gnn_0.jpg
MIMS Final Project 2023

GNNIE: GNN-Recommender-as-a-Service

GNNIE is a recommender system powered by graph neural networks (GNN) that can provide more diverse, more personalized, and more interpretable recommendations.

Our team seeks to build an API service that enables e-commerce operators to connect their data, select features, customize parameters, and plug in an API to generate better recommendations on their websites.

GNN as a Recommender System

Until recently, many recommendation systems were based on two major approaches: Collaborative Filtering and Content-based Filtering. The former suffers from popularity bias, where the most popular products or content are recommended, instead of niche products. The latter suffers from a lack of novelty and diversity, as it can only recommend based on existing interests of the user.

In recent years, Graph Neural Networks (GNN) have been shown to be effective in a variety of applications, including recommendation systems. [2] We believe that this third approach, GNN-based recommendation models, has the potential to make recommendations more diverse (i.e., not merely popular at large, or related to a specific search) and better personalized (i.e., more aware of interests and behaviors of the user). It is also highly scalable and more expressive than other neural network models.

Further, it can leverage rich contextual information (e.g., date, location, weather) to provide more personalized recommendations. In short, GNN recommendation models have the potential to make recommendations better, faster, and explainable. It can help people discover what they want and need, but didn’t know they did, in the sea of products and content online today.

Improving GNN Recommender Systems

In the past few years, research has demonstrated the efficacy of graph learning methods for recommendation tasks. [3] For example, Uber Eats has achieved a 12% increase in AUC compared to the baseline model by leveraging GNN. [4]

Separately, there has been research in the field of quantum chemistry that used edge contextual information in GNN. [5] Uber Eats recommendation model did not incorporate any contextual edge information in user-item bipartite graphs, only using ratings as the edge information. DiDi, the ridehailing company, has applied contextual information on physical environments to GNN and saw significant improvement in traffic predictions. [6]

We believe that rich contextual information could be applied to GNN-based recommendation engine for retail products in order to reduce the popularity bias and increase the representation of long-tail recommendations, resulting in more personalized and diverse recommendations.

Democratizing GNN Recommender Systems

Despite this potential, GNN recommendation models today are not yet accessible to most companies. Barring the biggest companies, most cannot develop customized GNN models for their own businesses. Making GNN recommendation models more widely accessible can not only help these companies compete, but more importantly, help people find the right product, person, or content at the right moment.

We believe one of the sectors that could benefit the most immediately is retail e-commerce. It has a vast number of smaller companies, and it has long been using recommendation engines. In the US, there are more than 4 million e-commerce companies, and by some measures, the top 10% accounts for the 90% of gross merchandise value. [7] GNNIE seeks to help the bottom 90% by leveraging GNN and rich contextual information to provide better product recommendations.

References

  1. Alejandro Bellogín, Pablo Castells ,Iván Cantador (2017). Statistical biases in Information Retrieval metrics for recommender systems. Information Retrieval Journal 20, 6 (2017), https://doi.org/10.1007/s10791-017-9312-z 
  2. Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, Bin Cui (2020). Graph Neural Networks in Recommender Systems: A Survey. ACM Computing Surveys. https://dl.acm.org/doi/10.1145/3535101 
  3. Vittal, B. (2022, March 16). Dissecting the $4.9 Trillion industry with 2022 data. PipeCandy Blog. https://blog.pipecandy.com/post/e-commerce-companies-market-size 
  4. Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton and Jure Leskovec. “Graph Convolutional Neural Networks for Web-Scale Recommender Systems.” KDD (2018)
  5. (2019, December 4). “Food Discovery with Uber Eats: Using Graph Learning to Power Recommendations.” Uber Blog. https://www.uber.com/blog/uber-eats-graph-learning/
  6. Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O. and Dahl, G.E. “Neural message passing for quantum chemistry.” International conference on machine learning, 1263-1272 (July 2017).
  7. Wenjuan Luo, Han Zhang, Xiaodi Yang, Lin Bo, Xiaoqing Yang, Zang Li, Xiaohu Qie, and Jieping Ye. 2020. “Dynamic Heterogeneous Graph Neural Network for Real-time Event Prediction.” KDD '20. 3213–3223.

Last updated:

May 15, 2023