Mobile Plan Classification: Smart vs. Ultra

1. The Problem

Megaline, a mobile carrier, needs to transition from manual or generalized upselling to an automated, data-driven approach. The company requires a robust system to analyze existing customer behavior and accurately recommend one of their two primary plans (Smart or Ultra) to personalize the user experience and optimize plan allocation.

2. Objective

The core objective of this project is to develop a classification model with a minimum accuracy threshold of 0.75 to automate and personalize plan recommendations for existing Megaline subscribers based on their usage behavior.

3. Methodology

• Data & Target: Analyzed behavior records from 3,214 users, featuring voice calls (frequency and total minutes), SMS sent, and monthly data consumption (mb_used). The target variable is is_ultra (1 for Ultra, 0 for Smart).
• Preprocessing: Applied feature scaling using StandardScaler to normalize consumption metrics, ensuring model convergence.
• Validation Strategy: Data was partitioned into Train (60%), Validation (20%), and Test (20%) sets, utilizing stratification to maintain strict class proportions.
• Modeling & Optimization: Evaluated four distinct architectures (Logistic Regression as the baseline, Decision Tree, Random Forest, and Gradient Boosting) to find the optimal balance between bias and variance. Implemented systematic hyperparameter tuning via GridSearchCV with 5-fold cross-validation.

4. Findings

Model Performance

Model	Test Accuracy
Random Forest	0.821
Decision Tree	0.804
Gradient Boosting	0.796
Logistic Regression	0.740

• Target Achievement: The Random Forest model emerged as the winning architecture, significantly exceeding the business target (0.75) by achieving an 82.11% accuracy on unseen data.
• Sanity Check: To ensure the model captured genuine behavioral patterns rather than noise, a permutation test was performed. After shuffling the feature indices, the accuracy dropped to 0.58 (close to random chance). This statistical confirmation supports the model's high predictive power and its ability to generalize to new subscribers.

5. Recommendations

• Automated Allocation: Deploy the optimized Random Forest model into Megaline's production environment to proactively automate subscriber plan upgrades based on their daily data and call limits.
• Data-Driven Marketing: Utilize the classification engine to filter and target existing Smart plan users whose behavioral metrics align with the Ultra profile, maximizing the ROI of promotional campaigns.

6. Technologies

• Language: Python
• Machine Learning Framework: Scikit-Learn
• Core Techniques: Classification Architectures, GridSearchCV, Feature Scaling, Stratified Splitting, Permutation Testing

---

Project Structure

├── data/                  # Local directory for raw dataset
├── notebooks/             # Jupyter Notebooks containing EDA and Modeling
├── README.md              # Project executive summary and documentation
└── requirements.txt       # Project dependencies

---

Installation & Usage

1. Clone the repository.
2. Install dependencies: pip install -r requirements.txt
3. Open and run the notebook: Project_Sprint_10.ipynb

---

Analysis by Fiorella Trigo M.
Part 2 of the Megaline Optimization Series