Multi-Channel Graph Neural Network for Recommender Systems
Recommender systems are one of the most powerful machine learning applications in e-commerce today as they are able to predict users’ interests and suggest relevant items to them. Knowledge graphs, which are models of a knowledge domain created by subject matter experts with the help of intelligent machine learning algorithms, play an important role in this. However, because different knowledge graphs can be structured differently, entity alignment is needed to identify their common entities in order for two or more knowledge graphs to be used concurrently. Current approaches to entity alignment include embedding-based models that learn entity embedding to capture the semantic similarities between entities in the knowledge graphs. These continue to be hindered by the heterogeneity of structures, as different knowledge graphs can differ so much that it misleads the representation learning and has limited seed alignments, which are existing alignments used as training data.
The ability to achieve effective entity alignment among different knowledge graphs can be a powerful tool in improving the prediction accuracy of recommender systems. This would help businesses capitalise on the global e-commerce market, which is expected to reach close to US$6 trillion by growing at a CAGR of 16.8% between 2017 and 2022.
This invention relates to a novel Multi-Channel Graph Neural Network (MuGNN) model that can encode different knowledge graphs to learn alignment-oriented embeddings. Comprising relation weighting, a multi-channel GNN encoder and an align model, MuGNN performs knowledge graph inference and alignment jointly to explicitly reconcile two main structural differences: missing relations and exclusive entities. Through this method, missing relations are completed by rule inference and transfer, while exclusive entities are pruned by cross-knowledge graph attention. Different channels are combined via pooling techniques, which enhances entity alignment with reconciled structures and the more efficient and effective use of seed alignments. Between knowledge graphs, each channel transfers structure knowledge using shared parameters.
In addition to powering e-commerce, MuGNN has application potential in digital advertising services where it can help predict audience preferences and click-through rates, as well as in business analytics applications to interpret specific user preferences.