README.md

Product Roadmap ROI Predictor

Business Problem

Product teams at banking apps face a critical challenge: prioritizing which features to build and bugs to fix from thousands of user reviews. With limited engineering resources, teams need to identify high-impact opportunities that will maximize return on investment (ROI) in terms of user satisfaction, retention, and app ratings.

Current Pain Points:

• Volume overload: Banking apps receive 25,000+ reviews annually across multiple platforms
• Subjective prioritization: Feature requests are often prioritized based on gut feeling or loudest voices, not data
• Unclear ROI: No systematic way to measure which fixes will have the biggest impact on user satisfaction
• Delayed insights: Manual review analysis is slow, causing teams to miss time-sensitive issues

Business Impact:

• A 0.5 star rating increase can improve conversion rates by 20-30% (industry benchmark)
• Fixing high-frequency negative issues can reduce churn by 10-15%
• Addressing the "right" complaints can increase App Store visibility and organic downloads

Goal: Build a data-driven system to analyze app reviews and predict which product roadmap decisions will deliver the highest ROI based on:

1. Issue frequency - How many users are affected?
2. Sentiment intensity - How unhappy/happy are users?
3. Impact correlation - Which issues correlate with low ratings and churn?

This project extracts 21+ features from banking app reviews (crashes, login issues, transfer problems, etc.) to identify and prioritize high-impact product improvements.

---

Project Status

Phase 1 Complete: Feature extraction, EDA, and baseline modeling finished

Current Phase: Building ROI prediction model to prioritize product improvements

---

Current Progress

Completed Work

1. Feature Extraction (21 features)

• Pattern matching: 10 issue categories (crash, login, mobile deposit, performance, frustration, etc.)
• Sentiment analysis: 4 VADER scores (compound, positive, negative, neutral)
• Text statistics: 7 metrics (length, word count, caps ratio, punctuation)

2. Exploratory Data Analysis

• Analyzed 25,000 reviews from 5 banking apps (Chase, Citi, BofA, Capital One, Wells Fargo)
• Key findings:
  • Top issues: Crashes (24%), Login (22%), Frustration (24%)
  • Highest negative ROI: Frustration (impact score: 997)
  • Sentiment vs star rating correlation: 0.85+
  • Identified data quality issues (VADER sarcasm detection, pattern overlap)

3. Baseline Classification Models

• Goal: Predict star ratings (1-5) from extracted features
• Models tested: Logistic Regression vs Random Forest
• Best model: Logistic Regression with class_weight='balanced'
  • Accuracy: 50%
  • Macro F1: 0.40
  • Beats naive baseline (35%) by 43%

Key Findings

What Worked:

• Sentiment features most predictive (34-35% total importance)
• Feature engineering captured meaningful signals
• Logistic Regression outperformed Random Forest despite lower accuracy
• Model learned ordinal relationships (predicts adjacent ratings when wrong)

What Didn't Work:

• Only 50% accuracy - not production-ready for exact rating prediction
• Cannot reliably predict middle ratings (2-4 stars) due to class imbalance
• Random Forest achieved higher accuracy (61%) but only by predicting 1 and 5 stars
• VADER sentiment analysis fails on sarcasm, mixed reviews, context

Error Analysis (363 big errors analyzed):

• 60% of errors: Model predicts too high (actual 1-star → predicted 4-5 stars)
  • Root cause: VADER scores misleadingly positive
  • Example: "After 2 years of fighting this bank" → sentiment +0.97
• 30% of errors: Model predicts too low (actual 5-star → predicted 1-2 stars)
  • Root cause: Mentions problems in passing even when overall positive
• 10% of errors: Middle ratings confused

Where Models Need Improvement

1. Class Imbalance (Primary Bottleneck)

• Classes 2-4 only represent 9-11% of data each
• Model defaults to predicting majority classes (1 and 5)
• F1-scores for middle ratings: 0.18-0.24 (very poor)

2. Sentiment Analysis Quality

• VADER cannot detect sarcasm: "Not the best app!" → +0.89 sentiment
• Mixed reviews averaged: "works great BUT crashes" → positive score
• Context blindness: "I changed banks" in review about NEW bank → neutral

3. Feature Quality

• has_satisfaction captures both "satisfied" and "NOT satisfied"
• Pattern features don't distinguish "had crash" vs "has crash"
• No n-grams or contextual features

4. Evaluation Limitations

• Single train/test split (no cross-validation)
• No hyperparameter tuning
• Default model parameters used

Possible Next Attempts

Immediate Improvements (Low Effort, High Impact):

1. Fix pattern matching

   • Split has_satisfaction into positive vs negative mentions
   • Use negative lookahead regex: (?<!not |un|dis)satisf

2. Address class imbalance

   • SMOTE (Synthetic Minority Over-sampling)
   • More aggressive class weights: {1: 1, 2: 5, 3: 5, 4: 5, 5: 1}

3. Cross-validation

   • 5-fold stratified CV to validate results

Future Improvements (Higher Effort, Higher Impact): 4. Better sentiment analysis

• Replace VADER with BERT or RoBERTa fine-tuned on app reviews
• Aspect-based sentiment (crash sentiment vs overall sentiment)
• Would fix 60% of big errors

5. Ordinal regression

   • Treat ratings as ordered (1 < 2 < 3 < 4 < 5) instead of independent classes
   • Penalize adjacent errors less

6. Advanced features

   • TF-IDF (top 100-500 words)
   • N-grams for phrases ("face id", "customer service")
   • Sentence-level sentiment
   • App version, device type metadata

7. Better models

   • XGBoost/LightGBM for better class imbalance handling
   • Ensemble methods

ROI Prediction Model (Current Focus):

• Pivot from exact rating prediction to ROI prediction
• Use feature importance + frequency to identify high-impact improvements
• Answer: "If we fix X, how much will ratings improve?"

---

Table of Contents

• Installation
• Usage
• Project Structure
• Methodology
• Results
• Future Work

---

Sections below to be completed...