Mining Intellectual In uence Associations

Abstract

Within the social system of science, citation practices characterize social functions like the conferral of recognition upon the work of others as well as the acknowledgement of one's intellectual debt. However, the structure of intellectual in uence is misrepresented when only the immediate citations and their cardinality are taken into consideration. Thus, in order to better understand the associative dissemination of in uence and approximately construe the anatomy of this structure, complex interactions in the convoluted network of authors and papers need to be probed. Our study aims at understanding these heterogeneous complex interactions. For the bibliographic dataset of authors and publications, we de ne proxy scores that attempt to determine the associative in uence of the cited author over the citing author. In order to harness structural connectivity of the network, we generate author vector representations using these in uence scores. Furthermore, with a view to assess the competence of our proposed scores, we evaluate these representations and provide an empirical study of the results obtained with our algorithm against the baseline and also present a qualitative analysis.

Keywords

• Citation Network
• In uence
• Representation Learning

INTRODUCTION AND RELATED

WORK

The contribution of an author in the form of publications holds an intrinsic value responsible for the e ective dissemination of knowledge. This knowledge based relay that is extended by and for the scienti c community results into establishment of conceptual relationships in the form of citations. Citations and references operate within a jointly cognitive and moral framework designed to provide the historical lineage of knowledge and to repay the intellectual debts through their open acknowledgement [ 15 ][ 7 ]. Thus, citation analysis has the potential in providing valuable insights about the social system of science spanning across a wide range of topics.

E ective digitization of bibliographic data has led to proliferation in the propositions of various quantitative bibliometric performance indicators for journals, papers, authors and institutions based on citation counts and graph-based ranking algorithms [ 6 ][ 17 ][ 22 ]. Studies dealing with in uence have examined various aspects like topic level in uence strength, in uence propagation and its indirect global e ect [ 8 ]; analysis of in uence evolution between communities [ 4 ]; time dependent estimation of in uence for evaluating pairwise community inuences [ 18 ]. Besides being extensively explored in literature, the concept of in uence has also surfaced in scholarly search engines such as Semantic Scholar. Many of these approaches focus on in uence analysis based on individual entities and their overall impact on the network structures. However, in uence relationships and associations do not emerge inherently considering just the global in uence. These in uence associations can be understood as the degree of in uence between a pair of nodes within the network. Our work aims at studying these pairwise in uence associations eventuated between authors within the scholarly network. The main contributions of the paper are as follows: { We propose an algorithm that simulates the in uence between the citing and cited authors and suggest in uence association scores. { Considering issues in quantifying and tracing of in uences that arise in scholarly communication and to harness structural connectivity of the network, we pro le authors and their interactions within the bibliographic network using representation learning and the proposed in uence scores. These representations thus form a generic result of the proposed in uence model and their e ectiveness in context of the problem statement is discussed. { For assessment of predictive capacity of the aforementioned scores and the thus obtained vector representations, experimental results subject to classi cation tasks are discussed along with comparative study against those obtained by measures such as citation counts. 2 2.1

PROBLEM STATEMENT

Problem Formulation

Consider a publication pi written by m co-authors ai1 , ai2 , .., aim , which cites a publication pj written by n co-authors aj1 , aj2 , .., ajn . Thus, a publication citation network can be de ned as a directed graph GP = (VP ; EP ) constructed from the list of references at the end of each publication, where VP represents vertices (publications) and EP represents set of all directed edges between the nodes denoting citations between publications. Consequently, an author citation network GA = (VA; EA) is de ned by projecting this publication citation network along the corresponding author(s) for each publication node in GP . For instance, the citation link pi ! pj is projected between their authors respectively, thus creating m n directed links from each of the m co-authors of the citing publication to each of the n co-authors of the cited publication. Accordingly, VA represents nodes (authors) and EA represents set of all such directed links between the authors. Let the directed pairwise author citation link between the citing author ai and the cited author aj be denoted as ai ! aj (8i = 1; ::; m and 8j = 1; ::; n). In the discussions that follow, we de ne and aim to quantify the associative intellectual in uence measure represented by I(ai; aj ) as the degree to which author aj in uences author ai when a citation is made from ai to aj . 2.2

