Jyi-Shane Liu, Ke-Chih Ning, Wan-Chun Chuang Department of Computer Science
National Chengchi University 64 Sec. 2 Zhih-Nan Rd., Wen-Shan District
Taipei, TAIWAN, R. O. C.
jsliu@cs.nccu.edu.tw
Abstract—Among many real world applications of social network analysis, political interaction and executive succession show some unique characteristics of dynamic community evolution and raise interesting research challenges. Interactions of political power among community members are mostly subtle and behind the scene. Visible relations are only nominal and are not readily apparent to key findings. Under such difficult circumstances of information deficiency, the research problem is to uncover the inner relations among some of the network entities and to discover the hidden network structure based on these inner relations. In this research, our objective is to identify the inner circles of government political power and bureaucracy underneath formal work relations and observe how the political elite groups form and change over time. A government official job change network in a time span of over twenty years is built to model synchronous post assignment and job promotion within a time window as entity relations. In each snapshot of network evolution, communities that exhibit strong association of synchronous job change are identified by the edge betweenness decomposition algorithm. Then, an event-based framework is used to characterize community behavior patterns in consecutive changes of network structures. The approach is effectually demonstrated on two scenarios: (1) identifying and tracking the inner circle of a leading political figure; (2) finding succession pool members in government agencies. We further propose two evolutionary community variation indexes to assess political executive succession. Experimental results with actual government personnel data provide evidence that government agency succession can be reasonably measured. This work also has the practical value of providing objective scrutiny on political power transition for the benefit of public interest.
Keywords- evolutionary community detection; political elite cliques; executive succession; opaque network
I. INTRODUCTION
Over the years, social network analysis has made great stride in discovering and understanding entity interactions in the context of static and homogeneous networks. Recent research interest, however, has gradually moved toward the analysis of dynamic and heterogeneous networks that are closer to real-world setting and bring functional richness. One of the most intriguing problems in dynamic social network analysis is the tracking of evolving communities, where each entity share more similarity or proximity with other entities within the community than with the entities outside, yet the community membership changes over time [1]. Such tasks involve capturing network data in a temporal sequence and determining appropriate clustering of entities in each time step such that
certain aspects of network structural changes are optimized or desired [2]. This allows the identification of evolving communities, observation of changing entity role, and discovery of emergent trends, thus developing great potential for real-world utilities.
While research progress in dynamic social network analysis has inspired new ideas in various application domains including organizational communications [3], and scientific collaborations [4], some of the less addressed problem areas involve a variety of networks that are not easily accessed from outside. These social networks are generally characterized by severe data limitations that network information from public sources are incomplete and may suffer from intentional or unintentional distortion. Members of these social networks tend to keep their membership private and conceal their interactions with each other. Among different motivations for imposing secrecy, network activities that are illegal or cause harm to the society call for research attention. Dark network or covert network was termed to describe such illegal or criminal networks as drug/arms trafficking and terrorist networks [5][6].
Later research attempted to derive a better understanding of the dark/covert networks with social network analysis techniques [7][8][9].
In between the dichotomy of overt (or bright) and covert (or dark) networks, we suggest the existence of a type of opaque (social) networks that share some of the characteristics of each end. Concerning the effects or purposes of network activities, opaque (social) networks are as admissible or legitimate as bright networks. In terms of data availability, opaque (social) networks sit somewhere in the middle of scarcity and abundance. Opaque network data are available to some extent but are typically inadequate to derive crucial information for the purposes of study. In most cases, network entities are discernible and relations among entities are visible. However, these visible relations are only nominal and are not readily apparent to the key finding of the problem domains. Similar to dark network under such difficult circumstances of information deficiency, research challenge in opaque (social) networks is to uncover the inner relations among some of the network entities and to discover the hidden network structure based on these inner relations.
