Predicting Group Performances Using a Personality Composite-Network Architecture During Collaborative Task

Full text

Predicting Group Performances using a Personality Composite-Network
Architecture during Collaborative Task
Shun-Chang Zhong1,3, Yun-Shao Lin1,3, Chun-Min Chang1,3, Yi-Ching Liu2, Chi-Chun Lee1,3
1Department of Electrical Engineering, National Tsing Hua University, Taiwan
2College of Management, National Taiwan University, Taiwan
3MOST Joint Research Center for AI Technology and All Vista Healthcare, Taiwan
flank03200@gmail.com, astanley18074@gmail.com, cmchang@gapp.nthu.edu.tw,
yichingliu@ntu.edu.tw, cclee@ee.nthu.edu.tw
Abstract
Personality has not only been studied at an individual level,
its composite effect between team members has also been in-
dicated to be related to the overall group performance. In
this work, we propose a Personality Composite-Network (P-
CompN) architecture that models the group-level personality
composition with its intertwining effect being integrated into
the network modeling of team members vocal behaviors in
order to predict the group performances during collaborative
problem solving tasks. In specific, we evaluate our proposed
P-CompN in a large-scale dataset consist of three-person small
group interactions. Our framework achieves a promising group
performance classification accuracy of 70.0%, which outper-
forms baseline model of using only vocal behaviors without
personality attributes by 14.4% absolutely. Our analysis further
indicates that our proposed personality composite network im-
pacts the vocal behavior models more significantly on the high
performing groups versus the low performing groups.
Index Terms: group interaction, personality traits, attention
mechanism, social signal processing
1. Introduction
Small group, which includes three to six people, is the most
common composite unit in forming a decision in workplaces.
Group scholars have been trying to understand what are the
right ingredients of the group members that would lead to a
better (effective) decision-making process collectively. Group
composition is the configuration of member attributes in a team,
and the composition of personality has been indicated to have a
direct effect on the group performance. In 2007, Bell’s meta
study shows that each of the Big-5 personality traits can di-
rectly impact team’s performance [1]. Furthermore, studies
have shown that it is more than the individual team member’s
personality that has an effect, the various configuration of per-
sonality attributes between members would bring a different im-
pact to the group [2, 3]. For example, team-level average of both
‘openness to experience’ and ‘emotional stability’ moderate the
relationship between team conflict and team performance [4];
the variability on ‘agreeableness’ and ‘neuroticism’ are nega-
tively related to the team’s oral presentation performance [5].
The right composition of team members is not only evident
in their group-level personality composition but also manifested
in ‘how each interacts with one another behaviorally’ during
a small group interaction. The behavioral patterns observed
at the group-level during each interaction session are uniquely
formed through time as each individual member expresses one-
self, exchanges ideas, gears toward consensus or conflictual sit-
uations [6, 7, 8, 9]. In fact, engineering researchers have al-
ready made extensive effort into computationally understanding
these different interaction processes through automated analy-
ses of verbal and non-verbal behaviors. For example, Okada
et al. computed co-occurrences of non-verbal behaviors of each
participant with one another to model the group impressions
[10]; Fang et al. designed intra- and inter-personal audio-video
behavior features to perform personality classification during
small group interactions [11]; Batrinca et al. recognized per-
sonality attributes using acoustic and visual nonverbal features
[12]. Most recently, Lin et al. proposed an interlocutor modu-
lated attention network to reach the state-of-the-art personality
recognition accuracy in small group interaction[13].
In this work, we focus on automatically predict group-level
performance during a three-person interaction on a collabora-
tive school policy task. There is very limited research on pre-
dicting group performance in the past. For example, Murray
et al. developed hand-crafted multimodal behavior features of
small group conversation to predict team performance on a col-
laborative task [14]; Avci et al. computed a large set of features
including nonverbal multimodal cues, personality traits, diverse
interpersonal perception to predict group performances [15].
While Avci et al. integrated personality attributes to demon-
strate an improved accuracy, their framework modeled person-
ality attribute simply as auxiliary independent inputs without
considering the intertwining effect of group composite person-
ality with individual members behaviors. In fact, group person-
ality is known to influence team performances in two ways, i.e.,
as an input factor that can increase or decrease the group’s over-
all resources and a modulating factor that shapes teamwork pro-
cesses [16]. Developing sophisticated frameworks that model
the intertwining effect is crucial in advancing the automated be-
havior modeling in small group interactions.