Issues and Caveats in Quantifying In uence

Citation networks are complex networks in which causal structure exists along the interactions between the nodes. However, consideration of just the citation counts and primary degree of interactions (immediate neighbours) within the citation network has its shortcomings as indicators for tracing intellectual inuences [ 23 ]. Without a subjective survey of authors in conjunction with their publications, it remains unknown as to what fraction of the work is cited that was indeed in uential directly or indirectly and whether the references exist which had no in uence of any kind yet cited due to other motivations. Such complex interplay of multiple citer motivations have been empirically studied and reported in previous studies as well [ 2 ][ 10 ].

Possible classes of errors in tracing scholarly in uences include: { The inclusion (or exclusion) of a reference in a bibliography does not completely indicate whether or not those references were directly or indirectly in uential for the proposition of the publication. { Citing bias in favor of elite scientists or highly cited papers i.e. over-citation described as the \Matthew e ect" [ 11 ]. { Under-citation of fundamental scienti c work is possibly noticed due to the obliteration (of sources) by incorporation (OBI) in the established knowledge [ 10 ]. { When a relevant piece of scienti c work is known through an intermediate publication, the intermediate publication serves as an intermediate in uence.

However, it may remain uncited. { Nature of citation types as they can be further categorized into organic or perfunctory, evolutionary or juxtapositional, con rmative or negational etc [ 14 ].

These certain and other such classes of errors exist within the citation data due to under-inclusion and over-inclusion of references [ 2 ][ 10 ][ 23 ]. This makes the task of tracing in uences more complex. 3

ASSOCIATIVE INFLUENCE MEASURES

We propose two associative in uence scores for e ectively capturing intellectual in uence of the cited author over the citing author. The underlying principle encapsulates the reasoning that the citations towards an in uential publication of the cited author and those from the nest works of the citing author are indeed signi cant. Further, the net in uential impact over a publication can be fairly attributed as an aggregation of in uences by all the cited publications. Considering citing author's temporal scholarly activity, associative in uences are instantiated for each such citing publication wherein references are made to other publications. Thus, for a particular author pair (ai; aj ), we consider each such instance wherein ai cites aj i.e. all such publications authored by ai wherein a citation has been made to a publication authored by aj . This instantiated in uence forms a component for the integral associative in uence between the author pair. Since the associative in uence is a directed mapping, the in uence scores resemble the same notion of directedness. Thus, for a citation relationship ai ! aj , the proposed in uence association score I(ai; aj ) represents the degree of in uence cited author aj has over the citing author ai. 3.1

Ranking Publications

In order to capture the collective nature of scholarly in uence, publication-level ranking is adopted (as opposed to researcher-level ranking). This mimics the spread of intellectual in uence among researchers via their publications. Quite a few studies [ 6 ][ 21 ] investigate extensively this key issue of scienti c credit diffusion by dissecting the credit di usion mechanism underlying both researcher level and paper level graph-ranking methods. Their ndings emphasize that scienti c credit is fundamentally derived from citation information between papers rather than the derived researcher network. Our model for the in uence dissemination within the heterogeneous network of authors and papers thus avoids the inaccurate allocation of scienti c credit among researchers that potentially arises in graph-ranking methods.

