Developed state-of-the-arts techniques in personalized recommendations.

In Oct 2019, NExT Center successfully completed the "Deep Learning for Recommender Systems" research project, in collaboration with Linksure Network, one of the top Internet Companies in China. Extending from prior works by Dr. Xiangnan He and Dr. Xiang Wang, this project aimed to design top features for deep recommender systems from various explicit and implicit users' feedback, as well as models that could optimize news recommendation.

The team had developed two state-of-the-art techniques on adversarial training and knowledge-aware reasoning, that enhances the pairwise ranking objectives and incorporated knowledge graph into the recommender system. As a result, the techniques outperformed the conventional models and further improved the robustness and explainability of the recommender systems.

In addition, the team also proposed a more expressive solution named High-order Attentive Factorization Machine to improve the precision of CTR (click-through rate) prediction by establishing expressive and informative cross features in recommender systems.

"Different personalized recommendation scenarios need tailored recommender models. It is important to analyze what data are available to us, what features are useful, and correspondingly design useful and practical recommender components." Explained Dr. Xiang Wang. "Thankfully, we have succeeded this problem statement and devised different recommender models that achieved state-of-the-art performance on our offline evaluations."

This project is funded under the AISG 100 Experiments Initiative and tremendous support by the management of Linksure Network.

Technical Papers published under this project:

-Adversarial Personalized Ranking for Recommendation

-Explainable Reasoning over Knowledge Graphs for Recommendation

-HoAFM: A Higher-order Attentive Factorization Machine for CTR prediction

Recent Posts

See All