Identifying Structure through Web Cartography

In 1998 Duncan Watts and Steven Strogatz discovered that certain characteristics of real world networks could be encapsulated by a relatively straightforward mathematical formula. More specifically, Watts and Strogatz (Reference Watts and Strogatz1998) found that networks exhibit properties of both complex and random systems, and that using a graphic model that introduced randomness into order revealed specific rules concerning network configuration, allowing them to discern a “universal class of networks; that is, a family of networks that share certain aggregate properties regardless of many of their individual details” (Watts, Reference Watts2004). Their discovery has become popularly known as the “small world phenomenon,” and has been replicated in the network configuration of social organizations, cities, the Internet, the World Wide Web, cerebral neurons, patterns of wealth accumulation, communication networks, power lines, and a myriad of other entities that are predicated on the small world architecture (Buchanan, Reference Buchanan2002; Johnson, Reference Johnson2001; Milo et al., Reference Milo, Shen-Orr, Itzkovitz, Kashtan, Chklovskii and Alon2002; Taylor, Reference Taylor2003).

This insight into complex networks has provided scientists with a set of global network properties that may be used to classify structure and understand how recurring patterns of organization condition network configurations. The pervasiveness of the small world phenomenon in dynamic social networks and human organization has the potential to trigger a paradigmatic shift in the social sciences. If, indeed, an empirically verified set of scientifically based rules is available to identify network structure, describe their evolution, and provide insight into the relational context among participants, then this discovery may be able to answer many of what were previously considered recalcitrant problems in the social sciences.

The application of small world networks for understanding the structure of online communities is associated with the popular theory of “six degrees of separation,” which measures the structural properties of the network through path lengths (the average separation between two vertices within the network). In the case of the Web, the degrees of separation, the average number of mouse clicks it takes to locate a specific document, is nineteen (Barabási, Reference Barabási2000). Considering the size of the Web, this nominal path length suggests that an underlying network is anything but random. The key attribute of random networks is that each node has an equal opportunity of connecting to another node, which produces a normal distribution curve that stops at the horizontal axis. In contrast, scale-free networks are characterized by a preponderance of scantily connected nodes and a restricted minority of highly connected nodes, which tend to grow throughout the life of the network (Watts, Reference Watts2003). Preferential node attachment results in clustering and non-uniform, continual network growth, suggesting some nodes will possess a higher number of linkages within the network (Barabási and Albert, Reference Barabási and Albert1999).

Examples of scale-free networks abound in the social world, with distributions derived from the power law or Pareto's law. The power law differs from a normal distribution curve in that it does not have a peak at its mean value, nor does it have a rapid decay (Watts, Reference Watts2003). One of the best-known examples of Pareto's law is associated with the distribution of wealth, where privileged nodes have a greater influence on the network. Just as privileged citizens have a greater influence in society, policy brokers have a greater influence on their domains, and social connectors have a greater influence on their circle of friends. The same holds true on the Web, with some pages possessing an anomalously large proportion of in-links. As Albert-László Barabási (Reference Barabási2000) explains,

As individuals surf the Web for information, search engines direct online activities through an independent indexing of the Web's content, assuming greater authority or legitimacy of information by the quantity of users endorsing the source. This human annotation or shadow produces two types of Web pages, hubs and authorities, which are systematically organized into topical communities (Kleinberg, Reference Kleinberg1999).

Hubs and authorities are determined by the intrinsic social organi-zation of the Web and are mutually reinforcing.

Consider the significance of a link p → q: with such a link p suggests, or even recommends, that surfers visiting p follow that link and visit q. This may reflect the fact that pages p and q share a common topic of interest, and that the author of p thinks highly of q's content. Such a link, called an informative link, is p's way to confer authority on q. (Lempel and Moran, Reference Lempel and Moran2000)

Authorities are those pages that are topically organized, provide information and guide the user to alternative sources of related knowledge. Hubs are centralized nodes that direct users to various authorities embedded within that network. Each individual Web page in the network has the potential to receive inbound links and to publish outbound links, and thus each node is characterized by both a hub and authority weight. Authority weights are calculated by the summation of all inbound links (a_p = Σh _q q such that q → p), while hub weights are determined by summing all outbound hypertext links (h _p = Σa _q q such that p → q).

