An agent-based model of interdisciplinary interactions in science
AAn agent-based model of interdisciplinaryinteractions in science
Juste Raimbault , , , ∗ CASA, University College London UPS CNRS 3611 ISC-PIF UMR CNRS 8504 G´eographie-cit´es ∗ [email protected] Abstract
An increased interdisciplinarity in science projects has been high-lighted as crucial to tackle complex real-world challenges, but alsoas beneficial for the development of disciplines themselves. This pa-per introduces a parcimonious agent-based model of interdisciplinaryrelationships in collective entreprises of knowledge discovery, to investi-gate the impact of scientist-level decisions and preferences on globalinterdisciplinarity patterns. Under the assumption of simple rules forindividual researcher project management, such as trade-offs betweeninvested time overhead and knowledge benefit, model simulations showthat individual choices influence the distribution of compromise pointsbetween emergent level of disciplinary depth and interdisciplinarity ina non-linear way. Different structures for collaboration networks mayalso yield various outcomes in terms of global interdisciplinarity. Weconclude that independently of the research field, the organization ofresearch, and more particularly the local balancing between verticaland horizontal research, already influences the final positioning of re-search results and the extent of the knowledge front. This suggestsdirect applications to research policies with a bottom-up leverage onthe interactions between disciplines.
The role of interdisciplinary projects in science has been highlighted ascrucial for the development of complexity approaches and an effectivetackling of real-world issues. Many aspects of knowledge production have1 a r X i v : . [ phy s i c s . s o c - ph ] J un role in enhancing interdisciplinary collaborations. [Hofstra et al., 2020]study the circular relationship between diversity and innovation, and showthat underrepresented groups have a higher likelihood of successfully in-novate in science. [Jang et al., 2019] use an agent-based model to studythe co-evolution between knowledge diffusion and the structure of knowl-edge. Each discipline has its own view on interdisciplinarity, as for exam-ple [Urbanska et al., 2019] unveil an asymmetry between social and hardsciences in the credit given to other disciplines within interdisciplinaryprojects. Other social or political factor are to be taken into account wheninvestigating the disciplinary structure of science: access to funding hasfor example a strong impact on the efficiency of knowledge production[Gross and Bergstrom, 2019]. [Akerlof and Michaillat, 2018] show that thediscrepancy between disciplines is intrinsic to the type of knowledge pro-duced, as they suggest that paradigms are more likely to persist in “low-power” sciences. The organisation of research is also an important factor,and teams and single authors produce different aspects of the commonknowledge [Pavlidis et al., 2014]. [Rouse et al., 2018] model probables tra-jectories according to the type of research environment. The link betweenopen access, which is a driver of increased collaborations and potentiallyincreased interdisciplinarity, and the quality of research, is investigated by[van Vlokhoven, 2019].Interdisciplinarity in itself has extensively been studied by quantita-tive studies of science. [Thurner et al., 2019] show that interdisciplinarypapers perform better in terms of citation on the long run than mainstreampapers. [Zeng et al., 2019] investigate the interdisciplinarity of scientiststhemselves and how it evolved in time, and show that more scientists haveswitched between topics recently. [Larivi`ere and Gingras, 2010] provide em-pirical evidence for an optimal intermediate level of interdisciplinarity interms of research impact.[Brown et al., 2020] study within the particularcontext of an interdisciplinary summer school the propensity of mixingwithin interdisciplinary projects, and find evidence consistent with randommixing. [Pluchino et al., 2019] show that randomness has an important rolein determining individual trajectories success in physics.2ollowing [Giere, 2010a], agent-based modeling is a privileged approachto simulate the behavior of scientists. [Shafiee and Berglund, 2019] use anagent-based model to simulate the impact of a workflow to process data underdifferent collaboration scenarios. [Bornmann et al., 2020] simulate citationdynamics, and more particularly the consequence of introducing a perfor-mance index on citation patterns. Agent-based modeling has extensivelybeen used for the evaluation of peer review practices. [Feliciani et al., 2019]surveys 46 simulation studies of peer review with numerous applications.