PageRank [ 1 ] takes into account the number and quality of links while measuring the importance of entities within a network. Using PageRank over GP , the importance of publications is thus measured considering the number of citations and reputation of the papers [ 9 ]. For the publication level ranking, we have:

P R(pi) = (1

N ) +

X P R(pj ) j!i Tout(pj ) (1) where j ! i implies paper pj referring paper pi, P R(pj ) denotes the PageRank of paper pj and Tout(pj ) denotes the number of outbound links from paper pj . With certain empirical studies, Chen et al. [ 3 ] showed that, scienti c papers usually follow a shorter path of about average two links. This is in opposition to six hyperlinks for the web considering the individual surfer illustration as mentioned in the original study [ 1 ]. Accordingly, we set the damping factor to 0.5 for the purpose of our studies as well. The basis for this score is that, citations from the signi cant works of the citing author (ai) are indeed relevant, considering the intellectual and cognitive in uences. The citing author's signi cant works can be regarded as those scholarly works which have a relatively high PageRank among other works of the same author. The instantiated associative in uence components resulting from citing author's citation activity sum up to form the net score. Thus, we calculate IS (ai; aj ) as follows: IS(ai; aj) =

P P R(pik) k Tout(pik) P R(pai ) 8pik 2 pik ! paj (2) where, pik denotes the kth publication of the citing author ai wherein a citation has been made to a publication authored by the cited author aj, P R(pik) represents the PageRank of the citing publication pik, Tout(pik) denotes the number of outbound links (references) from publication pik and P R(pai ) is the normalization factor which is the average PageRank value of all the publications authored by ai. The normalization factor accounts for the variance in the ranks of citing publications. So, higher the normalized weight of each such component (i.e. higher the normalized PageRank of citing publication) and higher the cardinality of such components exchanged between ai and aj, higher is the associative in uence. The underlying notion for devising this score is to calculate a measure of the extent to which the citing author tends to cite a set of authors weighted by their in uential scholarly contributions in his/her publications. Extending the in uence instantiation notion for ID, the association components of in uence comprise of the PageRank of the cited publication and the inbound links of that cited publication (distributing impact of a publication among the inbound citation references). Thus, we calculate ID(ai; aj) as follows:

ID(ai; aj) =

P P R(pjk) k Tin(pjk) P R(paj ) 8pjk 2 pai ! pjk (3) Here, pjk denotes the kth publication of the cited author aj wherein a citation has been made from a publication authored by the citing author ai, P R(pjk) represents the PageRank of the cited publication pjk, Tin(pjk) denotes the number of inbound links to publication pjk and P R(paj ) is the normalization factor which is the average PageRank value of the publications authored by aj. The disparity in the ranks of cited publications is accounted for by this normalization factor. So, more the normalized weight of each such association components and higher the cardinality of such components exchanged between ai and aj, higher is the associative in uence. 3.4

Qualitative Analysis

Author and publication ranks of prominent scientists and young researchers span across a wide spectrum. For example, a highly cited publication of a prominent scientist might be in uential but in a modest way to each individual researcher in the scienti c community at large. Also, the nest works of a young researcher possibly might have less inbound citations comparatively, however, the substantial in uence of cited publications within such publications should not possibly be overlooked. Considering the di erent degrees of author and publication ranks, it can be seen from Equation (2) that signi cant works of young as well as even mediocre researchers are instrumental in contributing towards highlighting the cited author's in uence. Thus, irrespective of the cited author's prominence, his/her in uence over the citing author is conspicuous and visibly pronounced. Also, from Equation (3) it is evident that even scientists not belonging to the higher order of ranks receive their due scholarly attribution pertinent to the their respective notable scienti c studies. Such notions of relative author impact and allocation of due credit persist across the spectrum for most of the researchers belonging to di erent degrees of author ranks. Normalization of associative inuence measures as illustrated in sections 3.2 and 3.3 account for such variances along with taking into consideration the e ect of accumulated advantage (i.e. Matthew e ect [ 11 ]).

