Cur Research

Forecasting Complex Time Series Dynamics with Generative Multimodal Models

My research focuses on generative frameworks for multimodal and single modal time series forecasting, unifying diverse streams, numerical counts, policy or textual annotations, graph‐structured mobility data, and sensor readings into coherent, uncertainty‐aware predictions. The work I do applies analogous principles to purely temporal and structured modalities.

1. Multimodal Fusion

  • Numerical Data: daily case counts, hospital occupancy, energy usage
  • Textual Signals: policy announcements, public‐health advisories, news sentiment
  • Graph Inputs: aggregated mobility/contact networks, supply‐chain topologies
  • Shared Latent Space: modality‐specific encoders feed into a unified representation, with cross‐attention mechanisms inspired by transformer architectures

“By treating each data modality as a sequence of tokens, we leverage attention to learn interactions across streams—much like in vision transformers, but wholly within the time‑series domain.”

2. Generative Forecasting & Uncertainty

  • Diffusion‐Style Sampling: a learnable perturb‐and‐denoise process that yields full trajectory ensembles
  • Quantile Conditioning: guide sampling toward specified quantiles (e.g., 95th percentile demand) to assess tail‑risk scenarios. This is similar to the way classifier free guidance work in diffusion model.
  • Flow‐Based Refinement: apply continuous normalizing flows to sharpen distributional outputs and correct sampling biases

“Our hybrid diffusion–flow pipeline produces well‐calibrated uncertainty bands, that offers decision‑makers clear confidence intervals instead of single‑point estimates.”

4. Evaluation & Case Studies

  • Generic Time Series Benchmarks:
    • Evaluate on standard forecasting datasets (M‑3, M‑4, NN5) covering industry, finance, and demographics.
    • Compare point‑forecast accuracy (sMAPE, MASE) and probabilistic metrics (CRPS, interval coverage) against classic and state‑of‑the‑art baselines (ARIMA, ETS, Prophet).
  • Epidemic Forecasting:
    • COVID‑19 and influenza‑like‑illness (ILI) weekly forecasts, benchmarked against CDC‑aggregated models with aligned dates and metrics (sMAPE, CRPS, coverage).
      Each study measures both point‐error (sMAPE, RMSE) and distributional accuracy (CRPS, quantile coverage).

5. Future Directions

Interpretability & Explainability

  • Attention Maps: visualize which modalities (e.g., mobility vs. text signals) drive forecast changes at each timestep

  • Local Attribution: adapt SHAP‐style methods to our generative paths, highlight key inputs behind any given prediction

“These explainability tools ensure transparency and build trust, critical when forecasts guide high‑stakes resource allocation.”

  • Dynamic Adaptation: implement online learning to update models as new data arrive.
  • Meta‐Learning Extensions: enable rapid adaptation to new domains (e.g., emerging pathogens).
  • Open‑Source Toolkit: package our multimodal encoders, generative samplers, and explainability modules—so that colleagues specializing in generative vision or interpretability methods can readily contribute their expertise.