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

Shun-Chang Zhong, Yun-Shao Lin, Chun-Min Chang, Yi-Ching Liu, Chi-Chun LeeInterspeech 2019

TL;DR

This paper introduces a novel "Personality Composite-Network" (P-CompN) that predicts group performance in collaborative tasks by jointly modeling individual acoustic behaviors and the composite effect of team members' personalities. The P-CompN significantly outperforms baseline models, demonstrating that understanding the interplay between personality and behavior is crucial for predicting team success.

What Problem Does This Paper Address?

Effective collaboration in small groups is vital across many domains, from workplaces to research teams. While individual personality traits are known to influence performance, the composition of personalities within a group and how it interacts with individual behaviors during collaboration is less understood. Existing research often treats personality as an auxiliary input or focuses solely on individual behaviors. This paper addresses the gap in computationally modeling the synergistic effect of group-level personality composition and individual acoustic behaviors to predict group performance in collaborative tasks.

Methods

The core innovation is the Personality Composite-Network (P-CompN), which fuses two main components:

  • Interlocutor Acoustic Network (IAN): This network processes individual acoustic features (extracted using eGeMAPS) from team members. It employs a Bidirectional Long Short-Term Memory (BLSTM) with an attention mechanism. Crucially, the attention mechanism is modulated by a "personality composite control weight." This weight is derived from the group's collective personality attributes (e.g., max, min, mean of Big-5 traits) and learned through a weight matrix. This allows the network to dynamically re-weight the importance of different parts of an individual's speech based on the group's personality composition.
  • Personality Network (PN): This is a standard Deep Neural Network (DNN) that directly takes the 20-dimensional group composite personality features as input to predict group performance.

The final prediction is a decision-level fusion (averaging) of the outputs from the IAN and PN. The authors also introduce a "functional pooling" layer in the IAN, which computes statistical properties (max, min, mean, median, std dev) of the BLSTM hidden layer outputs, offering a richer temporal representation than simple summation.

Datasets

The study utilizes the NTHULP Audio-Video Database. This dataset comprises 97 recorded sessions of three-person collaborative tasks focused on solving school policy problems. Each session includes audio and video recordings, individual personality trait metadata (Big-5 traits), and a group performance outcome score. The performance score is binarized into "high performing" (score >= 50) and "low performing" (score < 50) for classification.

Key Insights and Ideas

  • Personality is not just additive, it's compositional: The study strongly suggests that the combination of personalities within a group, not just the average or individual traits, significantly impacts how team members interact and perform.
  • Intertwining effect is key: The P-CompN's success highlights that modeling the interplay between group personality composition and individual acoustic behaviors is more effective than treating them as independent factors.
  • Attention modulation matters: Dynamically adjusting attention weights based on group personality composition leads to better performance, implying that certain behaviors become more or less important depending on the team's personality makeup.
  • High-performing groups exhibit different behavioral emphasis: Analysis of attention weights shows that personality-controlled mechanisms tend to assign higher weights to behavioral segments in high-performing groups, suggesting more collaborative and engaging interactions.

Key Results

  • The proposed P-CompN achieved an unweighted average recall (UAR) of 70.0% in classifying group performances.
  • This is a significant improvement, outperforming baseline models (e.g., using only acoustic behaviors without personality) by 14.4% absolutely.
  • The P-CompN also outperformed individual models (IAN at 63.1% UAR, PN at 63.6% UAR) by fusing their predictions.
  • Correlation analysis revealed that Agreeableness (max and mean) and Neuroticism (min) were positively correlated with group performance, aligning with existing theories on cooperation and conflict management.

Limitations and Future Work

  • Limited to acoustic features: The current model primarily relies on acoustic signals. The authors acknowledge that incorporating other non-verbal modalities like facial expressions, gestures, and linguistic content (e.g., conversation flow, question-answering patterns) could further enhance performance.
  • Focus on three-person groups: The architecture was evaluated on three-person interactions. Future work could explore its applicability to larger group sizes.
  • Sleeper cell exclusion: The study intentionally excluded the "sleeper cell" participant due to their passive behavior. Future research might investigate how such passive participants influence group dynamics and performance prediction.
  • Deeper interaction analysis: The authors aim to investigate the complex interaction effects between expressed behaviors and group-level personality traits to provide insights into strategies for effective communication.

Summary updated

BibTeX
@inproceedings{zhong2019predicting,
  author = {Shun-Chang Zhong and Yun-Shao Lin and Chun-Min Chang and Yi-Ching Liu and Chi-Chun Lee},
  title = {{Predicting Group Performances Using a Personality Composite-Network Architecture During Collaborative Task}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure