Revisiting the Modifiable Areal Unit Problem in Deep Traffic Prediction with Visual Analytics

Deep learning methods are being increasingly used for urban traffic prediction where spatiotemporal data is aggregated into sequentially organized matrices that then fed convolution-based residual neural networks. However, the widely known modifiable areal unit problem within such aggregation processes can lead to perturbations in network inputs. This issue significantly destabilize feature embeddings and predictions - rendering deep networks much less useful experts. paper approaches this challenge by leveraging visualization techniques enable investigation of many-to-many relationships between dynamically varied multi-scalar aggregations predictions. Through regular exchanges with a domain expert, we design develop visual analytics solution integrates 1) Bivariate Map equipped an advanced bivariate colormap simultaneously depict input errors across space, 2) Moran's I Scatterplot provides local indicators spatial association analysis, 3) Multi-scale Attribution View arranges non-linear dot plots tree layout promote model analysis comparison scales. We evaluate our approach through series case studies involving real-world dataset Shenzhen taxi trips, interviews observe geographical scale variations have important impact on performances, interactive exploration varying inputs outputs benefit experts development models.