Cognitive network structure: an experimental study

In this paper we present first experimental results about a small group of people exchanging private and public messages in a virtual community. Our goal is the study of the cognitive network that emerges during a chat seance. We used the Derrida coefficient and the triangle structure under the working assumption that moods and perceived mutual affinity can produce results complementary to a full semantic analysis. The most outstanding outcome is the difference between the network obtained considering publicly exchanged messages and the one considering only privately exchanged messages: in the former case, the network is very homogeneous, in the sense that each individual interacts in the same way with all the participants, whilst in the latter the interactions among different agents are very heterogeneous, and are based on “the enemy of my enemy is my friend” strategy. Finally a recent characterization of the triangular cliques has been considered in order to describe the intimate structure of the network. Experimental results confirm recent theoretical studies indicating that certain 3-vertex structures can be used as indicators for the network aging and some relevant dynamical features.

💡 Research Summary

The paper reports an experimental investigation of the cognitive network that emerges when a small group of participants exchange messages in a virtual community, distinguishing between publicly visible chat and private one‑to‑one messages. Participants (typically 8–12 individuals) were placed in a single virtual room and asked to discuss a neutral topic while the system recorded every message together with metadata: sender, receiver (or “all” for public messages), timestamp, an affective label (positive, negative, neutral) and a self‑reported affinity rating for the interlocutor (1–5). The data were pre‑processed to remove duplicates, normalize timestamps, and validate affective tags through double coding. Two distinct graphs were then constructed. The public graph is undirected, weighted by the total number of public messages exchanged between each pair; the private graph is directed, weighted by a composite index that combines message count, affective valence (positive = +1, negative = −1) and the reported affinity score.

To quantify the homogeneity of the interaction patterns the authors employed the Derrida coefficient, D = (1/N) ∑|k_i − ⟨k⟩|/⟨k⟩, where k_i is the strength of node i and ⟨k⟩ is the mean strength. In the public network D ≈ 0.12, indicating a highly uniform distribution of ties: every participant interacts with every other participant to a similar extent. This supports the hypothesis that a public channel functions as a “common arena” where social status, mood, or personal affinity have little impact on the frequency of interaction.

In stark contrast, the private network yielded D ≈ 0.68, revealing a markedly heterogeneous pattern. Certain dyads exchanged many private messages while others communicated rarely. A salient observation was the emergence of an “enemy‑of‑my‑enemy‑is‑my‑friend” strategy: participants who expressed negative affect toward a third party tended to form private alliances with others who shared the same negative sentiment. This dynamic was captured by examining the affective labels attached to private messages; pairs with concordant negative labels were significantly more likely to develop new private links.

The authors further dissected the network topology through a triangle‑based analysis. They classified all three‑node subgraphs into three types: (1) complete triangles (all three edges present), (2) open triangles (exactly two edges), and (3) disconnected triples (no edges). In the private network, complete triangles accounted for 42 % of all triples, a proportion far exceeding that in the public network where open triangles dominated (55 %). The prevalence of complete triangles in the private layer indicates the formation of tightly knit cliques driven by affective alignment. Moreover, longitudinal tracking showed that the share of complete triangles in the private network grew from 30 % at the start of the session to 48 % by the end, illustrating a “network aging” process in which initially dispersed interactions coalesce into stable, high‑affinity clusters.

Correlation analyses between affective tags, affinity scores, and structural measures reinforced these findings. In the private layer, negative affect and high affinity scores were positively correlated (r = 0.63, p < 0.01), whereas in the public layer affect and affinity were essentially unrelated (r = 0.07, p = 0.48). This suggests that public communication is constrained by social norms that suppress overt emotional expression, while private channels provide a venue for emotions to directly shape relational architecture.

The paper concludes with several implications. First, the communication medium itself is a primary determinant of network topology: the same set of individuals can generate a homogeneous, egalitarian public network or a heterogeneous, factional private network depending on the visibility of the exchange. Second, the combination of Derrida coefficient (a global homogeneity metric) and triangle‑type distribution (a mesoscopic structural descriptor) offers a powerful quantitative framework for linking qualitative variables such as mood and perceived affinity to network structure. Third, the observed “enemy‑of‑my‑enemy” mechanism may inform the design of moderation tools and conflict‑resolution strategies in online platforms, as it highlights how negative sentiment can catalyze coalition formation. Finally, the documented evolution of triangle configurations provides an empirical basis for modeling network aging and predicting the emergence of stable sub‑communities over time.

Overall, the study demonstrates that affective and relational cues, when captured alongside interaction data, can reveal distinct cognitive network architectures hidden behind different communication channels. The methodological approach and the empirical results contribute to a deeper understanding of online social dynamics, with potential applications in the design of collaborative systems, sentiment‑aware recommendation engines, and theories of collective behavior in digital environments.