Bodong Chen

Crisscross Landscapes

Notes: Andrade, A. (2015). Using Situated-Action Networks to visualize complex learning



Citekey: @andrade2015using

Andrade, A. (2015). Using Situated-Action Networks to visualize complex learning. In O. Lindwall, P. Häkkinen, T. Koschmann, P. Tchounikine, & S. Ludvigsen (Eds.), Exploring the Material Conditions of Learning: The Computer Supported Collaborative Learning (CSCL) Conference 2015, Volume 1 (pp. 372–379). Gothenburg, Sweden: International Society of the Learning Sciences.


This paper introduces a quite interesting analytical technique, named Situated-Action Network, which is theoretically informed by sociocultural learning theories and methodologically grounded in SNA. Basically, it puts both subjects and objects into one network, and use principle component analysis to reduce data dimensions to 2 to reveal progression of learner. A nice intro to the technique and an interesting case study of using SAN for understanding learning progression in a piloting task.

(I am still about confused about the PCA part.. On which data is this analysis conducted? Using which algorithm?)


The purpose of this paper is to show how Situated-Action Networks (SAN) is a viable visualization tool for representing the learning that happens as a movement from peripheral to more engaged forms of participation in a social activity. (p. 1)

In shifting the unit of analysis from cognitive processes to social activity, this form of visualization accounts for the social dynamics, tracing the learner’s trajectory of participation from the periphery to a more engaged form of participation (Lave & Wenger, 1991). Building on Social Network Analysis (Wasserman, 1994), various descriptive network indices can provide useful information about the learner’s level of participation in the social activity. (p. 1)

We take action as the unit of analysis. According to activity theory (Engeström, 1987; Leont’ev, 1974; Roth, 2007; Vygotsky, 1978), an action stands as the mediational link between a subject and an object. In this dialectical sense, actions are processes that connect two qualitatively different expressions of the same unit: in actions, subject and object are two constitutive elements that cannot be considered independently from each other (Roth, 2007). The idea that both subjects and objects are opposite but mutually constitutive parts of the same unit has important consequences for what, and how, subjects and objects are included in a representation of the learning process. In our analysis, subjects and objects are both represented as nodes in a network. This idea is not new, however, and some antecedents can be found in Latour’s Science and Technology studies (Latour, 1987) and some educational approaches inspired by his ideas (see for instance Shaffer & Clinton, 2006). Also, the theory of distributed cognition takes the system of cognition of both persons and objects as the unit of analysis (Hutchins, 1995; Hutchins & Klausen, 1996). (p. 1)

Two main kinds of actions are envisioned in SAN analysis: a) instrumental and b) communicative actions. Instrumental actions refer to the use of artifacts – as a transformative material process (e.g., hammering a nail) or as the consumption of a source of information (e.g., reading a ruler or a scale). Communicative actions refer to linguistically achieved (i.e., verbal or non-verbal) intersubjective actions. In this case, language is the object. Vygotsky (1986) conceived of language as a tool that shared the same ontological status as that of a physical tool. For Cole (1996), this ontological status is clarified by using the term “artifact” to refer to both linguistic communication and material tools. (p. 1)

The conception of learning that we bring to bear here can be defined as a micro-genetic change in the pattern of actions. A pattern of actions represents the regularities of the learner’s actions as well as the actions from all the other participants in the situation. That is, the connections between subjects and objects through actions in a regular manner. A micro-genetic change can be considered as a small developmental change that occurs as a consequence of an appropriation of the social practices (Cole, 1996; Roth, 2007; Vygotsky, 1978). (p. 1)

Situated-action networks (p. 2)

The first step is to create a raw data matrix where nodes and actions are recorded throughout segments of time (e.g., 1 minute or 10 second intervals). Second, a number of segments is subset depending on the starting and ending point of a socially meaningful activity (e.g., a 45 min class or a 10 min class activity). Third, an adjacency matrix is produced from the raw data matrix. And, fourth, a graphic representation is created. (p. 2)

In comparing several of these representations over various activities, these networks reveal informative changes in the pattern of actions. The thickness of the lines can be used to represent the density of the actions between pairs of nodes. (p. 2)

In addition, the novice’s learning trajectory toward more engaged, expert-like participation can be displayed by plotting each network’s center of mass in a 2-dimensional plot using the first two principal components. A principal component analysis is a statistical method that uses the eigenvalues from the adjacency matrix to create a linear combination of the nodes so as to cluster the most important ones (i.e., the nodes that carry the most weight) in the first dimension and a second linear combination with the second most important nodes for the second dimension. (p. 3)

the progression of the center of mass from our previous example would show that the density of actions of child B progresses (throughout the various iterations of the activity) towards that of child A in the first attempt (see Figure 3). (p. 3)

A brief example: Learning how to land an airplane (p. 4)

Situated-action networks (p. 6)

Thus far, we have been asking questions about changes in individual pieces of information for the learner’s actions. Now, we ask how the overall functional model of actions has changed. (p. 6)

In the pretest network (see Figure 6.a), when the novice played the role of the pilot for the first time and the expert the copilot, the pilot’s degree (i.e., the amount of the pilot’s actions) was 36, and the network density (i.e., the average degree across all nodes) was 27. In the posttest network (b), which was after five landing activities in which the novice played the alternate role of the pilot or the copilot, it is apparent that his node played a more engaged role in the network. The pilot’s degree for network (b) increased to 84, and the network density increased to 36.2. In addition, network © represents the activity when the expert played the role of the pilot. It shows a high level of engagement on the part of the pilot, with a high degree of 112 and high average network density of 40.6. (p. 7)

Finally, the participation progression can be seen in Figure 6.d, where center of masses are plotted along the first two principal components. The first component represents the amount of gaze at the instruments (and represents 35% of the variance) whereas the second component represents the amount of questioning (and represents 16% of the variance). It is apparent that the apprentice’s dependency on the instruments decreased from the preto the post-test activity as revealed by a lower posttest center of mass on the first dimension (p. 7)