🍽 NYC Dining Safety Prediction (Data & ML System)

A data-driven machine learning project that predicts NYC restaurant health inspection grades (A / B / C) using public city data, demographic indicators, and historical inspection records.

The goal is to identify high-risk restaurants before inspections occur, enabling more proactive public health interventions. Innocube screenshot

🎯 Project Overview

This project builds an end-to-end data science pipeline to predict restaurant inspection outcomes based on a combination of:

  • Restaurant inspection history
  • Neighborhood-level socioeconomic indicators
  • NYC 311 complaint patterns
  • Geographic and demographic context

The system is designed not just for prediction accuracy, but for policy-relevant decision making — specifically maximizing recall for Grade C (high-risk) restaurants, where missed cases are costly.

🎥 Demo & Presentation

  • ▶️ YouTube Presentation

🔧 Tech Stack

  • Python
  • pandas, NumPy
  • scikit-learn
  • XGBoost
  • SQL-style feature aggregation
  • Data visualization (matplotlib / seaborn)

📊 Data Sources

  • DOHMH NYC Restaurant Inspection Results
    Historical inspection grades and violation records

  • NYC 311 Service Requests
    Non-emergency complaints (noise, sanitation, heat/hot water, etc.)

  • NYC Census Data
    Population demographics and neighborhood-level economic indicators

These datasets are joined and aggregated at the restaurant and geographic level to construct predictive features.

🧪 Feature Engineering & Preprocessing

Feature Selection

  • Correlation analysis to identify redundant signals
  • Mutual Information (MI) used to measure feature relevance
  • Final numeric features include:
    • complaints_per_capita
    • income
    • total_population
    • census_tract
    • zipcode

Preprocessing Pipeline

  • Train / test split (~97k training, ~24k testing)
  • Standard scaling for numeric features
  • One-hot encoding for categorical variables
  • Label encoding for targets (A / B / C → 0 / 1 / 2)

⚖️ Handling Class Imbalance

The dataset exhibits strong class imbalance, with Grade C restaurants underrepresented.

To address this:

  • SMOTE (Synthetic Minority Oversampling Technique) is applied
  • Balances class distribution during training
  • Enables models to better learn high-risk patterns

🤖 Modeling Strategy

Models evaluated include:

  • Logistic Regression (baseline)
  • Random Forest (tuned)
  • Gradient Boosting (XGBoost, GBC)
  • Neural Network (MLPClassifier)

Evaluation Focus

  • Metrics emphasize recall for Grade C
  • Goal: catch as many risky restaurants as possible, even at some precision cost

📈 Results Summary

  • Tuned Random Forest outperformed other models
  • Achieved approximately 80% recall for Grade C restaurants
  • Consistently stronger than Logistic Regression, XGBoost, and Neural Networks
  • Demonstrated robustness across macro F1 and recall-oriented metrics

🧠 Insights & Interpretation

Model interpretation reveals that neighborhood-level factors dominate prediction power:

  • Income
  • Complaints per capita
  • Census tract
  • Total population

This suggests that sanitation risk is not random, but clusters geographically and socioeconomically — an important insight for public health planning.

🚀 Future Directions

  • Time-aware modeling to capture temporal trends
  • NLP on violation text and 311 complaint descriptions
  • Improved geographic modeling
  • Probability calibration for policy use
  • Integration into operational inspection workflows