● Implemented GPU version GraphSage (forward propagation part); achieved 10X speedup compared with CPU OpenMP code. ● Designed a method that utilizes graph embedding, clustering, and semi-supervised learning with label propagation model, to solve the multiple accounts per user problem in social networks; achieved AUC at 0.95 and good scalability with a small percentage of ground truth;● Proposed a novel approach using both static and dynamic user activity features, to analyze attackers’ temporal behaviors, and to identify malicious URLs on social networks; achieved >90% accuracy. ● Parallelized the minimum step number (MSN) algorithm, by concurrently searching each layer of the solution space; achieved 56X speedup, and expanded the practically computable problem size by 5000X.

Proposed a novel approach for both targeted and non-targeted attacks on Graph Convolutional Networks (GCNs) by adding fake nodes; reduced accuracy of GCN down to 0.10 for non-targeted attack and achieve success attack rate up to 0.99 for targeted attack.Designed Greedy-GAN method to update adjacency and feature matrices, which generate fake features similar to real ones.

• Apply clustering on social network, including mobile money and call-detailed data • Analyze connections in above data based on geo-locations• Visualize mobile money activities and identify money flows within and between areas• Build logistic models using call-detailed data to predict banking behavior

• Analyzed call-detailed data for customer profiling and basic statistic using SQL• Built logistic predictive models for credit risk assessment

Guangdong Provincial Key Laboratory of Intelligent Transportation System, Sun Yat-sen; Helping with data analysis and mathematical modeling for KLIT. In this job, I analysis the travelpattern of people in Guangzhou. Guangzhou is the third biggest city in China, the daily traffic data is about 2 million.