Multi-task learning on heterogeneous graph neural network for substitute recommendation

Substitute recommendation in e-commerce has attracted increasing attention in recent years, to help improve customer experience. In this work, we propose a multi-task graph learning framework that jointly learns from supervised and unsupervised objectives with heterogeneous graphs. Particularly, we propose a new contrastive method that extracts global information from both positive and negative neighbors. By feeding substitute signal data from different sources to learning tasks with different focuses, our model learns the representation of products that can be applied for substitute identification under different substitutable criteria. We conduct experiments on Amazon datasets, and the experiment results demonstrate that our method outperforms all existing baselines in terms of comprehensive performance among all metrics of interest.