[Kovanis et al., 2016] empirically calibrates an agent-based model of peerreview for more than 100 journals, and provides a tool to evaluate systemsof peer reviews. [Shneiderman, 2018] describes a theoretical model involvingvarious actors of science. Agent-based models are more broadly used to studysocial dynamics such as group organisation in [Dionne et al., 2019].Various works have dealt with microscopic modeling of knowledge produc-tion, among which for example the Nobel game introduced by [Chavalarias, 2016]which investigates the balance between falsification of previous theories andthe elaboration of new theories. [Giere, 2010a] also proposed an agent-basedmodel of science, consistently with the perspectivist approach developed in[Giere, 2010b]. We develop here a simple agent-based model of scientificresearch focusing on the interplay between disciplinary and interdisciplinaryresearch. The rationale relies on the basic assumption that scientists canchoose when starting a new project between interdisciplinary collaborationand a work within their discipline. How can the choice patterns at the micro-level influence the overall interdisciplinarity level ? The model is voluntaryparcimonious to test if even many simplification some structural effects stillhold. Many dimensions and processes are at play to shape collaborations betweenscientists and more broadly between scientific disciplines. These include for3xample social networks, governance and funding issues, or knowledge prox-imity (which can occur on various knowledge domains, from methodologicalto empirical or theoretical). Our rationale is to propose an agent-based modelgrasping some of this complexity from the bottom-up focusing on scientistbehavior, but simple enough so that it can be systematically explored. Weinclude thus in the model two basic antagonist processes, namely a propen-sity to collaborate mostly determined by knowledge proximity, and someresources constraints (time, funding) which affect negatively the possibilityto collaborate. Working with scientists outside one’s field has indeed a highcost, from finding common ground and research questions to an possibleconstruction of integrated knowledge [Frodeman, 2013].
Agents are N scientists A i , characterized by a probability distribution d ( x )representing their disciplinary positioning in an abstract way: research issummarized by a one dimensional variable R , and the disciplinary positioningon this axis is given by the distribution. The model is setup with normaldistributions of width σ with an average distributed uniformly in [0; 1].Scientists also have a time budget per day, that we will summarize as a futuretimetable T ( t ) : t > t (cid:55)→ p ( t ) ∈ P where P is the space of scientific projects.The central feature of the model is the utility function U ( d i , d j ) determiningan abstract utility for scientist i to collaborate with j for a given project. Itwill be a function of the disciplinary overlap o = (cid:82) x d i ( x ) · d j ( x ) dx and differentassumptions on the form of this cost function can be tested. We take a linearcost in the overlap and a varying benefit, expressing the fact that researchershave different strategies regarding their interdisciplinary positioning. Thisway, we have U ( d i , d j ) = o/i α − o , assuming a fat-tail distribution of individualpreferences for interdisciplinarity, given by a power law of parameter α . Adiscrete choice formulation gives the probabilities for a scientist i to chooseamong j collaborators by p j = exp ( βU ( d i , d j )) / (cid:80) k exp ( βU ( d i , d k )). Givena social network of relations, that we take for now as a fixed scale-free socialnetwork, the temporal evolution of the model goes as follows: (i) one scientist4ith no current activity is picked up at random, and starts a project with oneof its potential collaborators taken as its neighbors in the network that havefree time, chosen with the probability p j . The project has a random uniformduration and timetables are updated accordingly; (ii) current projects areupdated and finished if necessary. The outcome of the model if measuredby average depth across project, defined for one project as the overlappingareas between distribution, and average interdisciplinarity measured by totalarea covered. In order to give empirical support to the modeling choices for the ABM,we first study the properties of a large scientific corpus. We propose touse the Arxiv citation network, which represents a significant proportionof physics and computer science. An open dataset providing parsed au-thors and citations is made available by [Clement et al., 2019]. This allowsconstructing a citation network with | V | = 1 , ,
261 nodes (papers) and | E | = 6 , ,
633 citation links. This corresponds to 1 , ,
500 unique au-thors which we disambiguated by concatenating first name and last name.We then proceed to a community detection in the citation network, using aLouvain community detection algorithm. We obtain therein a modularityof 0 .
