AI-Powered Churn Prevention Pipeline

End-to-end machine learning pipeline with comprehensive data analysis and model deployment

Ready to execute

0 / 7 stages completed

Data
Sources

Data
Analysis

Feature
Engineering

Model
Training

Model
Evaluation

Deployment

AI Agents

Data Sources & Pipeline

Real-time data integration from multiple sources powering AI-driven churn prediction

👥

Customer Profile Dataset

Postpaid Subscriber Data

Records

Features

🎤

Call Transcription Data

AWS Transcribe + Bedrock NLP

Calls Processed

NLP Features

Data Analysis

AI-powered data profiling and quality analysis

98.5%

Data Completeness

Features Identified

Outliers Detected

0.87

Avg Correlation

🤖 AI-Powered Data Analysis Agents

📊

Structured Data Analyst

Numerical & Tabular

Capabilities:

Statistical analysis & correlation detection

Data quality assessment & cleaning

Outlier detection & handling

Temporal pattern analysis

50K

Records Analyzed

98.2%

Data Quality

🧠

Unstructured Data Analyst

Text, Audio & NLP

Capabilities:

Call transcript sentiment analysis

Complaint text classification

Intent & emotion detection

Competitor mention extraction

127K

Calls Processed

96.8%

NLP Accuracy

🏷️

Data Labeller Agent

AI-Powered Annotation & Classification

Active Learning

Automated Labelling Tasks:

Churn Labels

High/Medium/Low risk classification

Sentiment Tags

Positive/Neutral/Negative

Complaint Types

Network, Billing, Service

Intent Detection

Switch, Upgrade, Downgrade

Performance:

50K

Labels Generated

94.5%

Accuracy

Feature Engineering

Creating sophisticated engineered features from raw data

📞 Call Patterns (12 features)

Call duration, frequency, time-of-day patterns, weekend vs weekday usage

📊 Usage Metrics (10 features)

Data consumption, voice minutes, SMS counts, roaming usage

🎯 Behavioral Indicators (8 features)

Complaint rate, support call frequency, service requests, plan changes

🔧 Feature Engineering Agents

🔢

Encoding Agent

Categorical Transform

Techniques:
• One-Hot Encoding
• Label Encoding
• Target Encoding
• Frequency Encoding

8 categorical features encoded

⚖️

Scaling Agent

Normalization

Methods:
• StandardScaler
• MinMaxScaler
• RobustScaler
• MaxAbsScaler

23 numerical features scaled

📦

Binning Agent

Discretization

Strategies:
• Equal-Width Binning
• Equal-Frequency
• K-Means Clustering
• Custom Quantiles

6 continuous features binned

Model Training

Dual-model approach: Churn Prediction + Churn Attribution using structured & unstructured data

🎯

Model 1: Churn Prediction

Binary Classification (Will Churn / Won't Churn)

Input Features (47 total):

Structured Data (35 features):
• Customer demographics (age, tenure, plan)
• Usage patterns (voice, data, SMS)
• Financial metrics (ARPU, CLV, payment history)
• Behavioral signals (support calls, complaints)

Unstructured Data (12 features):
• Sentiment score (AWS Nova Pro)
• Intent classification
• Emotion vectors (Titan embeddings)
• Competitor mentions
• Urgency indicators

Model Architecture:

• Algorithm: Random Forest Classifier
• Estimators: 200 trees
• Max Depth: 12
• Features per split: sqrt(47) = 7
• Training Method: Stratified K-Fold CV (5 folds)

89.3%

Accuracy

0.91

Precision

0.87

Recall

0.89

F1-Score

Output: Churn Probability (0-100%)

🔍

Model 2: Churn Attribution

Multi-Class Classification (Root Cause Analysis)

Churn Attribution Categories:

Low NPS Score

Customer dissatisfaction & poor experience

Competition

Competitor offers & switching intent

Pricing Concerns

High bills & affordability issues

Quality of Service

Poor service quality & complaints

Connection Drops

Network reliability & dropped calls

Low Usage

Declining engagement & activity

Other Factors

Contract issues, relocation, miscellaneous

Model Architecture:

