1: %\documentclass[aps,pre,preprint,floatfix,nofootinbib,showpacs]{revtex4}
2: \documentclass[aps,pre,twocolumn,floatfix,nofootinbib,showpacs]{revtex4}
3: \usepackage{epsf,amsmath,amssymb,verbatim}
4: \begin{document}
5: \newcommand{\be}{\begin{equation}}
6: \newcommand{\ee}{\end{equation}}
7: \title{Structure and evolution of online social relationships:
8: Heterogeneity in warm discussions}
9: \author{K.-I. Goh$^*$, Y.-H. Eom$^{\dag}$, H.
10: Jeong$^{\dag}$, B. Kahng$^{*,\ddag}$, and D. Kim$^*$\\}
11: \affiliation{{$^*$School of Physics and Center for Theoretical
12: Physics, Seoul National University, Seoul 151-747, Korea}\\
13: {$^{\dag}$Department of Physics, Korea Advanced Institute of
14: Science and Technology, Daejon 305-701, Korea \\}
15: {$^{\ddag}$Center for Nonlinear Studies, Los Alamos National
16: Laboratory, Los Alamos, New Mexico 87545}}
17: \begin{abstract}
18: With the advancement in the information age, people are using
19: electronic media more frequently for communications, and social
20: relationships are also increasingly resorting to online channels.
21: While extensive studies on traditional social networks have been
22: carried out, little has been done on online social network. Here
23: we analyze the structure and evolution of online social
24: relationships by examining the temporal records of a bulletin
25: board system (BBS) in a university. The BBS dataset comprises of
26: 1,908 boards, in which a total of 7,446 students participate. An
27: edge is assigned to each dialogue between two students, and it is
28: defined as the appearance of the name of a student in the from-
29: and to-field in each message. This yields a weighted network
30: between the communicating students with an unambiguous group
31: association of individuals. In contrast to a typical community
32: network, where intracommunities (intercommunities) are strongly
33: (weakly) tied, the BBS network contains hub members who
34: participate in many boards simultaneously but are strongly tied,
35: that is, they have a large degree and betweenness centrality and
36: provide communication channels between communities. On the other
37: hand, intracommunities are rather homogeneously and weakly
38: connected. Such a structure, which has never been empirically
39: characterized in the past, might provide a new perspective on
40: social opinion formation in this digital era.
41: \end{abstract}
42: \pacs{} \maketitle \section{Introduction} With the advancement in
43: the information age, people are using electronic media for
44: communication more frequently, and social relationships between
45: people are also increasingly resorting to online communications.
46: For example, the advent of online bulletin board systems (BBS)
47: made it possible to develop a new type of online social
48: relationship and social consensus. Very similar to the Usenet
49: service, which was fairly popular during the earlier days of the
50: Internet, BBS is based on the communication between people sharing
51: common interests; the topic of interest is usually identified by
52: the board itself. People with common interests post messages on a
53: certain board and a response is conveyed by posting another
54: message, thereby forming a thread. Thus, a thread in the BBS
55: roughly represents a dialogue between people, and such a dialogue
56: constitutes the basic relationship among the people participating
57: in it. In the BBS, dialogues or discussions usually proceed with
58: little restriction on message writing and discrimination based on
59: personal information, thereby forming the so-called ``warm
60: discussions'' as described in psycho-sociology \cite{doise}.
61: Therefore, the pattern of such online social relationships may be
62: different from that of traditional social relationships based on
63: face-to-face contact or online communication involving exchange of
64: personal information, such as e-mail
65: transactions~\cite{bornholdt,huberman,dodds,eckmann,adamic} and
66: instant messaging \cite{messenger}. Thus, it would be interesting
67: to study the structure of online social relationship networks
68: constructed by people in warm discussions; this would be useful in
69: resolving diverse sociological and political issues and
70: understanding the manner in which social opinion is formed in the
71: digital era~\cite{axelrod,klemm,dodds_pnas,deffuant,weisbuch}.
72: Extensive studies on traditional social networks have been carried
73: out~\cite{milgram,granovetter,wasserman}; however, few studies
74: exist on online social networks. Here, we investigate the
75: structure of online social networks by studying BBS networks,
76: which are familiar to university students.
