My current research is mainly focused on designing friend suggestion / location recommendation system based on user location history in location-based social network

Our paper is accepted to SIGKDD'18 conference. Code is available HERE. Check our poster below:

Previously I worked on developing topic-based methods for academic search and recommendation, which is the theoretical part of our Academic Search Engine. I am also interested in other machine learning related topics (eg. deep learning), data mining, system design and implementation. 

An Interactive Topic Model for academic recommendation

I worked to construct a topic-based model to recommend scientific publications.

We have made such contributions:
  • We provide an interactive topic model with tree-structured priors and encode user feedback into the prior tree.
  • We significantly increase the computational efficiency by adopting similar mechanism like SparseLDA. 
  • We propose a crowdsourcing framework for recommending publications and further modify our interactive topic model     to a collaborative version. In this scenario, users with similar interests can fix a shared prior tree, which promotes article recommendation in related topics.


Android Online Malware Detection Based on Deep Learning Methods

I participated in the 8th National Information Security Competition in August 2015. My team members and I constructed an Android online malware detection website, which featured on deep learning methods. Our main contributions include:
  • Combined static and runtime detection and extracted features of both types with DroidBox  
  • Proposed a two-layered classification model, which included a recurrent neural network used for training run-time features with different dimensions, as well as a multilayer perceptron used for final classification
  • Constructed an online detection website which enables users to upload apk files and return analysis report (it's a pity that the website server stopped working now.)
Our classification method achieved higher accuracy for detecting Android malware. This is because typical patterns (sequences of behaviours) of malicious applications could be effectively detected by runtime detection and recognized or remembered by recurrent neural networks. 

In the final contest, our team won national second prize. Also, we have submitted patent application for our work. 
(Application No. 201510667016.2).