Hi everyone,
I need some advice/guidance on graph based neural architectures(non graph approaches are also welcome) for the following problem.
I’m working with neural recording data (specifically using Neuropixels probes), but I think my question could apply broadly to cases where multiple time series are recorded from spatially-distributed points with known spatial relationships.
I have time series data (electrophysiological recordings) from multiple recording sites distributed across a standardized spatial volume — in my case, the mouse brain.
This brain volume is hierarchically subdivided into anatomical regions. For example:
The top-level node is “root”.
Under root are major regions like Cortex, Thalamus, etc.
These are further subdivided, e.g. Cortex → Motor Cortex, Auditory Cortex, etc.
Each recording site is located at a known spatial point within this hierarchy.
I want to predict the region (leaf node in the anatomical hierarchy) corresponding to each recording site, based on the time series data.
Currently, I extract features from each site independently and train a classifier (e.g., XGBoost) to predict the region. But this completely ignores two important aspects:
- The anatomical hierarchy – some regions are subregions of others.
- Spatial consistency – if two nearby recording sites are known to be in the same region, this imposes constraints on their labels.
I think a Graph Neural Network (GNN) could help here, by incorporating both the spatial relationships between recording sites and the anatomical hierarchy as priors. Has anyone worked on something similar, or can point me to relevant GNN models, papers, or codebases that handle structured prediction with hierarchical labels and spatial dependencies?
Would really appreciate any leads or ideas!