| --- |
| license: mit |
| tags: |
| - tabular-regression |
| - tabular-classification |
| - spatiotemporal |
| - urban-mobility |
| - traffic |
| - onnx |
| library_name: sklearn |
| --- |
| |
| # MargaDrishti — Models |
|
|
| Trained artefacts for **[MargaDrishti](https://github.com/adarshcod30/MargaDrishti)**, |
| a spatio-temporal ML system over Bengaluru parking-violation and traffic-event |
| data (Nov 2023 – Apr 2024). |
|
|
| Live dashboard: **[adarshcod30/margadrishti](https://huggingface.co/spaces/adarshcod30/margadrishti)** |
|
|
| ## Read this before using the numbers |
|
|
| **These models rank predicted _enforcement demand_, not where illegal parking |
| occurs.** `corr(log patrol_hours, log total_captures) = 0.96` across all 2,534 |
| res-9 cells — about 93% of the |
| variance in per-cell violation counts is explained by how much patrolling |
| happened there. The project's own |
| [enforcement-bias audit](https://github.com/adarshcod30/MargaDrishti/blob/main/reports/fairness_audit.md) |
| returns **HIGH concern** on this system, and that is published rather than |
| buried: no Getis-Ord hotspot in the data survives exposure adjustment. |
|
|
| Two targets are **not** what a leaderboard row suggests: |
|
|
| - **B3 (priority) is a recovered rule, not a prediction.** Corridor status |
| predicts High priority at 99.84% (13 exceptions in 8,173 events). Scores near |
| 1.0 reflect that rule. |
| - **A3 (capture rejection) is not yet answerable.** The review process changed |
| regime mid-window — February 2024 rejection jumped to 0.790 from ~0.28 while |
| volume collapsed sixfold — so every family is miscalibrated. |
|
|
| ## What's here |
|
|
| | File | What it is | |
| |---|---| |
| | `onnx/b1_clearance_duration.onnx` | Event clearance duration (log1p minutes), RandomForest | |
| | `onnx/b2_road_closure.onnx` | Road-closure probability, RandomForest | |
| | `onnx/*.features.json` | **Exact input feature order** — ONNX takes a bare float array with no column names | |
| | `leaderboard.csv` | Module A: every family × A1/A2/A3 | |
| | `leaderboard_module_b.csv` | Module B: every family × B1–B4 | |
| | `model_cards/` | One card per target, generated from the leaderboards | |
| | `reports/` | Spatial statistics, CIS validation, enrichment ablation, fairness audit | |
|
|
| ### Why RandomForest and not the leaderboard winner |
|
|
| The browser ships RandomForest because **onnxmltools misroutes ~1% of LightGBM |
| rows at split boundaries** — median divergence is float32 noise (4.3e-07) but 5 |
| of 500 sample rows diverged by up to 0.09. skl2onnx converts RandomForest |
| exactly (max 1.0e-06; the classifier is bit-exact). |
|
|
| That trade is only defensible because the Module B leaderboard already put |
| random_forest at 0.5034 against xgboost's 0.4990 — a gap far inside the ±0.13 |
| fold standard deviation. Exact client/server agreement costs no measurable |
| accuracy. |
| |
| ## Headline results |
| |
| | Target | Best family | Metric | vs strongest baseline | |
| |---|---|---|---| |
| | A1 · parking intensity | xgboost | 0.4999 Poisson deviance | **+43.0%** | |
| | A2 · hotspot | random forest | 0.1365 PR-AUC | 46.9× base rate (0.291%) | |
| | A3 · capture rejection | xgboost | 0.3088 PR-AUC | *not answerable* | |
| | B1 · clearance duration | xgboost | 0.4990 ± 0.128 | +27.7% | |
| | B2 · road closure | catboost | 0.3615 ± 0.071 PR-AUC | 4.1× (8.7% prevalence) | |
| | B4 · cause from text | char n-gram + logreg | 0.478 macro-F1 | 10.8× | |
| |
| B4's notable result is equity rather than accuracy: **English 0.483 vs Kannada |
| 0.472**, essentially no language gap on a field that mixes both scripts. |
| |
| ## Reproduce |
| |
| ```bash |
| git clone https://github.com/adarshcod30/MargaDrishti |
| cd MargaDrishti && make data && make reproduce |
| ``` |
| |
| Seed 42. Splits are chronological, never random. |
| |
| ## Licence |
| |
| MIT. Source data are anonymised public releases and remain subject to their |
| original terms. |
| |
