Sequential decomposition and attribution for trajectory forecasting
A model-agnostic protocol that decomposes trajectory forecasting inputs into Past, Current, and Prior to quantify each one's standalone and marginal contribution.
Enhanced trajectory reconstruction from sparse and noisy GPS data: A progressive chunked transformer approach
ProChunkFormer, a progressive chunked transformer that reconstructs trajectories from sparse, noisy GPS data, matching an autoregressive transformer's accuracy while running …
Cross-Taxonomy Label Alignment in Transportation Networks With Prompt-Driven Large Language Models
LLM prompt-based pseudo-labels plus road-network structural embeddings align labels across transportation taxonomies, even with only 1% labeled data.
A Generalizable Semantic Graph Learning Framework for Road Network Representation
A heterogeneous graph framework (HetSemNet) that fuses LLM-derived POI semantics with road topology, generalizing across cities on road type classification.