77:
78: From the graph theoretical perspective, the BBS network offers
79: distinct features such as weighted and modular network structure.
80: Since the number of times a given pair of people exchange
81: dialogues can be counted explicitly, a weighted network is
82: naturally obtained~\cite{vespig_pnas}. Moreover, since people are
83: sharing a board corresponding to their common interests, BBS
84: provides an unambiguous way of defining modules or communities
85: \cite{gn}. This is unlike other examples of accessible protocols,
86: including the sibling/peer relationship in the online community
87: \cite{holme} and trackback in the blog system \cite{blog}. In
88: fact, the BBS network constructed by us differs in crucial aspects
89: from other affiliation networks such as the collaboration network
90: \cite{newman_coll} and student course registration network
91: \cite{course}. In these examples, the relationship between people
92: is not explicitly defined but is indicated indirectly by their
93: affiliation. Such an indirect definition generates several
94: cliques-completely connected subgroups-which may result in an
95: artifact particularly in the case of large-sized affiliations.
96: Thus, to obtain a network of people with explicit pairwise
97: interaction strength together with a distinct community definition
98: is crucial for an appropriate description of the social system.
99: The BBS network provides such ingredients.
100:
101: \begin{figure}\centerline{\epsfxsize=.7 \linewidth \epsfbox{fig1.eps}}
102: \caption{\sf Schematic network snapshots of the BBS network (a)
103: and traditional social network (b).} \end{figure}
104:
105: \section{Conclusions and discussion}
106: The BBS network has interesting structural features and
107: implications. It contains hub members who participate in dialogues
108: across a large number of boards, thereby connecting one group of
109: people at one board to another group at a different board.
110: Further, their degrees, which are the numbers of people they have
111: exchanged dialogues with, are large, thereby influencing other
112: people throughout different communities. As a result, the hub
113: members act as weak ties in connecting different communities;
114: however, their links are strong during on actual activity. On the
115: other hand, intraboard connections are rather homogeneous in
116: degree. Such a network feature is in contrast to traditional
117: social networks maintained by the ties bridging disparate
118: communities, which tend to be weak~\cite{granovetter}. The
119: difference is schematically depicted in Fig.~1. In the BBS
120: network, the strength $s$, i.e., the total number of dialogues
121: each individual participates in has a nonlinear relationship with
122: the degree $k$ as $s\sim k^{1.4}$. This implies that the hub
123: members tend to post messages at considerably more frequently than
124: the other people with small degrees. The neutrality in the
125: assortative mixing is another feature of the BBS network compared
126: with the assortativity in traditional social networks. Such a
127: behavior may originate due to the absence of personal information
128: on the partner during online social communication. Thus, hub
129: members are democratic in their connections to the remaining
130: people, and they are indeed ``ubiquitous persons." Since the hub
131: members play a dominant role in providing communication channels
132: across different boards, it might be more efficient to use a
133: BBS-like online media for persuading people and drawing social
134: consensus than traditional social networks based on
135: person-to-person relationships. We attempt to understand the BBS
136: network from the perspective of a simple network model. In the
137: model, we take into account the empirical fact that the BBS
138: network contains groups of which size are inhomogeneous. In
139: addition, the link density of each group is not uniform, however
140: decreases with increasing group size, which has been usually
141: neglected in constructing model.
142:
143: It would be interesting to implement the present work in the
144: context of a previous study involving a psycho-sociological
145: experiment on group discussions and the resulting
146: consensus~\cite{doise}, in which, group discussions are
147: distinguished into two types, ``warm'' and ``cold''. In the former
148: type, people express their thoughts freely without any
149: restriction, while in the latter, group discussions are restricted
150: by some constraint either explicitly or implicitly, for example,
151: the hierarchy in group members. The experimental study concludes
152: that the consensus measured after group discussions can be
153: different from that before the discussions depending on the type.
