Dynamic Offset Metric on Heterogeneous Information Networks for Cold-start Recommendation

The cold start problem is a significant challenge in recommendation systems. Traditional methods are ineffective when the amount of interaction data is small. Further, as meta-learning has achieved increasingly remarkablesuccess in few-shot classification, some studies in recent years has abstracted cold-start recommendations into few-shot problems and applied meta-learning-based approaches, but mostly, simple transplants of generic approaches have been adopted. Owing to the differences in problem definition, the more effective meta-learning methods represented by metric learning are not directly adaptable to the rating prediction problem. Furthermore, a heterogeneous information network (HIN), as a high-order graph structure, can capture more semantic information even in data-starved conditions. From the data perspective, efficient use of HINs can alleviate the cold-start dilemma. Hence, this study successfully combines metric learning and HIN to propose OMHIN, a Dynamic Offset Metric approach to Heterogeneous Information Networks for solving cold-start recommendations at both the model structure and data representation levels. We transform a direct similarity metric into an indirect metric to improve the robustness of our model. The flexible application of one-dimensional convolution enables OMHIN to integrate the rich information contained in the HIN effectively while avoiding the introduction of excessive noise. Experiments on two datasets validate that OMHIN achieves state-of-the-art performance in a variety of scenarios, with significant improvements in complex and difficult scenarios, and is therefore especially well-suited for use in sequence cold-start recommendations.