Associative pairwise in uences eventuated due to scholarly contributions over time semantically imply that there is certain form of in uence of the cited author over the citing author irrespective of the citation types. In this paper, we focus on the existence and extent of conceptual relationships formed between authors. However, the precise nature of in uence maybe hard to quantify without the factual ontological citation representations. For alleviating issues concerning under-inclusion of references (as discussed in Section 2.2), capturing the network neighborhood and harnessing the structural connectivity, we pro le authors and their interactions using representation learning and the proposed in uence scores. This e ectively maps the latent features within the citation network into a vector space. 4

AUTHOR REPRESENTATION LEARNING

Recent work in language modeling and representation learning such as Word2Vec [ 12 ] focuses on application of probabilistic neural networks which map words into vector spaces. The author vector representations are learned with a similar intuition as discussed in the following sections: 4.1

Random Walk on Weighted Directed Network

Using proposed in uence measures, we model weighted random walks over the author citation network. These walks can be approximated as sentences in the context of language modeling. Analogous to recent researches [ 16 ], it represents a network as a \document". The motivation behind converting a graph into a series of text documents is: Word frequency in a document corpus and the visited node frequency during a random walk for a connected graph, both follow the power law distribution [ 16 ].

We sample random walks over the weighted directed author citation network. Considering the author citation network (GA) with n author nodes (a1, a2, . . . , an); wij represents the weight of the edge connecting nodes ai and aj where the edge weight is the in uence score I(ai; aj) (as derived from Equation 2 or Equation 3) for the author pair and therefore we have, wij = I(ai; aj). Since the in uence scores are non-negative, we have wij 0. For the study of tracing in uence associations, we prune self-citation loops and thus wii = 0. Now, for a given source node ai, the transition probability that the author node aj is chosen from the direct successors of ai is proportional to the in uence measure I(ai; aj). This is computed as:

I(ai; aj) pai;aj = P I(ai; ak) k (4) where pai;aj represents the transition probability. For each source author node ai, we simulate a weighted directed random walk Wai . This sampling is a stochastic process consisting of author nodes wa1i , wa2i , . . . , wani as random variables such that waji is a vertex chosen with transitive probability pai;aj from the direct successors of ai. In our experiments we set the length of these walks to be xed. For each source vertex, the random walk generator samples author nodes based on respective transition probabilities until a maximum length l (= 40) is reached. For the purpose of our study, we generate such weighted random walks (= 15) times for each author. 4.2

Representation Learning Framework

Modelling social structures and relationships within networks can be aligned with the optimization techniques used to model natural languages [ 16 ][ 5 ]. With the ordered sequence of nodes constructed using weighted random walks, we learn representations using Skip-gram model. This represents authors and the citation relationship shared between a pair of authors in an unsupervised manner. Based on distributional hypothesis, these representations are latent features that capture neighbourhood as well as structural in uences in the citation network in a continuous low dimensional vector space. In e ect, these representations encapsulate more information and relationships between authors than using just the immediate citations.

Skip-gram is a language model that maximizes the conditional co-occurence probability of words occurring within a prede ned window [ 12 ]. Thereby, we have f : VA ! Rd as the mapping function from nodes to feature representations that we aim to learn. Here d speci es the number of dimensions of the feature representation and f represents a matrix of size jVAj d parameters. Now, we try to optimize the likelihood function as formulated in Equation 5: max f j=i+w

X j=i w;j6=i log Pr(aj j f (ai))

(5) where w is the size of the window, ai 2 VA and Pr(aj j f (ai) is de ned by the softmax function:

Pr(aj j f (ai)) =

exp(f (aj ) f (ai)) P exp(f (ak) f (ai)) k2V (6)

Skip-gram assumes that inside the context, all nodes are independent of each other and are equally important. However, as seen from Equations 5 and 6, update step per node is proportional to jVAj. This is computationally expensive for large networks (such as the author citation network in consideration). We approximate the optimization function using negative sampling [ 13 ].

