Ji Chen

Experience

1. Applied Scientist

   Amazon

   Apr 2021 — Present

   Trained a global machine learning (ML) model across marketplaces to predict customers’ likelihood of contacting customer service for unsubscribing Prime membership. Increased prediction accuracy through shared training data across marketplaces. Reduced future maintenance efforts by replacing one model per marketplace with a uniform model. Enabled day one launch for new marketplaces. – Drove the proof of concept analysis on enabling real time personalization for existing ML models with a team consisting of scientists and engineers. Enabled personalization for customers with limited visiting history. Evaluated the counterfactual performance of proposed ML driven policies leveraging only offline log data through off-policy evaluation (OPE). Improved estimated customer sign up rate by over 30 bps. – Trained a deep neural network model to personalize Prime plan offer recommendation. Worked with stakeholders to analyze data quality and evaluate statistical test results. Employed two phase learning to better fit highly imbalanced data. Used OPE to estimate performance of proposed ML model. Lifted estimated long term paid yield rate by over 10 bps comparing to benchmark business rule based policy. – Prepared a long term proposal for enabling real time inference in production within Prime. Explored possible ML architectures to best utilize offline and real time features. Designed experiments to verify the effectiveness of proposed architectures. Explored the deployment of ML models on Amazon SageMaker. – Mentored an intern working on integrating percentiled currency features to existing global ML models.