Projects

Working Projects

1. Suspicious Behavior Modeling in App Markets

Introduction

Mobile App Markets like App Store and Google Play involves many fraudsters like spammers, botnets and crowd workers. We investigate the underground market of trading app downloads and reviews, design classifers with multi-view and multi-source information according to the intention of the fraudsters. We also want to develop scalable and robust graph based classfication models.

Resources

My intro slides about graph based review spam detection, Meng Jiang’s survey on suspicious behavior modeling, Srijan Kumar’s survey on online false information study.

2. Securing Graphical Classfication Model

Introduction

The long lasting campaign between the fraudsters and online review platforms like Yelp and TripAdvisor makes the security of classifiers become very important. We aim to improve the robustness of graphical classfiers like message passing algorithm and graph convolutional neural network against various kinds of adversarial examples.

Resources

A survey from our lab on adversarial attack and defense on graph data, KDD18 best paper on adversarial attack on neural networks for graph data, ICML18 paper on adversarial attack on graph structured data.

Previous Projects

1. Robust Influence Maximization Based on Edge Parameter

Introduction

We evaluate the robust performance based on the edge parameter calculated by the node centrality. It shows that PageRank algorithm has the optimal influence spread and running time among four centrality measurement methods. We also evaluate the edge parameter in different experiment sets.

Resources

Paper and Code

2. CPS Model in Opinion Governance and Opinion Collection Model

Introduction

We design a Cyber-Physical-Society triple domain model to simulate real-world crisis responding situation. By the risk classification method, we have built an efficient CPS Model-based simulated emergency accident replying and handling system. It has been proved useful in some real accidents in China in recent years.

Resources

Paper and Slides

3. GBRT Model with Time Factor in Job Recommendation

Introduction

We introduce a job recommendation model based on Gradient Boosting Regression Tree and time factors (T-GBRT). The T-GBRT model aggregates the time factors into the GBRT to predict personal preferences and adds time factor weight to top-K rankings, with a neighbor based filtering trick in reducing the amount of calculation. The model performs the best in the experiment with four criterions, comparing to other three models, which proves the efficiency of the new model.

Resources

Paper

4. A Location-based Picture Recommendation Android App

Introduction

A programming project of Mobile App Development course. The app is developed based on Android 5.0 and Bomb cloud service.

Resources

Code