We focus on estimating a research area of a researcher/user by nding a unique identity in digital libraries and social networks and by analyse of public metadata of their publications and published information on social networks pro les. The lack of content of the metadata in some of the publications is solved by the information retrieval using techniques of NLP. We estimate the author's domain by extracting keywords from abstracts as well as by information published on social pro les. The result of this work is a design, an original algorithm and experimental veri cation of the algorithm.
There are situations in life when we need to nd works of a speci c researcher, for example when we organize a conference. One of the most common way to solve this problem is to search for the information about this researcher, either by looking at the institutions and his publications or by examining the topics he had on various conferences, and then create a pro le of the researcher manually. With the boom of social networking people began to publish more openly accessible data than before. Using the data may reveal an interesting complement to the true identity of a person. Unfortunately, the expansion and the emergence of various social networks caused a relatively large fragmentation and users publish speci c information about themselves to a social network focusing on the speci c topic. The fact that people can have the same name is another obstacle, therefore it is necessary to verify that it is actually a pro le of the right person and not of his namesake. The main objective of this work is to identify researchers on social networks and digital libraries. Based on the public information on these sites, we estimate the area of a person's research. The results are keywords that serve both as a description of the person and as an input for further research in nding suitable reviewers of publications presented at conferences and for detecting the violations of a copyright.
To nd the right pro les we used a technique which compares
speci c attributes by di erent weights. Details are described
in [
i=0 wi sim(ai;u; ai;p) if sim(aname) > thname 0 otherwise (1) where sim(aname) is similarity between author and user prole names, thname is threshold value to decide if names are the same or not, n is count of compared attributes, wi is weight of compared attributes, ap is set of user's pro le attributes, au is set of user's attributes by his publications, sim(ai;u; ai;p) is similarity between attributes. The text comparison is done by fuzzy matching to include potential typing errors in attributes.
As shown in the Algorithm 1 the input is the name of the
researcher. Then the search requests to all the digital libraries
are executed and it downloads the publications. Each
publication is then categorized by the de ned criteria. Initially
we eliminated all the articles that were similar or equal and
were occurring in multiple libraries. Then we categorize the
publications by a liations using the text similarity
algorithm and also by their co-authors. Now we have groups
of possible unique authors. There is an issue now with the
author publishing on his own or being active in multiple
afliations, because then the algorithm divides him into more
groups. To handle the situation, we included a
comparison with user's connections retrieved from social networks
and additional information about skills, experiences and so
on. After that we categorized the keywords by social pro le
similarity. We found all the pro les associated with the
researcher name. Then we tried to nd common connections
and a liations, and if there were at least one in each pair we
would assign them together with the compared social
proles. The process was repeated for every found co-authors
and referred publications with the input of the previously
found authors, so the results would be more accurate.
People with the same name are not merged into one identity,
because of the classi cation by connections and a liates. It
is highly unlikely that these people will have the same
coauthors, friends and jobs. More information about a unique
user identity is described in [
The research domain is obtained by analysing the keywords of all found publications and by extracting the additional information from the social pro les. Because of lacking and incompletely chosen keywords in many publications we had to use our original technique to get additional keywords from abstract. We do not go into detail describing our original technique, because of the page limit of this poster. Data: Author's rst name and last name Result: User's identities
rstName, lastName fuser raw inputg; for searcher in DigitalLibrariesSearchers do
publications SearchAuthor( rstName, lastName); end GroupByPublication(publications); GroupByA liates(publications); for searcher in SocialNetworkSearchers do
publications SearchAuthor( rstName, lastName); end for group in groups do for publication in publications do groupKeywords +=
AnalyzePublication(publication); end end
nalGroups = GroupBySocialPro les(groups); Algorithm 1: Finding unique author identity on digital libraries and social networks
From the digital libraries we chose IEEExplorer1, ACM Digital Library2 and SpringerLink3. In this work, the researchers are found on LinkedIn4 and Researchgate5 social networks. In the experiment we check if we can nd unique identities and research domains of 180 randomly selected researchers. The search of user identities in digital libraries has been tested by at least 180 researchers, by downloading and analysing about 3100 publications (Table 1). The researchers were chosen randomly and included people of different nationalities. Initially there were users grouped only by co-authors and a liates. There were 118 authors grouped correctly ("R") with rate 65 %. 3 authors had assigned other author's publications ("POA", 2 % error rate) because of fact that searched author had publications with namesake co-authors and it was poorly evaluated as same person, error rate in this case was 59 authors were not merged correctly ("NA"), there were too many created identities of which should be same one author. This was caused by publications with no or one co-author and di erent a liations, it was not possible to nd connection between them. Error rate of this category was 33 %. In the next step we included comparisons of authors by data found on their social pro les. 132 users were identi ed correctly (73 %) and to 3 same authors were again assigned wrong publications due to the same reasons, error rate remained 2 %. The only improvements were made in the case when one author was in two di erent groups ("NA") and when there were connections found in social pro les between them, so error rate decreased to 25 %.Finally we added comparisons by keywords between publications with a single author and publications with multiple authors. 166 users were identi ed correctly, correct rate increased to 92 %. There was no situation with a one author in more groups ("NA", error rate of this category decreased to 0 %). Unfortunately 14 users had assigned wrong publications ("POA"), error rate increased to 8 %. It was caused by errors in extracting of keywords and the associated bad detection of a similarity between researchers and publications.
The goal of our work was to create algorithm to estimate research area of users by nding their identities in digital library and social networks and by analyse found data. As the results from our experiment show, the algorithm for identifying research identities on digital libraries and social networks was successful in 92 % of all the attempts in nal. This work was the rst step in the research of recommending publications to authors and nding violations of copyrights. We would want to try to add comparing authors' domains detected from publications and information on the Internet to classic full-text search approach. This work is input for further research in nding suitable reviewers of publications presented at conferences and for detecting the violations of a copyright.
The following grant is acknowledged for the nancial support provided for this research: Grant of SGS No. SP2014/42, VSB - Technical University of Ostrava, Czech Republic.