Yang Xu

Recent Projects

PI: National Natural Science Foundation of China

Research on Hash Representation Learning for Large‐scale Recommendations

With the explosive growth of Internet data, efficient recommendation is highly desired by both academic and industrial communities. Hashing is an important technique to support efficient recommendation. In general, recommendation systems are usually accompanied with large-scale fast-growing data (such as user interaction records, reviews and product descriptions), and users’ interests are distributed in the multi-modal data. However, existing hashing-based recommendation methods failed to take these important characteristics into account. As a result, they suffer from three key problems, i.e. hash representation of user interest, online update of hash code and interpretability of hash code. To solve these issues and fully consider the actual requirements of the recommendation system, we here investigate effective hash representation learning methods for large-scale recommendations. First, we will propose an efficient hash representation learning method for user interest based on multi-modal data fusion, which can discover the potentially valuable user interest from multi-model data with efficient discrete optimization. Second, we will propose an online update method for hash code, which can update the hash codes in real time through online hash learning. Last, we will propose a learning method for interpretable hash code. Our method can simultaneously make recommendations using hash codes while generate semantic interpretations for the recommendation results. Our research can substantially advance the study of hashing-based recommendation methods both theoretically and practically. Moreover, it can also serve as the key techniques for large-scale recommendation system.

PI: Natural Science Foundation of Shandong Province

Research on Dynamic Modeling Method of User Interest Based on Multi‐source Data

User interest mining is the key to provide personalized service for users. Current studies mainly focus on user interest feature extraction, interest timeliness modeling, yet no study tries to model the law of user interest dynamic development and the law of user interest transfer from the angle of user interest self-development. Our project is combined psychological research achievements, and fuses multi-source data to analyze user behavior characteristics, then design a kind of modeling method for the law of user interest inverted U development and the law of user interest transfer. The method is designed mainly according to the following two ideas: 1) Bring in the deep learning framework to fuse multi-model data, design the interest recognition algorithm based on multi-view fusion clustering, and present the increment learning algorithm of user interest dynamic development model based on inverted u curve; 2) Design the behavioral clustering algorithm which is continuous in time to identify user interest segments and present the multi-model LSTM method which is based on attention mechanism to model the law of user interest transfer. Based on public datasets, we design the recommendation framework based on user model to verify the effectiveness of user model. What’s more, we design disassemble experiments to verify the performance of each module of our model. Our research can substantially propel the user interest temporal model learning techniques, theoretically and applicably. Moreover, it can also offer the key algorithms for personalized service system.

PI: Natural Science Foundation of Shandong Province

Research on Key Technologies for Multimodal Content-Aware POI Recommendation

With the rapid development of consumer social platforms like Meituan and Dianping, point of interest (POI) recommendation systems have emerged as a crucial gateway for brick-and-mortar retail consumption, becoming core technologies for major commercial platforms. In recent years, actual POI recommendation scenarios have been characterized by limited spatial mobility of users, faster turnover of shops, and more cautious consumer choices. These trends have led to severe challenges such as sparse user check-in data, difficulty in recommending new POIs, and a lack of interpretability in recommendation results. Current POI recommendation methods focus only on spatiotemporal context data, overlooking the rich multimodal content information surrounding POIs, making it difficult to address the aforementioned challenges effectively. Therefore, this project proposes an innovative approach by introducing multimodal content information of POIs. It will conduct in-depth research in three key areas: modeling POIs based on multimodal content and context, a cold-hot collaborative POI recommendation model, and a general multimodal explanation framework for POI recommendations. The research will tackle key scientific problems such as multimodal knowledge fusion representation, collaborative enhancement for cold and hot POIs, and the generation of multimodal explanations for recommendation results. This project will significantly advance both theoretical exploration and practical application of POI recommendation systems, improving consumer decision-making efficiency, stimulating consumption potential, and driving economic growth.

Research

1. MMPOI: A Multi-Modal Content-Aware Framework for POI Recommendations
   Yang Xu, Gao Cong, Lei Zhu, Lizhen Cui
   The Web Conference (WWW), 2024 (CCF-A) | code

2. FUMMER: A Fine-grained Self-supervised Momentum Distillation Framework for Multimodal Recommendation
   Yibiao Wei, Yang Xu*, Lei Zhu, Jingwei Ma, Jiangping Huang
   Information Processing & Management (IPM), 2024 (Corresponding author, Q1 Journal, IF: 8.6)

3. Multi-level Cross-modal Contrastive Learning for Review-aware Recommendation
   Yibiao Wei, Yang Xu*, Lei Zhu, Jingwei Ma, Chengmei Peng
   Expert Systems with Applications (ESWA), 2024 (Corresponding author, Q1 Journal, IF: 8.5)

