CoSSA: Correlation-Structure Shift Adapter for Cross-City Urban Forecasting

Urban forecasting rarely transfers across cities because sensor IDs, layouts, and metadata seldom align. Ontology mapping is brittle and does not scale. We present CoSSA, a lightweight adapter that transfers models by aligning latent correlation structure, without ontology or node alignment. CoSSA uses a Temporal CNN with a dynamic similarity graph and a Similarity-Structure Matching (SSM) loss to match pairwise correlation geometry between source and target latent states using unlabeled target data. This ontology-free criterion preserves relations (who moves with whom) rather than identities. On METR-LA (N =207) → PEMS-BAY (N =325), CoSSA improves over a source-only baseline by ≈8.2% MAE and ≈6.5% RMSE on held-out target tests, while remaining simple and scalable. The method is few-shot ready and robust to schema mismatch.