From Research Papers to Production Systems: Le Yu's Journey in Applied Machine Learning

With eight peer-reviewed publications spanning healthcare AI, privacy optimization, and advertising technology, Le Yu exemplifies the modern machine learning engineer who seamlessly transitions between academic research and industrial application. Currently serving as Senior Machine Learning Engineer at TikTok, Yu has transformed theoretical insights into systems that serve billions of users daily.

His journey from Peking University's Electronics and Communication Engineering program to TikTok's advertising platform illustrates how rigorous academic training translates into real-world impact. We spoke with Yu about his research contributions, industry applications, and the symbiotic relationship between academic investigation and commercial innovation.

Academic Foundations

Q: Your educational path includes Peking University, National TsingHua University exchange, and Southeast University. How did this diverse academic exposure shape your approach?

Each institution offered unique perspectives. Southeast University's automation engineering program gave me strong fundamentals in control theory and signals processing—concepts that directly apply to bidding optimization algorithms. The exchange at National TsingHua University exposed me to different research methodologies and international collaboration approaches.

At Peking University for my master's, the focus on machine learning and digital image processing was transformative. The IoT Operating Systems course might seem unrelated to advertising, but understanding distributed systems at the hardware level helps optimize large-scale machine learning deployments. Every course contributed to my holistic understanding of complex systems.

Research Portfolio and Real-World Applications

Q: You've published eight papers across diverse domains. How do topics like drug repurposing relate to advertising technology?

My paper on "Artificial Intelligence-Driven Drug Repurposing for Neurodegenerative Diseases" might seem distant from advertising, but the underlying challenge is similar: matching existing solutions to new problems based on complex feature interactions. In drug repurposing, we identify existing drugs for new therapeutic uses. In advertising, we match existing ads to new user contexts.

The computational analysis methods developed for drug-disease interactions directly influenced how I approach user-ad matching at TikTok. Both require understanding high-dimensional feature spaces and predicting outcomes based on limited historical data.

Q: Your differential privacy research seems directly applicable to your current work.

Absolutely. The paper "Dynamic Optimization Method for Differential Privacy Parameters Based on Data Sensitivity in Federated Learning" was almost prophetic for my TikTok work. When I wrote it, federated learning was primarily theoretical. Now, I'm implementing these exact concepts in production systems serving millions of advertisers.

The key insight from that research—that privacy parameters should adapt based on data sensitivity—is now embedded in our PACS framework. Highly sensitive user behaviors get stronger privacy protections, while aggregate metrics can be processed with lighter privacy budgets. This dynamic approach maintains model performance while ensuring robust privacy.

Industry Experience Across Tech Giants

Q: You've interned at Microsoft, Tencent, WeChat, and consulted for Roland Berger. How did these experiences prepare you for TikTok?

Each experience built different capabilities. At Microsoft, I worked on two separate internships—first optimizing N-gram models for search ranking, then using BERT for news classification. These taught me the evolution from traditional NLP to transformer-based approaches, knowledge I now apply to understanding ad content and user queries.

The WeChat internship on video copyright detection using ResNet models was particularly relevant. The same computer vision techniques that identify reposted videos now help us understand visual ad content for better targeting. The pattern recognition skills transfer directly.

At Tencent, analyzing global cloud computing trends taught me to think strategically about technical infrastructure. This business analysis perspective helps me understand not just how to build systems, but why certain architectural decisions create long-term value.

The Roland Berger consulting experience was unique—analyzing quality inspection at smartphone OEMs. This taught me to identify inefficiencies in complex processes, a skill I use daily when optimizing advertising pipelines.

Cross-Disciplinary Research Applications

Q: Your publication on restaurant seating layout using customer flow patterns seems unexpected for an ads engineer.

That paper demonstrates how machine learning solves optimization problems across domains. Restaurant seating optimization and ad placement optimization share surprising similarities—both involve predicting user behavior, managing limited resources, and maximizing value under constraints.

The customer flow patterns we analyzed are analogous to user browsing patterns. The algorithms that optimize table arrangements can be adapted to optimize ad inventory allocation. This cross-pollination of ideas often leads to breakthrough innovations.

Q: How does your healthcare AI research influence your advertising work?

The paper on "Personalized Medication Recommendation for Type 2 Diabetes Based on Patient Clinical Characteristics" deals with personalization under strict safety constraints—we can't recommend medications that might harm patients. Similarly, in advertising, we must personalize while respecting user privacy and platform policies.

The comparative study on cardiotoxicity prediction taught me ensemble methods for high-stakes predictions. When an advertiser spends millions on campaigns, our predictions must be equally reliable. The rigorous validation methods from healthcare AI ensure our advertising models meet production standards.

Collaborative Research and Knowledge Transfer

Q: Several papers have co-authors. How does collaborative research benefit your industry work?

Collaborative research teaches you to integrate diverse perspectives—crucial when working on cross-functional teams at TikTok. My paper with Zhu L and Sun M on personalized advertisement recommendation was directly collaborative, combining our expertise in context awareness and system design.

The paper with Weng G on industrial surface defect detection brought together computer vision and real-time processing expertise. These collaboration skills directly transfer to working with product managers, data scientists, and infrastructure engineers at TikTok.

From Internship to Senior Engineer

Q: How did your E-Funds internship on LSTM stock prediction influence your approach to advertising metrics?

Stock prediction and advertising metrics share temporal dependencies—both require understanding how past patterns influence future outcomes. The LSTM networks I developed for stock prices inspired our approach to predicting advertiser lifetime value. The hundred-plus proprietary dimensions we used for stock prediction taught me feature engineering at scale.

However, advertising has an advantage: we can influence outcomes through better targeting, while stock prediction is purely observational. This ability to intervene makes advertising ML both more complex and more impactful.

Technical Innovation in Production

Q: You enhanced ad models through cross-domain collaboration with recommendation systems. What was the technical challenge?

The main challenge was preventing information leakage while enabling knowledge transfer. Recommendation systems optimize for user engagement, while advertising optimizes for advertiser value. Directly sharing features could bias recommendations toward commercial content, degrading user experience.

We solved this through careful feature abstraction and knowledge distillation. The recommendation system teaches the advertising system about user preferences without sharing raw behavioral data. This maintains system independence while improving ad relevance.

Future Research Directions

Q: With your strong publication record, do you plan to continue academic research?

Absolutely. I'm currently investigating secure multi-party computation for privacy-preserving advertising auctions. The goal is enabling competitive bidding without revealing individual bid values—protecting advertiser strategies while maintaining auction efficiency.

I'm also exploring applications of causal inference to advertising attribution. Much of current advertising analytics is correlational, but advertisers need causal understanding to optimize spending. This requires new theoretical frameworks and practical implementations.

Q: What advice would you give to students interested in applied ML research?

Don't silo yourself into pure theory or pure application. The most impactful work happens at the intersection. Publish your research, but also build systems. Understand the math, but also understand the business context.

Most importantly, maintain intellectual curiosity across domains. My most innovative solutions often come from connecting seemingly unrelated fields. The breadth of knowledge from healthcare to finance to advertising creates unique perspectives that specialized experts might miss.