154: In the former, the consensus after discussions shifts to an
155: extreme opinions, while in the latter, it leads to a trade-off
156: average group consensus. From the perspective of the experiment,
157: we might state that the dialogues in the BBS are warm because no
158: restriction is imposed on posting messages and little information
159: on the personal background of the partner is provided. Thus, the
160: dialogues in the BBS may lead to radicalized consensus, violent
161: group behaviors, or imaginative and creative solutions to a given
162: issue. Since students still in the process of developing a value
163: system are vulnerable to negative influences, and have more
164: opportunities to be influenced by their peers through online
165: networks in this digital era than in the past, the proposed
166: network pattern we report here will be useful in guiding them in
167: the right direction. Moreover, the BBS network data will be
168: helpful in understanding the manner in which diverse opinions are
169: synchronized from the psycho-sociological perspective.
170:
171: \section{BBS network}
172: We mainly examined the BBS system at the Korea Advanced Institute
173: of Science and Technology; it is named as {\tt loco.kaist.ac.kr}.
174: The characteristics of the network structure obtained from this
175: BBS system also appear in another system-{\tt bar.kaist.ac.kr}.
176: The data comprises records of all the threads posted from March
177: 9, 2000 to November 2, 2004, thus corresponding to a duration of
178: around three and a half years. As of November 2004, the system
179: comprised 1,908 boards with a total of 7,446 participating
180: students. In order to ensure privacy, we are only allowed to
181: access the information on ``from,'' ``to,'' the date of posting,
182: and the name of the board it was posted on, for each message.
183: Based on this information, we constructed the network between
184: students such that for each message, an edge was assigned between
185: two students appearing as ``from'' and ``to.'' Alternatively, an
186: arc (a directed edge) can be assigned for each message; however,
187: we found that the communications are largely reciprocal:
188: Approximately a half of the postings are accompanied by another
189: one with its from and to fields reversed, for example, a ``Re:''
190: message. Subsequently, we shall consider the network as undirected
191: for simplicity.
192:
193: Our network construction naturally yields a weighted network in
194: which the weight $w_{ij}$ of the edge between two students $i$ and
195: $j$ is determined by the number of messages they exchanged during
196: the period. The detailed statistics of the BBS are listed in Table
197: I.
198:
199: \begin{table}[p]
200: \caption{\sf Statistics of the BBS network as of November 2004.
201: The numbers in parentheses are the statistics for non-self
202: dialogues.}
203: \begin{tabular}{ll}
204: \hline \hline
205: Number of students $N$ & 7446 (7421) \\
206: Number of links $L$ & 103498 (103473) \\
207: Number of dialogues $W$ & 1299397 (1267292) \\
208: Number of boards $G$ & 1908 (1872) \\
209: Size of the largest cluster $N_1$ & 7350 \\
210: Average size of the boards $\bar{S}$ & 32.0 (32.6) \\
211: Average board memberships of a student $\bar{B}$ & 8.2 \\
212: Average path length $D$ & 3.3 \\
213: Mean degree $\langle k\rangle$ & 27.8 (27.9) \\
214: \hline \hline
215: \end{tabular}
216: \end{table}
217:
218: \section{Structure of the BBS network}
219:
220: \subsection{Student network}
221: The global snapshot of the student network in Fig.~1 reveals the
222: inhomogeneity among the students. The degree $k_i$ of a student
223: $i$, which is the number of students he/she has exchanged
224: dialogues with, is distributed according to a power law with an
225: exponent of around $-1$ followed by an exponential cutoff, as
226: shown in Fig.~2(a). This feature is similar to that of the
227: scientific collaboration network \cite{newman_coll}. The strength
228: $s_i$ of a student $i$ is the sum of the weight of each edge
229: attached to $i$. Therefore, $s_i=\sum_{j}^N a_{ij}w_{ij}$, where
230: $a_{ij}$ is the component of the adjacent matrix; its value is 1
231: if an edge is connected between vertices $i$ and $j$ and 0
232: otherwise. $w_{ij}$ is the weight of the edge between $i$ and $j$.
233: The strength and degree of a student exhibit a scaling behavior
234: $s(k)\sim k^{\beta}$ with $\beta\approx1.4$; however, the
235: fluctuation is quite strong, particularly for a small $k$
236: [Fig.~2(b)]. The strength distribution exhibits a behavior that is
237: similar to that of the degree distribution; however, the value of
238: the cutoff is larger[Fig.~2(a)]. The nonlinear relationship
239: between $s$ and $k$ implies that the hub members tend to post
240: messages at considerably more frequently than the other people, as
241: is evident in Table II.
242:
243: \begin{table}[p]
244: \caption{\sf The fraction of the dialogues contributed by hub
245: members with a degree larger than 80 in the first ten longest
246: threads. The degree value of 80 is chosen approximately in
247: Fig.~2(a); beyond this degree, the power law for the degree
248: distribution fails.}
249: \begin{tabular}{lccccc}
250: \hline \hline
251: Rank~~~& Thread length~~~ & Number of dialogues ~~~ & Fraction \\
252: & & contributed by hub members & (\%) \\
253: \hline
254: 1 & 229 & 181 & 79 \\
255: 2 & 121 & 70 & 58 \\
256: 3 & 92 & 92 & 100 \\
257: 4 & 74 & 45 & 61 \\
258: 5 & 67 & 16 & 24 \\
259: 6 & 66 & 45 & 68 \\
260: 7 & 65 & 27 & 41 \\
261: 8 & 64 & 34 & 53 \\
262: 9 & 54 & 54 & 100 \\
263: 10 & 50 & 50 & 100 \\
264: \hline \hline
265: \end{tabular}
266: \end{table}
267:
268: \begin{figure}\centerline{\epsfxsize=.7\linewidth \epsfbox{fig2.eps}}
269: \caption{{\sf Structure of the BBS network.} (a) The degree
270: distribution $P_d(k)$ ($\circ$) and the strength distribution
271: $P_s(s)$ ($\diamond$) of the entire network. The straight line is
272: a guideline with a slope of $-1$. (b) The degree-strength scaling
273: relation $s(k)$. The straight line is a guideline with a slope of
274: $1.4$. (c) The clustering function $C(k)$ ($\circ$) and its
275: weighted version ($\diamond$). The straightlines are guidelines
276: with slope of $-0.5$ (lower) and $-0.3$ (upper), respectively. (d)
277: The average nearest-neighbor degree function $k_{\rm nn}(k)$ and
278: its weighted version ($\diamond$). (e) The correlation between the
279: degree and the membership number $B$. The dotted line is a
280: guideline with a slope of $1$. (f) The membership number
281: distribution of the vertices $P_B(B)$, where $B$ is the number of
282: boards that a student participates in. The straight line is a
283: guideline with a slope of $-1$.}
284: \end{figure}
285:
286: Other standard measures of network topology are also obtained. The
287: local clustering coefficient $c_i$ is the local density of
288: transitive relationships, defined as the number of triangles
289: formed by its neighbors, cornered by itself, $i$, divided by the
290: maximum possible number of these, $k_i(k_i-1)/2$. The average of
291: $c_i$ over vertices with a given degree $k$ is referred to as the
292: clustering function $C(k)$. For the student network, $C(k)$ decays
293: as $\sim k^{-0.5}$ for large $k$, and its weighted version defined
294: in Ref.~\cite{vespig_pnas}\footnote{ In Ref.~\cite{vespig_pnas},
295: the local weighted clustering coefficient was defined as
296: $c_i^{(w)}=\sum_{j,h}(w_{ij}+w_{ih})a_{ij}a_{ih}a_{jh}/[2s_i(k_i-1)]$.
297: $C^{(w)}(k)$ is the average of $c_i^{(w)}$ over vertices with
298: degree $k$. The weighted average nearest-neighbors degree of
299: vertex $i$ was defined as $k_{{\rm nn},i}^{(w)}=\sum_{j=1}^N
300: a_{ij}w_{ij}k_j /s_i$. $k_{{\rm nn}}^{(w)}(k)$ is the average of
301: $k_{{\rm nn},i}^{(w)}$ over the vertices with degree $k$.} behaves
302: as $C^{(w)}(k)\sim k^{-0.3}$, as shown in Fig. 2(c). The
303: clustering coefficient $C$, which is the average of $c_i$ over all
304: vertices with $k>1$, is $\approx0.48$. This is one order of
305: magnitude greater than $C_{\textrm{random}}\approx 0.04$ of its
306: typical randomized counterpart with an identical degree sequence
307: \cite{maslov}. The average nearest-neighbor degree function
308: $k_{\rm nn}(k)$, which is defined by the average degree of the
309: neighbors of vertices of degree $k$, is almost flat for the
310: student network; nevertheless, its weighted version defined in
311: \cite{vespig_pnas} shows a slightly upward curvature for large $k$
312: (Fig. 2(d)). The assortativity coefficient \cite{assort} for the
313: binary network and the Spearman rank correlation of the degrees
314: are measured to be close to zero, as $r \approx 0.011$ and $r_{\rm
315: Spearman} \approx 0.024$, respectively. This almost neutral
316: mixing, which is in contrast to the common belief that social
317: networks are assortative, has also been observed in another online
318: social network \cite{holme}.
