Shuichiro Ogawa
日本語

Notes · updated 2026-10-11

How Have Network Science and Design Studies Related to Each Other? (Academic Research)

This note examines, through the literature, the relationship between network science and design studies from three angles: how design research has used network methods, how network science has used the word "design", and where the two meet and where they reach their limits.

Contents (7)
  1. Two lines of work that started from the same question
  2. What has design research measured with networks?
  3. How has network science used the word “design”?
  4. Whose judgment is it to draw a network?
  5. Limits of the borrowed tools
  6. The relationship between the two fields
  7. What research has not yet answered

The citation network of this wiki draws the sources cited by the public notes as points and draws a line between two sources whenever they appear in the same note. The literature on how to draw it is summarized in Drawing Networks as Data Art: Literature on Placement, Bundling, Time, and Flow, and the way it is counted is described in The Citation Network of This Wiki: Analysis and Artworks Treating Each Note as a Hyperedge. One question was left out there. How have the field that built the drawing tools (network science) and the field this wiki mainly deals with (design studies) used each other? More specifically, what has design research measured with network methods, what has network science meant by the word “design”, and where have the two met and where have they reached their limits? Sources checked only through their abstracts are marked “abstract” in the citation parentheses. How far each source was read is recorded in the verification log of the corpus.

Two lines of work that started from the same question

The two lines of work start from the same question in the 1960s. The question is how to divide a complex whole into parts that can be handled. In 1962, Simon argued that hierarchic systems have the property of being nearly decomposable (Simon 1962). In a nearly decomposable system, the short-run behavior of each component is approximately independent of the other components, and in the long run each component depends on the others only in an aggregate way. Simon explained this with the temperature of a building divided into rooms, and he described it as a matrix of linkages in which all elements outside a string of square submatrices along the main diagonal are zero or small. Because linkages within components are stronger than linkages between them, representing such systems as hierarchies loses comparatively little information; this is Simon’s argument.

In Notes on the Synthesis of Form (1964), Alexander presented a method that represents design requirements as a graph and solves the design problem by dividing the graph into parts with few interactions between them (Gabriel 2023). Gabriel re-read the reports on the partitioning programs (HIDECS) that Alexander wrote for this method. He confirmed that the programs called the requirements “vertices” and the interactions between them “links”, and that they built a binary tree by repeatedly splitting the vertices into two sets with few links crossing the boundary (Gabriel 2023).

However, in “A City Is Not a Tree” the following year, Alexander moved away from this direction. Alexander distinguished cities that have arisen more or less spontaneously over many years (Siena, Liverpool, Kyoto, Manhattan) from cities deliberately created by designers and planners (Alexander 1965). In a tree, for any two sets, either one is wholly contained in the other or they are wholly disjoint; in a semilattice, the overlap of two overlapping sets is also a unit. Alexander argued that many designers have conceived cities as trees, whereas the natural structure of a city is in every case a semilattice. By Alexander’s count, a tree based on 20 elements can contain at most 19 further subsets, while a semilattice based on the same elements can contain more than 1,000,000. Alexander also wrote that the systems of friends and acquaintances form a semilattice, not a tree. Gabriel summarized that in this essay Alexander repudiated the very idea of definitely partitioning a complex design problem into disjoint, tree-like sets, and that his 1971 preface to the paperback edition of Notes repeated the repudiation (Gabriel 2023).

Harary and Rockey examined Alexander’s argument from the side of graph theory. They acknowledged the value of Alexander’s attempt to analyze the city in graph-theoretical terms, refuted some of his arguments and conclusions, and concluded that the city is neither a tree nor a semilattice (Harary and Rockey 1976, abstract). They also stated that, at this stage, the complexity of a city may be understood more readily through negative than through positive descriptions, because graphs with specified properties fail to characterize it.

