Stock Analysis and Clustering: Collected and analyzed stock trends, market concepts, technical indicators, and social media popularity. Utilized Dynamic Time Warping to accurately calculate the similarity of time-series data. Applied spectral clustering and Gaussian mixture models to cluster stocks. Developed alpha factors based on the principle of stock linkage using C++.Consensus Forecast Optimization: Combined time-weighted averaging and outlier exclusion. Assigned weights based on analysts’ historical forecast accuracy to optimize the calculation of consensus expectations. Developed alpha factors using C++.Performance Cycle Analysis: Analyzed the periodicity of companies’ historical performance. Combined this with analysts’ expectations to determine the current position in the performance cycle and predict future performance. Developed alpha factors based on performance cycle expectations using C++.Alpha Factor Development: Identified groups of stocks with similar financial performance through hierarchical clustering analysis and fundamental data. Developed alpha factors using C++.Automated News Aggregation System: Developed an automated system to aggregate company news sources. Used NLP techniques for keyword extraction, Named Entity Recognition, and news deduplication. Associated news with relevant stocks through Named Entity Recognition using Python.（Python）。

Microsoft - Applied Data ScientistProject Introduction: Enhancing the click-through rate (CTR) for new users on Bing's video streaming platform.Project Responsibilities: Content Quality Selection: Identified top content creators by analyzing multiple indicators such as YouTube vloggers' fan base, view counts, likes, comment volume, update frequency, and the speed of like growth on their videos, thereby improving the quality of top-tier content on the platform. Diversity in Video Recommendations: Developed and implemented a rule-based and collaborative filtering algorithm for initial ranking, significantly enhancing the diversity and novelty of recommended videos by analyzing user interaction data including click-through rates and actual video watch times. User Profile Enrichment: Created a cookie-based user matching system that integrates browsing data from the MSN platform, effectively enriching user profiles and improving recommendation accuracy. Video Fine-Tuning: Utilized the Light GBM model for the fine-tuning of recommended videos, optimizing model parameters based on the analysis of users' historical interactions and video characteristics to promote top video content. Real-time Feedback Mechanism: Designed and deployed a real-time feedback system capable of swiftly adjusting recommended content based on users' latest clicks, thereby enhancing user satisfaction. Optimization for News Videos: Developed a real-time crawler to monitor Google Trending for the latest trending search articles. Employed pre-trained NLP models to calculate embeddings for video titles and descriptions, and article embeddings to retrieve the most relevant videos through an ANN model. Applied the Light GBM model for precise ranking, significantly improving the timeliness and relevance of news videos.Outcome: Achieved an 8% increase in click-through rate (CTR).

Huolala - Recommendation Algorithm EngineerApp Popup Ad Recommendation SystemProject Introduction: Aimed to enhance ad click-through rate (CTR), conversion rate, and business revenue while ensuring balanced ad exposure across various business units.Project Responsibilities:Data Analysis and Feature Engineering: Conducted in-depth analysis of user, ad, and tracking data, extracting key features of users and ads through feature engineering and expanding dimensions with feature crossing techniques to bolster model predictive power.Data Management and Cold Start Mitigation: Leveraged Flink and Hive for real-time and historical data synchronization. Implemented trials and Lookalike strategies to improve new user and ad matching efficiency.Feature Optimization and Backend Development: Enhanced XGB model by selecting key features and managing ad exposure with dynamic thresholds. Led Java backend algorithm development for improved efficiency and stability.Monitoring and Collaboration: Ensured online-offline feature consistency and model updates for accuracy. Partnered with transaction and advertising departments to enhance ad conversions and recommendations, significantly increasing click-through and conversion rates.Ad Analysis and Project Execution: Conducted data analysis and developed models and algorithms to guide ad design with user preferences. Showcased project management and technical skills by independently executing these processes.Outcomes: Performance Improvement: Achieved a 3-percentage point increase in click-through rate (from 12% to 15%), along with a monthly increase in order revenue of approximately 8 million, significantly enhancing the company's profitability. Leadership Recognition: The project's success received high praise from the CEO, becoming a key example of driving the optimization of the company's advertising strategy.

Hong Kong University of Science and Technology - Research Assistant in Natural Language Processing Conducted in-depth research on over 30 natural language processing (NLP) models, including BERT, GPT-2, and Transformer, culminating in a comprehensive report. Leveraged this expertise to deliver educational lectures to laboratory students, enhancing their understanding of advanced NLP techniques.

+ Extend DeepMoji model to predict the positive/negative tendency of comments on Yelp/IMDB and predict sentences’ sentiment intensity.+ Build a serverless automatic workflow for the client-side inference. Deploy model with docker on AWS Sagemaker endpoint. Deploy AWS Lambda to automatically call endpoint to process newly uploaded data on AWS S3 bucket.

+ Supervisor: Prof. David Held, Work with: Edward Ahn+ Use TRPO algorithm to train state-based policy that lets agent move on a target trajectory within a simulation environment.+ Design and build a waypoints system for the trajectory.+ Render real-time agent trajectory graphs.+ Code Link: https://github.com/HyperionZhou/aa_simulation

+ Build websites for file uploading and management.+ Develop a search module to return files by keywords with Node.js and a server-side file manager to manage log files and documents uploaded by users.+ Improve response efficiency with asynchronous characteristic and use Jade to render multiple pages of the same style.+ Build web pages with the Express framework and design test cases for the system.