In this work, we propose a Personality Composite-Network
(P-CompN) architecture to predict group performance on a
large-scale dataset including 97 sessions of three-person inter-
actions. The P-CompN includes a fusion of two major net-
works, i.e., Interlocutor Acoustic Network (IAN) and Personal-
ity Network (PN). The PN network predicts team performance
based on the group-composite Big-5 personality attributes. The
IAN network is trained using bi-directional long short-term
memory network (BLSTM) with attention being modulated
by the group-composite personality. Our P-CompN considers
both group-level personality configuration and team member’s
acoustic behaviors with their intertwining effect being jointly
modeled. P-CompN achieves a promising unweighted aver-
age recall (UAR) of 70.0% in classifying group performances.
Our analysis further reveals that group-composite personality
Copyright © 2019 ISCA
INTERSPEECH 2019
September 15–19, 2019, Graz, Austria
http://dx.doi.org/10.21437/Interspeech.2019-20871676

Figure 1: A complete schematic of our Personality Composite-Network (P-CompN). It includes an Interlocutor Acoustic Network (IAN)
and a Personality Network (PN) with a decision-level fusion for the classification task. Specifically, we propose to learn a personality
composite control weight to modify the original BLSTMs attention mechanism that models the effect of group-level personality attributes
on participants acoustic behaviors jointly.
attributes alters significantly the IAN’s attention weights be-
tween the high versus the low performing groups.
2. Research Methodology
2.1. The NTHULP Audio-Video Database
Our NTHULP audio-video database is collected at the College
of Management of the National Taiwan University (NTU). Each
recording includes a session of three participants engaged in
a collaborative school policy task [17]. The three participants
play different roles chosen at random: vice president of uni-
versity, vice president of business school, and a member of the
business school teachers committee. They are asked to carry
out a task to solve school problems by discussing potential al-
teration on the school policy. Each of the participants would be
given a piece of relevant information that is different from oth-
ers in the team, and they are asked to work together by sharing
ideas and communicating collaboratively. However, one of the
three participants is a sleeper cell assigned by the experimental
personnel. While the sleeper cell knows about all the detailed
information to complete the task, he/she would only take part in
the task passively. The goal of this task is to study the interac-
tion of the other two participants to understand how they may
be influenced by the sleeper cell’s unresponsive behaviors and
its effect on the outcome of this collaborative task.
The NTHULP contains 97 recorded sessions with 194 sub-
jects total (age ranges from 19 to 51 years old, 95 males and 99
females). It includes audio and video recordings collected us-
ing two cameras and three separate wireless lapel microphones.
Additionally, the database contains the following metadata: in-
dividual personality trait and group performance outcome score.
Figure 2: A histogram of the group performance score.
Personality. Each participant’s Big-5 personality attribute, i.e.,
Extraversion, Agreeableness, Conscientiousness, Neu-
roticism and Openness, is measured using the Goldbergs
(1992) 10-item scale [18]. Participants are asked to eval-
uate how accurately each statement described to them on
a 5-point scale, with anchors of 1 = “very inaccurate”
and 5 = “very accurate”.
Group Performance. The performance of each team is evalu-
ated by two trained research assistants using the scoring
manual for the school policy task developed by Wheeler
and Menneck [17]. The scoring manual includes over
300 possible solution scores to this task scenario. The
scoring includes two distinct dimensions: a problem-
solving score for how well the solution solves the case
problem, and a feasibility score for how feasible the so-
lution is to the case problem. The two research assistants
independently code all of the 97 groups by identifying
the best match between the participants final decision
and the solution listed in the manual. Any disagreement
between the two coders is reconciled by the third coder.
In this work we use the binarized feasibility score as the
class label indicating group performance. We define class 1 as
high performing groups with a score greater or equal to 50, and
class 0 as low performing groups with a score less than 50. Fig-
ure 2 depicts the database distribution of the feasibility score.
