Skip to content

Models.md

Latest commit

 

History

History
37 lines (31 loc) · 2.17 KB

Models.md

File metadata and controls

37 lines (31 loc) · 2.17 KB
Raw

Machine Learning Models

The UIDAI Risk Engine employs four machine learning and mathematical models to compute component anomaly scores, which are then fused into a single risk score.

1. Isolation Forest (Behavioral Engine)

  • Library: scikit-learn
  • Objective: Identifies multi-dimensional out-of-distribution updates.
  • Inputs: Scaled features (bio_demo_ratio, rolling_mean, rolling_std, age_transition_skew).
  • Score Mapping: (1.0 - predictions) / 2.0 (maps anomalies to 1 and normal records to 0).

2. Dense Autoencoder (Behavioral Skew Engine)

  • Library: Keras / TensorFlow
  • Architecture: A fully connected feed-forward autoencoder (Input(4) -> Dense(8, relu) -> Dense(4, linear)).
  • Objective: Learns dependencies between features and flags structural discrepancies (e.g. unexpected age transitions).
  • Score Mapping: Mean squared reconstruction error, min-max normalized.

3. LSTM Network (Temporal Trend Engine)

  • Library: Keras / TensorFlow
  • Architecture: Sequential LSTM network (Input(14, 1) -> LSTM(8) -> Dense(1)).
  • Objective: Fits on sequence histories of length 14 days for bio_demo_ratio per pincode. Predicts the next ratio and compares the prediction with the actual value.
  • Score Mapping: Absolute error of prediction vs actual value, min-max normalized.

4. Nearest Neighbors (Spatial Clustering Engine)

  • Library: scikit-learn
  • Objective: Evaluates geographical clusters of updates.
  • Algorithm: For each date, groups records and fits a 1D NearestNeighbors model on bio_demo_ratio (k=5). Computes the absolute deviation of a pincode's ratio from the local neighborhood mean.
  • Score Mapping: Absolute deviation, min-max normalized.

5. Unified Risk Fusion

The components are aggregated using a weighted average: $$RiskScore = w_1 \times IF + w_2 \times AE + w_3 \times LSTM + w_4 \times Spatial$$ Defaults weights are 0.3, 0.3, 0.2, 0.2 respectively. Risk classification levels:

  • Normal: $<0.4$
  • Monitor: $0.4 - 0.7$
  • High Risk: $\geq0.7$
  • Early Warning: Assigned if risk trend rises by $>0.15$ over 7 days while the current risk remains under the High Risk threshold.