Using the obtained resultant author vector representations, we validate the e ectiveness of our proposed scores as discussed in the following sections. 5

PERFORMANCE MODEL

Limitations in the assessment of intellectual and cognitive in uences prevail due to its subjective nature. However, despite the shortcomings of citation data, studies [ 23 ] assert that citations can be used as approximate proxy indicators of in uence for the aggregates of authors and papers. Collectively, with measurable factors and practical limitations of the study, it can be fairly argued that if a publication proves to be relatively in uential in the scienti c work of an author, then he/she is quite de nitive to have a higher relative citation ratio towards the cited in uential authors. So, we utilize and bucket these relative citation ratios for author pairs as labels for classifying the extent of in uence. Representations obtained using in uence scores can be evaluated against baseline by a comparative study of citation prediction and its extent between author pairs. In our study, the purpose is to evaluate whether our proposed in uence measures capture meaningful relationships. To do so, it su ces to test their relative capacity in citation prediction, and the absolute predictive accuracy is not the criterion being assessed. 5.1

Validating In uence Associations

To capture the semantic relatedness between the citing and cited author's inuential relationship and in order predict weighted citation link between a pair of authors a1 and a2, we generate edge representation e(a1; a2). This is done by de ning a binary operator over the corresponding author feature vectors f (a1) and f (a2). Similar strategies have been successfully used in earlier studies for link prediction tasks [ 5 ]. We de ne ith component of the edge representation ei(a1; a2) as concatenation of author feature vector components denoted by fi(a1) t fi(a2). e(a1; a2) spans across R2 D as the author feature vectors 2 RD are concatenated.

These edge representations are now further used in training and evaluation for predicting the in uence between a pair of authors. However, predicting the presence of link between the author pair in the testset is necessary but insufcient to assess predictive capacity of the degree of in uence. For evaluating the extent of in uence more rigorously, we extend this binary classi cation link prediction evaluation to multi-label classi cation. Here, the edge representations are classi ed against labels viz., Nil In uence (NI), Slightly In uenced (SI) and Highly In uenced (HI) depending upon the relative citation ratio between the pairs of authors in the training and testing sets respectively. Thus, for a pair of authors ai and aj , e(ai; aj ) is classi ed based on the relative citation ratio (cr(ai;aj)) which is computed as: cr(ai;aj) =

c(ai; aj ) P c(ai; ak) k (7) where c(ai; aj ) represents number of citations from the citing author ai to the cited author aj in the author citation network during that speci c temporal segment. The calculated citation ratios are then mapped to aforementioned class labels. If cr(ai;aj) 2 (0.0, ], then e(ai; aj ) is classi ed as SI. When cr(ai;aj) > , then e(ai; aj ) is classi ed as HI. Based on repeated experiments to maximize discrimination among citation ratios, is set to 0.036. Lastly, cr(ai;aj) = 0 implies that there is no in uence of the cited author aj over the citing author ai, thus, classifying e(ai; aj ) as N I.

Baseline: For comparing the performance of our model and the in uence scores, we use citation counts as baseline for our evaluation. Author pro ling and generation of vectors for this baseline is achieved using weighted random walk over author citation graph. Here, weights are the number of citations from the citing author to the cited author (as opposed to in uence measures as discussed in 4.1). Edge representations for this baseline are then computed using the obtained author vector representations. 6 6.1

EXPERIMENTAL DESIGN

Dataset Description

The DBLP dataset used consists of papers published from the period 1960 to 2014 wherein the citation data is enriched by using bibliographic metadata from ArnetMiner [ 20 ]. The dataset contains information such as paper's title, its authors and their a liations, citation list, publication year, etc. For our experiments, pre-processing is performed over the dataset for pruning incomplete records where: 1) The publications with incomplete meta data (absence of year, authors, etc.) are removed. 2) Internal citations can be de ned as references to publications within the snapshot of dataset being considered. References other than these internal citations are removed.

