Identifying and forecasting research trends is of critical importance for a variety of stakeholders, including researchers, academic publishers, institutional funding bodies, companies operating in the innovation space and others. Currently, this task is performed either by domain experts, with the assistance of tools for exploring research data, or by automatic approaches. The constant increase of research data makes the second solution more appropriate, however automatic methods suffer from a number of limitations. For instance, they are unable to detect emerging but yet unlabelled research areas (e.g., Semantic Web before 2000). Furthermore, they usually quantify the popularity of a topic simply in terms of the number of related publications or authors for each year; hence they can provide good forecasts only on trends which have existed for at least 3-4 years. This doctoral work aims at solving these limitations by providing a novel approach for the early detection and forecasting of research trends that will take advantage of the rich variety of semantic relationships between research entities (e.g., authors, workshops, communities) and of social media data (e.g., tweets, blogs).
The research environment evolves rapidly: new potentially interesting research areas emerge regularly while others fade out, making it difficult to keep up with such dynamics. The ability to recognise important new trends in research and forecasting their future impact is however critical not just for obvious stakeholders, such as researchers, institutional funding bodies, academic publishers, and companies operating in the innovation space, but also for any organization whose survival and prosperity depends on its ability to remain at the forefront of innovation.
Currently, the task of understanding what the main emergent research areas are and
estimating their potential is usually accomplished by experts with the help of a
number of systems for making sense of research data. Systems such as Google Scholar,
FacetedDBLP [1] and CiteSeerX [
Currently, there are a number of approaches for detecting topic trends in a fully
automatic way [
The doctoral work presented here aims to solve the aforementioned limitations and produce a novel approach to detect and forecast research topics. This approach will be based on two main intuitions. First, I believe that by analysing the various dynamics of research it should be possible to detect a number of patterns that are correlated with the creation of new embryonic topics, not yet labelled. For example, the fact that a number of authors from previously unrelated research communities or topics are starting to collaborate together may suggest the emergence of a new interdisciplinary research area. Secondly, I theorize that taking into account the rich variety of semantic relationships between research entities (e.g., authors, workshops and communities) and analysing their diachronic evolution, it should become possible to forecast a topic impact in a much shorter timescale, e.g., 6-18 months. This holistic and semanticbased analysis of the research environment is today made possible by the abundance of both scholarly data and other sources of evidence about research, including social networks, blogs, and so on. 2
In many real-world contexts, being aware of research dynamics can bring significant benefits. Researchers need to be updated regularly on the evolution of research environments because they are interested in new trends related to their topics and potentially interesting new research areas. For academic publishers or editors knowing in advance new emerging topics is crucial for offering the most up to date and interesting contents. For example, an editor can gain a competitive advantage by being the first one to recognize the importance of a new trend and publish a special issue or a journal about it. Institutional funding bodies and companies need also to be aware of research developments and promising research trends. Thus, an automatic approach to detect novel topics and estimate their potential will bring significant advantages to a variety of stakeholders. Indeed support for this PhD project comes from SpringerVerlag, which is a global publishing company.
Several tools and approaches for the exploration of scholarly data already exist. From
the perspective of topic trend detection, we can classify these systems as either
semiautomatic or fully automatic. In particular, some systems for exploring the publication
space provide implicit support for semi-automatic trend detection, such as Google
Scholar, FacetedDBLP [1] and CiteSeerX [
In the context of providing a fully-automatic way to detecting topic trends, many
approaches assess the impact of a topic by simply using the number of publications or
patents directly associated with it. For example, Wu et al [
State of the art methods for forecasting trends in research take usually into
consideration the number of publications and authors associated with a topic [
Another important aspect that needs to be taken into account is how to represent a
topic. In literature, several ways to define a topic model can be found. The first is
characterised by the use of keywords as proxies for research topics. Systems like
MAS and Saffron [
Considering the gaps identified in the previous section, the main research question of the PhD will be: “How is it possible to detect the early emergence of new research topics and forecast their future impact?”.
This question entails two different challenges. The first one is how to detect very early research topics that may not even be labelled. The second one is how to forecast their impact with good accuracy. A specific set of sub-questions has been articulated in order to describe the process through which the doctoral work plans to answer the questions above.
Q1 – Finding the data. Understanding which data to integrate and exploit for the process is the first step. In particular, it is important to investigate the value of nonscholarly data (e.g., tweets, blogs, micro-posts, slides) in supporting trend detection and forecasting. As far as semantic technologies are concerned: how can research elements be gathered and connected by means of semantic relations?
Q2 – Detection of new emerging research topic. How is it possible to extract patterns in the evolution of research areas in order to predict the emergences of new ones? How can historical patterns be used to support the detection of future trends? Is it possible to develop a general approach able to consider the peculiarities of different fields (e.g., Computer Science, Business, Medicine and so on)? How emerging and unnamed research areas can be labelled? How social media can contribute in the detection of research trends?
Q3 – Forecasting of research trends. Can the impact of a research topic be measured just in terms of number of citations and publications? As soon as it has been defined, how the impact of research areas can be forecasted? What kind of forecasting approach should be adopted for research areas that do not yet exist? Which contribution can be given from the social media? 5
From a philosophical point of view, academic disciplines are specific branches of knowledge which together form the unity of knowledge that has been produced by the scientific endeavour. When two or more disciplines start to cooperate they share their theories, concepts, methods and tools. The results of this cooperation may lead either to the creation of a new interdisciplinary research area or simply to a contribution in knowledge from one area to another. The basic hypothesis is that the creation of a topic is thus anticipated by a number of dynamics involving a variety of research entities, such as other topics, research communities, authors, venues and so on. Therefore, recognizing these dynamics might enable a very early detection of emerging topics.
