Yansen Zhu

Designed and implemented a personalized live stream recommendation system, achieving offline AUCs up to 0.85 and online UV CTR lifts up to 8% by optimizing CTR and CVR models (XGB CTR, DeepFM, ESMM) across diverse scenarios.

Conducted in-depth analysis of existing online experiments, providing actionable insights and new experimental designs to align business unit goals and improve penetration.

Benchmarked competitor paywall strategies, recommending a 5-15% earlier paywall integration based on chapter count to optimize user retention and ARPU.

Performed post-hoc AB test analysis using causal inference techniques (Rule Grouping, K-prototype, CasualForest) on 7M DAU, identifying key user segments and increasing ad revenue by $0.75W/day with a 1.5% penetration uplift.

Developed and shared internal best practices for AB testing and causal inference, including a comprehensive MVP checklist, enhancing data-driven decision-making across the company.

Analyzed homepage search bar clickstream data and built CPV tables for offline model training, resolving critical data quality issues.

Iteratively refined sample selection and negative sampling strategies, along with multi-round feature engineering, to establish a robust baseline for search relevance.

Implemented and optimized deep learning models (MMOE, ESMM), achieving a 0.5% UV conversion rate increase and 0.8% Suggest usage rate increase for search scenarios.

Developed NER models (bilstm+crf, bert+bilstm+crf) for subway work order analysis, achieving a 5 percentage point improvement in average F1 performance and an overall F1 score of 0.85 for 24-category entity recognition.

Improved data annotation quality and consistency through comparative analysis and standardized guidelines, leading to significant performance gains with increased data volume.

Developed and encapsulated five core NLP modules (word segmentation, syntactic analysis, keyword extraction, named entity recognition, TFIDF), ensuring robust functionality and API integration.

Refactored module APIs using RegisterModel and BaseNLPModel classes, streamlining function calls and enabling unified external interfaces for parameter setting, training, and prediction.

Deployed the intelligent Q&A system and the NLP platform into Docker images, including Python/Java services, facilitating seamless integration and deployment for various business applications.