• Algorithm: XGBoost Multi-Class Classifier
• Classes: 7 churn categories
• Features: Same 47 features + churn probability
• Objective: multi:softmax
• Training: Only on churned customers

84.7%

Accuracy

0.82

Macro F1

Two-Stage Prediction Pipeline

STAGE 1
Churn Prediction Model
↓ Probability: 87%
IF Churn = Yes
(Probability > 50%)
STAGE 2
Attribution Model
↓ Reason: Pricing (73%)
ACTION
Retention Strategy
→ Discount Offer

Model Evaluation

Comprehensive model performance metrics and validation

Random Forest

89.3%

Accuracy

✓ SELECTED

XGBoost

87.1%

Accuracy

Model Deployment

Deploying trained model to production environment

Deployment Status

✓ Model Deployed

Production endpoint active

⚡ Real-time Inference

API endpoint ready

Customer Retention & Outreach

Proactive customer re-engagement using churn attribution insights from Model 2

CHURN ATTRIBUTION TRIGGERS - REAL-TIME MONITORING

7-category attribution model identifying root causes of customer churn

😔

Low NPS Score

Customers with NPS ≤ 6 (Detractors)

156

Detected

In Queue

71%

Retained

🔄

Competition

Competitor offers or port-out inquiries

Detected

In Queue

75%

Retained

💰

Pricing Concerns

Billing complaints or pricing sensitivity

112

Detected

In Queue

82%

Retained

⚡

Quality of Service

Service quality or speed complaints

Detected

In Queue

69%

Retained

📡

Connection Drops

Network reliability and connectivity issues

Detected

In Queue

73%

Retained

📉

Low Usage

Customers with >30% usage decline

342

Detected

In Queue

68%

Retained

📋

Other Factors

Miscellaneous churn indicators

Detected

In Queue

64%

Retained

🏆

TOP CHURN CATEGORY - HIGHEST IMPACT

Low Usage (342 Customers)

Root Cause: 45% average usage decline • Primary Action: Usage incentives + plan optimization • Revenue at Risk: $8.5M annually

PRIORITY CUSTOMER QUEUE - AI OUTREACH

Customers ranked by churn risk and attribution category

AI Dialer Integration

Dialer Status ● ACTIVE

AWS Connect Integration

234

Active Calls Today

3.2m

Avg Call Duration

Today's Performance

342

Calls Initiated

234

Retained

68.4%

Success Rate

$487K

Revenue Saved

AI Agent Workflow

18-agent autonomous workflow for churn prevention and customer retention

Agent Status

🎯 Real-time Monitoring

Monitor customer behavior patterns

📊 Feature Selection

Select optimal feature set

📞 Dialer Agent

Initiate retention calls

💬 Conversation Agent

Handle customer interactions

🎁 Offer Manager

Create personalized offers

📈 Analytics Agent

Generate insights and reports

Notebook Configuration

Notebook Type

ML Framework

Scikit-learn XGBoost PyTorch

Include Sections

Exploratory Data Analysis Feature Engineering Hyperparameter Optimization Model Explainability (SHAP)

Generated Code Preview

# Configure options above and click "Generate Notebook" to create code

AWS Nova-Powered Feature Importance Analysis

Advanced AI explanations using SHAP (SHapley Additive exPlanations) values + AWS Bedrock insights

Features Selected

Initial Features

61.7%

Reduction Rate

92.1%

Model Accuracy

3.2x

Speed Improvement

Top 10 Features - Ranked by SHAP Importance

usage_decline_percentage

Behavioral Signal · Structured Data

23.5%

SHAP Importance

AWS Nova Pro AI Explanation:

Why it matters: Declining usage is the #1 predictor of churn. Customers who reduce their monthly voice/data consumption by >20% are 4.7x more likely to churn within 90 days. This behavioral shift signals dissatisfaction, financial constraints, or migration to competitors.

Business Impact: Early detection allows proactive retention campaigns. A 10% improvement in identifying usage decline can save $2.4M annually by preventing high-value customer churn.