78 and 38 communities with a size larger than 1000. Working withthese main endogenous citation communities (which can be interpreted asscientific fields of citation practice), we construct probabilities for authors tobelong to each community. These are computed as p ik = N ik /N i for author i and community k , were N ik is the number of articles authored withinthis community and N i the total number of articles authored. This allowscomputing a cosine proximity between authors defined as s ij = (cid:126)p i · (cid:126)p j , andalso an interdisciplinarity measure as an Herfindhal diversity index givenby h i = 1 = (cid:80) k p ik . Finally, we also study co-authorship probabilities c i → j defined as the probability for author i to co-author with author j knowing5igure 1: Collaborations and interdisciplinarity within the Arxivdataset. (Top left)
Cumulative distribution function of the number of ar-ticles per author (these were disambiguated using first and last name only,statistics may not be accurate). We compare a log-normal and a power-lawfit. (Top right)
Distribution of interdisciplinarity per author, computed asan Herfindhal index of probabilities within endogenous citation communities. (Bottom left)
Distribution of positive author proximities, defined as cosinesimilarity between authors probability distribution within citation communi-ties. (Bottom right)
Distribution of co-authorship probabilities, conditionedby the number of articles. 6hat the author has written a paper (the matrix is thus non symmetric).We show in Figure 1 the empirical results obtained. The number ofpapers by author is close to a power-law with an exponent of 2.82, although alog-normal law seems to better fit the data. Regarding interdisciplinarity ofauthors, although a large majority of authors are mono-disciplinary, we find asecondary peak at 0.5 and a non negligible proportion of authors spanning theindicator range up to very high values of 0.8. This confirms the relevance ofour model with an active interdisciplinarity. When studying cosine similaritybetween authors using their probabilistic description within communities, wefind a broad range of values, also witnessing a high diversity (knowing thatmost authors are at a 0 proximity, since the plot is conditional for readability).Co-authorship probabilities follow rather symmetrical distributions with fattails on a log-scale, consistently when conditioning on the number of papersauthored. This is consistent with the power-law assumed for the propensityfor interdisciplinarity for authors.
The model is implemented in NetLogo [Tisue and Wilensky, 2004] and ex-plored with OpenMole [Reuillon et al., 2013]. Source code and results areavailable on the open git repository of the project at https://github.com/JusteRaimbault/Perspectivism . Data used in the paper is available onthe dataverse at https://doi.org/10.7910/DVN/GMQ5A8 .We run a basic grid exploration of the parameter space, both with randomand small-world social networks, for parameters α, β, σ with 50 repetitions ofthe model for each parameter points, corresponding to 158,400 model runs.Figure 2 shows indicators variation on a given subspace and the correspondingPareto front between depth and interdisciplinarity. We show a second orderinfluence of preference hierarchy α and non-linearity of model behavior as afunction of all parameters. Convergence properties are reasonable with thisnumber of repetitions. Large individual disciplinary width σ causes the choiceparameter β to have no influence, whereas low values give an increasinginterdisciplinarity and a decreasing depth as a function of β . Random7ehavior ( β = 0) leads to a constant depth of projects. When examiningthe Pareto front between the two contrary objectives, the optimal pointsoccur for intermediate β when σ is fixed, suggesting non-trivial behavioraloptima at a fixed disciplinary configuration. These first exploration show thecomplex dynamics of interdisciplinarity even with simple interaction rules andnetwork structure, and suggests further applications such as the explorationof policies by changing network structure or studying in a more refinedway the influence of α . Preliminary non-systematic model experiments, inparticular changing the type of network structure, suggest that it may alsohave significant effect on model outcomes. Beyond the simplifying opposition between fully constructivist and realisticapproaches to science, several alternatives have been developed, among whichPerspectivism [Giere, 2010b] is a way to tackle most of the issues opposingthese two by taking an agent-based approach to the production of scientificknowledge. The main feature of this viewpoint is to consider each scientificenterprise as a single perspective, in which an agent aims at understandingan aspect of the real world (the ontology) with the mean of a medium, whichis considered as a model. Constituted disciplines thus contains more or lesscompatible perspectives. The explicitation of this approach has been doneby [Raimbault, 2017] to embed it into knowledge domains, as a generalizationof knowledge domains introduced by [Livet et al., 2010].We postulate that this approach to science may be a powerful tool tofoster interdisciplinary collaborations, if used in a reflexive way in the con-struction of projects. [Ellemers et al., 2020] propose a similar framework.More precisely, we suggest to apply an “Applied Perspectivism”, in thesense of an explicit perpectivist positioning within a given collaboration, andassociated guidelines and protocols for collaboration. This would imply ahigh-level of reflexivity for each agent implied, a mapping of the different8igure 2:
Patterns of interdisciplinarity from model simulations.