In this paper, we characterize government officials as members of an opaque (social) network and apply evolutionary community detection to estimate the temporal change in network structure. The objective is to identify the inner circles of political power and bureaucracy underneath formal work
relations and observe how the political elite groups form and change over time. In an attempt to capture the political power partnership in a political elite clique, we model synchronous post assignment and job promotion within a time window as entity relations and build the government official job change network in a time span of over twenty years. In each snapshot of network evolution, communities that exhibit strong association of synchronous job change are identified by the edge betweenness decomposition algorithm [10]. Then, an event-based framework is used to characterize community behavior patterns in consecutive change of network structures.
The approach is effectually demonstrated on two scenarios: (1) identifying and tracking the inner circle of a leading political figure; (2) finding succession pool members in government agencies. We further propose two evolutionary community variation indexes to assess political executive succession.
Experimental results with actual government personnel data provide evidence that government agency succession can be reasonably measured. Clustering in political power circle is usually off record and behind the scene. By showing the discovery of crucial information in opaque (social) networks, our research represents significant progress in extending dynamic social network analysis to the application domain of political observation. This work also has the practical value of providing objective scrutiny on political power transition for the benefit of public interest.
II. AGOVERNMENT ELITE CLIQUE EVOLUTION MODEL Governments are typically characterized as bureaucratic systems with a hierarchical structure of functional division and vertical chains of authority. Government officials are composed of elected political executives and nonelected government bureaucrats with formal work relations as defined by the organizational hierarchy. We are interested in finding political elite cliques among government officials. A political or government elite clique is usually developed by a coalition formation process that involves recruitment, promotion, and departure [11]. Government elite cliques are typically small exclusive groups of elected political leaders and executive bureaucrats with size ranging from 2 to 30. Leaders of government elite cliques are either elected political executives or high-ranking executive bureaucrats with adequate authority to fill positions and grant promotion. While clique membership may be dynamic, clique cohesion is usually strengthened by multiple job advancement as a group over a long time.
The problem of identifying government elite cliques has the characteristics of dynamic data mining in which patterns are found over a sequence of domain data collected over a sliding time window. Government elite cliques are formed over a period of time and involve only a very small percentage of government officials. Static work association is not direct evidence of clique membership. Cue of political partnership, however, is found by repeated co-occurrence of job advancement of a small group of government officials.
Therefore, the problem is formulated as finding association patterns among entities over time where association is defined by synchronous job change.
Figure 1. Evolutionary Community Tracking Process
With access to a bureaucratic career database of Taiwanese government officials [12], we constructed a dynamic government official job change network and developed an evolutionary community tracking process for identification and observation of government elite cliques. The process is divided into three steps: (1) network construction; (2) community identification; (3) community evolution tracking.
As depicted in Fig. 1, the network construction step carries out data encoding that collects individual job change data and transforms them into association networks. Then, the community identification step examines strength of clustering to select candidate cliques. Finally, the community evolution tracking step characterizes changes of communities from one time step to the next and provides aggregate indexes to indicate the overall traits of the selected cliques.
A. Network Construction
The dynamic government official job change network is composed of a sequence of static job change network taken as a snapshot at each evolutionary time step. In each static job change network, entity nodes are government officials and links represent instances of synchronous job change within a specified time window. Data in the bureaucratic career database are read and translated into network data. The primary task is to establish the synchronous job change association links between government officials.
The network construction step involves three parameters, time interval of each evolutionary time step, △t, time window of synchronousness, w, and rank level difference, △r. The parameter of evolutionary time interval determines the frequency of observing community evolution. Considering the average terms of bureaucratic executive jobs, △t is set to 1, 2, and 4 years in various experimental observations. The parameter of synchronousness time window specifies a time interval boundary within which two instances of job change at different dates can be viewed as synchronous. In later experiments, the w parameter is reasonably set to 6 months based on common practices of political post assignments.
Lastly, the parameter of rank level difference restricts the relative range in the bureaucratic hierarchy where two government officials would be admitted for association of synchronous job change. Currently, we consider two levels of rank difference as appropriate range of superior and
Network
subordinate relation. Therefore, △r is set to 2 in later experiments. The combination of △t, w, and △r parameters determines the complexity of the dynamic government official job change network in terms of the number of nodes, the number of links, and the number of static networks.
