BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//UNIFR/WEBMASTER//NONSGML v1.0//EN
CALSCALE:GREGORIAN
BEGIN:VEVENT
DTSTART;VALUE=DATE:20231103T141500
DTEND;VALUE=DATE:20231103T141500
UID:14632@agenda.unifr.ch
DESCRIPTION:This thesis mainly introduces two topics in complexity science: Ranking of\nresearchers and optimal learning in complex networks, both of which help us \nunderstand the ever-increasing complexity of the world. First, we design a novel \nmethod for ranking or evaluating researchers in citation networks, which is very \nvital for the scientific community. A well-designed ranking method can be used \nto rank scientists in various practical tasks, such as hiring, funding application \nand promotion. However, a large number of evaluation methods are designed \nbased on citation counts which can merely evaluate scientists’ scientific \nimpact but can not evaluate their innovation ability which actually is a crucial \ncharacteristic for scientists. In addition, when evaluating scientists, it has \nbecome increasingly common to only focus on their representative works \nrather than all their papers. Accordingly, we here propose a hybrid method by \ncombining scientific impact with innovation under representative works \nframework to rank scientists. Our results show that the ranking performance of \nthe hybrid method is the best compared with other mainstream methods. This \nstudy can provide policy makers an effective way to rank scientists from more \ncomprehensive dimensions.\nBesides, taking into account the time bias issue in ranking or evaluating the \nlong-term or lifetime achievements of researchers, a competition-aware\nranking method for researchers is proposed. Since the number of scholars and \nthe number of scholarly outputs grow exponentially with time, a well-designed \nunbiased ranking metric for researchers is of prime importance. To rank \nscholars, it is important to put their achievements in perspective by comparing \nthem with the achievements of other scholars active in the same period. We \npropose here a particular way of doing so: by computing the evaluated \nscholar's share on each year's citations which quantifies how the scholar fares \nin competition with the others. Our results show that the new ranking method \nsignificantly outperforms other ranking methods in identifying the prize \nlaureates.\nSecond, we study the optimal learning in information networks. Similar to \nranking for researchers, optimal learning is also a significant topic in complexity \nscience. The ever-increasing complexity of the world around us challenges our \nability to understand it. We study a model where a single agent learns node \ntypes or say forms opinions about many interconnected topics. This model was \nshown to be challenging for a boundedly rational agent. We extend it by \nassuming that the agent combines a targeted study of some topics and various \nheuristics for the remaining ones. We find that the highest opinion accuracy is \ngenerally achieved neither when one topic is studied very well nor when many \ntopics are studied a little. Despite big differences in accuracy between the \nconsidered heuristics, the optimal number of topics in which the study budget \nis invested grows linearly with the budget for all of them. The study budget \nnecessary to achieve the desired opinion accuracy exhibits a simple scaling \nwith the total number of topics. In this way, we exemplify how to use limited \ncognitive effort for efficient learning in a complex system.
SUMMARY:Ranking and learning in citation and information networks
CATEGORIES:Soutenance de mémoire/thèse
LOCATION:PER 08\, 2.73\, Chemin du Musée 3\, 1700 Fribourg
URL;VALUE=URI:https://agenda.unifr.ch/e/fr/14632
END:VEVENT
END:VCALENDAR