Group Analytics Project

Course: BANA 320 - Predictive Analytics
Dataset: Yelp Open Dataset
Group Name: Three Gits
Team: Vincent Balalian, Sameer Patel, Arish Patel

Overview

Problem Statement: Does adding sentiment features from the first 90 days of reviews significantly improve the accuracy of predicting a restaurant's 1-year Yelp rating, compared to using non-text features alone?

Target Variable: Average star rating 12 months after first review

Result: No - sentiment features did not significantly improve prediction accuracy when geographic and user-based context features were already included.

Data Pipeline

Raw Yelp JSON data (hosted in Google Cloud Storage) is transformed via dbt + BigQuery:

1. Filtering: Business dataset filtered to restaurants only
2. Qualification criteria:
   • Minimum 1 year of review history
   • At least 3 reviews in first 90 days
   • At least 10 reviews in first year
3. Feature aggregation: Check-ins, user metrics, and zip-code comparisons aggregated to restaurant level
4. Time windowing: Separate datasets for 90-day (features) and 365-day (target) review periods

Feature Categories

Category	Description	Count
Early Review Metrics	Review count & average rating (first 90 days)	2
Check-in Patterns	Day-of-week and time-of-day distributions	11
User Characteristics	Reviewer reputation, engagement, experience	10
Zip Code Context	Restaurant metrics vs. local averages	4
Sentiment (VADER)	Positive, negative, neutral, compound scores	4

Methodology

Phase 1: Train models using all features except sentiment
Phase 2: Add sentiment features derived from VADER analysis of first-90-day reviews
Comparison: Paired one-tailed t-test on absolute prediction errors (α = 0.05)

Models Tested

• Support Vector Regression (linear kernel)
• Linear Regression
• Random Forest Regressor
• XGBoost Regressor
• Stacked Ensemble (average of all four)

Key Findings

1. Geographic context subsumes sentiment signal - Zip-code comparison features (rating vs. local average, etc.) already capture competitive positioning that sentiment would otherwise indicate.

2. Early rating is the dominant predictor - The 90-day average rating had the highest mutual information with 1-year rating by a wide margin.

3. Reviewer quality correlates with outcomes - User characteristics (average stars given, engagement metrics) ranked among top predictive features.

Tools & Technologies

• Data Warehouse: Google BigQuery
• Transformation: dbt
• Analysis: Python (pandas, scikit-learn, XGBoost, VADER)
• Environment: Google Colab
• CI/CD: GitHub Actions

Analysis

The complete analysis notebook is available here. It includes all code, visualizations, and statistical test results.

Note: The data pipeline requires access to a private Google Cloud project. The notebook is provided for review purposes; reproducing it would require setting up your own BigQuery environment with the Yelp Open Dataset.

BANA 320 - Fall 2025