The nal dataset includes 1277594 papers and 1003387 authors. The total count of internal references for publication citation network is 7962820 whereas edgelist for the author citation network sums up to 39713499. 105 ilscobunP#110024 ita103 101 In order to evaluate and assess the degree of in uence as quanti ed by the suggested in uence measures, we consider work of individual authors over a temporal scale. We divide the dataset into 3 segments as follows:

Profiling: This segment of dataset represents the activity of researchers and scientists within the bibliographic network up to 2006. Interactions between researchers by means of collaborations, citations and conceptual exchanges in the form of publications are signi cantly eventuated considering such a wide temporal span. This partition of dataset helps in the process of author pro ling by means of learning author vector representations using weighted random walks.

Training: This segment of dataset is used for learning in uence associations by generating edge representations using the author vector representations captured in the Profiling segment. The exact testset author pair is excluded from this segment to avoid over- tting of the classi er. Author pairs for whom citations have been eventuated between 2006 and 2010 are considered for this segment.

Testing: The proposed in uence scores and the baseline (citation count) are validated using the bibliographic interactions in this segment. An author pair is valid for testing as long as we have individual vector representations for both the authors. Citation exchanges resulted since 2011 are considered for this segment. 6.3

Evaluation and Results

Using the Training set, edge representations are constructed for each pair of authors between whom citations are exchanged during this segment. Consequently, relative citation ratios between these author pairs are calculated using Equation (7). The edge representations are then classi ed with the class labels namely, (N I, SI and HI) using RandomForest. For authors spanning across wide spectrum of ranks and extent of in uences, this enables us to capture what kind of authors cite what kind of authors. Since we aim at comparing the relative predictive capacity of these representations, we focus less on exact classi er settings and report results achieved by each representation using the same parameters.

On evaluation, precision, recall and f-scores for baseline and in uences measures are as shown in the tables (1), (2) and (3). From these results, we can see that even the resultant representations obtained using baseline citation count performs well enough for the multi-label citation prediction task. However, inuence measures IS and ID can be clearly seen as better performers almost throughout for each of the aforementioned classes, considering the reported precision-recall values. We observe that, for class N I, almost all the three measures perform equally. This can be attributed to the better accuracy of classi er for non-existent edges between author pairs, irrespective of the weights in consideration. It can also be theorized that certain in uence associations may traverse from a class to another over a span of time. For example, the citations of the citing author with higher in uence of the cited author possibly might get narrowed down (and vice versa) over a period of time. This can happen due to various possible reasons such as a shift in research trends, cultivation of interests in newer elds, etc. Thereby, we also witness recall values for class HI on lower sides for all the three measures.

CONCLUSION AND FUTURE WORK

In this paper, we present a model to trace intellectual in uences harnessing the structural connectivity within an academic network. Further, we generate author pro ling by mapping the latent features into a vector space using the proposed in uence scores. We also evaluate e ectiveness of the captured author relationships and the resultant author representations by performing experiments for classi cation tasks, such as citation prediction and the extent of citation. It is observed that results obtained using the suggested in uence scores perform better as compared to immediate citation counts. A future direction would be to incorporate types of citations (as mentioned in Section (2.2)) into the current model, possibly using ontological representation of citations [ 19 ]. This might help us in knowing and gaining insights on the nature of in uences. It would also be interesting to analyze e ects of research trends on in uences associations.

References

1. Brin , S. , Page , L. : The anatomy of a large-scale hypertextual web search engine . Computer networks and ISDN systems 30(1-7) , 107 { 117 ( 1998 )

2. Brooks , T.A. : Evidence of complex citer motivations . Journal of the Association for Information Science and Technology 37 ( 1 ), 34 { 36 ( 1986 )

3. Chen , P. , Xie , H. , Maslov , S. , Redner , S. : Finding scienti c gems with googles pagerank algorithm . Journal of Informetrics 1 ( 1 ), 8 { 15 ( 2007 )

