KR-GCN

.

Background

previous study:

1. error propagation: consider all paths between every user-item pair might involve irrelevant one
2. weak explainability

Model

image-20230426111943016

4 parts:

1. the Graph encoding module: learn the representations of nodes in the heterogeneous graph.
2. the Path Extraction and Selection module: extract paths between users and items from the heterogeneous graph and select higher-quality reasoning paths
3. the Path Encoding module: learn the representations of the selection reasoning paths.
4. the Preference Prediction module: predicts users’ preferences according to the reasoning paths.

Graph Encoding - GCN

1. propagation and aggregation

initialized randomly

weighted sum aggregator : the neighborhood nodes are aggregated via mean function.

image-20230426113233853

2. weight sum to merge every layers

image-20230426114004010

Path Extraction

we prune irrelevant paths between each user-item pair.

we extract multi-hop paths with the limitation that hops in every single path are less than l.