Measuring Team Creativity Through Longitudinal Social Signals
Peter A. Gloor, Adam Almozlino, Orr Inbar
MIT Center for Collective Intelligence
5 Cambridge Center, Cambridge MA 02139, USA pgloor@mit.edu Wei Lo
Computer Science and Technology department
Zhejiang University, Hangzhou, P.R. China
Shannon Provost
McCombs School of Business, University of Texas at Austin
Austin, TX, USA

Summary
Research into human dynamical systems has long sought to identify robust signals for human behavior. We have discovered a series of social network-based indicators that are reliable predictors of team creativity and collaborative innovation. We extract these signals from electronic records of interpersonal interactions, including e-mail, and face-to-face interaction measured via sociometric badges. The first of these signals is Rotating Leadership, measuring the degree to which, over time, actors in a team vary in how central they are to team’s communication network’s structure. The second is Rotating Contribution, which measures the degree to which, over time, actors in a team vary in the ratio of communications they distribute versus receive. The third is Prompt Response Time, which measures, over time, the responsiveness of actors to one another’s communications. Finally, we demonstrate the predictive utility of these signals in a variety of contexts, showing them to be robust to various methods of evaluating innovation.

Introduction
In this paper we introduce a series of longitudinal, network-based measures of social interaction patterns that predict collaborative innovation. Innovation is a universal, emergent human behavior. According to noted evolutionary biologist E.O Wilson “…it was necessary for the evolving populations to acquire an ever higher degree of intelligence. They had to feel empathy for others, to measure the emotions of friends and enemy alike, to judge the intentions of all of them, and to plan a strategy for personal social interactions” [1]. Innovation is a universal, emergent human behavior, one that rarely occurs through the actions of a single individual, but rather through collaboration among individuals [2]. Here we focus on the predictive utility of observing this collaboration at the level of interpersonal interaction events.
Recently, researchers have had success in identifying reliable, quantitative indicators of phenomena in human systems. Among these indicators are “honest signals” [3][4][5], which signify the presence of social influence. This name captures both the separation of these signals' from the subjectivity that often plagues other methods for measuring human behavior, and the robustness of these signals to a variety of behavioral contexts. Understanding these “honest signals” can convey a significant advantage. To quote
E.O Wilson again, “…social intelligence was therefore always at a high premium. A sharp sense of empathy can make a huge difference and with it an ability to manipulate, to gain cooperation, and to

deceive” [1]. Robust, quantitative measures for collective human behavior may serve as the quantitative, larger-scale analog for individual social intelligence.
Previous work studying collective creativity and innovation has been primarily qualitative, focused on the creativity of individuals, or both [6][7]. Other research has been restricted to a particular interpretation of creativity, studying for example patent production [8], or to a particular setting, studying for example large corporations [9]. Therefore, this research has failed to identify reliable signals of collective innovation. Part of the reason previous work has had limited success may lie in the difficulty of understanding innovation itself. A formal definition of innovation remains elusive, as does the boundary between incremental improvements and innovative change. If a certain dependent variable, such as creativity, is difficult to formally define, it may be difficult to identify a quantitative and reproducible independent variable that indicates the dependent.
Our Approach
We have attempted to work around this issue by evaluating several different proxies for creativity across several different scenarios, and identifying measures that reliably signal the presence of these proxies across the scenarios. Using a wide selection of proxies in a variety of context, we have identified reproducible independent variables that strongly correlate with the proxies. We term these variables (1)
Rotating Leadership, (2) Rotating Contribution, (3) Prompt Response Time.
From these variables, Rotating Leadership and Rotating Contribution show positive correlation in
“creative” work scenarios, but strong negative correlation with “non-creative” scenarios, suggesting that
Rotating Leadership and Rotating Contribution are a good “honest signal” for team creativity. This corresponds with the intuition that creative work requires innovation and breaking known patterns of thought and behavior, while breaking known patterns may disrupt non-creative work. Prompt Response
Time, on the other hand, shows positive correlation across all scenarios, suggesting that it is a better indicator of team productivity. This corresponds with the intuition that it is, in general, better to have a more promptly communicating team.
(1) Rotating Leadership (RL)
Rotating Leadership (RL) measures the degree to which, over time, the members in a team vary in how
“central” they are to the team’s communications. The advantage of centralized leadership for creative tasks was for instance observed among Wikipedians [10], where it was found that Wikipedia articles authored by more centrally communicating editors became articles of the highest quality (featured articles) more rapidly. RL can be observed in a visualization of a network when distinct nodes appear, over time, to oscillate between central and peripheral positions in the network. Intuitively, RL evaluates how much, across time and the team members, team members switch between being highly central to the overall communications of the team, and being peripheral to those communications. Formally, RL measures oscillations in Betweenness Centrality (BC) over time among actors in the team.
The effects of the centrality of team’s actors to the team’s performance was first observed among teams of Eclipse open source developers communicating electronically [12]. It was subsequently observed in a study of a marketing team in a bank communicating face-to-face [13], and in a study of nurses communicating in a hospital [14]. In this last scenario, quantitative measures were compared with personality characteristics such as openness, as measured by the Neo-FFI [15], and group creativity was

