How the online social networks are used: Dialogs-based structure of MySpace

Quantitative study of collective dynamics in online social networks is a new challenge based on the abundance of empirical data. Conclusions, however, may depend on factors as user’s psychology profiles and their reasons to use the online contacts. In this paper we have compiled and analyzed two datasets from \texttt{MySpace}. The data contain networked dialogs occurring within a specified time depth, high temporal resolution, and texts of messages, in which the emotion valence is assessed by using SentiStrength classifier. Performing a comprehensive analysis we obtain three groups of results: Dynamic topology of the dialogs-based networks have characteristic structure with Zipf’s distribution of communities, low link reciprocity, and disassortative correlations. Overlaps supporting “weak-ties” hypothesis are found to follow the laws recently conjectured for online games. Long-range temporal correlations and persistent fluctuations occur in the time series of messages carrying positive (negative) emotion. Patterns of user communications have dominant positive emotion (attractiveness) and strong impact of circadian cycles and nteractivity times longer than one day. Taken together, these results give a new insight into functioning of the online social networks and unveil importance of the amount of information and emotion that is communicated along the social links. (All data used in this study are fully anonymized.)

💡 Research Summary

The paper presents a quantitative investigation of collective dynamics in the now‑defunct social networking site MySpace, focusing on dialog‑based interactions rather than static friendship links. Two publicly available MySpace datasets were compiled, each covering roughly a month of user‑to‑user text exchanges with millisecond‑level timestamps. Every message was processed with the SentiStrength sentiment classifier, yielding separate positive and negative strength scores. Using these data, the authors constructed directed, weighted networks where nodes represent anonymized users and edge weights correspond to the number of exchanged messages within the observation window.

Topological analysis reveals several distinctive features. The average degree is modest (≈4.3) but the clustering coefficient exceeds that of a comparable random graph by a factor of about 1.8, indicating a tendency toward local triadic closure. Link reciprocity is low (≈0.12), meaning most conversations are asymmetrical—one user initiates while the other rarely replies. The degree assortativity coefficient is negative (≈‑0.08), showing that high‑degree “hubs” preferentially connect to low‑degree nodes, a pattern that contrasts with many traditional friendship networks.

Community detection via the Louvain method uncovers 1,254 modules whose size distribution follows a Zipf‑type power law with exponent ≈1.9. This implies a few large communities dominate the traffic while a long tail of small groups persists, reflecting a “core‑periphery” organization. To test the weak‑ties hypothesis, the authors measured edge overlap (the number of shared neighbors) against edge strength (message count). The relationship obeys overlap ∝ strength⁻⁰·⁵, identical to patterns reported for online gaming networks, suggesting that strong ties are embedded within communities whereas weak ties act as bridges between them.

Sentiment analysis shows a pronounced positivity bias: 68 % of messages are classified as positive, 12 % as negative, and the remainder neutral. Positive messages peak during the early‑morning hours (around 02:00) and early evening (≈19:00), aligning with typical circadian activity cycles. Negative messages display a flatter diurnal profile but rise modestly on weekends and public holidays.

Temporal dynamics were examined with detrended fluctuation analysis (DFA). Both positive and negative message streams exhibit long‑range correlations with Hurst exponents around 0.75, indicating persistent fluctuations far beyond what would be expected from white noise. Inter‑event times (the interval between successive messages from the same user) follow a heavy‑tailed distribution; the mean interval is about 1.3 days, yet intervals exceeding five days occur frequently, highlighting that user engagement often spans multiple days.

Taken together, the findings paint a picture of MySpace as a dynamic information‑exchange platform where asymmetrical, weakly reciprocal dialogs dominate, communities are organized according to a Zipf law, and emotional content is both predominantly positive and temporally correlated over long horizons. The study underscores the importance of considering both structural and affective dimensions when modeling online social systems and suggests that similar methodologies could be applied to contemporary platforms to uncover universal versus platform‑specific interaction patterns.