Weshow measures of depth and interdisciplinarity (top row) at fixed α = 0 . β as a functionof individual extent σ . On the bottom, the Pareto front of average pointbetween these two objectives. 9ayers of the enterprise and the positioning of each agent regarding the do-mains of knowledge. This way, in the particular case of model coupling, theexplicitation of positioning and of the structure of each knowledge impliedshould ease interactions. As Banos points out [Banos, 2013], transversalwork must alternate with deeper investigations in each discipline, in a kindof “virtuous circle” [Banos, 2017]. Fostering a synergy between complemen-tary knowledge is the core aspect more important than interdisciplinarityin itself [Leydesdorff and Ivanova, 2020]. This raises the issue of, beforeindividual researcher particularities, how a given collective structure of scien-tific knowledge production should balance between these disciplinary andinterdisciplinary knowledge. It is clear that this question is deeply endoge-nous to each studied subject, and even each particular approach taken, butwithin the applied knowledge framework described above, we have reasonsto believe that certain structural properties may be rather general. Indeed,each discipline is expected to bring components for each knowledge domain,and the co-evolving perspective is built on their interrelations. This paperproposed to investigate basic aspects of this issue, by means of agent-basedmodeling.This work aimed at providing quantitative evidence of the feasibilityof the epistemological point of view described above and inform potentialimplementation for some of its processes, more precisely how can certainlevel of coupling of perspectives (or overlap of ontologies) may be achievedgiven specializations of scientists and a given dynamic of interaction. Possible refinements of the model, towards a less stylized and more behavioraland micro-based model, could for example include the introduction of timebudgets, simultaneous projects and dynamical time investment for scientists.The assumption of two-person projects is also strongly constraining, andrelaxing it would require the extension of depth and interdisciplinaritymeasures that is not necessary straightforward. Furthermore, the absence oflearning and of evolution of the social network when completing a project10uggests a short time scale of application: further refinements should includedynamics of individual distributions and of individual relationships.
In conclusion, we show with a simple model that the individual choicesproduce an emerging structure of the research front, suggesting that appliedperspectivism requires a careful tuning of research structure and researcherbehaviors since Pareto-optimal configurations correspond to non-trivial pa-rameter points. Future developments should include more realistic behavioralassumption, and a formalisation of the applied perspectivism approach toinclude it in the agent-based model.
References [Akerlof and Michaillat, 2018] Akerlof, G. A. and Michaillat, P. (2018). Per-sistence of false paradigms in low-power sciences.
Proceedings of theNational Academy of Sciences , 115(52):13228–13233.[Banos, 2013] Banos, A. (2013).
Pour des pratiques de mod´elisation et desimulation lib´er´ees en g´eographie et SHS . PhD thesis.[Banos, 2017] Banos, A. (2017). Knowledge acceleratorin geography andsocial sciences: Further and faster, but also deeper and wider.
UrbanDynamics and Simulation Models , pages 119–123.[Bornmann et al., 2020] Bornmann, L., Ganser, C., Tekles, A., and Ley-desdorff, L. (2020). Does the h-index reinforce the matthew effect inscience? the introduction of agent-based simulations into scientometrics.
Quantitative Science Studies , 1(1):331–346.[Brown et al., 2020] Brown, J., Murray, D., Furlong, K., Coco, E., andDablander, F. (2020). A breeding pool of ideas: Analyzing interdisciplinarycollaborations at the complex systems summer school.11Chavalarias, 2016] Chavalarias, D. (2016). What’s wrong with science?
Scientometrics , pages 1–23.[Clement et al., 2019] Clement, C. B., Bierbaum, M., O’Keeffe, K. P., andAlemi, A. A. (2019). On the use of arxiv as a dataset.[Dionne et al., 2019] Dionne, S. D., Sayama, H., and Yammarino, F. J.(2019). Diversity and social network structure in collective decision making:Evolutionary perspectives with agent-based simulations.
Complexity , 2019.[Ellemers et al., 2020] Ellemers, N., Fiske, S. T., Abele, A. E., Koch, A.,and Yzerbyt, V. (2020). Adversarial alignment enables competing mod-els to engage in cooperative theory building toward cumulative science.
Proceedings of the National Academy of Sciences , 117(14):7561–7567.[Feliciani et al., 2019] Feliciani, T., Luo, J., Ma, L., Lucas, P., Squazzoni,F., Maruvsic, A., and Shankar, K. (2019). A scoping review of simulationmodels of peer review.
Scientometrics , 121(1):555–594.[Frodeman, 2013] Frodeman, R. (2013).
Sustainable knowledge: A theory ofinterdisciplinarity . Springer.[Giere, 2010a] Giere, R. N. (2010a). An agent-based conception of modelsand scientific representation.
Synthese , 172(2):269–281.[Giere, 2010b] Giere, R. N. (2010b).
