Overview

A hybrid recommendation system that combines content-based and collaborative filtering approaches to provide personalized product recommendations. The system analyzes product details, user ratings, and review data to suggest relevant items to users.

Features

• Content-based filtering using product details and reviews
• Collaborative filtering using user-item interactions
• Hybrid recommendation approach combining both methods
• Extensive data preprocessing and cleaning
• Detailed exploratory data analysis (EDA)
• Performance evaluation metrics
• Interactive Dashboard with streamlit to generate recommendations based on user-id

Dataset

The system uses an Amazon Sales Dataset with the following key features:

• Product information (ID, name, category, price [actual and discounted])
• User reviews and ratings
• Product descriptions
• Category hierarchies

Usage

Step 1

You can clone this repository locally using CLI or download it as a zip file.

git clone https://github.com/Aakarsh204/Recommender-System.git

Step 2

To run the Streamlit App Dashboard, you need to create an Anaconda environment

conda create --name Recommender-System --file requirements.txt

This will install all the dependencies that you require to get started.

Step 3

After they have finished installation, download the Collaborative Filter Model and Content Based Similarity Matrix from my GDrive

Link to Cosine Matrix

Link to CF Model

Extract the models to the main working directory

Also extract the python scripts to the main working directory

Step 4

Download the dataset from this link

Extract this to the main working directory as well. So far your workspace should look like this:

Recommender-System
 |
 +-- cfmodel.pkl
 |    
 +-- cosine_matrix.npz
 |    
 +-- amazon.csv
 |    
 +-- requirements.txt
 |    
 +-- app.py
 |
 .... Rest of your scripts
                  

Step 5

Once you are done setting up, activate your conda environment

conda activate Recommender-System

Now inside the environment, run this command

streamlit run app.py

This should start up the streamlit app in your local browser. Happy Shopping!

Technical Details

Dependencies

pandas
numpy
matplotlib
seaborn
scipy
scikit-learn
nltk
surprise
streamlit
plotly

Data Preprocessing

• Text cleaning and normalization
• Handling missing values
• Price formatting
• Category hierarchy splitting
• Rating weight calculation
• Label encoding for user and product IDs

Recommendation Approaches

Content-Based Filtering

• Used TF-IDF vectorization for product details
• Computed cosine similarity between products
• Recommended products based on item-item similarity

Collaborative Filtering

• Implemented SVD (Singular Value Decomposition)
• Used the Surprise library for model training
• Cross-validation for model evaluation

Hybrid System

• Combines content-based and collaborative filtering scores
• Weighted recommendation scores
• Configurable weights for each approach

Performance Metrics

• RMSE (Root Mean Square Error)
• MAE (Mean Absolute Error)
• Average Precision Score

Challenges

• The dataset had a unique format, where each product had comma separated user values
• Splitting the user values often led to many predictions being the same item, but from different users
• Going from experimenting in Colab to modularizing the code
• Debugging the errors in the streamlit dashboard

Screenshots

Author

Aakarsh Pathak
I’m a passionate developer focused on building and sharing machine learning and data science projects to advance my knowledge and real-world applications.
My linkedIn: link