stanley-jeffrey-attributesConflation

Overview

Our project aims to develop an automated decision logic for attribute conflation, selecting the most accurate and up-to-date attributes when multiple data sources represent the same real-world place.
This process supports the broader Overture initiative to unify place data from multiple providers into clean, high-quality records.

---

Check Hybrid model: https://github.com/project-terraforma/stanley-jeffrey-attributesConflation/tree/after-normalization1

Repo Structure

.
├── scripts/
│   ├── rule_based_selectionV1.py
│   ├── ML_based_selectionV1.py
│   ├── place_id_matches.py
│   ├── generate_ground_truth_dataset.py
│   ├── feature_generator.py
│   ├── ML_train.py
│   ├── ML_infer.py
│   ├── ML_eval.py
│   └── extract_*.sql
├── src/
│   ├── data_preprocessing/
│   ├── matching/
│   ├── conflation/
│   └── utils/
├── data/
│   ├── raw/
│   ├── interim/
│   └── processed/
├── models/
├── notebooks/
├── tests/
├── requirements.txt
├── environment.yml
├── README.md
└── LICENSE

Summary of workflow:

• Raw data is in data/raw (original JSON/GeoJSON from sources).
• Normalization happens in src/data_preprocessing → produces cleaned datasets in data/interim.
• Triplet matching is performed by scripts in scripts/, producing yelp_triplet_matches.csv.
• Attribute conflation:
• Rule-based in rule_based_selectionV1.py, evaluated with rule_based_accuracy.py.
• ML-based in ML_train.py, ML_infer.py, ML_eval.py, producing ml_predictions.csv.
• SQL scripts in scripts/extract_data/ are used to pull city-specific datasets.
• Final processed datasets (ground truth, conflated, ML predictions) live in data/processed/.

Project Roadmap

Phase 1 – Data Preparation

• Gather data: Collect place data from multiple sources (Overture, Yelp, Google, etc.) to produce Raw JSON/CSV files.
• Normalization: Clean text, standardize casing, remove duplicates, and validate ZIP codes & coordinates to produce Clean DataFrames per source.

Phase 2 – Entity Matching / Place ID Assignment

• Blocking / Candidate generation: Use city, ZIP, or proximity to reduce comparisons and generate candidate record pairs.
• Matching logic: Use fuzzy name/address, category, and optional lat/lon proximity to assign a unified place_id.
• Evaluate matches: Spot-check and compute precision/recall against known matches, producing a Matched dataset with place_id.

Phase 3 – Attribute Conflation

• Group by place_id: Combine all source records for each place to create a Grouped dataset.
• Decide attribute logic: Implement rule-based or ML-based selection for name, address, etc., to define a unified "golden record".
• Implement conflation algorithm: Apply logic or a trained model to produce a Conflated dataset.
• Evaluate accuracy: Compare to Yelp ground truth to generate metrics and reports.

Phase 4 – Results & Reporting

• Performance evaluation: Quantify accuracy, precision, and recall of rule vs. ML selection.
• Recommendations: Suggest a scaling strategy and create a summary report or slide deck.
• Final deliverables: Clean labeled dataset, code notebooks, and an evaluation report for the full POC.

---

Challenges

• Data availability and coverage: Some datasets do not cover all geographic areas, limiting evaluation scope.
• Overpass API limitations: Encountered query timeouts for large OSM data batches; mitigated by narrowing geographic scopes and using Python tools.

---

OKRs

Objective 1: Build a high-quality labeled dataset for attribute conflation

Key Results:

• Collect and preprocess ≥2000 pre-matched place records from ≥3 distinct data sources.
• Label ≥90% (≥1800) of collected records with verified ground-truth attributes using Yelp Academic Dataset.
• Achieve ≥95% data consistency across the final labeled dataset via automated field validation.
• Ensure ≤2% missing or invalid attribute values (≤40 records) after final cleaning.
• Complete dataset preprocessing and labeling within 6 weeks of project start.

Objective 2: Develop and compare algorithms for optimal attribute selection

Key Results:

• Implement 2 algorithms: one rule-based and one machine learning model.
• Achieve ≥80% validation accuracy on attribute prediction using a 70/15/15 train/validation/test split.
• Identify and document the top 3 most impactful features or rules contributing to ≥60% of total model importance.
• Limit model inference time to ≤200 ms per record on 2000 records.
• Conduct ≥5 cross-validation folds for reliable performance estimation.

Objective 3: Evaluate performance and recommend a scalable solution

Key Results:

• Evaluate model performance using precision, recall, and F1-score on a test set of 1,500 records.
• Demonstrate ≥17% relative improvement in F1-score of the best-performing model compared to the baseline rule-based approach.
• Conduct 1 live presentation demonstrating algorithm performance and recommending next steps for Overture Places deployment.
• Provide ≥3 recommendations for scaling the solution to datasets exceeding 1 million records.

---