Recommender Systems For Instacart: A Collaborative Filtering Approach

Collaborators

• Nakul Camasamudram
• Guiheng Zhou
• Rahul Verma
• Rosy Parmar

Exploratory data analysis is performed in src/eda.ipynb.

We have implemented three collaborative filtering methods in independent Jupyter Notebooks under src/

1. tfidf.ipynb: Neighborhood based method that uses cosine similarity on a tf-idf weighted matrix to recommend products from similar users.
2. svd.ipynb: Matrix factorization using SVD. Computes the largest K singular values/vectors for a sparse matrix. Based upon the largest K singular values, we find top K recommended items for users.
3. imf.ipynb: Matrix factorization using Alternating Least Squares by representing the utility matrix as a confidence matrix. Based on the paper Collaborative Filtering for Implicit Feedback Datasets.

More details are in documents/final/report.pdf

Video presentation here.

Usage

• Install dependencies: pip install -r requirements.txt.
• Download all *.csv's from here into the data/ directory.
• Run any of the Jupyter Notebooks under src/
• You can view the evaluation results in data/eval. [It is currently empty. Will be populated if the notebooks are run again.]