On the design side, a reaction against the formal methods of this period followed. In Cross’s account, the design methods movement of the 1960s culminated in Simon’s plea for “a science of design”, but in the 1970s some of the early pioneers of the movement turned against it (Cross 2001). Cross quotes Alexander saying that there is so little in what is called “design methods” that has anything useful to say about how to design buildings, and “I would say forget it, forget the whole thing.” In the same period, Rittel and Webber argued that planning problems are wicked problems, which have no definitive formulation and no stopping rule that tells when a solution has been found (Rittel and Webber 1973). Cross positions this argument as separating design and planning problems from the “tame” problems that the methods of science and engineering deal with. He then notes that Schön explicitly challenged the positivism underlying the “design science” movement, and he argues for design as a discipline based on the reflective practice of design (Cross 2001).

What has design research measured with networks?

Design research has used networks mainly in four places: the structure of products and organizations, the design process, urban streets, and teams and the field itself. Chen and colleagues divided network research in engineering design into three areas: the architecture of products and systems, design decisions and the connections among designers, and the markets and institutions in which products and design processes are embedded (Chen et al. 2018). The same editorial states that applications of network science have so far focused largely on understanding the mechanisms of complex systems, and to a lesser extent on using this understanding as an engineering tool.

The structure of products and organizations

Methods that represent design dependencies in a matrix take the same form as Simon’s matrix. In 1981, Steward pointed out that the determination of interdependent design variables involves circuits that conventional planning techniques cannot handle (Steward 1981, abstract). He therefore proposed organizing the dependencies in a matrix and deciding where to use estimates and how to handle iterations. Browning reviewed the uses of this design structure matrix (DSM) and divided them into four types: component-based architecture, team-based organization, activity-based schedule, and parameter-based design processes (Browning 2001, abstract).

This line of work later incorporated measures from graph theory and social network analysis. Sosa, Eppinger, and Rowles treated a product as a network of components, defined three measures of component modularity based on the notion of centrality, and applied them to a large commercial aircraft engine (Sosa et al. 2007, abstract). They reported that the relationship between component modularity and component redesign depended on the type of interfaces connecting the components (Sosa et al. 2007, abstract). In the development of the same engine, the same authors examined how the design interfaces of the product mapped onto the interactions among development teams (Sosa et al. 2004, abstract). They reported that organizational and system boundaries moderated the impact of interface strength and indirect interactions on this mapping, and that these effects were contingent on system modularity (Sosa et al. 2004, abstract). The idea that the structure of a product and the structure of an organization correspond like a mirror image has been tested as the mirroring hypothesis. Colfer and Baldwin reviewed 142 empirical studies and reported that industry and firm studies showed mirroring to be prevalent but not universal, whereas studies of open collaborative projects, mostly in software, did not support the hypothesis (Colfer and Baldwin 2016, abstract).

Some studies applied the statistical tools of network science directly to product development. Braha and Bar-Yam reported that networks of people engaged in distributed problem solving resemble information, biological, and technological networks in being sparse, having small-world properties, and showing scaling regimes (Braha and Bar-Yam 2004, abstract). The distributions of incoming links had sharp cutoffs that were substantially lower than those of outgoing links, and the authors related this asymmetry to the difference between each actor’s capacity to process information from others and the capacity to transmit information. They also analyzed the information-flow networks of product development organizations and stated that their structure has properties similar to those of other social, biological, and technological networks (Braha and Bar-Yam 2007, abstract).

The design process

The design process itself has also been recorded as a network. Goldschmidt’s linkography is a notation that divides a designer’s thinking into small steps called design moves and records the connections between moves as links, focusing on the “critical moves” that are rich in links (Goldschmidt 2014, publisher’s description). According to Smith and colleagues, traditional linkography is time-consuming because a human coder must annotate both the design moves and the connections between them, and as a result it has not been much applied at scale. Smith and colleagues presented a method that constructs a linkograph automatically from a record of design moves by using a computational model of semantic similarity, and they applied it to traces such as text-to-image prompting journeys (Smith et al. 2025).

Kelly and Gero presented a method that uses an LLM to extract concepts from think-aloud protocols of 17 designers working on two design problems, clusters the concepts with word vectors, constructs graphs, and computes network measures (Kelly and Gero 2025). They suggest that these measures are useful for studying how designers understand a problem (design framing) and for aiding qualitative analysis. Across engineering design research, the survey by Han and colleagues shows that researchers have mainly relied on semantic networks built from non-engineering data, such as WordNet and ConceptNet, and that efforts to build engineering-specific semantic networks from technical publications and patents are only emerging (Han et al. 2021).

