Aristotle University Of Thessaloniki (AUTH)
Complex networks and machine learning in scientometrics
Abstract
dc:descriptionComplex networks constitute a structure of interconnected entities with non-trivial topological properties and are often represented as graphs. In real world social processes, as is scientific activity for instance, multiple such networks are formed and the complexity of their underlying dynamics rises exponentially with the data volume that can be collected about them. Scientometrics, on the other hand, entails the systematic quantification and evaluation of the scientific process and has given birth to a new research avenue referred to as “the science of science”. Hitherto, the scientific enterprise has been studied from various perspectives both qualitative and quantitative. In this dissertation, we adhere to the data mining approach to extract data intelligence from the graphs formulated around scientific entities and unravel their underlying interrelations. In this context, we propose novel analytics approaches to quantify, evaluate and finally predict scientific evolution. The contribution of this dissertation is three-fold:First, we propose a versatile evaluation process for the scientific community based on correlation analysis amongst the plethora of existing scientometric indicators. From an adjustable set of criteria, that capture different aspects of scientific activity, Pareto frontiers are produced in the form of scientific rankings, to assist tenure committees, grant allocation, award giving and other decision support scenarios. Additionally, new metrics were introduced to incorporate network and evolutionary properties of scientific entities into the evaluation process. As scientometric indicators are utilized in decision making, the need to improve them and broaden their focus is imminent. Taking the quantification of scientific evolution one step further we reach out to identify scientific profiles and trends. The fractal dimension of citation acquisition process is introduced to characterize publishing patterns and incorporate complex network properties (Matthew effect, power laws, etc.) into a single number metric. Moreover, citation pattern mining is employed to identify trends and shifts in the citation accumulation process. Since scientific activity is not a static process, methods for evaluation ought to capture the evolution patterns and comparisons need to be performed on a common basis. Therefore, we created profiles and extracted temporal patterns, while we identified their correlation with overall scientific impact to discover if a pattern for success exists. We proposed a relative ranking and impact characterization of scientists amongst their group of peers to distinguish prominent members of the scientific community (e.g. award winners).But it is not only vital to assess the past behavior; we need to gain insights on the future status of scientific enterprise. In this direction, we created a taxonomy of existing approaches to predict future scientific impact, identified their shortcomings and proposed a unified adjustable framework to mitigate them. In this approach we incorporated dynamic clustering of scientists with their impact cohort and profiling methods using a graph based predictive model to estimate future citation counts. Through extensive experimentation across time, age and impact levels, while combining scientists’ relative position in their surrounding graphs, we extracted meaningful conclusions on the mechanism with which science evolves over time. This way data intelligence can be obtained from the abundance of bibliographic data and provide assistance to publishers, scholars, institutions, policy makers and other science stakeholders.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Gogoglou, Antonia
- Γώγογλου, Αντωνία
Subjects
dc:subject × 6Rights
dc:rights- Statement dc:rights
-
- BY_NC
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/42463