An algorithm for topic modeling of researchers based on their interests from Google Scholar's profiles is proposed. As topics for modeling, we took research groups from research classification system ANZSRC - Australian and New Zealand Standard Research Classification. Researchers' distribution to research groups is found based on their interests' usage statistics in categorized publications from Dimensions. Topic modeling is conducted accordingly to principles of statistical support, multi-labeling, noise filtering, ignoring stopwords, solidarities, focusing, compactness and research groups' interactions. We compare topic modeling based on data with low level of information from researchers' profiles in Google Scholar with topic modeling based on a few dozen authored publications categorized by Dimensions. Comparison is made by modified Czekanowski metric that takes into account the interaction between research groups. By comparing the results of topic modeling based on different sources of initial information a good match was found. It allows to use the proposed algorithm as the intellectual core of information technology in regards to scientific staff, in particular, for the selection of candidates as opponents of a dissertation, as reviewers for research projects, for forming a team to collaborate on mutual research projects etc.
Google Scholar aggregates the most volumetric collection of researchers’ profiles. The most used information from Google Scholar profiles is citations. It, for example, is used as primary information for university ranking in Webometrics. Several studies, in particular [1, 2], are concerned with comparing concordance of Google Scholar citations with different scientometrics systems such as Scopus, Web of Science, Dimensions and others, that use only meta-information from publishers. A researcher’s profile in Google Scholar contains not only publications and their citations, but also other information. In particular, a researcher provides his or her interests. A researcher chooses the interests in a loose manner without any limitations. Google Scholar provides a web interface to search researchers by an interest. However, the results are formed by literal coincidence. That is why the results for fuzzy set and fuzzy sets are different; the same applies for synonymous interests such as fuzzy evidence and fuzzy inference. Moreover, Google Scholar does not take into account an interconnection of the interests, that is the search by an interest is done independently and isolated. Given that, the search and analytical services that provide information about many researchers in Google Scholar are relatively straightforward.
The goal of this paper is topic modeling of researchers based on their interests from Google Scholar. Methods that process a researcher’s interests from Google Scholar profile are not studied well. We identified only two relevant publications. One of them is [3], it describes a recommendation system that recommends supervisors based on some information and interests of candidates from Google Scholar profiles as well. Another paper [4] presents an information technology that synthesizes a research profile of institute or research laboratory. It also uses interests of researchers from their profiles at Google Scholar. Articles [3, 4] are based on pairwise comparison using cosine similarity metric between researcher and a set of keywords from a given topic. Such a topic in [3] is an article at Wikipedia. Unlike these methods, we strive to categorize researchers by a given research classification system, that is to assign a research group to each of them.
Automatic researchers’ categorization is usually done as a result of generalizing the topics of their publications. One of the methods for this is presented at [5]. The authors present a statistical model “Author-Topic” that is based on topic modeling model Latent Dirichlet Allocation [6]. This model represents a researcher as a distribution over some abstract topics. The topics are clusters of similar words. One of the drawbacks of this model is low interpretation of the topics because they are formed by words frequency in a document. To improve the interpretation another model “Author-SubjectTopic” is proposed in [7]. This model additionally uses a research specialty that is defined by journal in which an analyzed publication is published. In [8] another improvement of “Author-Topic” model is presented – “Author-Persona-Topic” model. In this model rather than representing all researcher’s documents as single topic distribution, authors group all documents into different clusters, each with its own topic distribution. These clusters represent “personas” under which an author writes.
Apart from topic modeling methods there are also methods based on word embedding. They generally perform better than topic models because they can incorporate semantic relationships. One of the most popular models of embedding is word2vec [13]. It is used in [9] as a part of similarity metric between researchers using their publications. They assess the similarity between words that comprise publications of different researchers using representation of words defined by word2vec. The authors of [10] use publications’ titles as source information for solving the problem of collaboration recommendation. The words from the titles are represented as vectors using word2vec. These vectors are then clustered using k-means to partition researchers into different academic domains. The representation of a researcher is further improved by using his co-authorship and the random walk method to find his influence in different domains. In [11] authors represent a researcher as a set of documents he/she has written. The words of the documents are defined as vectors trained by word2vec model. These representations are then used to solve the problem of expert finding by utilizing a restricted convolutional neural network. In [12] a researcher is represented as a concatenation of all his/hers abstracts. Each word in the concatenation is then represented as a vector from word2vec model to solve the problem of reviewer recommendation.