2.2. Personality Composite-Network (P-CompN)
Figure 1 shows our Personality Composite-Network (P-
CompN) architecture. We model only the two actual partici-
pants within the session ignoring the sleeper cell’s behaviors
due to his/her consistent non-engaging behaviors in this group
performance prediction task. Specifically, our proposed P-
CompN architecture is composed of a fusion between two sub-
networks, Interlocutor Attention Network (IAN) and Personal-
ity Network (PN). We will first describe the two different fea-
tures inputs to P-CompN and then the details of our framework.
2.2.1. Feature Inputs: Acoustics and Personality Attributes
The audio signals are first segmented into speaker utterances
automatically. We extract the extended Geneva minimalistic
acoustic parameter set (eGeMAPS) for each utterance [19] as
acoustic inputs. eGeMAPs computes 88 dimensional features
1677

Table 1: Model performances using the metric of unweighted average recall (UAR) for high and low performing groups. The overall
result shows that the P-CompN outperforms all other methods in group performance classification task achieving 70.0% UAR.
Individual Models Group Models
Model
(talkative/less) Overall Low High Models Overall Low High
Model 0 55.6/54.4 47.2/41.7 64.0/67.2 Model 3 58.1 55.6 60.7
Model 1 55.6/53.2 47.2/50.0 63.9/56.5 Model 4 60.1 61.1 59.0
Model 2 58.5/59.1 69.4/44.4 47.5/73.8 IAN 63.1 63.9 62.3
PN 63.6 58.3 68.5
P-CompN 70.0 77.8 62.3
including statistical properties of mel-frequency cepstral coeffi-
cients (MFCCs), associated delta, and prosodic information.
In terms of personality attributes, since each of the inter-
locutors has different traits, in order to measure personality
composition characteristics within the group, an intuitive man-
ner is to compute statistics. Specifically, each member has 5
personality scores, and we compute the maximum, minimum,
mean and difference value of the group members to derive a
20-dimensional features as inputs of personality attributes.
2.2.2. Interlocutor Acoustic Network (IAN)
The core of IAN uses BLSTM with an attention mechanism
trained on the acoustic inputs. Each utterance is a time step
t. For each session, we first assign the interlocutors as either a
talkative or a talk-less subject; the talkative subject is the per-
son that speaks the most and often takes the leading role in the
interaction, and talk-less subject tends to look quieter and much
more tolerant to the existence of an assertive person. We train
a typical BLSTM for each subject with attention weight, αt,
defined as:
αt=exp[(uTyt)]
Ptexp(uTyt)(1)
where ytis the hidden layer of time step t.
In this work, we design a novel personality control mecha-
nism that integrates the effect of group personality composition
into the attention weight. Specifically, we take the 20 dimen-
sional personality composite features multiplies by a learnable
weight matrix W20×τto derive the personality control weight
for the i-th sample as below:
ctrli×t=Pi×20 ×W20×t(2)
where Pi×20 indicates the group composite personality inputs
mentioned in section 2.2.1, Wis normalized for summing to 1
by softmax. We can then reweight the original attention weight:
α0
t=αt+ctrlt(3)
With this personality reweighted attention mechanism, we
further derive the representation of the IAN, zIAN, by concate-
nating the BLSTM hidden layer output from both the talkative
and the talk-less subject :
zIAN = [z00
talkative z00
talk-less](4)
z00
{talkative,talk-less}=Gα0
t×y{talkative,talk-less},t(5)
where Gindicates a functional pooling layer over time, i.e.,
computing the maximum, minimum, mean, median, standard
deviation of the hidden layer output for the BLSTM. After ob-
taining zIAN, we feed it into the prediction layer consists of five
fully-connected (DNN) layers to perform binary classification.
2.2.3. Personality Network (PN)
Another sub-network is the personality network (PN). PN is
based on a 8-layer DNN that takes the input of 20 dimensional
group composite personality features to predict the group per-
formance directly.
The final prediction using our P-CompN in the binary group
performances is based on the average softmax output probabil-
ity of IAN and PN.
3. Experiment Setup and Results
3.1. Experiment Setup
In this section we briefly describe different comparison meth-
ods, model parameters, and our evaluation scheme.
3.1.1. Model Comparison
•Model 0-Baseline
Using a standard talkative-only or talk-less-only subject’s
BLSTM with attention (without DNN layers) to perform
recognition directly.
