KEYNOTES

ROGER GUIMERA Rovira i Virgili University, Tarragona, Catalonia
Social networks: From description to prediction using network inference
Social actors interact with each other through complex networks that are neither perfectly regular nor completely random. Social scientists have studied these networks for over half a century, but only recently, with the advent of information and communication technologies, have we been able to systematically and quantitatively explore the large-scale structure social systems. So far, most of this research has been descriptive, that is, it has focused on describing the structure of social networks. In the last few years, however, we have developed network models that are amenable to statistical treatment and that enable us to be predictive, that is, to make quantitative predictions about the structure and evolution of social systems. In our talk, we will discuss some of these approaches to network inference. We will also discuss how these approaches can be applied to problems as diverse as the prediction of conflict within work teams, the prediction of decisions of judges, or the prediction of user ratings on movies and books.

LUDO WALTMAN Leiden University, Leiden, Netherlands
Comparing scientific performance across disciplines: Methodological and conceptual challenges
Quantitative measurement of scientific performance has become a pervasive phenomenon. Research institutions are benchmarked in university rankings, individual researchers are compared using h-indices and other similar types of metrics, and scientific journals compete with each other to have the highest impact factor. What is the scientific basis of these different approaches to the measurement of scientific performance? Do these approaches really allow us to make meaningful statements on the performance of journals, institutions, and individual researchers, even if they are active in different scientific fields with different publication andcitation practices? In my talk, I will critically reflect on commonly used metrics of scientific performance, and I will illustrate some important pitfalls in the use of these metrics. I will also show a number of more advanced approaches to the measurement of scientific performance. These approaches have been developed by bibliometric research centers, including my own center at Leiden University. In particular, I will focus on the CWTS Leiden Ranking, a worldwide ranking of universities based on a sophisticated bibliometric methodology, and the SNIP indicator, an alternative to the journal impact factor developed at Leiden University and available in the Scopus database. I will point out the advantages of these advanced bibliometric approaches to the measurement of scientific performance, but I will also highlight their intrinsic limitations. Finally, based on the recently published Leiden Manifesto for research metrics, I will discuss a number of good practices in the use of quantitative metrics in assessing scientific research.

IGOR MOZETIČ Jožef Stefan Institute, Ljubljana, Slovenia
Social media analytics: The role of sentiment
We present several studies of sentiment analysis, applied to different media (Twitter, Facebook, news and blogs), in different languages, and to different domains. The main issue is to find and quantify a relationship between a social media and another complex system. We combine text mining, network analysis and standard statistical methods to uncover and highlight important relations between different systems. In particular, we will demonstrate the role of sentiment in monitoring of political elections, detecting abnormal returns of stocks, spreading of conspiracy theories, identifying influential communities and their leaning towards environmental issues, ranking and mapping emojis by their sentiment, comparing everyday to major news events, and finally, in the evolution of the recent refugees crisis in Europe.

MARKUS ABEL Ambrosys GmbH, Potsdam, Germany and Potsdam University, Potsdam, Germany
Machine Learning for power production forecast
In recent years big data and machine learning have buzzed around the world. Methods have been designed by global players, in particular for semantic analysis of data. These methods can be used for numerical data, too, which are often not big, but meaningful. The presented approach is to use machine intelligence paired with physical insight in the systems considered. As an example, it is shown how a forecast system can be 
designed based on good features and suitable methods. A special challenge is the spatio-temporal analysis of data: one has to reduce the possible “words”, i.e. temporal sequences from different locations to useful “words”, i.e. few sequences which characterize a certain property of the system considered – a kind of clustering is required. We use facts known from turbulence theory to identify good features for wind energy data. Our focus lies on symbolic regression methods and stochastic modeling, but a set of other methods are used for comparison.