Analyzed methods assume to have enough number of publications for a given researcher with selected keywords. At the same time, they do not account for the fact that co-author contribution is sometimes relative to a small subset of the paper keywords. A researcher, especially a young one, may have only a few publications that may not be enough for a valid categorization. On the other hand, the researcher can manually specify at the profile a set of keywords that describe his (or her) activities. As the time goes on a researcher may change his research direction, for example, move to another laboratory or another project. Given that there is no change when a researcher is categorized based on his publications, categorization based on his keywords may find these changes. By taking that in mind, we study the topic modeling based on actual interests that a researcher specified by himself (or herself) at the current moment.
We use the following notations:
W w1, w2 ,..., wn is a set of keywords that are equal to researcher interests in Google Scholar profile;
T t1, t2 ,...,tm is a set of research topics from a research classification system; D1, D2 ,..., Dm is a set of topic-marked collections of texts; each collection contains only publications from topics t1, t2 ,..., tm , respectively;
B D1 D2 ... Dm is the general collection of topic-marked texts; each element of B belongs to one or more topics from the set T ;
R(D,T ) D T is a relation that describes membership of a publication to topic-marked collections.
The problem is to find out topics from T that correspond to the set of interests W . The results of ~ mapping W T is a fuzzy set W defined on the universal set of topics T as follows:
W~ Wt1(t1) , Wt2(t2 ) ,..., Wtm(tm ) ,
the cardinality of the fuzzy set support must be small 1 supW~ Tmax , for example, with Tmax 2,3,4a researcher will be assigned only to a few topics; W t p 1 , which is equivalent to the topic modeling regularization condition.
____ p1,m
We use a researcher’s profile from Google Scholar to get the keywords. For example, in Figure 1 we have a researcher’s profile with three keywords that are marked with blue color. For this researcher: w1 "Computational Intelligence" ; w2 " Fuzzy Logic" ; w3 " Artificial Intelligence" . The order of keywords in the set W is not important, this corresponds to the bag of words model. Interests often complement each other thus making their research topics more focused. To take that into account we synthesize additional keywords defined as pairs of initial interests. Interests in a pair are combined by a logical operation AND. For researcher from Figure 1 additional keywords are defined as follows: w4 " Computational Intelligence" AND " Fuzzy Logic" ; w5 "Computational Intelligence" AND " Artificial Intelligence" w6 " Fuzzy Logic" AND " Artificial Intelligence"
If a researcher’s profile has 3 interests, additional 3 keywords are synthesized, if it has 4 interests then 6 additional are synthesized etc.
For researchers’ topic modeling, we need to choose a research classification system. There are a lot of them, but when choosing we take into account not only their semantic advantages and disadvantages, but also that there is an information system with this research classification system that has available search services. In addition, we require that the information system must indexes a large number of categorized publications over all research. The information system that satisfies these requirements is Dimensions.
Dimensions indexes more than 110M of publications. All publications are categorized by the twolevel variant of Australian and New Zealand Standard Research Classification (ANZSRC) with 22 research divisions and 154 research groups (Table 1). In this work we use the research groups to model a researcher’s interests. P1 ‐ Anthropology; P2 ‐ Criminology; P3 ‐ Demography; P4 ‐ Human Geography; P5 ‐ Policy and Administration; P6 ‐ Political Science; P7 ‐ Social Work; P8 ‐ Sociology; P9 ‐ Other Studies In Human Society Q1 ‐ Psychology; Q2 ‐ Cognitive Sciences; Q3 ‐ Other Psychology and Cognitive Sciences; R1 – Law; R2 ‐ Other Law and Legal Studies S1 ‐ Art Theory and Criticism; S2 ‐ Film, Television and Digital Media; S3 ‐ Journalism and Professional Writing; S4 ‐ Performing Arts and Creative Writing; S5 ‐ Visual Arts and Crafts; S6 ‐ Other Studies In Creative Arts and
Writing Language, T1 ‐ Communication and Media Studies; T2 ‐ Cultural Studies; T3 ‐ Language Communication and Studies; T4 – Linguistics; T5 ‐ Literary Studies; T6 ‐ Other Language, Culture Communication and Culture History and U1 ‐ Archaeology; U2 ‐ Curatorial and Related Studies; U3 ‐ Historical Studies; Archaeology U4 ‐ Other History and Archaeology Philosophy and V1 ‐ Applied Ethics; V2 ‐ History and Philosophy of Specific Fields; V3 ‐ Religious Studies Philosophy; V4 ‐ Religion and Religious Studies; V5 ‐ Other Philosophy and
Religious Studies
A query to Dimensions is formed separately by each element of the set W. If an element is a phrase, then it is surrounded by quotes. As a search scope, we use Title and Abstract and we search only the last 5 years – 2016 – 2020. An example of a search result for the query “fuzzy logic” is presented in Figure 2. For each research division and research group there is a number of publications that has the query mentioned in either the title or abstract. The results are sorted by the number of publications descending. We can also find the overall number of publications for each research division and group, that is without any query.
