MetaGC-MC: A graph-based meta-learning approach to cold-start recommendation with/without auxiliary information

Abstract

Listen

Translate article

Collaborative filtering-based methods have achieved distinctive performance in ordinary recommendation tasks. However, they suffer from a cold-start problem when historical interaction is sparse. To carry out cold-start recommendations, many methods assume auxiliary data, such as user/item information and heterogeneous network information, being available. However, considering auxiliary data are not always available in practice, we explore a novel approach, MetaGC-MC, to alleviate cold-start recommendation issues, which can provide more effective cold-start recommendations, regardless of whether auxiliary data is available. MetaGC-MC is based on two emerging techniques, namely, graph convolutional network and meta-learning. In MetaGC-MC, a mechanism of stochastic enclosing subgraph sampling is introduced to randomly sample h-hop enclosing subgraphs as the meta-learning tasks for new users or new items. According to the γ-decaying theory, low-hop enclosing subgraphs contain enough information to learn good high-order graph structure information. Without relying on auxiliary data, MetaGC-MC can capture various subgraph structure information under a meta-learning framework, and encode learned information as a meta-prior that makes rapid adaptions in new subgraphs. MetaGC-MC can also take advantage of auxiliary data to enhance model performance. Extensive experimental results in three real-world datasets illustrate that MetaGC-MC is competitive with other state-of-the-art methods for user and item cold-start scenarios.