In the realm of energy data management and user engagement, a team of researchers from ByteDance’s Recommendation Systems and Graph Learning team has developed a novel framework that could potentially revolutionize how energy-related data is processed and utilized. The team, comprising Rui Xue, Shichao Zhu, Liang Qin, Guangmou Pan, Yang Song, and Tianfu Wu, has introduced E2E-GRec, an end-to-end training framework that integrates Graph Neural Networks (GNNs) with recommender systems.
Traditionally, GNNs are used to model complex relationships within data, such as user-item interactions in recommender systems. However, the conventional approach involves a two-stage process where GNNs are pre-trained offline to generate node embeddings, which are then used as static features for downstream recommender systems. This method has two main drawbacks: it requires repeated, large-scale GNN inference to refresh embeddings, leading to high computational overhead; and it lacks joint optimization, as the recommender system’s feedback cannot directly influence the GNN learning process, resulting in suboptimal embeddings for the recommendation task.
E2E-GRec addresses these limitations by unifying GNN training with the recommender system in a single, end-to-end framework. The framework consists of three key components: efficient subgraph sampling from large-scale, cross-domain heterogeneous graphs to ensure scalability and efficiency; a Graph Feature Auto-Encoder (GFAE) that serves as an auxiliary self-supervised task to guide the GNN in learning structurally meaningful embeddings; and a two-level feature fusion mechanism combined with Gradnorm-based dynamic loss balancing, which stabilizes graph-aware multi-task end-to-end training.
The researchers evaluated E2E-GRec through extensive offline evaluations and online A/B tests on large-scale production data. The results demonstrated significant improvements across multiple recommendation metrics, such as a 0.133% relative improvement in user stay duration and a 0.3171% reduction in the average number of videos a user skips. These findings suggest that E2E-GRec could potentially enhance user engagement and satisfaction in energy-related platforms and services.
The research was published in the Proceedings of the ACM Web Conference 2023, a prestigious conference in the field of web and data sciences. The practical applications of E2E-GRec in the energy sector could include improving energy consumption recommendations, optimizing energy management systems, and enhancing user engagement in energy-related platforms. By providing more accurate and personalized recommendations, E2E-GRec could help users make more informed decisions about their energy consumption, ultimately leading to more efficient energy use and reduced carbon emissions.
This article is based on research available at arXiv.