Urban streets

In architecture and urban design, the space syntax approach of Hillier and colleagues has measured the configuration of street networks as graphs. Hillier and colleagues presented evidence that the configuration of the urban grid, rather than the attraction of land uses, is the main generator of pedestrian and vehicular movement, and that the space syntax measure of “integration” was consistently the most important among the configurational measures (Hillier et al. 1993, abstract). The abstract states the implication that, to design well-used urban space, the configurational relations of a space to the larger urban system matter more than its local properties.

The same street networks were also analyzed from the side of network science. Porta, Crucitti, and Latora pointed out that the idea that some places are more central than others originated in structural sociology before its use in urban studies. They proposed a method that measures the centrality of street networks with multiple indices together, within a metric rather than a topological framework (Porta et al. 2006). Crucitti and colleagues analyzed the street networks of world cities as graphs whose nodes are intersections and whose edges are streets, using four centrality indices. They reported that planned and self-organized cities belong to two different universality classes, and that self-organized cities exhibit scale-free properties similar to the degree distributions of non-spatial networks (Crucitti et al. 2006). Alexander’s 1965 distinction between cities that arose spontaneously and cities that were deliberately created is measured here as a difference in centrality distributions.

Teams and the field

The relationships within design teams have also been studied as social networks. Sosa conducted a sociometric study in the development department of a software firm (Sosa 2011, abstract). He reported that ties transmitting a wide set of knowledge domains favor creative ideas when the exchanges are frequent enough, and that the cohesion of a dyad (its connections to common contacts) becomes detrimental when it exceeds its average level. Ball and Lewis proposed a simulation framework that evaluates the design potential of mass collaboration teams from individual attributes and from team network structure, such as centrality and density (Ball and Lewis 2018).

The field of design research itself has also been drawn as a network. Chai and Xiao analyzed 12,107 citations in 459 papers published in Design Studies from 1996 to 2010 with a co-citation network at the level of individual publications (Chai and Xiao 2012, abstract). They showed that the core themes centered on the design process and design cognition, and that protocol analysis had become more popular in recent years.

How has network science used the word “design”?

Network science refers to a line of research that spread from the late 1990s, as Chen and colleagues wrote in 2018 that it had emerged as an interdisciplinary field “in the last two decades” (Chen et al. 2018). In his 2003 review, Newman named the comparative study of networks from different branches of science, inspired particularly by the 1998 paper of Watts and Strogatz, as one of the principal thrusts of this line of work (Newman 2003). The same review states that graph theory has a long history going back to Euler and that sociology has reconstructed relationships among people from questionnaires, but that the focus has recently shifted from single small graphs and the properties of individual vertices to the large-scale statistical properties of graphs. This shift was driven by the ability to gather and analyze data on networks with millions or billions of vertices.

In this field, “design” was first used for properties of systems that have no designer. Milo and colleagues defined network motifs as patterns of interconnections that occur significantly more often than in randomized networks, and they searched for such motifs in biochemical, neurobiological, ecological, and engineering networks to uncover the structural design principles of complex networks (Milo et al. 2002, abstract). Kashtan and Alon asked why biological networks are modular, with a design that can be separated into units that perform almost independently. Using evolutionary simulations in which the goal switched among combinations of subgoals, they showed that modular structure and network motifs evolve spontaneously (Kashtan and Alon 2005, abstract). The abstract states that this result may also offer ways to improve the evolutionary design of engineered systems. Csete and Doyle argued that advanced technologies and biology, despite very different physical implementations, are alike in system-level organization, and that both converge on modular architectures with hierarchies and layers of feedback driven by the demand for robustness to uncertain environments (Csete and Doyle 2002, abstract).

“Design” was also used to distinguish designed systems from self-organized ones. Carlson and Doyle named HOT (highly optimized tolerance) a mechanism in which systems optimized by natural selection or engineering design face tradeoffs among yield, the cost of resources, and tolerance to risk, which lead to designs that allow occasional large events and therefore to power-law distributions (Carlson and Doyle 1999). They contrasted this mechanism with self-organized criticality, which had been widely suggested as the origin of power laws. From the opposite direction, Solé and colleagues pointed out that biological networks, which should result from tinkering, appear to settle into a limited repertoire of forms that optimize communication, whereas technology graphs, although designed, show strong effects of tinkering and internal constraints (Solé et al. 2002, abstract).

