Session-based Recommendation with Graph Neural Networks

Zongyue Qin

Oct 4 2022

Background

• When online shopping, people usually click several items before they decide which items to purchase.
• Many recommender systems rely on user profiles to do recommendation, which might be unavailable in practice.
• Therefore, recommending items solely based on previously visited items is very useful for online shopping platform.

Problem Definition

Non-GNN Solutions

• Conventional Methods
• Markov Chains: Past components are combined independently, such independent assumption is too strong in practice.
• K-NN: Strong baselines. Competitive performance with SOTA models.
• RNN-based Methods (GRU4REC, STAMP, etc.)
• Tend to overlook the complex transitions among distant items.
• The unidirectional assumption is too strong, people can go back-and-forth in a session.

GNN Solutions

• SR-GNN (Session-based Recommendation with Graph Neural Networks, AAAI-19)
• GCE-GNN
• DHCN

SR-GNN: Session Graphs

SR-GNN: Learning node Embeddings on Graph

• As each node represents an item, the input node features are the corresponding item embeddings.
• Then the node embeddings will be updated using a Gated Graph Neural Network (GGNN). Formally, the update function is shown on the right.

SR-GNN: Generate Session Embeddings

SR-GNN: Making Recommendation and Training

SR-GNN: Sum-up

• One of the earliest papers to use GNN for session recommendation
• Use GGNN and attention mechanism to get session embeddings.
• Use cross entropy to optimize the model.

GNN Solutions

• SR-GNN
• GCE-GNN
• DHCN

Global Context Enhanced Graph Neural Networks for Session-based Recommendation

• SR-GNN heavily rely on the relevance of the last item to the user preference of the current session.
• Previous studies only consider the current session while ignoring the useful item-transition patterns from other sessions.
• Therefore, they propose a novel approach to exploit the item-transitions over all sessions explicitly.

Session Graph and Global Graph

• GCE-GNN has two types of graphs to capture item transition information.
• The session graph is basically same to SR-GNN, except that they add a self-loop for each item and there are four type of edges: in, out, in-out, self.

Global Graph

GCE-GNN: Overview

Global Level Item Representation

Session Level Item Representation

Session Representation

Prediction Layer

• The prediction layer is the same as in SR-GNN

Experiments

Experiments

Experiments

• GCE-GNN-NP: GCE-GNN with forward positional encoding
• GCE-GNN-SA: with self-attention instead of the position-aware attention.

GNN Solutions

• SR-GNN
• GCE-GNN
• DHCN

Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation

• GNN-based methods regard item transitions as pairwise relations, which neglect the complex high-order information among items.
• Hypergraph provides a natural way to capture beyond-pairwise relations.
• To overcome data sparsity, they introduce line graph channel and integrate self-supervised learning.

Definition 1. Hypergraph

Definition 2. Line graph of Hypergraph

Hypergraph Construction

• Each session is represented as a hyperedge.

DHCN: Overview

Hypergraph Convolutional Networks

• The node representation is updated by aggregating its neighbors

​

​

• It can be written in matrix form with row normalization

​

​

• The final node representation is the average of node representation of each layer.
• The positional embeddings are added into item representation in the same way as in GCE-GNN.
• The session embedding is obtained in the same way as in SR-GNN.

Enhancing SBR with Self-Supervised Learning

• Maximize the mutual information between hypergraph representations and line-graph representations.
• The line graph can be seen as a simple graph which contains the cross-session information and depicts the connectivity of hyper-edges.
• Since each node in the line graph represents a session, the input features are the average of item embeddings in the session.
• The embeddings is computed using GCN.

Enhancing SBR with Self-Supervised Learning

Experiments

Ablation Studies

• DHCN-P: without positional encodings
• DHCN-NA: without self-attention mechanism

Ablation Studies

Refenrence

• Wu, Shu, et al. "Session-based recommendation with graph neural networks." Proceedings of the AAAI conference on artificial intelligence. Vol. 33. No. 01. 2019.
• Wang, Ziyang, et al. "Global context enhanced graph neural networks for session-based recommendation." Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 2020.
• Xia, Xin, et al. "Self-supervised hypergraph convolutional networks for session-based recommendation." Proceedings of the AAAI conference on artificial intelligence. Vol. 35. No. 5. 2021.