Project topics and materials
Get Complete Project Material File(s) Now! »
Epistemic communities
In this chapter, we present the existing works concerning epistemic community appraisal and representation, and we introduce a formal framework along with various definitions.
Context
Several works ranging from social epistemology to political science and economics have given an account of the collaboration of agents within the same epistemic framework and towards a given knowledge-related goal, namely knowledge cre-ation or validation. For social epistemologists, it is a scientist group, or epistemic community, producing knowledge and recognizing a given set of conceptual tools and representations — the “paradigm,” according to Kuhn (1970) — possibly work-ing in a distributed manner on specialized tasks (Schmitt, 1995; Giere, 2002). Con-sidering a whole knowledge field as a huge epistemic community (e.g. biology, linguistics), one can see subdisciplines as smaller, embedded, and more specific epistemic communities — subfields within a paradigm. Haas (1992) introduced the notion of epistemic community as “a network of knowledge-based experts (…) with an authoritative claim to policy-relevant knowledge within the domain of their expertise.” Cowan, David and Foray (2000) added that an epistemic community must share a subset of concepts. To them, an epistemic community is “a group of agents working on a commonly acknowledged subset of knowledge issues and who at the very least accept a commonly understood procedural authority as essential to the success of their knowl-edge activities.” The “common concern” aspect has been emphasized by Dupouet, Cohendet and Creplet (2001) who define an epistemic community as “a group of agents sharing a common goal of knowledge creation and a common framework allowing to understand this trend.” These authors nevertheless acknowledge the need of a notion of authority and deference.
On the other hand, scientists have shown an increasing interest for methods of knowledge community structure analysis. Several conceptual frameworks and automated processes have been proposed for finding groups of agents or docu-ments related by common concepts or concerns, notably in knowledge discovery in databases (KDD) (Rocha, 2002; Hopcroft et al., 2003) and scientometrics (Ley-desdorff, 1991a; Lelu et al., 2004). Dealing with and ordering categories automati-cally has indeed become central in data mining and related fields (Jain et al., 1999), along with the massive development of informational content. Besides, since a large amount of data is freely and electronically available, the study of scientific communities in particular has attracted a large share of the interest — especially biologist communities: biology is a domain where the need for such techniques is also the most pressing because article production is so high that it becomes hard for scientists to figure out the evolution of their own community.
Yet, existing approaches in community finding are often either based on so-cial relationships only, with community extraction methods stemming from graph theory applied to social networks (Wasserman & Faust, 1994), or on semantic simi-larity only, namely clustering methods applied to document databases where each document is considered as a vector in a semantic space (Salton et al., 1975). There have been few attempts to link social and semantic aspects, although the various characterizations of an epistemic community insist on its duality, i.e. the fact that such a community is on one side a group of agents who, on the other side, share common interests and work on a given subset of concepts. By contrast, only sci-entometrics have developed a whole set of methods for characterizing specifically such communities, working on both scientists and the concepts they use. Cate-gorization has been notably applied to scientific community representation, using inter alia multidimensional scaling in association with co-citation data (McCain, 1986; Kreuzman, 2001) or other co-occurrence data (Callon et al., 1986; Noyons & van Raan, 1998), in order to produce two-dimensional cluster mappings and track the evolution of paradigms (Chen et al., 2002).
Along with this profusion of community-finding methods, often leaning to-wards AI-oriented clustering, an interesting issue concerns the representation of communities in an ordered fashion. On the whole, many different techniques have been proposed for producing and representing categorical structures including, to cite a few, hierarchical clustering (Johnson, 1967), Q-analysis (Atkin, 1974), for-mal concept analysis (Wille, 1982), information theory (Leydesdorff, 1991b), block-modeling (White et al., 1976; Moody & White, 2003; Batagelj et al., 2004), graph theory-based techniques (Newman, 2004; Radicchi et al., 2004), neural networks (Kohonen, 2000), association mining (Srikant & Agrawal, 1995), and dynamic ex-ploration of taxonomies (Sacco, 2000). Here, the notion of taxonomy is particularly relevant with respect to communities of knowledge. A taxonomy is a hierarchical structuration of things into categories, as such an ordered set of categories (or taxons), and is a fundamental tool for representing groups of items sharing some properties. Taxonomies are useful in many different disciplinary fields: in biol-ogy for instance, where classification of living beings has been a recurring task (Whittaker, 1969; Simpson & Roger, 2004); in cognitive psychology for modeling categorical reasoning (Rosch & Lloyd, 1978; Barthélemy et al., 1996); as well as in ethnography and anthropology with folk taxonomies (Berlin, 1992; Lopez et al., 1997; Atran, 1998). While taxonomies have initially been built using a subjective approach, the focus has moved to formal and statistical methods (Sokal & Sneath, 1963; Benzécri, 1973).
