Recommendation System with Collaborative Filtering

This project implements a recommendation system using collaborative filtering. The model computes user similarities based on various attributes retrieved from a Firebase Realtime Database. The output is a JSON file containing each user's similarity with other users, ranked from highest to lowest.

Setup and Installation

1. Clone the repository

git clone https://github.com/alexchans/Recommendation_System_The_Social_Panacea.git

2. Install required dependencies

pip install firebase-admin gensim numpy scikit-learn

3. Download required files

• Firebase Admin SDK Key: Download your key.json from Firebase (this is your private key).
• GloVe Pre-trained Embeddings: Download the glove.6B.50d.txt file from the GloVe website and place it in your working directory.

4. Configure Firebase

In the script, ensure the Firebase Realtime Database URL is properly set:

firebase_admin.initialize_app(cred, {
    'databaseURL': 'https://your-database-url.firebaseio.com/'
})

How It Works

User Data:

The script retrieves user data from the Firebase Realtime Database, containing user attributes like age, university, preferences, and personality types.

Preprocessing:

1. Weighted Attributes:
   Each attribute is processed and given a specific weight based on its importance.
2. Text-based Fields (University, Personality Types):
   Text fields are converted into word embeddings using GloVe embeddings and normalized.
3. Numerical Fields (Age, Preferences):
   Numerical values are normalized to fit between 0 and 1, then scaled by their respective weights.

Collaborative Filtering:

• After preprocessing, the system uses cosine similarity to measure how similar each user is to others based on the weighted attributes.
• The result is a similarity matrix that shows the similarity score between each pair of users.

Output:

• The final output is a JSON file (user_similarities.json) that lists each user and their similarity with others, sorted from the most similar to the least.

---

Data Preprocessing

The script uses different preprocessing techniques based on the type of data:

1. Text Fields (e.g., University, Personality Type):
   Processed using word embeddings and normalized.
2. Numerical Fields (e.g., Age, Preferences):
   Scaled between 0 and 1, then multiplied by predefined weights.
3. Boolean Fields (e.g., Cooking with Friends):
   Converted to binary (0 or 1) and weighted accordingly.

---

Weights

Each attribute is assigned a weight to reflect its importance in the similarity calculation. For example:

• Age: 10%
• University: 15%
• Personality Type: 14%
• Sports Preferences: 8%

---

Collaborative Filtering

The model uses cosine similarity to compute how similar one user is to another based on their processed attributes.

Running the Script

Once everything is set up, simply run the script to generate the similarity scores between users:

python recommendation_system.py

The similarity scores between users will be saved in a file named user_similarities.json.

Output

The output JSON file will contain each user as a key, and a sorted list of other users with similarity scores as the value. Each user's own similarity score is excluded.

Example structure of the output:

{
    "user1": {
        "user3": 0.923,
        "user4": 0.806,
        ...
    },
    "user2": {
        "user4": 0.973,
        "user1": 0.776,
        ...
    }
}

Notes

• The script is flexible and can be easily modified to accommodate additional user attributes or adjust the weights used in similarity calculations.
• The data preprocessing approach can be adjusted depending on the type of attribute being processed.