Lahore POI Recommender

A hybrid recommendation system built using the LightFM library to recommend Points of Interest (POIs) in Lahore, Pakistan.

This repository contains the source code, data, and documentation for a project that processes spatial place data, simulates user interaction demographics, and trains a hybrid collaborative filtering model.

---

📁 Repository Structure

lahore-poi-recommender/
├── data/
│   ├── pakistan_gazetteer_data.xlsx # Raw Pakistan Gazetteer dataset
│   ├── lahore_locations_data.csv   # Filtered POI data from the Pakistan Gazetteer
│   └── generated_users_data.csv    # Simulated user demographics and interaction history
├── docs/
│   ├── presentation.pdf            # Project presentation slide deck
│   └── report.pdf                  # Detailed project report
├── lahore_poi_recommender.ipynb    # Jupyter Notebook containing the full pipeline
└── README.md                       

---

🛠️ Project Overview

The project is structured into three main phases, all contained within lahore_poi_recommender.ipynb:

1. Data Extraction & Filtering:
   • Filtered national spatial data to extract Lahore's coordinates.
   • Categorized locations into relevant feature classes (e.g., Parks, Mosques, Monuments, Shrines, Libraries).
2. User Simulation:
   • Generated synthetic profile data for 5,000 users.
   • Assigned demographics (age, gender) and category preferences.
   • Built simulated visit histories (10–20 visited locations per user) to train the model.
3. Model Training & Evaluation:
   • Built a Hybrid Recommender System using the LightFM algorithm.
   • Incorporated user features (gender, binned age groups) and item features (feature classes/codes)
   • Evaluated using ranking metrics: Precision@k, Recall@k, and AUC Score.

---

🚀 Getting Started

Prerequisites

To run the notebook locally, install the following Python packages:

pip install pandas numpy lightfm scikit-learn openpyxl

Running the Notebook

1. Clone this repository.
2. Open lahore_poi_recommender.ipynb in Jupyter Notebook or Google Colab.
3. Run all cells in order. The notebook is fully self-contained and runnable from start to finish, including the initial coordinate filtering of the raw gazetteer data (data/pakistan_gazetteer_data.xlsx).

---

📊 Evaluation Results

The hybrid model achieves the following typical validation metrics on the simulated interaction data:

• Precision@5: ~0.33
• Recall@5: ~0.11
• AUC Score: ~0.70