dispatch-desk-risk / README.md
saiyko's picture
Upload 3 files
5f8e8b7 verified
|
Raw History Blame Contribute Delete
3.43 kB
---
library_name: sklearn
license: mit
tags:
- onnx
- tabular-classification
- sklearn
- education
- game
pipeline_tag: tabular-classification
---
# Dispatch Desk · costly-decision risk
Predicts, when a crisis scene opens, the probability that this player will pick the costly option, from how they played so far. The game uses it to offer extra support.
## Intended use
- **Use:** inside the Dispatch Desk teaching game, to decide whether to show a support card before a crisis decision.
- **Not for:** grading students, judging real dispatchers, or any decision about a person outside the game.
- **Users:** students playing the game in a browser; the model runs on their device (onnxruntime-web).
## Model
- scikit-learn `Pipeline(StandardScaler -> LogisticRegression)`, seed 13
- Trained with Python 3.13.17, scikit-learn 1.9.1
- Exported to ONNX (opset 17, zipmap off). Input `input` float32 `[n, 8]` in the order below, output `probabilities` `[n, 2]` (column 1 = costly).
- Decision threshold tau = 0.5 (fixed in the spec, not tuned on the test split)
## Features
| name | meaning | example |
|---|---|---|
| `seviye` | Level (0 Just starting, 1 Know a bit, 2 Experienced) | 1.0 |
| `sahne_no` | Crisis number (1-3) | 2.0 |
| `canli_bekleme_ort_dk` | Average call wait in the previous live phase (sim. minutes) | 3.11 |
| `canli_en_yakin_orani` | Share of assignments to the closest crew (0-1) | 0.6 |
| `canli_bekleyen` | Calls still waiting at the alarm | 1.0 |
| `gorev_yanlis` | Field tasks answered wrongly so far | 1.0 |
| `onceki_maliyetli` | Costly choices in earlier crises (0-2) | 0.0 |
| `onceki_karar_suresi_sn` | Seconds spent on the previous crisis (0 for Crisis 1) | 3.5 |
## Coefficients (standardized features)
| feature | weight |
|---|---|
| `seviye` | -0.146 |
| `sahne_no` | -0.138 |
| `canli_bekleme_ort_dk` | +0.754 |
| `canli_en_yakin_orani` | +0.198 |
| `canli_bekleyen` | +0.294 |
| `gorev_yanlis` | +0.563 |
| `onceki_maliyetli` | +0.452 |
| `onceki_karar_suresi_sn` | -0.710 |
| `intercept` | +0.491 |
## Training data
`data/oturumlar.csv` from the GitHub repository: 648 rows, 16 real sessions (48 rows) and 600 synthetic rows from a seeded rule (`data/generate_synthetic.py`, seed 13). No personal data: the session ID is random and no name, location or device field is collected.
## Evaluation (stratified 30 % hold-out, seed 13)
Acceptance gate: ROC AUC >= 0.7 (committed in `spec.md` before training).
| metric | value |
|---|---|
| roc_auc | 0.761 |
| accuracy | 0.7179 |
| f1 | 0.7737 |
| brier | 0.1926 |
| roc_auc_real_only | 0.2188 |
| n_test_real | 16 |
| n_rows | 648 |
| n_rows_real | 48 |
| n_sessions_real | 16 |
| positive_rate | 0.5802 |
| onnx_max_abs_diff | 8.49e-08 |
## Limitations and risks
- Most rows are synthetic, so the weights partly reflect the generator's rule. The real-only AUC above is the honest number, and it rests on few sessions.
- Real players were friends of the author, which is a narrow and self-selected group.
- Players who know the model exists could play differently. The model only adds support; it never removes help or changes the outcome.
- The weights are public in this file, so anyone can read them. That is fine here: nothing about the model is secret (see `RELEASE.md`).
## Use
```
https://huggingface.co/<user>/<repo>/resolve/<commit>/model.onnx
```
Always pin a commit, never `main`, so a new upload cannot change a running app.