•Model 1-Individual Personality Network
Using a 8-layer DNN to model talkative or talk-less subject’s
five personality attributes only (not the composite statistic
measures) to perform recognition directly.
•Model 2-BLSTM + Individual Personality Network
Combining Model 0 and Model 1 using decision-level fusion
by averaging the output probability to perform recognition.
•Model 3-Dual-BLSTM
Concatenating output of each interlocutor’s BLSTM using a
summation pooling layer (not the functional pooling layer)
and feeding it to a five-layer DNN to perform recognition.
•Model 4-Dual-BLSTM + Personality Control
Integrating Model 3 with the personality control mechanism
to the BLSTM attention weight to perform recognition.
•Interlocutor Acoustic Network (IAN)
Using the method detailed in section 2.2.2, which modifies
Model 4 by replacing summation layer with the functional
pooling layer to perform recognition.
•Personality Network (PN)
Using the method in section 2.2.3, which uses 5-layer DNN
on personality composite features to perform recognition.
•Personality Composite-Network (P-CompN)
Using our proposed architecture to perform recognition.
3.1.2. Other Experimental Parameters
We pad sentences to equal length before training (224/147 time-
steps for talkative/talk-less respectively), then each BLSTM is
trained with a fixed length step. The number of hidden nodes in
the BLSTM is 64. IAN has 5 fully-connected layers with node
1678

size of: 1280, 640, 256, 256, 128, 2. PN has 8 fully-connected
layers with node size of: 20, 64, 64, 64, 32, 32, 32, 16, 2. We
use ReLU as activation function, drop out layer for first and last
layers, and batch normalization is also applied. Batch size is
set at 16, learning rate is set at 0.0005 using ADAM optimizer.
Cross-entropy is our optimized loss function, and we train our
network using 40 epochs. The experiment is carried out using
5-folds cross validation using the metric of unweighted average
recall (UAR). We adjusted to make the distribution of 5 folds
data consistent and reduce the bias.
3.2. Results and Analyses
3.2.1. Analysis on Model Performance
Table 1 summarized our complete prediction results. Our pro-
posed P-CompN obtains the best overall UAR (70.0%), which is
15% higher than baseline Model 0. Model 0 and Model 1 mod-
els acoustic behavior and personality attribute using individual
participant only (talkative-only or talk-less-only). The accu-
racy obtained with these two models are only around 55%, and
by using complementary information from individual model of
acoustic behaviors and personality attribute, i.e., Model 3, it
increases slightly to around 59%. We observe that by simply
modeling a single participant within a small group collabora-
tive task is not sufficient to obtain a sufficient predictive power
of the group performance. Generally, by comparing model ac-
curacy obtained in the ‘Group Models’, i.e., modeling both par-
ticipants, is better than the ‘Individual Models’.
Furthermore, Model 3 and Model 4 differs by whether the
participants acoustic BLSTMs have attention mechanism being
modulated by a personality composite control weight. Model
4 improves about 2% over Model 3 in predicting group perfor-
mance, which indicates that indeed the group personality infor-
mation affects jointly the behavior manifestation when complet-
ing this collaborative task. IAN replaces the conventional sum-
mation part of BLSTM attention mechanism with a functional
pooling layer, this method computes statistical properties on
the time-series output of the BLSTM weighted by personality-
controlled attention mechanism. The functional pooling pro-
vides another 3% improvement indicating the need of a more
complex temporal dynamics characterization of the participants
acoustic behaviors, which are shown to be beneficial in this
group performance recognition task.
Finally, we also note an interesting observation that PN
by itself achieves 63.6% UAR in group performance predic-
tion task. Our experiments demonstrate that the group mem-
ber’s personality configuration carry significant information on
the team performance, which corroborates with past literature
in group studies [2, 3]. In summary, our P-CompN architecture
that fuses the prediction output of IAN (63.1%) and PN (63.6%)
to obtain the best performing model of 70% UAR.
3.2.2. Analysis of Personality and Group Performance
Our experiments demonstrate that personality composite fea-
tures computed within a group can be used to predict team per-
Table 2: The bolded number indicates a statistically significant
correlation between group performance and each of the Big-5
personality composite attribute.
