Human learning of hierarchical graphs.

Area of Science:

• Cognitive science
• Network science
• Machine learning

Background:

• Real-world networks often exhibit hierarchical structures.
• Human ability to learn complex hierarchical graph topologies is not well understood.
• Understanding how humans learn sequential event probabilities is crucial.

Purpose of the Study:

• To investigate human learning of hierarchical graph structures.
• To determine if humans can learn transition probabilities at different hierarchical levels.
• To explore potential trade-offs in learning hierarchical graph structures.

Main Methods:

• Utilized the surprisal effect (slower reaction to unexpected events) to probe mental estimates of transition probabilities.
• Employed mean-field predictions and numerical simulations.
• Conducted a serial response experiment with 100 human participants.

Main Results:

• Surprisal effects were stronger for finer-level hierarchical transitions compared to coarser levels.
• A surprisal effect was detected at the finer level but not the coarser level in human participants.
• Evidence suggests a trade-off exists where better learning at one hierarchical level may impair learning at another.

Conclusions:

• Humans can learn finer-level hierarchical structures in event sequences, but coarser-level learning is more difficult.
• Learning efficiency may be constrained by a trade-off between different hierarchical levels.
• This research provides insights into human graph learning and suggests future directions for neuroscience and behavioral studies.