319:
320: The number of boards that a student participates in is likely to
321: be larger for students with a larger degree, as shown in
322: Fig.~2(e). Its distribution follows a skewed functional form in
323: Fig.~2(f). These results imply an important fact that a group of
324: people with a large degree tend to participate in diverse
325: dialogues on different boards and will play a dominant role in
326: drawing social consensus on diverse issues. Moreover, they work as
327: mediators between different groups in an online social community.
328:
329: The betweenness centrality (BC) or load \cite{bc,bc_newman,load},
330: which is defined as the effective number of paths or packets
331: passing through a given vertex when every pair of vertices gives
332: and receives information, is also measured. The BC distribution
333: follows a power law with an exponent $\approx 2.2$, as shown in
334: Fig.~3(a) and the BC of a given vertex $\ell$ is strongly
335: correlated to its degree $k$ as $\ell \sim k^{1.6}$ as shown in
336: Fig.~3(b). This implies that the hub members have a large BC and
337: have a strong influence on the remaining people.
338:
339: \begin{figure}\centerline{\epsfxsize=.7 \linewidth \epsfbox{fig3.eps}}
340: \caption{(a) The betweenness centrality (BC) distribution of the
341: BBS network. The dotted line is a guideline with a slope of
342: $-2.2$. (b) The relation between BC ($\ell$) and degree ($k$) of
343: the BBS network. The dotted line is a guideline with a slope of
344: $1.6$.} \end{figure}
345:
346: In other words, the student network is extremely heterogeneous,
347: highly clustered, and yet, almost neutrally mixed, thereby
348: exhibiting a strong nonlinear relationship between the strength
349: and degree.
350:
351: \subsection{Board network}
352: The procedure for constructing the board network is similar to the
353: usual projection method of the bipartite affiliation network. We
354: create a link between two boards if they share at least one common
355: member. In other words, each student participating in more than
356: one board contributes a complete subgraph---a clique--- to the
357: board network. Thus, the board network is the superposition of
358: cliques, each of which originates from the crossboard activities
359: of a student. Such crossboard activities will provide channels for
360: information transmission across the boards. In order to assign
361: meaningful weights to these channels, all the links in each clique
362: are assigned a weight that is equal to the inverse of the number
363: of vertices in that clique. In other words, the communication
364: channels created by the students posting on fewer boards are
365: stronger. Therefore, the weight of an edge between two boards
366: increases with the number of co-members; however, the
367: contributions of ``ubiquitous persons'' would only be moderate.
368: The strength of a board is the sum of the weights of its edges.
369: Such a strength distribution along with the degree distribution,
370: which does not account for the weight, is shown in Fig.~4(a). The
371: relation between the strength and degree is shown in Fig.~4(b).
372:
373: \begin{figure}\centerline{\epsfxsize=.7 \linewidth \epsfbox{fig4.eps}}
374: \caption{{\sf Structure of the board network.} (a) The degree
375: distribution $P_d(k)$ ($\circ$) and strength distribution $P_s(s)$
376: ($\diamond$) of the board network. (b) The degree-strength
377: relation in the board network. The straight line is a guideline
378: with a slope of $1$. (c) The clustering function $C(k)$ ($\circ$)
379: and its weighted version ($\diamond$). (d) The average
380: nearest-neighbor degree function $k_{nn}(k)$ ($\circ$) and its
381: weighted version ($\diamond$).} \end{figure}
382:
383: The board network is quite highly clustered with a clustering
384: coefficient of $\approx 0.61$, and the clustering function
385: decreases with $k$ [Fig.~4(c)]. However, it is noteworthy that
386: such a high clustering may result from the generation of cliques
387: by the projection procedure. Moreover, even the randomized board
388: network has a clustering coefficient as high as $\approx 0.48$.