Big-5 Max Min Mean Difference
Extraversion -0.06 -0.04 -0.06 0.07
Agreeableness 0.17 0.06 0.16 0.09
Conscientiousness -0.01 0.01 -0.01 0.01
Neuroticism -0.01 0.20 0.08 -0.14
Openness 0.10 -0.05 0.01 0.05
formance classification in this school policy collaborative task.
To understand the influence of group personality on team per-
formance, we compute spearman correlation between each of
the 20 dimensions of composite group personality measures and
our target group performance label.
Table 2 includes the correlation results. The number in bold
indicates significant correlation at α= 0.05 level. We observe
that the maximum, the average of Agreeableness and the min-
imum of Neuroticism are positively correlated with group per-
formance. Previous study has also shown that Agreeableness
is one of the most important personality traits for team perfor-
mance due to its emphasis on cooperation and facilitation, if the
group members could treat others more friendly (maxAgree ),
show patience (minNeur ), and keep a collaborative atmosphere
(meanAgree ), it could facilitate a more engaging and comfort-
able interaction and help finish the task collaboratively with
quality [20, 21, 22, 23].
3.2.3. Analysis of Attention Weights
In section 3.2.1, we demonstrate that personality controlled
reweighting of the BLSTM attention network help improve the
overall prediction accuracy. We would like to further ana-
lyze this modified attention weights, which help re-emphasize
the important interlocutors behavior regions in the session, as
a function of the group performances. Specifically, we com-
pute the ratio within each session that these modified attention
weights have positive values, and compare the ratios between
the high performing groups versus the low performing groups
using t-test (α= 0.05). We find that the high performing group
sessions have a larger percentage of weights being positive than
the low performing group (p= 0.015). Our personality con-
trolled weights operates by shifting up and down the original
attention weights. Personality mechanism tends to add more
weights to the high performing group’s behavior segments. This
results seems to be intuitive that for those groups that have the
right composition of personality configuration would behave
more collaboratively, e.g., willing to communicate more, share
more ideas, and be more engaging. This is evident in a having
larger attention weights placed on their behaviors.
4. Conclusion and Future Work
Personality attribute is not only related to individual behavior
pattern during interaction, the personality composition within
the group also affects the overall team performance, especially
in small group collaborative task solving interactions. In this
work, we propose a novel Personality Composite-Network (P-
CompN), which includes a personality network (PN) and an in-
terlocutor acoustic network (IAN) that jointly integrate the ef-
fect of group members personality attributes into the attention
mechanism. We evaluate our P-CompN on a large three-person
interaction of school policy task and achieve a promising 70%
accuracy in predicting the group performance. Our analyses re-
veal several important personality attribute configurations to the
group performance and demonstrate the effect of higher empha-
sis on behaviors for groups with higher collaborative effort. We
will continue to advance our technical framework by including
other non-verbal modalities (e.g., facial expressions and ges-
tures), linguistic contents, and conversation flow (e.g., question
answering patterns). Furthermore, by continuously collaborat-
ing with group scholars, we would like to investigate the com-
plex interaction effect between the behaviors expressed and the
personality traits at the group-level and bring insights about the
specific interaction strategy that can help better achieve effec-
tive communication within a group discussion.
1679

5. References
[1] S. T. Bell, “Deep-level composition variables as predictors of team
performance: a meta-analysis.” Journal of applied psychology,
vol. 92, no. 3, p. 595, 2007.
[2] A. Kramer, D. P. Bhave, and T. D. Johnson, “Personality and
group performance: The importance of personality composition
and work tasks,” Personality and Individual Differences, vol. 58,
pp. 132–137, 2014.
[3] R. L. Moreland, J. Levine, and M. Wingert, “Creating the ideal
group: Composition effects at work,” Understanding group be-
havior, vol. 2, pp. 11–35, 2013.
[4] B. H. Bradley, A. C. Klotz, B. E. Postlethwaite, and K. G. Brown,
“Ready to rumble: How team personality composition and task
conflict interact to improve performance.” Journal of Applied Psy-
chology, vol. 98, no. 2, p. 385, 2013.
[5] S. Mohammed and L. C. Angell, “Personality heterogeneity in
teams: Which differences make a difference for team perfor-
mance?” Small group research, vol. 34, no. 6, pp. 651–677, 2003.
