Yijia Wang

Yijia Wang, Ph.D.
Applied Scientist
Amazon

Contact

Applied Scientist in SCOT (Supply Chain Optimization Technologies) at Amazon, Seattle, WA., May. 2022 - Now

Utilized SQL, Python, AWS technologies (S3, Sagemaker, and Athena) and ETL processes to conduct a comprehensive analysis and visualization of customer search data, product demand data, and the interrelatedness of substitutable products.
Implemented Machine Learning and Causal Inference methods to estimate the demand for products that were suppressed in search ranking results due to inventory unavailability.
Risk Control: Developed group-level metrics of substitutable products that reflect the risk of out-of-stock for search ranking. This information will be taken into consideration by the business team to take appropriate action.
Major Tools and Languages: Python, SQL, AWS, ETL, EMR, Redshift, Spark, Scala

Applied Scientist Intern at Amazon, Seattle, WA., May. - Aug. 2021

Applied regression models and Causal Inference methods to analyze the product substitute effect on demand shaping (suppress out-of-stock products in search results), with the result being used to correct the estimated lost demand for out-of-stock products.
Major Tools and Languages: Python, SQL, AWS, ETL, Redshift

Research Engineer Intern at SigOpt, an Intel company, San Francisco, CA., May. - Aug. 2019

Applied noisy EI (a method of Bayesian optimization) to improve the performance of the API for noisy problems.
Applied additive Gaussian processes for optimizing high dimensional problems and improved the performance of the API by 20% on average.
Utilized data analysis and data visualization tools to prove the performance of the applied algorithms.
Major Tools and Languages: Python, AWS

Research Assistant at University of Pittsburgh, 2016 Fall – 2022 Spring

Utilized various programming languages such as Python, R, and Matlab to effectively analyze and visually represent data across multiple fields, with specific emphasis on topics including, but not limited to, opoid overdose and naloxone (an overdose reversal drug) distribution, as well as energy pricing and demand analysis.
Deployed the fast-slow structure of Markov decision processes, proposed approximate dynamic programming methodologies and algorithms to efficiently solve the problems. The algorithms can be applied to real-world problems including the service allocation for a multi-class queue, multi-armed bandits, energy demand response, and so on. Algorithms were implemented in Python.
Focusing on discrete inventory and dispensing problems of public health with stochastic demand, developed a structural actor-critic algorithm, which was implemented in Matlab and Python and performs at least 20% better than benchmarks within limited CPU time. The algorithm was applied to naloxone dispensing problem in a case study.
Studied the exploration problem with expensive interactions in reinforcement learning, developed subgoals with intrinsic rewards to efficiently solve it and proposed a Bayes-optimal algorithm. Algorithms were implemented in Python.

Approximate Hierarchical Value Iteration For Fast-Slow MDPs, INFORMS 2021.
Structured actor-critic for managing and dispensing public health inventory, INFORMS 2020.
Exploration via sample-efficient subgoal design, INFORMS 2019.
Structured actor-critic for managing public health points-of-dispensing, INFORMS 2019.
Exploration via sample-efficient subgoal design, ICLR workshop (Task-Agnostic RL) 2019.
Structured actor-critic for managing and dispensing public health inventory, INFORMS 2018.
Data-driven models for naloxone distribution amidst the opioid crisis, NIPS workshop (WiML) 2017.
Data-driven models for naloxone distribution amidst the opioid crisis, INFORMS 2017.