Shen Fan(范深)
I am a Staff Algorithm Engineer, at Alibaba Group.
Graduated from School of Computer Science and Engineering,
University of Electronic Science and Technology of China(UESTC), Chengdu, Sichuan, P.R.China. Supervised by Prof.Tao Zhou.
My expertise lies in network analysis and graph mining, especially in user modeling, recommendation.
Currently, my interests focus on Multimodal LLMs and AI Agents towards AGI(Artificial General Intelligence).
Email: ofanshen at gmail.com
Research Interests
Graph Mining: Link Prediction, Graph Representation
AGI: Multimodal Large Language Model, AI Agents
Experiences
Alibaba Data, Hangzhou | August 2017 - Present
Lead a team of algorithm engineers in creating a cross-platform user identification and linking system, OneID.
Build a large-scale social network and conducted its applications in social recommendation, user growth and marketing.
R&D on AI4System, build large foundation models on meta-graph of massive data work-flow to optimize offline-database(ODPS) management.
Tmall Tech, Hangzhou | March 2014 - August 2017
Developed algorithms for item sales prediction and online personalized product ranking.
Created a Topic-based Recommendation System.
Projects
One-World: Multi-level cross-platform user identification in E-Commerce Platform
Related Techniques: Link Prediction, Community Detection
Real-world graphs often exhibit noise and incompleteness due to limitations in data collection.
In order to gain a deeper understanding of our users, we propose a tri-level identity resolution approach known as 'Device-Person-Household’, supporting for a accurate targeting and effective user engagement strategy.
Link prediction techniques like CN, AA, RA(Zhou et al., EPJ’09) and SEAL(Zhang et al., NeuIPS’18) are employed to establish missing connections and classify relations.
Datasets and Codes:[EasyLink]
InfoNet: Modeling Information Diffusion in Social Networks
Related Techniques: Graph Neural Networks, Social Recommendation
Our research has demonstrated that social relations on Taobao exhibit homophily(the tendency of individuals to associate and bond with similar others), which can serve as an additional data source for user behavior modeling (Huang et al., SIGIR’21) and item sales modeling (Yang et al., TKDE’22).
To further boost the power of social relations in the recommendation, we employ graph structure learning(GSL) techniques on the social graph to generate superior recommendations for inactive users[Comming Soon], and introduce GraphTranslator(Zhang et al., WWW’24) to conduct open-ended tasks on understanding users interests in social context with GNN(Graph Neural Network) and LLM(Large Language Model).