We perform the topic modeling of researchers based on the following principles: the principle of statistical support – the more publications from a specific research group a given keyword contains, the more membership degree of this keyword to this research group is; the principle of multi-labeling – a keyword can belong to a few research groups; the principle of noise filtering – we ignore research groups with low membership degree to a given keyword; the principle of ignoring stop-words – we ignore keywords that appear in a very large number of publications; the principle of solidarities – the more keywords belong to the same research group the larger the chance that the researcher belongs to this research group; the principle of focusing – if a topic-marked collection of publications contains a few keywords of a researcher at once then the chances to assign this researcher to the respected topic increase; the principle of compactness – a researcher can only be assigned to a few research groups; the principle of research groups interaction – when cutting the tail of topic distribution, the contribution of minor research groups is redistributed on leaders by taking into account their similarity.
We propose an algorithm to implement the proposed principles that consists of 3 stages. On the first stage the set of queries based on keywords and their combination is formed. We use only pairs of keywords because the results using triples of keywords are often empty and increase the processing time. The second stage performs topic modeling by each query separately. Research groups are chosen by the frequency of mentions at a topic-marked collection. Stop-words and noise are filtered by the frequency of mentions in research groups at all topic-marked collections. The minor research groups are left out using cumulative principle, by cutting the tail of the distribution. On the third stage all membership degrees of queries are averaged, the resulted distribution is cut and research groups with the low membership degree are dropped. To ensure compactness we only allow 1 to 4 research groups. %Topic modeling algorithm % #1 – creating the set E of search queries from the keywords E=W for i=1:length(W) for j=i:length(W)
E={E; [‘“’ W(i) ‘AND’ W(j) ‘”’] }
end
% #2 – compute membership degrees to research groups by each query
< Find the number of publications at each topic‐marked collection
N=[N(
Mu(Counter)=Gamma end If Counter==0 return (‘Unsuccessful’) end % #3 – compute membership degrees using all queries Mu_mean=mean(Mu) % averaging all successful queries < Form leaders of research groups that have cumulative contribution Mu_mean not lower than Tail_2. We restrict the number of leaders to be at most 6 with the largest cumulative contribution. If we have more than 6 leaders their numbers will be in the vector Rejected > % Ignore research groups with contribution lower than Tail_2: Mu_mean(Rejected)=0 Mu_mean= Mu_mean./sum(Mu_mean) % norm to be in [0, 1] % Current number of research groups: Current_N_fields=sum(Mu_mean>0) % Set the max number of research groups for a researcher: T_max=min(4, Counter+1) < Find Mu_worst – the smallest membership degree among the leaders > while (Current_N_fields>T_max OR Mu_worst<Tail_3) < Drop the minor group and redistribute its contribution to others based on their similarity > Current_N_fields=Current_N_fields‐1; Mu_mean= Mu_mean./sum(Mu_mean) % norm to be in [0, 1] < Find Mu_worst – the smallest membership degree among chosen research groups >
On the last stage of the algorithm when dropping a minor research group its contribution is redistributed to other research groups based on the similarity defined at [14, 15]. For example, let us say that on an intermediate stage a researcher is assigned to research groups in the following way: ~ 0.5 0.2 0.2 0.1 W , , , . Let us drop the minor group O4. For this, first using method from [14, 15] H 6 O5 O6 O4 we compute Jaccard indexes between O4 and other research groups. For the data from 2016 – 2020 they are:
J O4, H 6 0 ; J O4,O5 0.13 ; J O4,O6 0.22 .
By taking into account the similarity, the contribution of the minor specialty O4 is redistributed in the following way: ~ 0.5 0 0.1 0.2 0.13 0.1 0.2 0.22 0.1 W , , .
H 6 O5 O6 As a result, we get: ~ 0.5 0.213 0.222 W , , .