Scholarly data can be used to analyse a huge amount of research elements such as
papers, authors, affiliations, venues, topic and communities [
These explicit connections can be used to derive a number of second order connections, e.g. a topic is also associated with publication venues through relevant papers published in venues. These relationships can be analysed diachronically to derive the dynamics that led to the emergence of a topic and to estimate how they affect its future impact. For example, if two communities start to share research interests or authors, this may lead to the fact that a common new topic is developing. In a nutshell, the fundamental hypothesis at the basis of this PhD is that by exploiting the large variety of scholarly data which are now available, as well as modelling their semantic relationships, it will be possible to perform detection and forecasting of research trends even in a relative small interval of time. In addition, since many researchers are actively involved on social networks, I believe that analysing data from social media can also provide an effective support for the detection of research dynamics. 6
The approach is structured according to the proposed research questions. Basically, it is organised in four main steps.
Data integration. In this first phase I plan to integrate a variety of heterogeneous
data sources, including both scholarly metadata and less traditional sources of
knowledge, such as tweets, blogs post, slides and so on. The output will be a
comprehensive knowledge base containing both the research entities from Figure 1 and
entities from social media (authors’ profiles, number of followers, analytics, etc.). I will
identify topics and communities by extending state of the art techniques. In particular,
I plan to treat topics semantically, by describing their relationships using the topic
networks produced by the Klink algorithm [
The rich network of semantic relationship between the research elements will be described by an ontology and it will be populated by semi-automatic statistical methods. To build it, I plan to extend the topic network created by Klink with the research entities discussed in section 5 and their relationships. The analysis of these relationships and how they change in time will support the next steps of the approach.
Exploration of the Research Dynamics. In this step, the dynamics involving research elements correlated with the emergence of new topics will be investigated. To do so, I plan to verify empirically a number of hypotheses about these dynamics. In particular, I will analyse a number of topics which appear in the 2000-2010 interval and verify if their emergence is correlated with a number of dynamics, such as the raise of co-publications of related research areas, the increase of collaborations between authors of related areas, shifts of interests or migration phenomena in related communities, transfer of topics between related venues, and so on.
The output of this analysis will be a collection of patterns of knowledge flows associated with the creation of a new research area.
Early topic detection. This step aims to exploit the previously defined patterns for early research trend detection. To this end, I will build a number of distinct graphs, in which nodes represent a kind of research entity (e.g., topics) and the links are one of the elements of the dynamics, which were found in the previous phase – e.g., the increase in the number of collaborations between authors from two distinct topics. Highly connected sub-graphs, representing the area in which multiple entities exhibit the identified dynamics could thus suggest that a new discipline is emerging. In order to produce more robust evidence, I will use the semantic network of research entities to confirm that the emergence of a new topic is supported by a number of different ‘traces’ and research entities. For example, if a set of topics suggests that a correlated research area is emerging, the dynamics of the set of communities and venues related to these topics will also be checked. The intuition is that, while the evidence coming from a single dynamics or a single kind of entity could be biased or noisy, their combination should yield a more accurate result. The result will be a number of sets of linked entities, each one anticipating the emergence of a new topic. Different kinds of combination of entities and metrics will be tested, aiming to find the best approach to derive sets that are strongly correlated with the creation of new topics. At this stage, another challenge will be the definition of a method for labelling future research topics.
Trend forecasting. Initially, I will investigate different techniques to estimate the
impact of a topic, taking in consideration both basic metrics, such as the number of
publications and citations, and more complex indexes. As mentioned before, in
contrast with current approaches, [
A set of different machine learning methods, such as Artificial Neural Networks, Support Vector Machines and Deep Belief Networks, will exploit the extracted features in order to forecast the performance of a topic. 7
I plan to conduct an iterative evaluation during the different phases of my work using both quantitative and qualitative approaches.
From a quantitative point of view, I will evaluate both the ability of the system to identify novel topics and its accuracy to assess their impact in the following years. The discussed approaches will be compared with current methods and the difference between their performances will be measured via statistical tests. I will evaluate the detection of emerging trends in terms of recall, precision and F-measure using crossvalidation on historical data. Similarly, I will assess the agreement between the estimated and the real impact of a research area.
In the qualitative evaluation, the achieved results will be compared with experts’ opinions in order to measure its reliability. I will prepare a number of surveys for domain experts with questions both about the past - such as the main topics recently emerged in their area of expertise - and about the future - such as the research areas which seem on the verge of being created and an estimation of their likely impact. 8
This paper presents the goal of my doctoral work, which is currently at an early stage (month 6). As discussed, I intend to produce a new approach for detecting and forecasting research trends, which is based on a semantic characterization of research entities, on the statistical analysis of research dynamics and on the integration of scholarly and social media data.
Currently I am investigating a number of knowledge sources for selecting the ones more apt to support my approach. At the same time I am using an initial dataset to test the hypotheses about research dynamics discussed in section 6. The next step will be the creation of an approach for extracting highly connected sub-graphs of entities exhibiting dynamics associated with the emergence of new topics.
I would like to acknowledge my advisors, Enrico Motta and Francesco Osborne, as well as my sponsors, Springer-Verlag GMBH for supporting my research. 1. Diederich J, Balke W-T, Thaden U Demonstrating the semantic growbag: automatically creating topic facets for faceteddblp. In: Proceedings of the 7th ACM/IEEE-CS joint conference on Digital libraries, 2007. ACM, pp 505-505