron dataset 4 is used to show the functionality of the developed module. This dataset consist of emails collected from Enron employees during the Enron scandal. Figure 3 shows the top-level clusters of this dataset. Important facts on the Enron scandal can be grasped at a glance without further knowledge on the background of this dataset. The system identifies Enron’s Executive Vice President Steven J. Kean as one of the key players in the scandal in which attorneys and New York played a crucial role. A combination of information retrieval means and social network analysis techniques is introduced in this paper. The aim is to reveal discussed topics in social networks like email archives and their relationships among each other. Instead of relying only on information retrieval techniques the structure of the underlying social network is taken into account. Foundations of information retrieval and social network analysis are described, techniques of both fields are combined to obtain taxonomies of the topics discussed in the communication of social networks. Instead of solely relying on methods of IR when determining concepts discussed in the communication archives, the structure of the underlying network is respected by factoring in SNA key measures. Unsupervised clustering algorithms are used to scrutinize the content of communication in social networks. Classification of terms with those algorithms is a new approach to gain insights on communication networks. The resulting taxonomies of terms can be used to obtain an overview on the whole communication network at a glance. A temporal analysis module allows assessing the development of discussed topics over time. Finally, the design and details of the implementation are presented. The aim of this paper was to provide users of social network analysis packages like Condor with an automated method to reveal topics discussed in analyzed networks and their hidden relations among each other. First steps were made in this paper to allow users to gain an impression on the nature of discussions in analyzed networks. This work can only serve as a step in the right direction of automatically revealing discussed topics in social networks. Supporting users with enhanced automated or semi-automated methods to analyze the content of social networks is crucial with more and more data available on social ...