Using a Web-based algorithm that retrieves large quantities of data from the Internet, the topographical structures of the online communities are calculable through measures of link information. Manipulating link information allows the identification of various online communities and traces incoming and outgoing links to determine connectivity and identify network structure. Hypertext links, generated by human judgment, are the ties that bind the Web together, and when users add links to their homepages, they are revealing both an interest in a specific topic, as well as conferring trust in another related Web page. Extraction and identification of those links most relevant to the topical community provides identification of the key nodes (Kleinberg, Reference Kleinberg1999). Informative links are thus critical to network identification and are the foundations upon which the community or topical cluster is defined.

Despite the heterogeneity of pages, online hyperlinked collections are self-organizing systems that evolve naturally (Flake et al., Reference Flake, Lawrence, Giles and Coetzee2002). Nodes tend to cluster around topical interests, with new members joining through preferential attachment to already well-connected pages. The absence of a central organizing body online allows Web communities to evolve collectively without a predetermined organizing principle, suggesting that the network is robust and resistant to erosion (Tu, Reference Tu2000). Distilling virtual networks is thus achieved by identifying the relationships between highly connected nodes operating as information hubs that direct users to trusted online sources. The status of these nodes, and the density of the linkages among them, reveal the ideological, political and policy-orientated propensities of the community.

Tracing virtual networks from the Web is achieved through a methodology referred to as link-structure analysis, which applies various techniques to study linkage patterns (Foot et al., Reference Foot, Schneider, Dougherty, Xenos and Larsen2003; Park, Reference Park2003; Park and Thelwall, Reference Park and Thelwall2003; Rogers, Reference Rogers2002). The largest hubs on the Web are search engines that use electronic spiders that crawl through the Web space, mapping the network through hyperlinked ties. Using a search engine, researchers may enter specific search criteria that, when processed through a search engine, will distill critical nodes (hubs and authorities) in the communities under investigation. Virtual networks are identified using an algorithm that maps the networks using a focused sub-graph of the Web (Rogers and Marres, Reference Rogers and Marres2000). The method was first developed as a search engine optimization procedure used to analyze the cartography of the social ties between Internet-mediated communities publicizing their affiliations, partnerships and belief systems in a hyperlinked environment. The architects who originally created this algorithm based their engine on hub and authority weight/rank measures by using the citations of the Web graph (Brin and Page, Reference Brin and Page1998). Following from Jackson (Reference Jackson1997), this project combines the technologically embedded human annotation between Web sites and hyperlink text with traditional social network analysis.

The computational complexity of collecting and analyzing a large number of nodes is achieved through the use of a specific algorithm designed to crawl through the Web's link-structure extracting informative links that are recorded in a large database. The algorithm used for data capture in this project was designed to determine a ratio for each source point and provide a weighted ranking. The engine used in this project was the Issue Crawler, which is designed to map a Web space using seeded URLs. Developed by Richard Rogers and the govcom.org foundation in Amsterdam, the Issue Crawler generates raw data and provides network visualization (Rogers, Reference Rogers2002). The issue space in each network is defined broadly using Boolean criteria that identifies domain (in this case .ca) and names the policy field or sector. For this project, all initializing nodes were gathered from Google with various content filters applied, disallowing linkages from Amazon.com and eBay, for instance. To locate an online community, “seeds” or initializing nodes (URLs) are entered into the Issue Crawler, which then begins mapping the Web graph from the outgoing links of the seeds (Rogers and Marres, Reference Rogers and Marres2000; Rogers, Reference Rogers2002). Network membership is determined through co-link analysis, which only includes those nodes that are bidirectionally paired (p ↔ q) and that have received a minimum of two inbound links from a third node in the sample population. The engine crawls two layers of hyperlink text, beginning with the pages derived from the seeds and tracing two more levels thereafter.

After the parameters of the network are determined and unrelated links are discarded, a second iteration occurs mapping the link structure of the networks by measuring linkage occurrences. Using the Issue Crawler for data collection, the link structure of the four networks under consideration (Aboriginal policy, agriculture, banking and women-centred policy) were located and mapped. Each of the four networks did constitute a VPN, as each was characterized by growth through preferential attachment and scale-free distribution.

Modeling Virtual Policy Networks

Knowing that VPNs exist allows for consideration of the extent to which these Web-based networks mimic their offline counterpart policy communities. Drawing insight from existing empirical studies and Canadian policy scholars, predictions concerning the structural configuration of these four networks are provided in this section. However, prior to hypothesizing structure, a conceptual model designed to empirically validate network configuration is necessary. The model must be dexterous enough to be applied in both real world and virtual settings.

One of the most recent studies of policy network configurations in Canada was undertaken by Howlett (Reference Howlett2002), who sought to determine the empirical validity of network analysis. Howlett's evaluation of network structures was conducted through an in-depth study of longitudinal policy change in four sectors: banking, education, trade and transportation. This study produced evidence that network configurations provided reliable, predictive measures concerning the pace and type of policy change.

