DSC180B Capstone Project on Graph Data Analysis

Project Website: https://nhtsai.github.io/graph-rec/

Amazon Product Recommendation using a graph neural network approach.

• dask
• pandas
• torch
• torchtext
• dgl

Amazon Product Dataset from Professor Julian McAuley (link)

• Product Reviews (5-core)
• Product Metadata
• Product Image Features

The graph is a heterogeneous, bipartite user-product graph, connected by reviews.

• Product Nodes (ASIN)
  • Features: title, price, image representation
• User Nodes (reviewerID)
• Edges (user, reviewed, product) and (product, reviewed-by, user)
  • Features: helpful, overall

We use an unsupervised PinSage model (adapted from DGL).

• name: model configuration name
• random-walk-length: maximum number traversals for a single random walk, default: 2
• random-walk-restart-prob: termination probability after each random walk traversal, default: 0.5
• num-random-walks: number of random walks to try for each given node, default: 10
• num-neighbors: number of neighbors to select for each given node, default: 3
• num-layers: number of sampling layers, default: 2
• hidden-dims: dimension of product embedding, default: 64 or 128
• batch-size: batch size, default: 64
• num-epochs: number of training epochs, default: 500
• batches-per-epoch: number of batches per training epoch, default: 512
• num-workers: number of workers, `default: 3 or (#cores - 1)
• lr: learning rate, default: 3e-4
• k: number of recommendations, default: 500
• model-dir: directory of existing model to continue training
• existing-model: filename of existing model to continue training, default: null
• id-as-features: use id as features, makes model transductive
• eval-freq: evaluates model on validation set when epoch % eval-freq == 0, also evaluates model after last training epoch
• save-freq: saves model when epoch % save-freq == 0, also saves model after last training epoch

• GraphSAGE Homepage
• GraphSAGE Research Paper
• PinSage Article
• PinSage Research Paper