GaECCL: Graph-assisted Edge-Cloud Collaborative Learning for Click-Through Rate Prediction

Click-through rate (CTR) prediction is widely studied for its commercial value in recommendation systems. CTR models have evolved from shallow to deep to capture high-order feature interactions and user behaviors. Recently, with the development of mobile devices, several works try to deploy models from cloud to device with concerns of privacy, real-time response and personalization. However, traditional edge-cloud collaborative paradigm suffers from performance degradation due to out-of-date embeddings during cloud updates. To solve this problem, we propose graph-assisted edge-cloud collaborative learning (GaECCL) method which uses heterogeneous information network(HIN)-based graph auto-encoder to force CTR model to learn predicted user-item interactions in advance. By reconstructing meta-paths in the HIN, embeddings with better quality are generated. This method is compatible with most deep CTR models. Experiments on three public datasets show that our method not only outperform the state-of-the-art model but also greatly alleviate the performance degradation.