Yangyang Hong

Recommendation System

• Enhanced end-to-end recommender system by employing advanced techniques (DeepFM, hash embedding) and fine-tuning, resulting in 4.3% increase in CTR, up from previous 3.5%.
• Integrated multi-head attention (MHA) to model user behavior sequences, achieving an additional 1.5% CTR uplift.

Search Engine

• Built sophisticated search engine by integrating Elasticsearch-powered matching, relevance model, and DeepFM-based ranking, improving CTR by 4% from a 76.5% baseline.
• Built a keyword clustering solution using SBERT embeddings, Faiss retrieval, and graph merging (with Jaro-Winkler / Longest Common Subsequence refinement) to cluster noisy user search keywords; Classified clusters with LLMs to support merchant discovery.

Engineering & Deployment

• Implemented feature system in Java that supports incremental updates and time tracking for both recommender system and search engine, reducing the time to generate features by 25%.
• Supported a framework to deploy both recommender system and search engine in a fully configurable manner across 9 markets.

Data Scientist, 01/2020 - 12/2021

• Built a repurchase prediction model using lightGBM with an AUC of 0.832 to forecast user repeat purchases, applying Optuna for hyperparameter tuning to enhance performance and reduce manual tuning time.
• Conducted extensive analysis, providing valuable insights to stakeholders about user behavior and interests.

• Collaborated with engineering team to design and build data warehouse using dimensional modeling, enabling more efficient data analysis.
• Provided comprehensive data support to various business units, enabling data-driven decision-making and ensuring the quality and timeliness of data requests.