389: The average nearest-neighbor degree initially increases with $k$
390: but decreases for larger $k$. However, its weighted version
391: increases monotonically with $k$, as shown in Fig.~4(d).
392:
393: \section{Student network within a board}
394: Upon examining the networks within a board, we were presented with
395: a different scenario. As shown in Fig.~5(a), the degree
396: distributions of the student networks within the boards are rather
397: homogeneous. They exhibit a peak followed by an exponential tail,
398: which overall fits well into the Gamma distribution. Here, the
399: degree $k$ must be specified in further detail. Consider a case
400: where two students A and B on a given board who do not communicate
401: directly with each other. However, this communication between A
402: and B can occur on a different board. In this case, the two
403: students are regarded to be connected for the definition of degree
404: in Fig.~5(a). When such a pair is regarded to be disconnected, the
405: degree $k_0$ is redefined and its distribution exhibits fat tails,
406: as shown in Fig.~5(b); this was also observed in another BBS
407: system.
408:
409: \begin{figure}\centerline{\epsfxsize=.7 \linewidth \epsfbox{fig5.eps}}
410: \caption{{\sf Properties of the board sub-network.} (a) The degree
411: distributions of subnetworks within the five largest boards.
412: Symbols used are ($\circ$), ($\triangle$), ($\diamond$), ($\Box$),
413: and ($\triangledown$) in the decreasing order of board size. The
414: fitted curves with the Gamma distribution
415: $k^{a-1}e^{-k/b}/[\Gamma(a)b^a]$ are shown. (b) The degree
416: distributions of subnetworks within the five largest boards with
417: degree redefined as discussed in the text. (c) The size
418: distribution of the boards $P_M(M)$. The straight line is a
419: guideline with a slope of $-0.7$. (d) The link density
420: $\Lambda(M)$ within a board as a function of its size $M$. The
421: straight line is a guideline with a slope of $-0.65$.}
422: \end{figure}
423:
424: The size of the board, which denotes the number of students
425: posting messages on it, has a broad distribution [Fig.~5(c)]- a
426: power law followed by a rapidly decaying tail. The edge density
427: $\Lambda$ inside a given board scales with its size $M$ as
428: $\Lambda(M)\sim M^{-0.65}$, as shown in Fig.~5(d). Such a behavior
429: cannot be observed in the random sampling of populations of
430: different sizes, thereby indicating that the communications
431: between students are indeed strongly constrained within each board
432: rather than across them. Further, the power-law scaling behavior
433: suggests that the BBS network is organized in a self-similar
434: manner. From this result, it is evident that the usual projection
435: method involving the creation of cliques by bipartite affiliation
436: graphs cannot provide an appropriate description of the BBS
437: system. Moreover, such a size-dependent scaling of edge density
438: within groups has not been realized thus far in a simple model of
439: a clustered network~\cite{newman_cluster}.
440:
441: \section{Evolution of the BBS network}
442:
443: The daily record of the BBS network also allows us to examine the
444: temporal evolution of the network. The number of vertices
445: (students) $N$ grows exponentially after the transient period;
446: however, the continuously moderated growth rate appears to attain
447: a steady state [Fig.~6(a)]. Similar behavior is observed in the
448: case of the number of links $L$ and the number of dialogues $W$.
449: The number of boards $G$ grows at a rather steady rate over the
450: period.
