Use custom CSV or Parquet data¶

Use this guide when you already have sequence data and want to run the app on it.

Required shape¶

Input file must be long format with one row per (entity, timestep).

You need:

• one entity identifier column
• one time-order column
• one binary target column (0 or 1)
• one or more numeric feature columns

Steps¶

1. Start the app:

uv run python -m streamlit run timeshap_app.py

1. In sidebar Data:
2. set Dataset to Upload file

3. upload .csv or .parquet

4. Map columns:

5. select Entity column
6. select Time column

7. select Target column

8. Verify feature detection:

9. app auto-detects features as all non-selected columns

10. if no features are detected, adjust column selections

11. Set model options in Model:

12. choose Sequence length small enough that entities have enough timesteps
13. choose model source (Train in app or checkpoint load)

Common pitfalls¶

• No windows could be created:
• sequence length is too large, or
• entities are too short
• bad time sorting:
• ensure the time column sorts in real sequence order
• non-numeric feature columns:
• convert features to numeric before upload