“Design” as balancing cost and performance under constraints was measured in networks embedded in space. Gastner and Newman showed that real distribution networks, such as gas pipelines and sewer systems, are close to optimal both in total edge length (cost) and in the directness of routes (efficiency) (Gastner and Newman 2006). They noted that these networks were not designed from the outset for global optimality but formed by growing outward from a root, and they proposed two models of such growth. Tero and colleagues placed the slime mold Physarum polycephalum at the location of Tokyo and placed food sources at the major cities of the surrounding region (Tero et al. 2010). They showed that the network the slime mold formed had efficiency, fault tolerance, and cost comparable to those of the real rail network, and they captured the mechanism of this adaptive network formation in a mathematical model. The title of the paper refers to rules for biologically inspired adaptive network design.

In research on the Internet, models that ignored design constraints were criticized. Li and colleagues criticized the view that treats the router-level topology of the Internet as a member of an ensemble of random networks rather than as an engineering system driven by economic and technological constraints plus some randomness (Li et al. 2005). They showed that, among graphs with an identical degree sequence, only a careful design process that explicitly incorporates technological constraints, traffic demands, and link costs yields high-performance networks.

In the network science literature collected here, “design” refers to optimization that balances cost, efficiency, and robustness under constraints. Whether the optimization was carried out by natural selection, by designers, or by a growth process differs from study to study. This view is close to the “science of design” that Simon called for, and it deals with something other than the design problems that Rittel and Webber described, whose very formulation is not settled.

Whose judgment is it to draw a network?

The two fields come closest when networks are drawn. Newman wrote that, for networks of tens or hundreds of vertices, drawing a picture of points and lines and reading the structure from it had been one of the primary methods of network analysts (Newman 2003). For networks of a million vertices, however, this approach is useless, and statistical methods are required. Even after network science moved toward statistics, the practice of drawing and reading pictures remained, and how to make such pictures has been studied as a design problem for human readers.

In the field of graph drawing, algorithm designers have assumed that layouts that satisfy certain aesthetics (fewer edge crossings, fewer bends, more symmetry) are easier to understand. Purchase tested this assumption in experiments with human readers (Purchase 1997, abstract). Among five aesthetics, reducing edge crossings was by far the most important, minimizing bends and maximizing symmetry had a lesser effect, and the effects of the minimum angle between edges and of an orthogonal grid were not statistically significant. Ware and colleagues pointed out that these aesthetics have mostly been taken as axiomatic without empirical tests. In a task of finding shortest paths in spring-layout graphs, they showed that after path length, the most important factors were continuity (keeping paths as straight as possible) and edge crossings (Ware et al. 2002, abstract).

The developers of force-directed layout tools also describe layout as a matter of design choices. Jacomy and colleagues stated that ForceAtlas2, the default layout algorithm of Gephi, is not a theoretical advance but an integration of existing techniques (Jacomy et al. 2014). They explained that it was designed for the Gephi user experience as a continuous algorithm, and they described which constraints led to which choices. Venturini, Jacomy, and Jensen noted that the practice of reading the visual qualities of force-directed drawings as proxies for topological features remains based on intuition, and they offered guidelines for reading such drawings (Venturini et al. 2021, abstract). They argued that the ambiguity that remains in a drawing can be an asset for exploratory analysis if properly interpreted (Venturini et al. 2021, abstract). At the same time, they stated that their attempts to separate the ambiguity inherited from the represented phenomenon from the distortion caused by fitting a multidimensional object into two dimensions were only partially successful.

