Sanoj111/mobile_prediction

📱 Mobile Price Prediction — ML Pipeline

Predict mobile phone prices in Nepal (NPR) from specs using an end-to-end machine learning pipeline.

---

📋 Project Overview

This project builds a regression model to predict smartphone prices in Nepalese Rupees (NPR) using device specifications parsed from a real-world, messy scraping dataset (all_mobile_data5.csv). The pipeline covers the full ML lifecycle: data parsing, feature engineering, EDA, model training, hyperparameter tuning, and inference on new data.

---

🗂️ Project Structure

mobile-price-prediction/
│
├── mobile_price_prediction.py   # Main pipeline (all steps)
├── requirements.txt             # Python dependencies
├── mobile_price_model.pkl       # Saved best model bundle
├── eda_dashboard.png            # EDA visualisation (12 plots)
├── model_evaluation.png         # Model evaluation charts
└── README.md                    # This file

---

📦 Dataset

Field	Description
Model	Phone model name (e.g., Galaxy S25 Ultra)
Price	Raw price string (e.g., "Rs. 184,999 (12/256GB)")
Brand	Brand name or brand-slug string
Price_clean	Partially cleaned price (noisy, not directly usable)

Raw records: 127 rows — heavily messy with header rows, duplicate entries, malformed prices.
Clean records after parsing: 109 usable rows.

---

⚙️ Pipeline Steps

1️⃣ Data Parsing & Preprocessing

The raw dataset is extremely messy — prices are embedded in strings like "Rs. 184,999 (12/256GB)" and "NPR 199,999", RAM/Storage specs are inside price strings, and brand names are sometimes URL slugs.

Custom regex-based parsers handle:

• Price extraction — extracts 4–7 digit numbers within the realistic range (5,000–500,000 NPR)
• RAM extraction — handles 8GB, 12/256, 12+256, 8GB+256GB patterns
• Storage extraction — handles TB→GB conversion, paired patterns, standalone GB values
• Brand normalization — maps slugs like samsung-galaxy-a36-price-nepal → Samsung

Missing RAM/Storage values are imputed with the median of available values.

---

2️⃣ Feature Engineering

Feature	Description
RAM_GB	RAM in gigabytes
Storage_GB	Internal storage in gigabytes
Is_5G	Binary flag: device supports 5G
Is_Ultra	Binary flag: model name contains "Ultra"
Is_Pro	Binary flag: model name contains "Pro"
Is_Foldable	Binary flag: Fold or Flip form factor
RAM_x_Storage	Interaction term: RAM × Storage
Log_Storage	log(1 + Storage) — captures diminishing returns
Premium_Score	Weighted sum: Ultra×3 + Pro×2 + Foldable×4 + 5G×1
Brand_Encoded	Ordinal: Apple=3, Samsung=2, Xiaomi=1, Other=0

---

3️⃣ EDA (Exploratory Data Analysis)

The EDA dashboard (eda_dashboard.png) includes 12 plots:

• Price distribution histogram
• Average price by brand (bar chart)
• Brand share (pie chart)
• Price distribution by tier (box plots)
• RAM vs Price scatter (by brand)
• Storage vs Price scatter (by brand)
• Feature presence vs absence price comparison
• Correlation heatmap
• 5G vs Non-5G price comparison
• Count by price tier
• RAM distribution
• Storage distribution

Price Tiers:

Tier	Price Range (NPR)
Budget	< 20,000
Mid-Range	20,000 – 59,999
Upper-Mid	60,000 – 119,999
Flagship	≥ 120,000

---

4️⃣ Model Training

Four models were trained and evaluated with 5-fold cross-validation:

Model	Test R²	MAE (NPR)	RMSE (NPR)	CV R²
Ridge Regression	0.6776	41,141	48,503	0.7409
Decision Tree	0.5677	38,241	56,167	0.7428
Random Forest	0.6913	33,623	47,464	0.7703
Gradient Boosting	0.7487	31,371	42,819	0.7860

---

5️⃣ Hyperparameter Tuning

GridSearchCV was applied to the Random Forest model across:

param_grid = {
    'n_estimators':      [50, 100, 200],
    'max_depth':         [None, 5, 10, 15],
    'min_samples_split': [2, 5],
    'min_samples_leaf':  [1, 2],
    'max_features':      ['sqrt', 'log2'],
}

Best params: max_depth=None, max_features='sqrt', min_samples_leaf=1, min_samples_split=5, n_estimators=50
Best CV R²: 0.7690

---

6️⃣ Best Model

🏆 Gradient Boosting Regressor was selected as the best model.

Metric	Value
R²	0.7487
MAE	NPR 31,371
RMSE	NPR 42,819
CV R²	0.7860

---

7️⃣ Inference on New Data

The saved model bundle (mobile_price_model.pkl) includes the model, scaler, feature list, and brand encoder. Example predictions:

Phone Specs	Predicted Price	Tier
Budget Android (4GB/64GB)	NPR 12,218	Budget
Mid-Range Samsung 5G (8GB/256GB)	NPR 56,312	Mid-Range
Samsung Ultra (12GB/512GB, 5G, Ultra)	NPR 163,363	Flagship
Apple iPhone Pro Max (12GB/1TB)	NPR 270,262	Flagship
Samsung Z Fold (12GB/512GB, Foldable)	NPR 128,310	Flagship

---

🚀 Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Run the full pipeline

python3 mobile_price_prediction.py

3. Load the saved model for inference

import joblib
import pandas as pd
import numpy as np

bundle = joblib.load('mobile_price_model.pkl')
model  = bundle['model']
cols   = bundle['feature_cols']

# Example: Apple iPhone Pro, 8GB RAM, 256GB, 5G, Pro
sample = pd.DataFrame([{
    'RAM_GB': 8, 'Storage_GB': 256,
    'Is_5G': 1, 'Is_Ultra': 0, 'Is_Pro': 1, 'Is_Foldable': 0,
    'RAM_x_Storage': 8 * 256, 'Log_Storage': np.log1p(256),
    'Premium_Score': 0*3 + 1*2 + 0*4 + 1*1,
    'Brand_Encoded': 3,   # Apple=3
}])[cols]

predicted_price = model.predict(sample)[0]
print(f"Predicted Price: NPR {predicted_price:,.0f}")

---

📊 Key Insights

• Brand is the strongest price driver — Apple phones command a premium across all storage tiers
• Storage is more predictive than RAM alone; the interaction term RAM × Storage improves accuracy
• Foldable devices carry the highest premium feature score
• 5G alone has a moderate effect; combined with Ultra/Pro it pushes prices significantly higher
• The dataset is small (109 records), which limits model generalisation — more data would improve R²

---

🛠️ Requirements

pandas>=2.0.0
numpy>=1.24.0
scikit-learn>=1.3.0
matplotlib>=3.7.0
seaborn>=0.12.0
joblib>=1.3.0

---

📝 Notes

• Dataset source: Nepal mobile price listings (scraped), prices in NPR
• The raw data contains significant noise — header rows mixed into data, malformed price strings, URL slugs as brand names. All cleaned via custom regex parsers.
• Model performance is limited by dataset size; with more diverse data, tree-ensemble models would perform significantly better.