First local explanation run¶

This tutorial walks through the first successful end-to-end run using demo data.

Outcome¶

At the end you will:

• train a GRU model in-app,
• inspect one validation window and its hidden activations,
• run TimeSHAP and read pruning, event, and feature attributions.

Prerequisites¶

• Python environment set up for this repo
• Dependencies installed (uv sync or equivalent)

Step 1: Start the app¶

uv run python -m streamlit run timeshap_app.py

Step 2: Keep demo data selected¶

In the sidebar:

• Dataset: choose Demo synthetic data
• Keep default columns:
• Entity column: entity_id
• Time column: time_idx
• Target column: target

Step 3: Train the baseline model¶

In Model:

• Model source: Train in app
• Keep defaults:
• Sequence length = 18
• GRU hidden size = 32
• Epochs = 8
• Batch size = 64
• Learning rate = 0.005
• Click Train / Retrain

Wait for training to finish, then verify:

• Model Diagnostics shows training and validation curves
• Sequence Inspector shows predicted probability and hidden-state heatmap

Step 4: Select a validation sample¶

• Set Validation sample to any index
• Confirm Predicted probability, Predicted class, and True class are displayed
• Review the raw window table under the heatmap

Step 5: Run TimeSHAP¶

In TimeSHAP Local Explanations:

• Keep defaults:
• Pruning tolerance (tol) = 0.03
• Kernel SHAP samples = 200
• Click Run TimeSHAP

Expected outputs:

• Pruning Output table/object
• Event-Level (Timestep) Attributions chart and table
• Feature-Level Attributions chart and table

Next steps¶

• Try How-to guides -> Use custom CSV or Parquet data
• Try How-to guides -> Load a Lightning checkpoint