Model Interpretation: SHAP values show this feature contributes most to churn probability. When usage drops >30%, model confidence increases by 18 percentage points.

complaint_count_90d

Service Quality · Structured Data

18.2%

SHAP Importance

AWS Nova Pro AI Explanation:

Why it matters: Customer complaints are direct signals of dissatisfaction. Each additional complaint in a 90-day window increases churn probability by 12%. Customers with 3+ complaints have a 67% churn rate vs. 8% baseline.

Business Impact: Complaint resolution within 48 hours reduces churn risk by 45%. Prioritizing high-complaint customers for proactive outreach can recover 32% of at-risk revenue.

Model Interpretation: This feature shows strong interaction effects with NPS score and sentiment. Combined low NPS + high complaints = 89% churn prediction accuracy.

sentiment_score_avg

NLP-Derived · Unstructured Data (AWS Nova)

15.3%

SHAP Importance

AWS Nova Pro AI Explanation:

Why it matters: NLP sentiment analysis from call transcripts reveals emotional state. Negative sentiment (score < -0.5) correlates with 73% churn probability. This unstructured data captures dissatisfaction that structured metrics miss.

Business Impact: Combining sentiment with structured features improves model accuracy by 8.4%. Real-time sentiment monitoring enables same-day intervention for highly negative interactions.

Model Interpretation: AWS Nova Pro extracts emotional context from 127K call transcripts. SHAP shows sentiment is critical for borderline cases where traditional metrics are neutral.

Features 4-10: Summary

4. customer_tenure_months 14.1%

Demographic · New customers (<6 months) churn 3.2x more

5. payment_delay_days 12.3%

Financial · Payment issues indicate affordability problems

6. competitor_mention_count 10.7%

NLP-Derived · Mentions of du, etisalat indicate switching intent

7. nps_score 9.8%

Service Quality · Detractors (NPS<7) have 54% churn rate

8. connection_drop_rate 8.9%

Network Quality · Drop rate >5% drives service quality complaints

9. avg_bill_amount 7.6%

Financial · High bills (>500 AED/mo) correlate with pricing churn

10. support_call_frequency 6.4%

Behavioral · >3 support calls/month indicate unresolved issues

Multi-Method Feature Selection Approach

SHAP Values

Game theory-based feature importance

Random Forest

Feature importance from tree splits

XGBoost Gain

Feature gain and coverage metrics

Mutual Info

Statistical dependency analysis

Consensus-Based Selection

Features selected only if ranked in top 20 by at least 3 out of 4 methods, ensuring robust and reliable feature set with minimal redundancy and maximum predictive power.

Omantel AI Churn Retention

AI-Powered Churn Prevention Pipeline

Data Sources & Pipeline

Customer Profile Dataset

Call Transcription Data

Data Analysis

🤖 AI-Powered Data Analysis Agents

Structured Data Analyst

Capabilities:

Unstructured Data Analyst

Capabilities:

Data Labeller Agent

Automated Labelling Tasks:

Performance:

Feature Engineering

🔧 Feature Engineering Agents

Encoding Agent

Scaling Agent

Binning Agent

Model Training

Model 1: Churn Prediction

Input Features (47 total):

Model Architecture:

Model 2: Churn Attribution

Churn Attribution Categories:

Model Architecture:

Two-Stage Prediction Pipeline

Model Evaluation

Model Deployment

Deployment Status

Customer Retention & Outreach

CHURN ATTRIBUTION TRIGGERS - REAL-TIME MONITORING

PRIORITY CUSTOMER QUEUE - AI OUTREACH

AI Dialer Integration

Today's Performance

AI Agent Workflow

Agent Status

Dataset Profile

Information

Code Generator

Notebook Configuration

Generated Code Preview

Code Ignite

Interactive Code Viewer

Execution Output

Jupyter Integration

SHAP Feature Analysis - AI-Powered Explainability

AWS Nova-Powered Feature Importance Analysis

Top 10 Features - Ranked by SHAP Importance

usage_decline_percentage

complaint_count_90d

sentiment_score_avg

Features 4-10: Summary

Multi-Method Feature Selection Approach