Dipartimento di Informatica

Descrizione del progetto

What makes a team effective has been widely studied across disciplines and applicative contexts. Several factors such as communication, coordination, distinctive roles, interdependent tasks, shared norms, personality traits, and diversity have been found to be relevant aspects.
Yet, our understanding of teams as evolving systems of interacting individuals as well as of the relation between team composition and performance is still partial.The heterogeneity of interaction patterns in co-present, virtual or distributed teams and the widespread adoption of multi-team and multilevel systems in which nested groups cooperate to perform functions at different scales, add additional challenges to capture the complexity of modern organizations. In contrast, the large majority of the relevant literature mainly dealt with self-reported surveys, in-depth interviews, or observations in scenarios with limited complexity, scale, scope, and general applicability. Recently, the access to digital traces on human activities is providing unprecedented opportunities to mine the emergent behavior of large-scale, loosely coupled, teams and task-oriented multilevel groups.
In this context, machine learning pipelines were proposed to predict team performance that however was mainly:
i) focused on specific aspects while overlooking the holistic picture underpinning team dynamics
ii) used a static snapshot to characterize composition and interactions and
iii) adopted hand-designed features that often fail to capture the complexity of high-order functional relationships.
To tackle these issues, this project aims to provide a novel methodological framework – called Dteam2vect – to predict team performance leveraging representational learning on temporal graphs. Our approach is based on the hypothesis that graphs embeddings offer a natural way to infer high-order representations of interacting systems, and state-of-the-art embedding methods can be extended to account for the temporal dynamics introduced by teams that continuously form, evolve, and grow over time.
Since representational learning comes with a well-known trade-off between accuracy and interpretability of the prediction outcome, our approach aims at carefully exploring the dimensions of explainability and underlying biases, to ensure that results are not a mere effect of the statistical tendencies of benchmarks.

Note