However, taxonomy building itself is generally poorly investigated; arguably, taxonomy evolution during time has been fairly neglected. Our intent here is to address both topics: build a taxonomy of epistemic communities, then moni-tor its evolution — as such a work which shares the aims of history of science. At the same time while taxonomies have long been represented using tree-based structures, we wish to produce taxonomies which deal with sub-communities af-filiated with multiple communities (such as interdisciplinary groups) or of di-verse paradigmatic statuses (i.e., rendering equally communities centered around methods, processes, fields of application, given objects, etc.); therefore introducing lattice-based structures.
Definitions
Basically, we are first trying to know (i) which agents share the same concerns and work on the same concepts, and (ii) which these concerns or concepts are. We are thus farther from the epistemological point of view and need not characterize authoritative groups and their role. Hence, the definitions of an “epistemic com-munity” introduced in the previous section seem to be too precise with respect to authoritative and normative properties, while they lack the ability to formalize community boundaries and extents accurately. Obviously, an epistemic commu-nity that is simply characterized by common knowledge concerns should not nec-essarily be a social community, with agents of the same communitiy enjoying some sort of social link: it is neither a department nor a group of research. In addition, we want a definition that allows some flexibility in the sense that an agent or a semantic item (or concept) can belong to several communities. Therefore, we adopt the following definition, keeping the notion of common “knowledge issues”, to which we add maximality:
Definition EC-1 (Epistemic community). Given a set of agents S, we consider the concepts they have in common and we call epistemic community of S the largest set of agents who also use these concepts.
In other words, taking the epistemic community (EC) of a given agent set ex-tends it to the largest community sharing its concepts. This notion is to be com-pared with the structural equivalence introduced in sociology by F. Lorrain and H. White (1971). Structural equivalence describes a community as a group of peo-ple related in an identical manner to a set of other people. When extending this concept to a group of people related identically to the same concept set, ECs are groups of agents related in an equivalent manner to some concepts.
Definition EC-1 is based on an agent set, and we could define correspondingly an epistemic community as the largest set of concepts commonly used by agents who share a given concept set. We will at first focus on agent-based epistemic com-munities, keeping in mind that concept-based notions are defined strictly equiva-lently and in a dual manner. In order to set up a comprehensive framework allow-ing to work on these notions, we now introduce a few basic definitions:
Definition 1 (Intent). The intent of a set of agents S is the set of concepts which are used by every agent in S.
Definition 2 (Epistemic group). An epistemic group is a set of agents provided with its intent, i.e. a group of agents and the concepts they have in common.
Consider for instance that some given agents s1, s2 and s3 work on “linguis-tics” (Lng), while “neuroscience” (NS) is being used by s2, s3 and s4 (Fig. 1.1). Therefore, the intent of {s1, s2, s3} is {Lng}, that of {s2, s3, s4} is {NS} and that of {s2, s3} is {Lng, NS}. Some epistemic groups of this example are thus ({s1, s2, s3}; {Lng}), ({s2, s3}; {Lng, NS}) and ({s1, s4}; {∅}).
For a given set of agents S, knowing its epistemic community comes to identi-fying the largest group of people who share the same knowledge issues as those of agents of S (this largest group thereby includes S) — notably, for a group of agents prototypic of a field, this amounts to know the whole set of agents of the field.
Definition 3 (Hierarchy, maximality). An epistemic group is larger than another epis-temic group if and only if (i) their intents are the same and (ii) the agent set of the former contains that of the latter.
An epistemic group is said maximal if there exists no larger epistemic group.
This statement enables us not only to compare epistemic groups but also and more significantly to expand a given epistemic group to its maximal social size. Interpreting definition EC-1 within this framework leads to the following refor-mulation:
Definition EC-2 (Epistemic community). The epistemic community based on a given agent set is the corresponding maximal epistemic group.
The epistemic community based on {s4}, for instance, is thus ({s2, s3, s4}; {NS}), and the one based on either {s1} or {s1, s2} is ({s1, s2}; {Prs, Lng}).1
Notice that we can similarly define an EC based on a concept set as the largest set of concepts sharing a given agent set. We introduce the concept-based notions, defined symmetrically to the agent-based notions, and thus, in the remainder of the thesis we will equivalently denote an EC by its agent set S, its concept set C or the couple (S, C).
Definition 4 (Extent, concept-based notions). The extent of a set of concepts C is the set of agents using every concept in C. A concept-based epistemic group is a set of concepts provided with its extent. A concept-based epistemic group is larger than another one if and only if (i) their extent are the same and (ii) the concept set of the former contains that of the latter. A concept-based epistemic community is a maximal concept-based epistemic group.
Formal framework
In order to work formally on these notions, we need to bind agents to concepts through a binary relation R between the whole agent set S and the whole concept set C. R expresses any kind of relationship between an agent s and a concept c. The nature of the relationship depends on the hypotheses and the empirical data. In our case, the relationship represents the fact that s used c (e.g. in some article).
Sets and relations Let us consider R ⊆ S × C binding S to C. We introduce the operation “∧” such that for any element s ∈ S, s∧ is the set of elements of C which are R-related to s. Extending this definition to subsets S ⊆ S, we denote by S∧ the set of elements of C R-related to every element of S, namely:
s∧ = { c ∈ C | sRc } (1.1a)
S∧ = { c ∈ C | ∀s ∈ S, sRc } (1.1b)
Similarly, “⋆” is the dual operation so that ∀c ∈ C, ∀C ⊆ C,
c⋆ = { s ∈ S | sRc } (1.2a)
C⋆ = { s ∈ S | ∀c ∈ C, sRc } (1.2b)
By definition we set (∅)∧ = C and (∅)⋆ = S.
Definitions 1, 2 and 4 mean that if S is a set of agents, S∧ denotes its intent, the set of concepts used by every agent in S (“∀s ∈ S”). Similarly if C is a concept set, C⋆ is its extent, the set of agents who use every concept in C. Thus, epis-temic groups are couples of kind (S, S∧) or (C⋆, C). On the sample community described on Fig. 1.1, we have for instance {s1, s3}∧={Lng} and {NS, prs}⋆={s3}. As Wille (1997) points out, this formalism constitutes a robust and rigourous way of dealing with abstract notions (in a philosophical sense), characterized by their extent (physical implementation) and their intent (properties or internal content). Here, concepts are properties of authors who use them (they are skills in scientific fields, i.e. cognitive properties) and authors are loci of concepts (concepts are im-plemented in authors).
Table of contents :
General introduction
I Knowledge Community Structure
Introduction
1 Epistemic communities
1.1 Context
1.2 Definitions
1.3 Formal framework
2 Building taxonomies
2.1 Taxonomies and lattices
2.2 Galois lattices
2.3 GLs and categorization
2.3.1 About relevant categorization
2.3.2 Assumptions on EC structure
2.3.3 GLs and selective categorization
2.4 Comparison with different approaches
3 Empirical results
3.1 Experimental protocol
3.2 Results and comparison with random relations
3.2.1 Empirical versus random
3.2.2 Rebuilding the structure
4 Community selection
4.1 Rationale
4.2 Selection methodology
5 Taxonomy evolution
5.1 Empirical protocol
5.2 Case study, dataset description
5.3 Rebuilding history
5.3.1 Evolution description
5.3.2 Inference of an history
5.3.3 Comparison with real taxonomies
6 Discussion and conclusion
II Micro-foundations of epistemic networks
Introduction
7 Networks
7.1 Global overview
7.2 A brief survey of growth models
7.3 Epistemic networks
8 High-level features
8.1 Empirical investigation
8.2 Degree distributions
8.3 Clustering
8.4 Epistemic community structure
9 Low-level dynamics
9.1 Measuring interaction behavior
9.1.1 Monadic PA
9.1.2 Dyadic PA
9.1.3 Interpreting interaction propensions
9.1.4 Activity and events
9.2 Empirical PA
9.2.1 Degree-related PA
9.2.2 Homophilic PA
9.2.3 Other properties
9.2.4 Concept-related PA
9.3 Growth- and event-related parameters
9.3.1 Network growth
9.3.2 Size of events
9.3.3 Exchange of concepts
10 Towards a rebuilding model
10.1 Outline
10.2 Design
10.3 Results
10.4 Discussion
Conclusion
III Coevolution, Emergence, Stigmergence
Introduction
11 Appraising levels
11.1 Accounting for levels
11.2 Emergentism
11.3 What levels are not
11.4 Observational reality of levels
11.4.1 Different modes of access
11.4.2 Illustrations
12 Complex system modeling
12.1 Complexity and reconstruction
12.1.1 Objectives
12.1.2 Commutative decomposition
12.1.3 Reductionism failure
12.1.4 Emergentism
12.2 A multiple mode of access
12.2.1 The observational viewpoint
12.2.2 Introducing new levels
12.2.3 Rethinking levels
13 Reintroducing retroaction
13.1 Differentiating objects
13.2 Agent behavior, semantic space
13.3 Coevolution of objects
13.4 “Stigmergence”
Conclusion
General conclusion
Version française abrégée
Introduction
Partie I — Structure des communautés de savoirs
1.1 Cadre formel
1.2 Treillis de Galois: des relations aux taxonomies dynamiques
1.3 Etude de cas
Partie II — Micro-fondations des réseaux épistémiques
2.1 Réseaux
2.2 Caractéristiques de haut-niveau
2.3 Dynamique de bas-niveau
2.3.1 Mesure du comportement d’interaction
2.3.2 AP empirique
2.3.3 Paramètres liés à la croissance et aux événements
2.4 Modèle de reconstruction
Partie III— Coévolution, émergence, stigmergence
3.1 Niveaux de description
3.2 Modélisation des systèmes complexes
3.2.1 Complexité et reconstruction
3.2.2 Réintroduire la rétroaction
Conclusion
List of figures
References
