UBER-GNN: A User-Based Embeddings Recommendation based on Graph Neural Networks

Bo Huang, Ye Bi, Zhenyu Wu, Jianming Wang, Jing XiaoarXiv 2020 · arXiv

Introduction

Most session-based recommendation systems rely exclusively on the immediate sequence of user clicks to predict the next action. While this approach is effective for low-stakes impulse purchases, it often fails in "high-stakes" environments like financial services. In applications such as Ping An Jinguanjia, users frequently navigate between insurance, investment products, and physical commodities simultaneously. In these contexts, a user's long-term profile—including demographics, financial history, and health status—is often a more reliable predictor of intent than their last few interactions.

UBER-GNN (User-Based Embeddings Recommendation with Graph Neural Networks) addresses this by bridging the gap between long-term user profiles ("who" the user is) and short-term session behavior ("what" the user is doing).

The UBER-GNN Framework

The architecture of UBER-GNN is designed to capture both static characteristics and dynamic transitions through three primary components:

  1. User Context Embedding (DeepFM): The model utilizes a DeepFM (Deep Factorization Machine) architecture to process "portrait data." This includes 306 features (212 categorical and 94 continuous) such as age, financial behavior, and consumption habits. DeepFM captures both low-order and high-order feature interactions to create a robust latent representation of the user.
  2. Session Graph Representation (GGS-NN): Rather than treating a session as a linear list, UBER-GNN converts interactions into a directed graph. This captures complex transitions and item revisits. A Gated Graph Sequence Neural Network (GGS-NN) processes this graph to learn sophisticated item embeddings.
  3. Hybrid Attention Mechanism: An attention layer weighs session items against the user's long-term profile. This mechanism effectively filters "noise" by determining which recent clicks are most relevant to the user's established persona.
The detailed technical architecture of UBER-GNN, illustrating the integration of DeepFM for user features and GGS-NN for session graphs.

Experimental Setup and Datasets

The model was evaluated using 12 months of real-world transaction data from the Ping An Jinguanjia app. The researchers developed two distinct datasets to test the model's versatility across different product types:

  • financial-include: A comprehensive dataset containing both virtual financial products (e.g., insurance) and physical commodities.
  • financial-exclude: A subset focused strictly on physical commodities to evaluate performance in a traditional e-commerce context.
Dataset statistics for the financial-include and financial-exclude benchmarks.

Key Results and Insights

UBER-GNN demonstrated superior performance over several strong baselines, including RNN-based (GRU4REC) and previous state-of-the-art graph-based models (SR-GNN).

  • Accuracy Gains: On the financial-include dataset, UBER-GNN achieved a Precision (P@20) of 77.91%, significantly outperforming SR-GNN (74.64%).
  • Ranking Performance: The model reached an MRR@20 of 26.66%, the highest among all tested frameworks.
  • Rationality in Finance: The results suggest that financial consumers are more "rational" than casual shoppers. Static user features proved highly predictive, and the hybrid approach successfully used these features to guide session-based predictions.

Conclusion and Future Work

UBER-GNN proves that integrating long-term user portraits with graph-based session modeling creates a more resilient recommendation engine for complex domains. While the results are promising, future work is needed to address the computational overhead of training dual architectures (DeepFM and GGS-NN) and to improve performance for "cold start" users who lack extensive historical portrait data.

Summary updated

BibTeX
@misc{huang2020uber,
  author = {Bo Huang and Ye Bi and Zhenyu Wu and Jianming Wang and Jing Xiao},
  title = {{UBER-GNN: A User-Based Embeddings Recommendation based on Graph Neural Networks}},
  year = {2020},
  eprint = {2008.02546},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2008.02546},
}
Figure