[2401.14973] Discovering group dynamics in synchronous time collection by way of hierarchical recurrent switching-state fashions

[Submitted on 26 Jan 2024]

Obtain a PDF of the paper titled Discovering group dynamics in synchronous time collection by way of hierarchical recurrent switching-state fashions, by Michael Wojnowicz and eight different authors

Obtain PDF

Summary:We search to mannequin a set of time collection arising from a number of entities interacting over the identical time interval. Latest work centered on modeling particular person time collection is insufficient for our meant functions, the place collective system-level habits influences the trajectories of particular person entities. To deal with such issues, we current a brand new hierarchical switching-state mannequin that may be educated in an unsupervised vogue to concurrently clarify each system-level and individual-level dynamics. We make use of a latent system-level discrete state Markov chain that drives latent entity-level chains which in flip govern the dynamics of every noticed time collection. Suggestions from the observations to the chains at each the entity and system ranges improves flexibility by way of context-dependent state transitions. Our hierarchical switching recurrent dynamical fashions may be realized by way of closed-form variational coordinate ascent updates to all latent chains that scale linearly within the variety of particular person time collection. That is asymptotically no extra pricey than becoming separate fashions for every entity. Experiments on artificial and actual datasets present that our mannequin can produce higher forecasts of future entity habits than present strategies. Furthermore, the supply of latent state chains at each the entity and system degree allows interpretation of group dynamics.