Instacart Market Basket Recommender System

This project aims to build and compare several recommendation algorithms to predict which products a user will reorder in their next shopping cart. The project uses the publicly available Instacart 2017 dataset from a Kaggle competition.

Project Goal

The primary objective is to develop a model that, given a user's purchase history, can accurately rank the products they are most likely to reorder. This project explores a progression of models, starting with a strong Factorization Machine baseline and moving towards more complex deep learning architectures.

---

Dataset

This project uses the "Instacart Market Basket Analysis" dataset, which contains over 3 million anonymized grocery orders from more than 200,000 users. The data is relational and split across six CSV files detailing users, products, orders, and their relationships.

---

Methodology

The recommendation task is framed as a supervised, pointwise Learning-to-Rank (LTR) problem.

1. Data Preprocessing:

   • The raw CSV files (orders.csv, products.csv, order_products__prior.csv, etc.) were merged to create a unified DataFrame of all historical user-product interactions.
   • Positive samples were identified as products a user reordered in their final train order.
   • Negative samples were created from products a user had purchased in the past but did not reorder in their final train order.

2. Train/Validation Split:

   • A user-based split (80% train, 20% validation) was performed.
   • This ensures that no user in the validation set appears in the training set, providing a robust test of the model's ability to handle the "user cold-start" problem.

---

Models Implemented

1. Baseline: Factorization Machine (Pointwise LTR)

A Factorization Machine was implemented as the baseline model. This model is designed to learn the interaction effects between users and products in a high-dimensional, sparse feature space.

• Framework: PyTorch
• Architecture:
  • torch.nn.Embedding layers were used to create dense latent vectors for users and products.
  • The model predicts a score for each (user, product) pair based on the FM formula (global bias + linear terms + dot product of latent vectors).
• Training Objective: The model was trained as a binary classifier using BCEWithLogitsLoss to predict the probability of a product being reordered.

---

Current Results

The baseline Factorization Machine was trained for 5 epochs and evaluated on the held-out validation set of users.

Metric	Score
Mean Precision@10	0.2089
Mean Recall@10	0.1528
MAP@10	0.1183
NDCG@10	0.2082

These results serve as the benchmark for all future, more complex models.

---

Future Work

The next steps in this project are to implement and compare more advanced models:

• Pairwise FM: Adapt the current model to use a pairwise loss (like BPR) to directly optimize ranking.
• Two-Towers Model: Build a two-towers architecture for efficient retrieval.
• DIN (Deep Interest Network): Implement an attention mechanism to model user interests based on their purchase history.
• Transformer-based Model: Use a sequential model like SASRec to capture the order of user purchases.

---

Setup and Usage

1. Clone the repository:

   git clone https://github.com/Dhyanesh18/instacart-recommender-benchmark
   cd instacart-recommender-benchmark

2. Install dependencies:

   pip install -r requirements.txt

3. Download the data:

   • Download the Instacart dataset from Kaggle.
   • Unzip and place the six CSV files into a folder named ./archive/.

4. Run the scripts:

   • Training: python train.py (This will train the baseline model and print training metrics).
   • Validation: python validation.py (This will run inference on the model and print evaluation metrics).