Univariate Channel Fusion for Multivariate Time Series Classification

The core insight here is not just efficiency: treating channels independently before fusion sidesteps the curse of dimensionality that plagues joint multivariate models, where adding sensors multiplies the feature space rather than enriching it. The three fusion strategies (mean, median, dynamic time warping) represent meaningfully different assumptions about channel relationships, and the choice between them is a modeling decision, not a tuning knob.

The compression-through-aggregation logic here rhymes with what appeared in the K-Token Merging paper from April 16, where consecutive token embeddings are collapsed into single representations to reduce sequence length before model processing. Both papers are working on the same underlying problem: reducing the dimensionality of sequential data at ingestion rather than at the model architecture level. That said, the deployment context differs sharply. K-Token Merging targets LLM inference on server hardware, while Univariate Channel Fusion is explicitly aimed at IoT and wearable devices where memory and compute ceilings are hard constraints, not soft optimization targets.

The real test is whether dynamic time warping fusion holds its accuracy advantage over mean or median fusion on datasets with highly asynchronous or phase-shifted channels, since that is the scenario where the added compute cost of DTW would actually justify itself on constrained hardware.