As Howlett's model is grounded in network theory, the application of this logic to virtual policy settings is reasonable. According to the model, the policy community is comprised of the set of all actors and institutions engaged in the policy process in any given domain. Howlett identifies two primary constellations of participants: a discourse community and an interest network. The first subset of participants he distinguishes is the discourse community, which is comprised of various actors associated through relationships premised on identifiable sectoral/topical affiliation. Alternatively, an interest network refers to actors engaged in information exchange relations, whereby participation is motivated through self-interest. Howlett's model (Reference Howlett2002: 248–9) posits that an important aspect of the “cohesiveness” or “closedness” of policy subsystems, identified in earlier studies as a key dimension of subsystem structure related to policy change, is the nature of the relationship existing between the discourse community and the interest network. Subsystems which feature closely integrated communities and networks will be more cohesive, and better able to resist the entrance of new ideas and actors into policy processes than will those which feature sizable distances between the two subsets of actors.

Howlett recognizes four ideal-type networks, each characterized by the policy subsystem's permeability to new actors and ideas. Policy communities with tightly fused membership will display strong cohesion and be insulated from outside influences. An open structure suggests a network configuration more exposed to external information circulation and the entrance of new policy actors, while closed subsystems are protected from exogenous influences and thus remain stable and insulated. Howlett also identifies resistant and contested networks, which are similarly identified by degrees of insulation and extent of symmetry between network and community (Table 1).

Table 1 Howlett's Operationalization of Policy Subsystem Configurations

Howlett's model was substantiated with evidence from four sub-sectors (airline deregulation, continental free trade, funding for federal-provincial post-secondary education, and the deregulation of federal banking). Considerations of sub-system policy goals, instrumental types, and programme specifications were collected to observe the policy dynamics of each sector and to determine the insulation and symmetry of the core of each policy network. The hypotheses concerning VPN configuration are premised on Howlett's model and are expected to reveal a relationship between network hyperlink structure and internal network dynamics. To test this, the two independent variables (symmetry and isolation) will be measured separately and then combined to determine where each VPN is located in Howlett's model. Applying this framework to the four VPNs under consideration should determine whether Web-based networks have similar structural configurations of real world policy communities.

Applying Howlett's model to the Aboriginal VPN, it is hypothesized that this network will be characterized by a closed structural configuration with a closely controlled relationship between the federal government and First Nations. Jonathon Malloy's (Reference Malloy2001) research suggests that this network should be characterized by contestable and competitive policy dynamics, as the bond between provincial institutions and Aboriginal populations is not as collaborative as that found between First Nations and Ottawa. However, at the national level, the constitutionally entrenched and fully institutionalized relationships with First Nations and the federal government would be expected to produce a set of dominant actors with flows concentrated at the core (Abele and Prince, Reference Abele, Prince, Bakvis and Skogstad2002). These state actors tend to have privileged status in decision-making outcomes, with First Nations and complex bureaucratic structures managing land claims, justice, and economic and social programme coordination.

The agriculture VPN is hypothesized to feature a structural configuration that would be classified as a resistant network. Although the offline agriculture policy community has traditionally been considered a closed policy community, this sector has been experiencing various exogenous shocks in the recent past as a result of environmental concerns (Montpetit, Reference Montpetit2002), the policy overlap with the biotechnology sector (Montpetit and Garson, Reference Montpetit and Garon2004), and increased media attention arising from such focusing events as the mad cow scare, avian flu fears and genetically modified seed. Furthermore, the VPN has significant connections to the American government and various international policy institutes, suggesting that the entrance of new international actors into the domestic agricultural policy community may be conditioning network behaviour (Coleman, Reference Coleman2003).

On the other hand, I predict that the banking VPN possesses a contested structural configuration. Research focused on this sector's real world policy activities suggests that the community will be influenced by high levels of internationalization and extensive interdependencies in the global policy domain (Coleman, Reference Coleman1996). The Canadian banking policy community has gone through various stages of adaptation in the last decade, including significant changes in the global financial structure (Coleman, Reference Coleman1996), new regulatory processes and service reforms, internal lobbies seeking mergers, and various attempts to expand eligibility criteria so that national banks may compete in the global market (Harris, Reference Harris2004b). Challenges arising from the larger policy universe and interest articulations outside the core have not toppled policy monopolies but instead seem to have triggered venue shopping (Harris, Reference Harris2004a). Drawing from real world analysis, it is hypothesized that the banking network will feature limited information flows beyond the interest network and higher levels of permeability.

