Yulin Wu Email & Phone Number

Menlo Park, Ca, Us

Beijing, China

Kwai–Recommender System Data ScientistReinforcement learning TD3 for tunning weights of queues (e.g. VTR, CMR) in Ranking phase :• Enhanced MDP Model:Introduced a cache state for real-time vs. cached recommendations, and developed an EL algorithm to address critic dependency, balancing computational load and recommendation performance,. (Accepted by WWW’ 24.)• Long-Term Reward Design:Penalized the reward of requests when users exit to enhance user’s returning rate.Added a second-order function (e.g. user playtime times vv) to action reward - a more nuanced definition of business goals.business goals.Rough Ranking Mechanisms and Exploration of replacing mechanisms with models.• Developed Composite Scores: Integrated VTR, SVR, CMR, and LTR to filter out low performers, leading to increased user playtime and interaction, and reduced negative rate in A/B testing.• Developed a model with Transformer to replace the existing human-constructed rough ranking mechanism, +2.6% user playtime in long-term AB testing. • Developed a transformer model with user feedback as supervised labels to learn personalized ranking formulas for each video metric (e.g., VTR, LTR), replacing fixed ranking functions (e.g., logarithmic, hyperbolic).• Incorporated Reinforcement Learning Framework with the model above to enhance long-term user experience.

Shenzhen, Guangdong, China

Weishi Recalling Team – Recommender System Applied ScientistReinforcement LearningDeveloped REINFORCE algorithm with off-policy correction to address data biases, reducing variance and improving recall channel acceptance rate.Multi-Interest ModelEnhanced model diversity with a between-interest penalty.Upgraded interest aggregation module (from MIND) to Self-Attentive Network.Implemented dynamic interest modeling (based on Octopus).Introduced an interest importance model.Applied contrastive learning, enhancing model robustness and generalizability.Multi-task Learning for new and existing users: Utilized H-MTL with shared bottom layers and later upgraded to PLE with Transformer Expert Network, boosting model accuracyAlso some Samples & Feature Engineerin, Data Analysis.

Master Of Science, Electrical And Computer Engineering

Bachelor Of Applied Mathematics, Mathematics