451:
452: \begin{figure}\centerline{\epsfxsize=.7 \linewidth \epsfbox{fig6.eps}}
453: \caption{{\sf Evolution of the BBS network.} (a) The temporal
454: evolution of the number of students $N$ (solid), number of links
455: $L$ (dashed), total number of dialogues $W$ (dotted), and number
456: of boards $G$ (dot-dashed). (b) The same plot as (a) in the double
457: logarithmic scale. (c) The evolution of the degree distribution
458: $P_d(k)$ of the student network. The degree distribution for each
459: year is shown. The symbols ($\circ$), ($\diamond$), ($\triangle$),
460: and ($\triangledown$) correspond to each year from 2001 to 2004,
461: respectively, and ($\Box$) represents the final configuration. (d)
462: The clustering function $C(k)$ for each year. The same symbols as
463: those in (c) are used.} \end{figure}
464:
465: Despite its continuous evolution, the structural properties of the
466: network seem to be in a stationary state. In other words, the
467: overall network characteristics such as the degree distribution
468: and clustering function achieve their forms in the initial period
469: (after $\sim$1 year), and do not change considerably with time, as
470: shown in Figs.~6(c) and (d). The crossover time scale of
471: approximately 1 year can also be observed in terms of the
472: evolution of the number of vertices $N$: Their growth patterns
473: change qualitatively after $\sim$10 months, as seen in Figs.~6(a)
474: and (b).
475:
476: \section{Simple model}
477:
478: Having identified the main statistical characteristics of the BBS
479: network, we attempt to understand them from the perspective of a
480: simple network model. First, we consider a simple extension of the
481: model of a clustered network introduced by
482: Newman~\cite{newman_cluster}. The original model of Newman is
483: specified with two fundamental probability distributions, $r_m$
484: and $s_M$. $r_m$ represents the probability that an individual
485: belongs to $m$ groups [$P_B(B)$ in our notation; (see Fig.~5(d))]
486: and $s_M$, the probability that the group size is $M$ [$P_M(M)$ in
487: our notation]. By assuming that the link density within the groups
488: is given by a constant parameter $p$, it is possible to obtain
489: several of formulae for the network structure using the generating
490: function method. For example, the degree distribution of the
491: network can be written as follows: \be P_d(k) =
492: \left.\frac{1}{k!}\frac{d^k}{dz^k}f_0[g_1(pz+q)]\right|_{z=0}, \ee
493: where $f_0(z)$ and $g_1(z)$ are appropriate generating functions
494: defined as $f_0(z)=\sum_{m=0}^{\infty}r_mz^m$ and $g_1(z)=\langle
495: M\rangle^{-1}\sum_{M=0}^{\infty}Ms_Mz^{M-1}$, and $q=1-p$.
496:
497: However an obvious shortcoming of the model is that in real data,
498: the link densities are not uniform across the boards and they
499: strongly depend on the board size, as shown in Fig.~5(d). In fact,
500: by simply applying this model with the average link density
501: $p\approx0.3$ along with $r_m$ and $s_M$, directly measured from
502: the data, the degree distribution of the BBS network cannot be
503: reproduced. Therefore, we modify the model by allowing $p$ to vary
504: across the group, based on the empirical formula $\Lambda(M)\sim
505: M^{-0.65}$. Such a modification complicates the mathematical
506: formulae and they must be solved numerically. The resulting degree
507: distribution of the modified model along with that of the real
508: data is shown in Fig.~7. Although it is imperfect, the agreement
509: improved significantly. Thus, it is crucial to incorporate the
510: nonuniform link density into the realistic modeling of the BBS
511: network.
512:
513: The manner in which the group size distribution, group membership
514: distribution, and group density scaling, which are the input
515: parameters of the model, achieve their present forms, as shown in
516: Figs.~5(c) and (d), is a topic for future study. \\
517:
518: \begin{figure}\centerline{\epsfxsize=0.7\linewidth \epsfbox{fig7.eps}}
519: \caption{Comparison of the degree distributions of a simple model
520: of the BBS network of Newman (dotted) and its modification
521: (solid), with that of the real network (circle).} \end{figure}
522:
523: This work was supported by KRF Grant No. R14-2002-059-010000-0 of
524: the ABRL program funded by the Korean government MOEHRD and
525: the CNS research fellowship from SNU (BK).\\
526:
527: \begin{thebibliography}{99}
528: \bibitem{doise} Moscovici, S. \& Doise, W. (1992) {\em Dissensions et consensus}, Presses Universitaires de France, Paris, (1994) English edition as {\em Conflict and consensus}, Sage, London.
529: \bibitem{bornholdt} Ebel, H., Mielsch, L.-I., \& Bornholdt, S. (2002) Phys. Rev. E {\bf 66}, 035103.