Limits of the borrowed tools

Among the findings that design research borrowed, the universality of scale-free structure has been seriously questioned. Broido and Clauset applied a severe statistical test to nearly 1,000 social, biological, technological, transportation, and information networks, and showed that strongly scale-free structure is empirically rare and that, for most networks, log-normal distributions fit the data as well as or better than power laws (Broido and Clauset 2019). Keller argued that power-law distributions are easy to generate by a variety of mechanisms, so they do not signify a universal architecture independent of the system’s specifics, and that the architecture is determined by the actual constraints on the system (Keller 2005, abstract). Li and colleagues also argued that there is no consistent definition of scale-free graphs and that some of the most celebrated claims regarding the Internet and biology are false (Li et al. 2005). The scaling regimes that Braha and Bar-Yam found in product development networks and the scale-free properties that Crucitti and colleagues found in self-organized cities were reported before these severe tests appeared.

Network measures depend on the representational choice of what to treat as points and lines. Ratti showed that applying space syntax to regularly gridded urban textures makes its claim to model pedestrian choice hard to accept, and that distorting two ideal textures produces a topological discontinuity in which one configuration produces two conflicting outcomes (Ratti 2004, abstract). The abstract concludes that the topological representation on which space syntax is based discards metric information and has difficulty taking building height and land use into account. The method of Porta and colleagues, which measures centrality within a metric framework, can be read as a response to this representational choice (Porta et al. 2006). The correspondence between products and organizations is also, as the review of the mirroring hypothesis showed, prevalent but not universal (Colfer and Baldwin 2016, abstract).

The word “network” itself refers to different things in different fields. Latour wrote that the “network” of actor-network theory is not a technical network in the engineer’s sense (sewage, trains, telephones), and that a technical network is only one of the possible final and stabilized states of an actor-network (Latour 1996). In the same essay, Latour stated that actor-network theory has very little to do with the study of social networks. Therefore, when design research refers to actor-network theory, its “network” needs to be treated as a concept different from the points and lines of the graphs discussed in this note.

The relationship between the two fields

The literature reviewed above suggests that the relationship between the two fields is not symmetric. What design research borrowed from network science was mainly a set of tools for representing and measuring structure. The dependencies of products and organizations, the links between design moves, the configuration of street networks, the ties within teams, and the co-citations of a field all map things that people made, or records of design processes, onto points and lines for measurement. On the network science side, within the literature collected here, “design” refers to optimization under constraints, its agent is often natural selection or a growth process, and how a designer defines a problem lies outside its scope.

After starting from the same question in the 1960s, the two fields moved in different directions. Simon’s idea of decomposition, with strong linkages within parts and weak linkages between them, continues in a changed form in engineering design through DSM and modularity research, and it also connects with research on modularity and robustness in network science. Alexander himself repudiated the method of decomposition, and in Cross’s account, design studies moved through Rittel and Webber’s wicked problems and Schön’s reflective practice toward the study of forms of knowledge peculiar to designers. Within the literature collected here, the questions on this side of design studies (how to define a problem and when to stop) appear to have only begun to be addressed with network methods. The attempt by Kelly and Gero to measure design framing with semantic networks of design protocols can be positioned as an effort to enter this gap (Kelly and Gero 2025).

The most concrete point of contact is the drawing of networks. Layout aesthetics are tested against human reading, layout algorithms are designed for the user experience, and the ambiguity of a drawing is left to the reader’s interpretation. This is where the judgment of designers and readers returns within the tools of network science. The /network page of this wiki, which produces many different pictures from the same co-citation relations by changing only the arrangement, is a use of these tools that shows this judgment on the side of the works.

What research has not yet answered

Network measures include the structure of a DSM, the shape of a linkograph, and measures of semantic networks. In the literature collected here, research that tests on the same tasks how such measures relate to the quality of designers’ judgment or of design outcomes was limited to the method proposed by Kelly and Gero. No study was found that re-examined the scale-free properties reported in design research (product development networks, self-organized cities) with the test of Broido and Clauset. For co-citation analyses of the whole field of design research, no reliable candidate other than Chai and Xiao’s study of Design Studies from 1996 to 2010 was found in this collection.

Unverified items

No unverified statements remain in the main claims. Simon 1962 was checked against a copy of the text (a transcription of a scan), Alexander 1965 against Part I of the reprint edited by Thackara (1988), and Latour 1996 against the author’s English manuscript. Sources checked only through their abstracts are marked “abstract” in the citation parentheses, and only what the abstracts state is reported.

References

Author: Shuichiro Ogawa (Design Researcher / Consultant) About me →