H 6 O5 O6 After norming to be in [0,1] we have: ~ 0.535 0.228 0.237 W , , .
H 6 O5 O6
Let us illustrate how the algorithm works using as an example topic modeling of the researcher from Figure 1. Using three interests, we form six queries. Figure 3 shows frequency of queries at topic-marked collections. Figure 4 shows the results after cutting the first tail of the distribution. Next, we average by all queries (Figure 5) and cut the tail of the distribution (Figure 6). The resulting distribution is overfilled due to the broad usage of interests for the given researcher. To make the results more focused the final stage of the algorithm reduces the number of research groups to 2 (Figure 7). As a result, we get that a researcher with interests at artificial intelligence and neural networks has the largest membership degree in research groups H1 – Artificial Intelligence and Image Processing with membership degree 0.441 and H2 – Computation Theory and Mathematics with membership degree 0.559. Such a categorization of the researcher does not contradict with the authors’ viewpoint. The example shows that even with two initial keywords the proposed algorithm finds a good enough membership relation between the researcher and research groups.
Figure 7: The result of topic modeling for the researcher from Figure 1
Comparing the results, we see that topic modeling based on interests from Google Scholar – laconic subjective information, with the proposed algorithm categorizes researchers good enough. For quantitative assessment of the results, we used Czekanowski metric. For the case when membership degrees are in [0, 1], Czekanowski metric between two researchers W1 and W2 is computed in the following way:
FitW1,W2 min( tp (W1),tp (W2 )) . (
____ p1,m
The metric (
Based on the data from Table 2 we get the following assessments using metric (
Using metric (
v1,M p1,M
where J tv ,t p denotes Jaccard index between research groups tv and t p ;
tv W1 min0, tv W1 tv W2 denotes residual of membership degree to research group tv in
~
W1 after taking into account the matching between tv W1 and tv W2 in (
tp W2 min0, tp W2 tp W1 denotes residual of membership degree to research group t p in
~
W2 after taking into account the matching between tp W1 and tp W2 in (
To filter information noise, we use the formula (
J D6, I 9 0.083 ; J H1, H 6 0.071.
Substituting the indexes in (
By taking into account the similarity of research groups the level of matching the topic modeling results takes the following values:
Fitsim Yager 0.441 0 0.441 ;
Fitsim Bodyanskiy 0.498 0.022 0.52 .
We proposed researchers’ topic modeling based on their interests in Google Scholar profiles. Interests in profiles researchers specify based on their discretion without using any vocabulary of keywords. In the paper, we propose an approach to categorization of such researchers using the research classification system ANZSRC. A mapping “researcher – research groups” is done using information system Dimensions that contains more than 110 millions of publications categorized according to ANZSRC.
The algorithm of researchers’ topic modeling has 3 stages. The first stage forms a set of queries based on keywords and their combination. On the second stage we perform topic modeling using each query separately with filtering stop-words and underused words. On the third stage we average membership degrees of all queries and cut the distribution to a few research groups. When dropping minor research groups their contribution is redistributed to the leaders. As a result, we get membership degrees for a researcher to a few research groups that correspond to the set of his interests the most. Such mapping of interests can be viewed as an analog to the word2vec procedure.
We compared topic modeling based on small amount of information from researchers’ profiles at Google Scholar with topic modeling based on a few dozens of authored publications categorized by Dimensions. As a result of comparison, we get a good matching of topic modeling results based on different sources of initial information. It allows using the proposed algorithm as the intellectual core of information technology in regards to scientific staff, in particular, for the selection of candidates as opponents of a dissertation, as reviewers for research projects, for forming a team to collaborate on shared research projects etc.
[9]. C. Sun, T. J. King, P. Henville, R. Marchant, Hierarchical Word Mover Distance for
Collaboration Recommender System, Springer 996 (2018) 289-302. doi:
10.1007/978-981-136661-1_23.
[10]. K. Xiangjie, J. Huizhen, Y. Zhuo, A. Tolba, X. Zhenzhen, X. Feng, Exploiting Publication
Contents and Collaboration Networks for Collaborator Recommendation, PlosOne 11(
Dimensions, CEUR Workshop Proceedings 2533 (2019) 117-128. [15]. S. Shtovba, M. Petrychko, An Informetric Assessment of Various Research Fields Interactions on Base of Categorized Papers in Dimensions, CEUR Workshop Proceedings 2845 (2021) 159169.