measured through peer and management/instructor assessment, based on the premise that experts can identify creativity [7]. Note that teams composed of highly intelligent individuals are not necessarily intelligent as a whole [16], while measures such as RL were dependably correlated with team creativity.
Betweenness centrality [11] (BC) is a global measure of how centrally located a node is in the structure of a network. For a given node, it is measured by evaluating the shortest paths in the network, specifically, the proportion of all shortest paths in the network that pass through the node of interest. Mathematically,
BC of a node v is defined as:

where is the total number of shortest paths from node paths that pass through v.

to node

and

is the number of those

In order to calculate RL, it is necessary to aggregate measures of BC, which occur at the scale of an individual actor at an individual time step, to the scale of the whole network over the full time frame. In order to do this in a fashion that indicates variation in BC we counted the number of local maxima and minima in the vector of each actor’s BC over time, and then summed this number across the actors in a team. Formally, we count the local maxima of function f(t)=g(t) within time interval [t1,t2]. There is a local maximum for time t at point t*, if there exists some ε > 0 such that f(t*) ≥ f(t) when |t – t*| < ε. Similarly, we count the local minima at t*, if f(t*) ≤ f(t) when |t – t*| < ε.
RL for actor i over time window ws is therefore:
RLi = #local minimaiws + #local maximaiws
1
=

!

!
!!!

Figure 1: RL visualized through oscillations in BC over time [17]
This figure illustrates Rotating Leadership (RL) for two teams, one with high RL, and one with low RL.
Actors are placed along the Y-axis, while the X-axis encodes time, and the Z-axis the Betweenness
Centrality (BC) of actors for each hour, sorted, each hour, by the decreasing BC of actors. The back plane, which rises and falls, represents the set of actors who rotate taking the lead in the team’s communication. The left picture illustrates an example of a team with high RL. This example was drawn from a 6-day long graduate student seminar, and communications were measured using sociometric badges. This image includes 15 actors, and has had BC oscillation computed every hour using a time window of 12 hours, with a date range 6/13/2010 12:37 pm to 6/19/2010 23:37 pm.
The right picture illustrates an example of a team with low RL. This example was drawn from the customers and employees of a large service provider serving one customer, and communications were measured using the email archive of the service provider. This image includes 2857 actors, and has had
BC oscillation computed every day using a time window of 7 days, with a date range between 6/13/2012 to 12/30/2012. The high back represents the key account managers who are consistently taking the lead in team communication.

(2) Rotating Contribution (RC)
Rotating Contribution (RC) measures the degree to which, over time, actors in a team vary in how much they broadcast communications versus listen to communications. RC can be observed in a visualization of a network when distinct nodes appear, over time, to vary widely in how many incoming versus outgoing links they have. Intuitively, RC evaluates how much, across time and the team members, team members switch off between broadcasting many communications and listening to may communications. Formally,
RC measures oscillations, over time, of the Contribution Index (CI) of actors in a team.
Contribution Index (CI) is a measure of how much an actor disseminates versus receives communications.
For a given node, it is equal to ratio of incoming to outgoing links incident upon that node. An actor that only sends messages will have a CI of 1, an actor that sends and receives an identical number of messages will have a CI of 0, and an actor that only receives messages will have a CI of -1 [18]. Formally, the CI of an actor over a given time frame is:

CI =

messages _ sent − messages _ received messages _ sent + messages _ received

In order to calculate RC, it is necessary to aggregate measures of CI, which occur at the scale of an individual actor at an individual time step, to the scale of the whole network over the full time frame. In order to do this in a fashion that indicates variation in CI we counted the number of local maxima and minima in the vector of each actor’s CI over time, and then summed this number across the actors in a team. Formally, we count the number of local maximum points of function f(t)=c(t) within time interval [t1,t2].
There is a local maximum for time t at point t* if there exists some ε > 0 such that f(t*) ≥ f(t) when |t – t*|
< ε. Similarly, we count the local minima at t*, if f(t*) ≤ f(t) when |t – t*| < ε. RCiws for actor i and time window ws is therefore
RCiws = #local minimaiws + #local maximaiws
RC=

!
!