530: \bibitem{huberman} Tyler, J. R., Wilkenson, D. M., \& Huberman, B. A., (2003) in {\em Communities and Technologies}, edited by M. Huysman, E. Wenger, and V. Wulf (Kluwer, Deventer).
531: \bibitem{dodds} Dodds, P. S., Muhamad, R., \& Watts, D. J. , (2003) Science {\bf 301}, 827--829.
532: \bibitem{eckmann} Eckmann, J.-P., Moses, E. \& Sergi, D. (2004) Proc. Natl. Acad. Sci. U.S.A. {\bf 101}, 14333-14337.
533: \bibitem{adamic} Adamic, L. \& Adar, E. (2005) Social Networks {\bf 27}, 187--203.
534: \bibitem{messenger} Smith, R. D. (2002) e-Print Archive, http://xxx.lanl.gov/abs/cond-mat/0206378.
535: \bibitem{axelrod} Axelrod, R. (1997) J. Conflict Res. {\bf 41}, 203--226.
536: \bibitem{klemm} Klemm, K., Egu\'{\i}luz, V. M., Toral, R., \& San
537: Miguel, M. (2003) Phys. Rev. E {\bf 67} 026120.
538: \bibitem{dodds_pnas} Dodds, P. S., Watts, D. J., \& Sabel, C. F.
539: (2003) Proc. Natl. Acad. Sci. U.S.A. {\bf 100}, 12516--12521.
540: \bibitem{deffuant} Amblard, F. \& Deffuant, G. (2004) Physica A {\bf 343}, 725--738.
541: \bibitem{weisbuch} Deffuant, G., Amblard, F. \& Weisbuch, G. (in prees) Journal of Artificial Societies and Social Simulation.
542: \bibitem{milgram} Milgram, S. (1967) Psychology Today {\bf 2}, 60--67.
543: \bibitem{granovetter} Granovetter, M. (1973) Am. J. Sociol. {\bf 78}, 1360--1380.
544: \bibitem{wasserman} Wasserman, S. \& Faust, K. (1994) {\em Social Network Analysis}
545: (Cambridge University Press, Cambridge).
546: \bibitem{vespig_pnas} Barrat, A., Barthelemy, M., Pastor-Satorras, R. \& Vespignani, A. (2004) Proc. Natl. Acad. Sci. U.S.A. {\bf 101}, 3747--3752.
547: \bibitem{gn} Girvan, M. \& Newman, M. E. J. (2002) Proc. Natl. Acad. Sci. U.S.A. {\bf 99}, 7821--7826.
548: \bibitem{holme} Holme, P., Edling, C. R. \& Liljeros, F. (2004) Social Networks {\bf 26}, 155--174.
549: \bibitem{blog} Adar, E., Zhang, L., Adamic, L. A., \& Lukose, R. M. (2004) in {\em Workshop on the Weblogging Ecosystem, 13th International World Wide Web Conference}.
550: \bibitem{newman_coll} Newman, M. E. J. (2001) Proc. Natl. Acad. Sci. U.S.A. {\bf 98}, 404--409.
551: \bibitem{course} Holme, P., Park, S. M., Kim, B. J. \& Edling, C. R. (2004) e-Print Archive, http://xxx.lanl.gov/abs/cond-mat/0411634.
552: \bibitem{maslov} Maslov, S. \& Sneppen, K. (2002) Science {\bf 296}, 910--913.
553: \bibitem{assort} Newman, M. E. J. (2002) Phys. Rev. Lett. {\bf 89}, 208701.
554: \bibitem{bc} Freeman, L.C. (1977) Socimetry {\bf 40}, 35--41.
555: \bibitem{bc_newman} Newman, M. E. J. (2001) Phys. Rev. E {\bf 64,} 016132.
556: \bibitem{load} Goh, K.-I., Kahng B. \& Kim, D. (2001) Phys. Rev. Lett. {\bf 87,} 278701.
557: \bibitem{bbs_old} Zhougbao, K. and Changshui, Z. (2003) Phys. Rev. E {\bf 67,} 036117.
558: \bibitem{newman_cluster} Newman, M. E. J. (2003) Phys. Rev. E {\bf 68}, 026121 (2003).
559: \end{thebibliography}
560:
561:
562: \end{document}
563: