This research addresses the challenge of modeling temporal events at multiple resolutions by proposing HiCMT, a hierarchical framework that employs multi-scale temporal decomposition and learnable cross-resolution interaction modules for information exchange between different temporal granularities.
Key findings
HiCMT employs a learnable decomposition strategy for adaptive segmentation of event sequences into hierarchical temporal scales.
Introduces a bidirectional cross-attention module for information exchange between adjacent resolution levels.
Develops a fusion mechanism that dynamically weights contributions from different resolutions based on their relevance to the prediction task.
Limitations & open questions
The proposed method's performance on diverse temporal event datasets is yet to be empirically validated.
The architecture's scalability and efficiency for very large datasets remain to be tested.