Kai Yuan

ASE Foundation Modeling

• Build Large Language Model (LLM) pre-training and fine-tuning tool.
• Research on domain adaptation to improve the performance of domain-specific tasks.

LLM-Based Query Auto-Completion Modeling

• Build LLM-based Query Auto-Completion suggester with RAG, multi-objective alignment & distillation;
• Yielded +3.20% query auto-completion usage rate lift and -5.37% reduction on average typed keystrokes per search;

Search Widget Ranking (2021-Current)

• Formulate search widget ranking as a multi-agent markov decision process (MAMDP), and tackle it by the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) model. Off policy evaluation (OPE) method shows +5.7% improvements on the doubly robust estimator; Online experiment is on-going.
• Led 5 experienced engineers to productionize the MADDPG model and a mix-offline multi-agent RL system.
• Requested a national patent on applying multi-agent RL to search ranking and related paper (Publication #1 below) got accepted by Amazon Machine Learning Conference (AMLC) 2022 as oral presentation with 8% acceptance rate.
• Designed mission-aware deep bandits model to rank widgets in search. Offline yields +3% CTR and +2.3% GMV prediction accuracy lift; Online experiment yielded +$42MM GMV gain.
• Led 1 scientist and 3 engineers to productionize the mission-aware deep bandits system.
• Related paper (Publication #3 below) got accepted by AMLC 2021 as oral presentation with 9% acceptance rate.

Search Query Autocomplete Ranking (2019-2020)

• Proposed metrics-weighted mean reciprocal rank (MW-MRR) as offline ranking evaluation metrics for Query Autocomplete ranking task, adopted by 10 scientists onwards and driven 12+ ranking iterations.
• Launched three tree-based and one deep learning-based pairwise learning to rank models, yielded +276bps suggestion acceptance rate lift and +$1.3B annualized GMV gain.
• Partnered with 2 senior engineers to enable deep learning online inference to serve 300K TPS at 15ms @p99.
• Related paper (Publication #4) got accepted by AMLC 2020 as oral presentation with 10% acceptance rate.

Series Search (2015-2017)

• Designed 3 query understanding models to link customer queries to book/movie series entities.
• Drove 9 online experiments and 3 were successfully launched, yielded +$4.3MM GMV gain.

Baidu News Recommendation

• News feed ranking using GBDT.
• Customer modeling and feature engineering: 1) mining interests using Ebbinghaus Forget Curve; 2) mining commuting route using DBSCAN; 3) cluster customers using K-means, and recommend news from belonging clusters.