What is it about?
This paper is an intel submission for ACM Recsys2022 Challenge. This year's challenge is focusing on fashion recommendations, dataset is 1 million online retail sessions resulted in a purchase. Each session is given as a sequence of item views with additional file of item features. This is a session-based recommendation problem, and our solution is based on Heterogeneous Graph Neural Network.
Featured Image
Photo by Bastian Riccardi on Unsplash
Why is it important?
Our solution is based on heterogeneous GNN and single model(before ensemble) score is around 0.2064 MRR. Final submission won 4th place. Different from other solutions, our solution is: 1) based on heterogeneous graph neural network(HGNN); 2) Embedding Fusion of item sequence and Knowledge graph. 3) A transfer-learning inspired pre-training + fine-tuning training strategy 4) Feature Engineering with anonymous fashion features 5) Intel provided HPO toolkit
Perspectives
Heterogeneous Graph Neural Network reveals underlying relation between nodes, which is showing a strong compatibility with session-based recommendation problem. There are bunch innovative works shared in this paper.
Chendi Xue
Intel Corp
Read the Original
This page is a summary of: SIHG4SR: Side Information Heterogeneous Graph for Session Recommender, September 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3556702.3556852.
You can read the full text:
Contributors
The following have contributed to this page