!
!!!

!

Figure 2: RC visualizing though CI oscillations over time [17]
This figure illustrates Rotating Contribution (RC) for two teams, one with high RC, and one with low RC.
Actors are placed along the Y-axis, while the X-axis encodes time, and the Z-axis the Contribution Index
(CI) of actors for each hour, sorted, each hour, by the decreasing CI of actors. The back plane, which rises and falls, represents the set of actors who rotate taking the lead as most vocal contributors.
The left picture illustrates an example of a team with high RC; RC oscillates highly among time steps and the actors of the team. This example was drawn from a 6-day long graduate student seminar, and communications were measured using sociometric badges. This image includes 15 actors, and has had

BC oscillation computed every hour using a time window of 12 hours, with a date range 6/13/2010 12:37 pm to 6/19/2010 23:37 pm.
The right picture illustrates an example of a team with low RC; CI oscillates relatively little among time steps and the actors of the team. This example was drawn from the customers and employees of a large service provider serving one customer, and communications were measured using the email archive of the service provider. This image includes 2857 actors, and has had CI oscillation computed every day using a time window of 7 days, with a date range between 6/13/2012 to 12/30/2012. The high back represents the key account managers who are consistently the most vocal by sending more emails than they receive.

(3) Prompt Response Time (PRT)
Prompt Response Time (PRT) measures the degree to which, over time, actors are prompt at communicating to those who have communicated to them. Intuitively, PRT corresponds with how fast, across actors in a network, actors are at “getting back” to each other’s communications. In order to capture this formally, PRT is defined in terms of the Communication Frame (CF), which groups communication events between pairs of actors which may “get back” to each other, and Frame Nudges, which measure the number of communication events in a CF, and Elapsed Time, which measures the time duration of a CF.
A Communication Frame (CF) groups a set of time-adjacent communications between a pair of actors.
Suppose a pair of actors X and Y in a network, with a set of communication events, or directional, timestamped edges, between them. A single CF defines all communication events from X to Y, prior to and including a communication event from Y to X. Intuitively, this is all the messages your colleague has sent you since you last messaged her, prior to and including the first message you send back to your colleague.
In this framework, you, actor X, are the “source” actor in the CF, corresponding with the origin of the first communication in the CF, and your colleague, actor Y, is the “target” actor in the CF, corresponding with the origin of the last communication in the CF. The Elapsed Time (ET) for this CF is the difference between the first communication in the frame and the last communication in the frame. The Frame
Nudges (FN) for this CF is the number of communication events in the CF, intuitively FN is the number of “pings” X sends until Y responds.
To get the network-level measure of PRT from the edge-level measure of ET in CFs it is necessary to aggregate measures. We accomplished this by using an intermediate actor-scale measure, where we evaluated the “responsiveness” of actors through their Responsiveness in Communication Frames (RCF).
Intuitively, we measure how quickly actors got back to people who messaged them.
This can be accomplished either by measuring the ET or the FN of CFs. We define RCF via ET (RCFET) for an actor as the mean ET for all CFs in which the node is the “target” node. We define RFC via
FN (RCF-FN) for an actor as the mean FN for all CFs in which the node is the “target” node. For actor i, where is a given CF in the set of CF denoted as ∈ F, and ∆T! is the time elapsed for frame ∗ , such that ∗ ∈ F∩i, where F∩i represents the set of all frames that actor i is a member, RCF -ET is:
RCF-ET i =

|!∩!|
!!! ∆!!

! ∩!

For actor i, where is a given CF in the set of CF denoted as ∈ F, and FN is the number of edges in frame ∗ , such that ∗ ∈ F∩i, where F∩i represents the set of all frames that actor i is a member, RFC-FN is: RCF-FN i =

|!∩!|
!!! !!!

! ∩!

We then aggregate this actor-scale measure to the networks-scale by averaging RCF for all actors in the network. This procedure is the same for RCF-ET and RCF-FN. For a network, where ! denotes the number of communications of each actor i, PRT is therefore:
PRT=

!
!!! !"#! ∗ !!
!
!!! !!

Analysis and Results
We extracted signals of team creativity and productivity from electronic records of interpersonal interactions, including e-mail, and face-to-face interaction measured via sociometric badges [28]. Some of our samples have quite a small N (