4. Chikhaoui , B. , Chiazzaro , M. , Wang , S. , Sotir , M. : Detecting communities of authority and analyzing their in uence in dynamic social networks . ACM Transactions on Intelligent Systems and Technology (TIST) 8 ( 6 ), 82 ( 2017 )

5. Grover , A. , Leskovec , J.: node2vec: Scalable feature learning for networks . In: Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining . pp. 855 { 864 . ACM ( 2016 )

6. Jiang , X. , Sun , X. , Zhuge , H.: Graph-based algorithms for ranking researchers: not all swans are white! Scientometrics 96 ( 3 ), 743 { 759 ( 2013 )

7. Kaplan , N.: The norms of citation behavior: Prolegomena to the footnote . Journal of the Association for Information Science and Technology 16 ( 3 ), 179 { 184 ( 1965 )

8. Liu , L. , Tang , J. , Han, J. , Yang , S.: Learning in uence from heterogeneous social networks . Data Mining and Knowledge Discovery 25 ( 3 ), 511 { 544 ( 2012 )

9. Ma , N., Guan , J. , Zhao , Y. : Bringing pagerank to the citation analysis . Information Processing & Management 44 ( 2 ), 800 { 810 ( 2008 )

10. MacRoberts , M.H., MacRoberts , B.R. : Problems of citation analysis: A critical review . Journal of the American Society for information Science 40 ( 5 ), 342 { 349 ( 1989 )

11. Merton , R.K. : The matthew e ect in science, ii: Cumulative advantage and the symbolism of intellectual property . isis 79(4) , 606 { 623 ( 1988 )

12. Mikolov , T. , Chen , K. , Corrado , G. , Dean , J.: E cient estimation of word representations in vector space . arXiv preprint arXiv:1301.3781 ( 2013 )

13. Mikolov , T. , Sutskever , I. , Chen , K. , Corrado , G.S. , Dean , J. : Distributed representations of words and phrases and their compositionality . In: Advances in neural information processing systems . pp. 3111 { 3119 ( 2013 )

14. Moravcsik , M.J. , Murugesan , P.: Some results on the function and quality of citations . Social studies of science 5(1) , 86 { 92 ( 1975 )

15. Nicolaisen , J.: Citation analysis . Annual review of information science and technology 41(1) , 609 { 641 ( 2007 )

16. Perozzi , B. , Al-Rfou , R. , Skiena , S. : Deepwalk: Online learning of social representations . In: Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining . pp. 701 { 710 . ACM ( 2014 )

17. Pinski , G. , Narin , F. : Citation in uence for journal aggregates of scienti c publications: Theory, with application to the literature of physics . Information processing & management 12(5) , 297 { 312 ( 1976 )

18. Rakoczy , M.E. , Bouzeghoub , A. , Gancarski , A.L. , Wegrzyn-Wolska , K. : In uence in time-dependent citation networks . In: 2018 12th International Conference on Research Challenges in Information Science (RCIS) . pp. 1 { 11 . IEEE ( 2018 )

19. Shotton , D. : Cito, the citation typing ontology . In: Journal of biomedical semantics . vol. 1 , p. S6 . BioMed Central ( 2010 )

20. Tang , J. , Zhang, J., Yao , L. , Li , J. , Zhang, L. , Su , Z. : Arnetminer: extraction and mining of academic social networks . In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining . pp. 990 { 998 . ACM ( 2008 )

21. Wang , H. , Shen , H., Cheng, X., et al.: Scienti c credit di usion: Researcher level or paper level? Scientometrics 109 ( 2 ), 827 { 837 ( 2016 )

22. Zhou , Y.B. , Lu  , L. , Li , M. : Quantifying the in uence of scientists and their publications: distinguishing between prestige and popularity . New Journal of Physics 14 ( 3 ), 033033 ( 2012 )

23. Zuckerman , H.: Citation analysis and the complex problem of intellectual in uence . Scientometrics 12 ( 5-6 ), 329 { 338 ( 1987 )