4. Temporal Social Graph Hashing for Efficient Recommendation
   Yang Xu, Lei Zhu, Jingjing Li, Fengling Li, Heng Tao Shen
   IEEE Transactions on Knowledge and Data Engineering (TKDE), 2023 (CCF-A, Q1 Journal, IF: 9.235)

5. Explainable discrete Collaborative Filtering
   Lei Zhu, Yang Xu*, Jingjing Li, Weili Guan, Zhiyong Cheng
   IEEE Transactions on Knowledge and Data Engineering (TKDE), 2022 (Corresponding author, CCF-A, Q1 Journal, IF: 9.235) | paper

6. Deep Discrete Hashing for Label Distribution Learning
   Zhen Zhang, Lei Zhu, Yaping Li, Yang Xu*
   IEEE Signal Processing Letters, 2021 (Corresponding author, CCF-C, Q1 Journal, IF: 3.201) | paper

7. Binary multi-modal matrix factorization for fast item cold-start recommendation
   Chengmei Peng, Lei Zhu, Yang Xu*, Yaping Li, Lei Guo
   Neurocomputing, 2021 (Corresponding author, CCF-C, Q1 Journal, IF: 5.779) | paper | matlab code

8. Multi-modal discrete collaborative filtering for efficient cold-start recommendation
   Yang Xu, Lei Zhu, Zhiyong Cheng, Jingjing Li, Zheng Zhang, Huaxiang Zhang
   IEEE Transactions on Knowledge and Data Engineering (TKDE), 2021 (CCF-A, Q1 Journal, IF: 9.235) | paper | matlab code

9. Temporal representation learning for time series classification
   Yupeng Hu, Peng Zhan, Yang Xu, Jia Zhao, Yujun Li, Xueqing Li
   Neural Computing and Applications (CCF-C, Q1 Journal, IF: 5.102) | paper

10. Long short-term memory with sequence completion for cross-domain sequential recommendation
    Guang Yang, Xiaoguang Hong, Zhaohui Peng, Yang Xu
    APWeb-WAIM'20 (CCF-C) | paper

11. Multi-feature discrete collaborative filtering for fast cold-start recommendation
    Yang Xu, Lei Zhu, Zhiyong Cheng, Jingjing Li, Jiande Sun
    AAAI'20 (CCF-A) | paper

12. Deeply fusing reviews and contents for cold start users in cross-domain recommendation systems
    Wenjing Fu, Zhaohui Peng, Senzhang Wang, Yang Xu, Jin Li
    AAAI'19 (CCF-A) | paper

13. Feature-based online segmentation algorithm for streaming time series
    Peng Zhan, Yupeng Hu, Wei Luo, Yang Xu, Qi Zhang, Xueqing Li
    CollaborateCom'18 (CCF-C) | paper

14. SARFM: A Sentiment-Aware Review Feature Mapping Approach for Cross-Domain Recommendation
    Yang Xu, Zhaohui Peng, Yupeng Hu, Xiaoguang Hong
    WISE'18 (CCF-C) | paper

15. Cross-Domain Recommendation for Mapping Sentiment Review Pattern
    Yang Xu, Zhaohui Peng, Yupeng Hu, Xiaoguang Hong, Wenjing Fu
    KSEM'18 (CCF-C) | paper

16. HOMMIT: A Sequential Recommendation for Modeling Interest-Transferring via High-Order Markov Model
    Yang Xu, Xiaoguang Hong, Zhaohui Peng, Yupeng Hu, Guang Yang
    WISE'17 (CCF-C) | paper

17. Efficient snapshot KNN join processing for large data using mapreduce
    Yupeng Hu, Chong Yang, Cun Ji, Yang Xu, Xueqing Li
    ICPADS'16 (CCF-C) | paper

18. Temporal recommendation via modeling dynamic interests with inverted-U-curves
    Yang Xu, Xiaoguang Hong, Zhaohui Peng, Guang Yang, Philip S Yu
    DASFAA'16 (CCF-B) | paper

19. Relevance search on signed heterogeneous information network based on meta-path factorization
    Min Zhu, Tianchen Zhu, Zhaohui Peng, Guang Yang, Yang Xu, Senzhang Wang, Xiangwei Wang, Xiaoguang Hong
    WAIM'15 (CCF-C) | paper

Awards

2022, ACM ICMR Outstanding Reviewer Award

2017, Scientific Research Outstanding Achievement Award

2016, The First Prize Academic Scholarship

2011, National Scholarship for Graduate Students

2010, National Encouragement Scholarship for Graduate Students