Scientific perspectivism . University ofChicago Press.[Gross and Bergstrom, 2019] Gross, K. and Bergstrom, C. T. (2019). Contestmodels highlight inherent inefficiencies of scientific funding competitions.
PLoS biology , 17(1).[Hofstra et al., 2020] Hofstra, B., Kulkarni, V. V., Munoz-Najar Galvez,S., He, B., Jurafsky, D., and McFarland, D. A. (2020). The diver-sity–innovation paradox in science.
Proceedings of the National Academyof Sciences , 117(17):9284–9291. 12Jang et al., 2019] Jang, J., Ju, X., Ryu, U., and Om, H. (2019). Coevo-lutionary characteristics of knowledge diffusion and knowledge networkstructures: A ga-abm model.
Journal of Artificial Societies & SocialSimulation , 22(3).[Kovanis et al., 2016] Kovanis, M., Porcher, R., Ravaud, P., and Trinquart,L. (2016). Complex systems approach to scientific publication and peer-review system: development of an agent-based model calibrated withempirical journal data.
Scientometrics , 106(2):695–715.[Larivi`ere and Gingras, 2010] Larivi`ere, V. and Gingras, Y. (2010). On therelationship between interdisciplinarity and scientific impact.
Journal ofthe Association for Information Science and Technology , 61(1):126–131.[Leydesdorff and Ivanova, 2020] Leydesdorff, L. and Ivanova, I. (2020). Themeasurement of interdisciplinarity and synergy in scientific and extra-scientific collaborations.
Available at SSRN .[Livet et al., 2010] Livet, P., M¨uller, J. P., Phan, D., Sanders, L., and Au-atabu, T. (2010). Ontology, a mediator for agent-based modeling in socialscience.
Journal of Artificial Societies and Social Simulation , 13(1).[Pavlidis et al., 2014] Pavlidis, I., Petersen, A. M., and Semendeferi, I. (2014).Together we stand.
Nature Physics , 10(10):700.[Pluchino et al., 2019] Pluchino, A., Burgio, G., Rapisarda, A., Biondo,A. E., Pulvirenti, A., Ferro, A., and Giorgino, T. (2019). Exploringthe role of interdisciplinarity in physics: Success, talent and luck.
PloSone , 14(6).[Raimbault, 2017] Raimbault, J. (2017). An applied knowledge frame-work to study complex systems.
Forthcoming in CSDM2017 proceedings.arXiv:1706.09244 at https://arxiv.org/abs/1706.09244 .[Reuillon et al., 2013] Reuillon, R., Leclaire, M., and Rey-Coyrehourcq, S.(2013). Openmole, a workflow engine specifically tailored for the distributed13xploration of simulation models.
Future Generation Computer Systems ,29(8):1981–1990.[Rouse et al., 2018] Rouse, W. B., Lombardi, J. V., and Craig, D. D. (2018).Modeling research universities: Predicting probable futures of public vs.private and large vs. small research universities.
Proceedings of the NationalAcademy of Sciences , 115(50):12582–12589.[Shafiee and Berglund, 2019] Shafiee, M. E. and Berglund, E. Z. (2019).Agent-based modelling approach to evaluate the effect of collaborationamong scientists in scientific workflows.
Journal of Simulation , 13(1):1–13.[Shneiderman, 2018] Shneiderman, B. (2018). Twin-win model: A human-centered approach to research success.
Proceedings of the National Academyof Sciences , 115(50):12590–12594.[Thurner et al., 2019] Thurner, S., Liu, W., Klimek, P., and Cheong, S. A.(2019). The role of mainstreamness and interdisciplinarity for the relevanceof scientific papers. arXiv e-prints , page arXiv:1910.03628.[Tisue and Wilensky, 2004] Tisue, S. and Wilensky, U. (2004). Netlogo: Asimple environment for modeling complexity. In
International conferenceon complex systems , volume 21, pages 16–21. Boston, MA.[Urbanska et al., 2019] Urbanska, K., Huet, S., and Guimond, S. (2019).Does increased interdisciplinary contact among hard and social scientistshelp or hinder interdisciplinary research?
PloS one , 14(9).[van Vlokhoven, 2019] van Vlokhoven, H. (2019). The effect of open accesson research quality.
Journal of Informetrics , 13(2):751 – 756.[Zeng et al., 2019] Zeng, A., Shen, Z., Zhou, J., Fan, Y., Di, Z., Wang, Y.,Stanley, H. E., and Havlin, S. (2019). Increasing trend of scientists toswitch between topics.