B. Community Identification
For each job change network at each evolutionary time step, we use the edge betweenness algorithm [10] to identify community with high density of job change association links.
In this approach, betweenness of an edge is defined as the number of shortest paths between pairs of nodes that run through the edge. The algorithm starts by calculating the betweenness of all edges in the network, and then repeats the process of removing the edge with the highest betweenness and recalculating betweenness for all edges affected by the removal. This decomposition strategy provides an iterative network reduction process that gradually increases the number of communities from one, i.e., the network, to many, i.e., all isolated nodes. For the purposes of deriving meaningful division, the process can be terminated when a quality function is maximized or the number of identified communities has reached a certain value.
The communities identified in each static job change network consist of government officials who were strongly associated among them for synchronous job change, and hence, are considered as candidate cliques for further review.
Government officials may be involved in synchronous job change as a batch and the number of officials in the batch may be large with the synchronousness time window of 6 months.
In many practical domains, clique formation is usually a continuous and prolonged process for cohesion development.
Political elite cliques are tightly associated groups of smaller sizes and are also gradually developed through multiple times of job advancement as effective cohesion. Therefore, in order to extract real elite cliques from candidate cliques, we need to acquire stronger evidence by tracking community evolution over time.
C. Community Evolution Tracking
For the purpose of analyzing compositional difference of the same community in two consecutive time steps, we consider the event-based framework of community evolution in dynamic social networks [13]. In previous work, Asur et al. [14]
formulated a set of critical events, i.e., continue, merge, split, form, and dissolve, to characterize community evolution between two consecutive time steps. Takaffoli et al. [15]
expanded the set of critical events by adding two event types, shrink and reform, to cover more forms of community evolution. These events define the changes of a community that involve at least a portion k of the community over two time intervals. Our critical event set mostly follow that of [15] but replace reform with expand for the need to detect new recruitment in a community in our problem domain. We denote a community i at time t as Cti and its node set as Vti. The parameter k is a portion of community population that defines event threshold. The set of critical events are defined as follows.
• Form: A new community, Ct+1x, is considered as formed when at least k portion of its members is recently appeared. Thus, Ct+1x is formed if
, | |
• Continue: A community, Ctx, is considered to be continued as Ct+1x when 1) members of the two communities overlap by more than k, 2) their size ratio falls between r1 and r2. Note that we do not require all such replacement is determined by k. Thus, Ctx is expanded if
| |
1 and
• Split: A community Ctx is considered as split into a set of new communities Ct+1* = { , … , when 1) members of each of the successive communities and members of the originating community overlap by more than k, 2) members of the union of the new communities and members of the originating community overlap by more than k. Thus, Ctx is split if
1
| |
• Merge: A set of communities, Ct* = { , … , is considered as merged into a new community, Ct+1x, when 1) members of each of the successive communities and members of the originating community overlap by more than k, 2) members of the
union of the previous communities and members of the new community overlap by more than k. Thus, Ct*, is merged if
, | |
| |
• Dissolve: A community, Ctx, is considered as dissolved when no community at time t+1 overlaps with Ctx by more than k. Thus, Ctx is dissolved if
such that 1
1, | |
As in previous work on event-based community evolution, these critical events are not mutually exclusive, except for continue, shrink, and expand. In other words, a community transition may match multiple events simultaneously. This event-overlapping issue presents conflicts that need to be resolved. Previous work has presented a number of approaches to tackle this issue, such as a similarity measure with adjustable parameters [17], and a decision-tree based event assignment [18]. In this paper we choose a convenient way to resolve potentially multiple matches by setting priority. Our event priority is form, merge, split, {continue, shrink, expand}, dissolve. During the life cycle of a community, form and dissolve occur only once, while all other events may recur. Fig.
