Session-Based Transfer Learning for MOOC Recommendation

This repository contains code and experiments for exploring transfer learning techniques in the context of session-based recommendation for MOOCs (Massive Open Online Courses). The goal is to leverage large-scale interactions from public e-commerce and book datasets to improve recommendation performance on a smaller, data-scarce MOOC dataset (MARS) using techniques like pretraining, fine-tuning, adapters, and meta-learning (Reptile).

Project Overview

• Objective: Address the cold-start and data scarcity problem in MOOC recommendation.
• Approach:
  • Pretraining: Train a session-based recommender (SASRec) on large source datasets.
  • Transfer: Fine-tune the pretrained model on the target MARS dataset.
  • Methods: Standard Fine-tuning, Adapter Modules, and Meta-Learning (Reptile).

Datasets

Source Datasets (Pretraining)

• YOOCHOOSE: content from the RecSys Challenge 2015, representing e-commerce clickstreams.
• Amazon Books: User-item interaction data from the Amazon dataset.

Target Dataset

• MARS: A MOOC dataset used for evaluating session-based recommendations.

Repository Structure

Notebooks

1. Exploratory Data Analysis (EDA)

• 01_eda_yoochoose.ipynb: Load and inspect YOOCHOOSE dataset.
• 02_eda_amazon_books.ipynb: Load and inspect Amazon Books dataset.
• 03_eda_mars.ipynb: Load and inspect MARS dataset.

2. Data Processing

• 04_session_gap_and_timeline_analysis.ipynb: Analysis of session temporal gaps.
• 05_sessionize_and_prefix_target.ipynb: Consolidate interactions into session sequences and generate prefix-target pairs for training.
• 05B_build_tensor_dataset.ipynb: Prepare PyTorch-compatible TensorDatasets.

3. Modeling & Transfer Learning

• 06_build_sasrec_model.ipynb: Implementation and pretraining of the SASRec (Self-Attentive Sequential Recommendation) model.
• 07_transfer_to_mars.ipynb: Experiments on transferring the pretrained model to the MARS dataset.
• 08_baselines.ipynb: Comparison with baseline models.

4. Advanced Techniques

• 08_reinit_emb_grid.ipynb: Grid search experiments on embedding initialization strategies.
• 09_adapters.ipynb: Implementation of Adapter modules for parameter-efficient transfer learning.
• 10_reptile_meta.ipynb: Application of the Reptile meta-learning algorithm for finding better model initializations.

Key Directories

• src/: Shared utilities and model definitions.
• data/: Raw and processed data storage.
• models/: Checkpoints for pretrained and fine-tuned models.

Setup

pip install -r requirements.txt