Laura Steinert

Where to Begin? Using Network Analytics for the Recommendation of Scientific Papers

This paper proposes a network analytic approach for scientific paper recommendations to researchers and academic learners. The proposed approach makes use of the similarity between citing and cited papers to eliminate irrelevant citations. This is achieved by combining both content-related and network-based similarities. The process of selecting recommendations is inspired by the ways researchers adopt in literature search, i.e. traversing certain paths in a citation network by omitting others. In this paper, we present the application of the newly devised algorithm to provide paper recommendations. To evaluate the results, we conducted a study in which human raters evaluated the paper recommendations and the ratings were compared to the results of other network analytic algorithms (such as Main Path Analysis and Modularity Clustering) and a well known recommendation algorithm (Collaborative Filtering). The evaluation shows that the newly devised algorithm yields good results comparable to those generated by Collaborative Filtering and exceeds those of the other network analytic algorithms.

Incremental Clustering of Dynamic Bipartite Networks

This paper deals with the problem of identifying clusters in evolving bipartite networks over time. In bipartite networks there exist two types of nodes while ties can only occur between nodes of different types. Hence, a cluster in a bipartite network consists of two node sets for the two node types each. A major challenge regarding the evolution of those clusters over time is that the two parts of a bipartite cluster may evolve independently. While there is already an increasing amount of research on the identification of clusters in dynamic unipartite networks, the bipartite case is still underrepresented. After a clear motivation of the problem, an adaptation of an existing method for optimising modularity in unipartite networks is extended to dynamic bipartite networks. The method is evaluated on computer generated as well as real world networks.

Implementation of a Decision Making Algorithm Based on Somatic Markers on the Nao Robot

Decision making is an essential part of Autonomous Mobile Systems. Research shows that emotion is an important factor in human decision making. Therefore an increasing number of approaches using modelled emotions for decision making are developed for artificial intelligent systems. Often those approaches are only evaluated in simulated environments in which dummies are used to represent actions. However, the realisation of a real robot application also requires the handling of problems which may not occur in a simulated environment, such as long execution times. Furthermore, the adaption of existing approaches to variant applications often includes several time-consuming adjustments to the system. In this paper the implementation of an emotional decision making algorithm for the Nao robot is presented. The implementation design is based on the human brain structure and models different brain parts which are included in the decision making process. Beside the fact that the chosen structure is closer to the human model, the modular architecture allows an easy implementation of enhancements or different approaches. A key point is the easy adaption of the approach to different applications, suitable even for users without technical expertise or programming skills. As an example, a possible real life scenario is used, in which the robot is embedded in a social environment.

Relational Patterns in Cross-Media Information Diffusion Networks

This paper describes an approach for identifying patterns of information diffusion across different media types on the web. In this context, a novel sampling and crawling strategy for social media content has been applied to extract contributions relevant to certain news events. Contributions can be any kind of published content in different information channels, including tweets (Twitter), web pages, or revisions of Wikipedia articles. Contributions can be interlinked by hyperlinks, revision links, or retweet relationships, and thus constitute a diffusion network with unidirectional links indicating influence. Our approach reduces the original possibly very large and complex diffusion network to its basic underlying structure by applying non-negative matrix factorization to group the nodes with similar positions in the diffusion network. Beyond focusing only on the spread of a news item through different channels, also the temporal aspect, especially delay, is explicitly taken into account.

A Comparative Analysis of Network-Based Similarity Measures for Scientific Paper Recommendations

In this paper three similarity measures for scientific papers are compared: bibliographic coupling, co-citation coupling and cosine similarity. All three measures are based on the connections of papers in citation networks. The comparison is conducted both on a mathematical as well as an empirical level. The latter is performed on a real citation network as well as artificially generated networks. The mathematical comparison shows that some measures are structurally very similar, yet if node pairs are ordered according to their similarity, the two measures do not always produce the same rankings. The empirical evaluation shows that bibliographic coupling and one variant of cosine similarity tend to produce the same rankings. The same holds for co-citation coupling and another variant of cosine similarity. Therefore, if only rankings are considered, these measures are interchangeable. The rankings produced by co-citation coupling and bibliographic coupling on the other hand are very different. This also applies to the two cosine similarity variants. Therefore, these measures are not interchangeable.

What Makes a Good Recommendation

In this paper we propose several new measures to characterize sets of scientific papers that provide an overview of a scientific topic. We present a study in which experts were asked to name such papers for one of their areas of expertise and apply the measures to characterize the paper selections. The results are compared to the measured values for random paper selections. We find that the expert selected sets of papers can be characterized to have a moderately high diversity, moderately high coverage and each paper in the set has on average a high prototypicality.