Finally, I expected the women's-centred VPN to be very porous, open to flows of competing information and fluid membership. Feminist organizational efforts aimed at achieving greater inclusion in policy making and democratic activity may be partially responsible for the configuration of this community (Burt, Reference Burt, Bickerton and Gagnon1999). There have also been several concerted efforts to increase women's online political mobilization using the Internet as a mechanism of empowerment (Davis, Reference Davis, Alexander and Pal1998; Harcourt, Reference Harcourt and Gauntlett2000). In the last decade, traditional state control over the Canadian women's policy community (Pal, Reference Pal1993) has been dramatically restructured, partially through the loss of state funding and partially through the efforts of feminist policy scholars, women's community activity, femocrats and critical policy analysts (Brodie, Reference Brodie1996; Burt, Reference Burt, Burt, Code and Dorney1993, Reference Burt and Gingras1995, Reference Burt, Bickerton and Gagnon1999; McLaren, Reference McLaren1998). Thus this VPN is hypothesized to be an open network with free flows of information.

Do Virtual Policy Networks Matter?

Howlett's model (Reference Howlett2002) suggests that the organizational structures of real world policy communities will condition policy dynamics and propensity towards change. He identifies two primary variables: network symmetry and core insulation from exogenous forces. To determine the levels of symmetry and permeability of the four VPNs studied here, each network was tested for density and information flows. The policy universe of each VPN was comprised of the full set of nodes culled from the Web crawl. To test for symmetry, the top 20 percentile of pages were chosen to represent the core or interest network. The universe of VPN nodes was then subdivided into two sets that paralleled Howlett's discourse community and interest network typology. Next, the core or interest network's hub weights (outbound links) were calculated to determine the degree of insulation the interest network enjoys from the discourse community.

If the pages published in the discourse community are recognized, and linked to, the core, then the network is open. If the core is insulated from Web pages advocating alternative policy positions, then the network is closed. Determining structural configuration of VPNs is in the earliest stages of development, and as a result the cutoffs for determining symmetry and insulation presented in this research note are experimental, with 72.05 being the mean average of 40 VPNs' core hub weights previously tested by the author. Thus any VPN with a hub weight of less than 72 is considered less insulated or more open, while a network with a core hub weight of greater than 72 is considered more insulated and closed. The Aboriginal policy VPN represented a closed, symmetric network, as did banking. Alternatively, the women's and agriculture VPNs were less symmetric, with agriculture particularly revealing lower degrees of cohesion (Table 2).

Table 2 Levels of Insulation

The second key characteristic to be tested was permeability or resistance to new ideas and actors, which was accomplished by simply studying the density of information flows. As already discussed, VPNs will reference other sources of information and provide lists of advocacy partners, providing researchers with strong evidence of information coalitions. Following Howlett's findings, circulation of informational sources and ideas is variable across networks, and should be evident in the amount and concentration of information circulating in the online policy community. To test for this, only the hyperlinks connecting the core pages are considered in determining the information density among the interest network.

To measure information flow, the links hosted by the core are subtracted from the universe of network linkages and normalized into percentages. The higher the concentration of linkages in the interest network, the lower the flow of competing information, suggesting that the network's core belief system is well established and resistant to alternative policy views. Similarly, higher flows of information would indicate that a network is more porous, providing opportunities for alternative agendas or competing perspectives. A mean average of information flows was calculated again using data from 40 previously tested VPNs. Within the context of this sample, networks characterized by information flows of less than 65 percent are considered more nebulous and open to competing policy alternatives and ideas. Alternatively, networks with a core concentration of information flows greater than 65 percent are more prone to lower levels of symmetry, and more protected interest networks. Of the four cases, the Aboriginal and agriculture VPNs are characterized by controlled (lower) information flows, while women's policy and banking are distinguished by higher degrees of information flows (Table 3).

Table 3 Symmetry of Information Flow

After measuring the two variables, degrees of insulation and extent of symmetry are combined so that each of the four networks may be located in Howlett's model (Table 4).

Table 4 Virtual Policy Network Configurations

The structural configuration of the women's VPN features low levels of insulation and symmetry, and thus represents an open network, amenable to new actors and discursive online dialogue. The Aboriginal VPN is characterized by higher levels of symmetry and insulation, suggesting that it would be closed to new actors and ideas. The agriculture VPN reveals attributes of a resistant network in that it features lower degrees of insulation combined with higher flows of information concentrated at the core. Finally, banking was found to be a contested network with low symmetry and higher insulation.