Time-Series Forecasting
Why text helps a forecaster — and where it hurts. Multimodal modeling over numeric series plus repo / document text, with closed decomposition and ablations that show which input actually moved the number.
TIME-SERIES · AGENT SYSTEMS · OPEN SOURCE
@LuYuan-Zjyy · github.com/LuYuan-Zjyy
Time-Series Forecasting
Two things I care about: time-series that actually models its inputs, and agents that get real work done. Text should make a forecast explainable, not just better.
Why text helps a forecaster — and where it hurts. Multimodal modeling over numeric series plus repo / document text, with closed decomposition and ablations that show which input actually moved the number.
Agents that finish the job: orchestrator + specialist workers, evidence-grounded reranking instead of free-form generation, RAG and tool-calling, shipped as containerized services.
Cold-start recommendation with multi-prototype + uncertainty transfer — cutting negative transfer while keeping the signal.
Multimodal forecasting over 16 OpenDigger metrics + repo text, 32-month horizon, with an LLM explainer layer.
Cloud-native fact-consistency agent for long collaborative docs: extract → cross-check → trace the source, with image-text comparison and a containerized deploy.
Image-text sentiment in 3 classes: dynamic-weighted IIF fuses modalities per sample, sentiment-prototype SIE prunes noise.
Multi-prototype + uncertainty bridge for cold-start cross-domain recommendation; negative transfer cut to 1/3 across 4 Amazon domains.
Studying when repository text actually improves a numeric forecast, and how to attribute the gain to a specific input instead of a black-box score.
84.7% acc / 84.0% F1, +3.5pt over the strongest baseline, with full ablation and 11 visualization scripts.
Dual-branch Conformer with envelope + resonance spectral fusion; lower EER on CFAD.
13 public repositories — dataset pipeline, model, service and evaluation harness, all readable end to end.