Abstract: In order to overcome the limitation of the cold start problem in the recommendation process on the performance of new users or new project scenarios, a meta-learning based graph neural network cold start recommendation model, namely MetaNGCF, is proposed to improve the accuracy and diversity of user-to-recommendation. Firstly, a perceptual meta-learning structure with adaptive properties is proposed to construct a model with a hybrid user-project interaction graph and neural graph, which expresses user behavior and project knowledge in a unified way. This structure incorporates an adaptive weighted loss strategy to correct the meta-learning paths in real time, in order to avoid the damage caused by noisy tasks on the model. Secondly, a clustering algorithm is applied to transform the high-dimensional feature space into a low-dimensional low-rank feature space. User preference learning is utilized to task aggregation layer gradient to encode the collaborative signals and automatically gene- ralize the higher-rank connectivity between users and projects, which in turn captures the NGCF general knowledge semantics. Finally, the results are validated in comparison with existing MetaHIN algorithm, and the results show that MetaNGCF has better performance on Recall@20 and NDCG@20.

Key words: meta-learning, cold start recommendation, collaborative filtering, graph neural network