[6] K. A. Jehn and E. A. Mannix, “The dynamic nature of conflict: A
longitudinal study of intragroup conflict and group performance,”
Academy of management journal, vol. 44, no. 2, pp. 238–251,
2001.
[7] C. Beyan, V.-M. Katsageorgiou, and V. Murino, “A sequen-
tial data analysis approach to detect emergent leaders in small
groups,” IEEE Transactions on Multimedia, 2019.
[8] P. Dhani and T. Sharma, “Emotional intelligence and personality
traits as predictors of job performance of it employees,” Inter-
national Journal of Human Capital and Information Technology
Professionals (IJHCITP), vol. 9, no. 3, pp. 70–83, 2018.
[9] N. Attia, Big Five personality factors and individual performance.
Universit´
e du Qu´
ebec `
a Chicoutimi, 2013.
[10] S. Okada, L. S. Nguyen, O. Aran, and D. Gatica-Perez, “Model-
ing dyadic and group impressions with intermodal and interperson
features,” ACM Transactions on Multimedia Computing, Commu-
nications, and Applications (TOMM), vol. 15, no. 1s, p. 13, 2019.
[11] S. Fang, C. Achard, and S. Dubuisson, “Personality classification
and behaviour interpretation: An approach based on feature cate-
gories,” in Proceedings of the 18th ACM International Conference
on Multimodal Interaction. ACM, 2016, pp. 225–232.
[12] L. Batrinca, N. Mana, B. Lepri, N. Sebe, and F. Pianesi, “Mul-
timodal personality recognition in collaborative goal-oriented
tasks,” IEEE Transactions on Multimedia, vol. 18, no. 4, pp. 659–
673, 2016.
[13] Y.-S. Lin and C.-C. Lee, “Using interlocutor-modulated attention
blstm to predict personality traits in small group interaction,” in
Proceedings of the 2018 on International Conference on Multi-
modal Interaction. ACM, 2018, pp. 163–169.
[14] G. Murray and C. Oertel, “Predicting group performance in task-
based interaction,” in Proceedings of the 2018 on International
Conference on Multimodal Interaction. ACM, 2018, pp. 14–20.
[15] U. Avci and O. Aran, “Predicting the performance in decision-
making tasks: From individual cues to group interaction,” IEEE
Transactions on Multimedia, vol. 18, no. 4, pp. 643–658, 2016.
[16] J. E. Driskell, R. Hogan, and E. Salas, Personality and group per-
formance. Sage Publications, Inc, 1987.
[17] B. Wheeler and B. Mennecke, “The school of business policy task
manual,” 1992.
[18] L. R. Goldberg, “The development of markers for the big-five fac-
tor structure.” Psychological assessment, vol. 4, no. 1, p. 26, 1992.
[19] F. Eyben, K. R. Scherer, B. W. Schuller, J. Sundberg, E. Andr´
e,
C. Busso, L. Y. Devillers, J. Epps, P. Laukka, S. S. Narayanan
et al., “The geneva minimalistic acoustic parameter set (gemaps)
for voice research and affective computing,” IEEE Transactions
on Affective Computing, vol. 7, no. 2, pp. 190–202, 2016.
[20] W. G. Graziano, E. C. Hair, and J. F. Finch, “Competitiveness
mediates the link between personality and group performance.”
Journal of Personality and Social Psychology, vol. 73, no. 6, p.
1394, 1997.
[21] G. A. Van Kleef, A. C. Homan, B. Beersma, and D. van Knip-
penberg, “On angry leaders and agreeable followers: How leaders
emotions and followers personalities shape motivation and team
performance,” Psychological Science, vol. 21, no. 12, pp. 1827–
1834, 2010.
[22] T. Sy, S. Cˆ
ot´
e, and R. Saavedra, “The contagious leader: impact
of the leader’s mood on the mood of group members, group affec-
tive tone, and group processes.” Journal of applied psychology,
vol. 90, no. 2, p. 295, 2005.
[23] R. E. De Vries, B. Van den Hooff, and J. A. De Ridder, “Ex-
plaining knowledge sharing: The role of team communication
styles, job satisfaction, and performance beliefs,” Communication
research, vol. 33, no. 2, pp. 115–135, 2006.
1680