ASIND algorithm predicts network dynamics without prior knowledge

A team led by Mingyu Kang introduced ASIND (Alternating Sparse Identification of Network Dynamics) in a paper accepted at the IFAC World Congress 2026. The algorithm tackles the challenge of understanding how complex social systems evolve over time—like predicting information spread or opinion shifts—without needing any prior knowledge about the system's internal rules or connections. Traditional methods either require detailed knowledge (self-dynamics, interaction functions, and network structure) to sparsely identify dynamics, or use simple universal approximators that are black-box and suffer from vast search spaces. ASIND bridges this gap by alternately sparsely identifying the self-dynamics function, interactive function, and interactive network, making the process interpretable and efficient.

In extensive experiments, ASIND demonstrated state-of-the-art performance for both identification and 100-step prediction compared to baseline methods. The researchers also revealed a key insight: the interactive network itself is weakly identifiable—different network structures can produce nearly identical trajectory dynamics. This has implications for real-world applications where we may infer dynamics correctly but remain uncertain about the underlying network. The code is publicly available, encouraging further research and application in fields like epidemiology, social networks, and power grids.