2 shows a scenario in which multiple merge and split events occur between two consecutive community snapshots.
Figure 2. A scenario of event-based community evolution
III. DYNAMIC OBSERVATION ON GOVERNMENT ELITE CLIQUES
The evolutionary community tracking process, as shown in Fig. 1, is designed to identify communities of government officials that exhibit strong association of job change and to follow community evolution by characteristic events. However, communities of synchronous job change may come from regular personnel advancement in hierarchical organizations.
Actual government elite cliques are most likely enveloped within these communities and become obscure to outsiders.
This identity blur exemplifies the information inaccessibility of opaque networks. Our approach of tracking community change
in opaque networks provides a potential solution to extract critical essence in a noisy context. In order to verify the performance of our approach, we consider two types of government elite cliques and observe how the targeted government elite cliques may appear in the evolutionary communities.
A. Clique Observation On Seed Core
The first type of government elite cliques is formed by a leading political figure as the seed core and the inner circle of this political power as members. We use the reigning President (Mr. Ma) of Taiwan in 2012 and his political associates as a prominent case of government elite cliques.
The observation period starts from 1990 and ends at 2009 for a total of 20 years, and is divided into five snapshots of four years. Each snapshot covers the second half of a previous President’s term and the first half of the next President’s term.
This is to capture major instances of personnel change before and after a new President’s tenure of office. A set of 35 government officials was selected as Mr. Ma’s possible political associates. The selection criteria are: (1) he/she has served in the same government agency with Mr. Ma for more than 6 months at least once; (2) he/she is in the top list of the most reported names with Mr. Ma in a news archive.
Figure 3. An elite clique evolution with a seed core
Fig. 3 shows the evolutionary process of the community that exhibits synchronous job change with Mr. Ma. The community in each snapshot S1, S2, S3, S4, and S5, is labeled as C1, C2, C3, C4, and C5. From 1990, Mr. Ma’s political career started as deputy of mainland affairs council, which is a ministry level post. He was appointed as minister of justice in 1993 and was elected as mayor of Taipei city (capital city) in 1999. After a re-election victory in 2003, he served the second term as Taipei city mayor. In 2008, he won the presidential election and became the reigning president at the end of our observation period. Mr. Ma is coded as P1 in Table I and is centered in each community snapshot C1, C2, C3, C4, and C5.
The initial size of the community is 11 and slightly increases to 13. The size substantially increases to 20 at C3, slightly reduces to 16 at C4, and reaches 20 again at C5. The critical events that characterize the community evolution are marked as continue, expand, continue, and expand based on the relative size ratio and the remaining portion between two consecutive community snapshots. Out of the 35 possible political associates 30 executive bureaucrats appear in at least one of the community snapshots. The distribution of the number of persons with appearance from one to five times is 9, 7, 6, 6, and 3 (including Ma himself). Half of the population participates in less than two community snapshots. In other words, Mr. Ma’s government elite clique is operated with more
S1
Cont.
S2 S3 S4 S5
11 13 20 16 20
Expand Cont. Expand
C1 C2 C3 C4 C5
recruitment and discharge and less lasting associates. This observation provides a concise view on an actual government elite clique that gradually develops, fortifies, and sustains as the leading figure’s political power grows.
Table I summarizes career path and community snapshot appearance of representative clique members that exhibit different patterns of association with Mr. Ma.
• Superior-turn-subordinate. P2 and P3 were more established and higher-rank government officials than Mr. Ma in S1, S2, and S3, during which time a positive and close relationship among them seemed to develop.
However, P2 and P3 left government offices at later part of S3 and all of S4 due to their political party’s presidential election loss. In the meantime, Mr. Ma (P1) built up greater political power through Taipei (capital city) mayoral election victories in S3 and S4.
However, P2 and P3 left government offices at later part of S3 and all of S4 due to their political party’s presidential election loss. In the meantime, Mr. Ma (P1) built up greater political power through Taipei (capital city) mayoral election victories in S3 and S4.