Doing AI at Meta discovery recommender systems.

Lead machine learning team for Notification Retrieval & Ranking @Nextdoor.

Build large-scale recommender systems and develop machine learning algorithms for Ads ML and Notification ML @Nextdoor.Notification ML Team • Lead the push and email volume personalization - improved user-engaged sessions by 12%+ and weekly active users (WAU) by 5%+ (most significant user engagementwin in the company's history). • Developed the very first deep learning models (Transformers, DCN, etc.) for notification pCTR rankingAds ML Team • Lead the first effort to convert and migrate rule-based ads ranking logic to ML-based ads ranking, leading to +5% SMB total ads clicks in Feed and +12% ads clicks for other surfaces combined.

Research and Apply large-scale deep Graph Neural Networks (GNN) at LinkedIn.

I worked as a Machine Learning Engineer in Knowledge Graph and Data Standardization and Knowledge Graph Team at LinkedIn.

Data Standardization and Knowledge Graph TeamI am working on "Automated Java Code Flaw Detection, Modification and Suggestion for Code Review Process" project using NLP, machine learning and deep learning techniques. It is a brand new subject in NLP area and I've done all things starting from scratch, so far including:• Generated over 2 billion records of training data (over 130 GB) from scratch in one week, and normalized the highly-noisy data to generate 3 high-quality datasets for research topics;• Designed the storage and access schema of data and improved the efficiency of data access by over 5 times;• Caught deep insight of data by running Latent Dirichlet Allocation (LDA) model and calculating inclusive statistics;• Implemented and evaluated various Machine Learning and Deep Learning models including Bag-of-Words, N-Gram, Logistic Regression, CNN, LSTM + Self-Attention, Sequence-to-Sequence + Attention, and improved the top-10% recall over 20%.

I worked as a software development intern in Laboratory of Artificial Intelligence and Machine Learning Institute of Automation, Chinese Academy of Sciences, Beijing, China.I got involved into the development of DeepQA system in Medical Application based on NLP Technology.During the internship, I focused on the following things:1. Investigated implementation process and method of NLP, analyzed and compared practical solutions of IBM Watson to optimize design scheme for our development project.2. Explored and tested different word representation methods including one-hot feature, TF-IDF, and word embedding.3. Implemented classification models including SVM, Naive Bayes, CNN, LSTM and GRU, and evaluated the models on datasets containing 24,420 records of texts, and finally improved the multi-class classification accuracy from 61% to 73%.

I worked as a software Engineering intern in State Key Laboratory of Software Development Environment, Beihang University, Beijing, China.I participated in the algorithm research and application status of feature extraction and pattern matching in iris recognition.During the internship, my duties are as followings:1. Investigated and analyzed advantage and disadvantage of each algorithm, and then make development scheme.2. Directed Java development of iris gray scale image preprocessing and image intensity normalisation3. Implemented matching algorithm after iris feature extraction 4. Demonstrated the feasibility of the configuration and the effectiveness of the algorithm

I worked as a research assistant in Laboratory of Computer Graphics and VisualizationDepartment of Computer Science and Technology, Tsinghua University, Beijing, China.My subject is "fully automated macular pathology detection in retina OCT images" found by National Natural Science Foundation of China (NSFC).During the research, my work includes following items: 1. Proposed a novel framework for ML-based automated detection of retinal diseases from retina OCT images, by using SVD-based sparse coding and dictionary learning as well as methods deep-learning based methods (CNN);3. Evaluated the model in Duke Spectral Domain OCT (SD-OCT) dataset and clinical SD-OCT dataset and Improved the classification accuracy of state-of-the-art methods by nearly 7%4. Published the work "Fully automated macular pathology detection in retina optical coherence tomography images using sparse coding and dictionary learning." in Journal of biomedical optics in 2017

I participated in the frontend and backend Development of O2O Sale and the Service Website of Elevators and Components1. Developed static and dynamic, interactive web pages with HTML5, JavaScript and jQuery 2. Implemented back-end elevator components intelligent recommendation module by employing Naïve Bayes model fed with data obtained by tracking the service life of elevators and components purchased by customer