Research map / Probabilistic and scientific modeling / Molecules and proteins

Explainable Origin-Destination Crowd Flow Interpolation via Variational Multi-Modal Recurrent Graph Auto-Encoder

AAAI 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Flow matching generative modeling · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UApex, an uncertainty-aware OD crowd flow interpolation framework with a variational multi-modal recurrent graph auto-encoder (VMR-GAE) and a Shapley-value-based spatiotemporal explainer.

Problem

Interpolating missing values in sparse, noisy origin-destination (OD) crowd flow matrices reliably and with explanations.

Approach

Formulates OD flow as a dynamic directed graph and combines a VGRNN-style variational recurrent diffusion-GCN encoder with distribution alignment of context-modality representations to the prior, a Poisson-likelihood MLP edge decoder, and a Shapley-value explainer over spatiotemporal subgraphs with an efficient search pruned by urban geography.

Key results

• On Beijing weekday rush hour, VMR-GAE reaches MAE 27.400 vs 29.450 for DeepEnsembles (about 6.96% MAE reduction), with MAPE 0.674 vs 0.782.
• On NYC holiday rush hour, VMR-GAE reaches MAE 6.932 vs 7.382 for DeepEnsembles (about 6.1% reduction), with MAPE 0.548 vs 0.625.
• Ablation on RMSE: removing distribution alignment raises Beijing RMSE from 39.437 to 42.308, and replacing the deep decoder with an inner product decoder raises it to 45.425.

Measured effects reported in the paper

Surprising findings

• Uncertainty quantification trades off RMSE: VMR-GAE improves MAE and MAPE consistently but only marginally or not at all on RMSE in some settings (e.g., NYC weekday rush hour RMSE slightly worse than DeepEnsembles).

Limitations

• Exact Shapley computation is O(n^2 * 2^n), requiring heuristic pruning and Monte Carlo approximation for the explainer.
• Uncertainty quantification can sacrifice RMSE improvement because it produces conservative outcomes.
• Evaluated only on two datasets from two cities.

Experimental setup

Datasets: Beijing OD crowd flow (anonymized navigation app) NYC taxi OD data (NYC TLC), with Green Taxi as context

Benchmarks: Beijing OD flow dataset NYC taxi OD dataset

Metrics: MAE RMSE MAPE

Baselines: SVD GAE 2D-GCN NGCF T-NGCF VGAE VDGNN DropOut BNN DeepEnsembles

Builds on: VGRNN diffusion convolution recurrent network variational graph auto-encoder SubgraphX

Closest papers by meaning

• Modeling Heterogeneous Relations across Multiple Modes for Potential Crowd Flow Prediction (AAAI 2021 · similarity 0.87)
• Generic and Dynamic Graph Representation Learning for Crowd Flow Modeling (AAAI 2023 · similarity 0.87)
• Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.84)
• Satellites Reveal Mobility: A Commuting Origin-destination Flow Generator for Global Cities (NeurIPS 2025 · similarity 0.83)
• Social Physics Informed Diffusion Model for Crowd Simulation (AAAI 2024 · similarity 0.83)
• SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark (NeurIPS 2026 · similarity 0.83)