YongboYu/pmf-tsfm-demo

Process Model Forecasting Explorer

An explorer for the CAiSE 2026 paper Process Model Forecasting with Time Series Foundation Models (arXiv:2512.07624). It makes the paper's result tangible: off-the-shelf Time-Series Foundation Models (Chronos-2, Moirai-2.0, TimesFM-2.5) can forecast how the directly-follows (DF) relations of a process evolve over time.

What you can do

Bundled explorer. Pick one of the four bundled event logs (bpi2017, bpi2019_1, sepsis, hospital_billing) and a model. You see:

• —the forecast DFG (predicted next week of process behaviour) beside the actual-future DFG (what really happened) — the bundled path is a holdout backtest: the real last week is held out and forecast from the rest, so genuine ground truth exists;
• —a Side-by-side / Diff toggle, with the Diff view offering Absolute (forecast | actual counts) and Relative (signed change %) labelling;
• —an ER / MAE / RMSE accuracy strip, with the truth-DFG ER baseline for context.

Live upload (your log). Upload a custom XES log and forecast its genuine next week (forecast origin = the log end) on ZeroGPU. Because an upload has no future ground truth, this tab reports drift — the DF relations the forecast adds or drops vs the last-known window — and never an accuracy metric. Live forecasts run with Chronos-2 (Moirai-2 and TimesFM-2.5 are available in the Bundled explorer tab); oversize logs are rejected so a call stays under the GPU time limit.

How it runs

The bundled path is served from precomputed, committed assets (incl. pre-rendered SVGs) — so it is instant, infinitely concurrent, needs no GPU, and cannot be DoS'd into a bill. The live path runs preprocessing + a Chronos-2 forecast under @spaces.GPU on ZeroGPU and renders its DFGs at request time (so the Space carries graphviz + the model libs; see packages.txt / requirements.txt).

Run it locally:

uv run --with gradio python app.py

Self-host & agent API

This folder is the hosted HF Space GUI only. To run the core forecasting pipeline on your own data — a self-host Docker image (GUI + Hydra CLIs, no caps) or an MCP agent server — use the dedicated docker/ and mcp/ artifacts at the repo root, which wrap pmf_tsfm.api directly (ADR-0008). See docker/README.md and mcp/README.md.