Oct 2019 – Dec 2019 · 2 mos · San Diego

• Developed replay tools on ROS, to visualize self-driving trucks' localization and camera pose algorithm performance.
• Write python scripts to evaluate the localization system.

Jun 2019 – Aug 2019 · 2 mos · Hangzhou, Zhejiang, China

• Interned with Taobao’s search and recommendation business unit, a key revenue center in Alibaba Group. Modified online and offline categories classification models for long-tail queries.
• Conducted research on distinguishing target words and context words in query text. Trained a Word2Vec model to analyze words similarity. Visualized results on heatmap to present words similarity.
• Extracted similar and dissimilar queries from user session log and external knowledge to enrich data sets.
• Polished deep learning model for categories classification. Added multi-task, self-attention, convolutional neural network, and LSTM mechanism to better represent the relationship of words in queries.
• Increased query classification system precision from 0.67 to 0.70, recall from 0.71 to 0.76, AUC from 0.74 to 0.78

Mar 2019 – Jun 2019 · 3 mos · Greater Los Angeles Area

• Conducted research on building Knowledge Graph based on Wikidata and Wikipedia. "Wikified" and constructed Knowledge Graph for U.S. government's poverty data sets.
• Worked on child cancer text mining project collaborated with USC Keck School of Medicine, detecting protocols from prescriptions to provide guidance for doctors.

May 2018 – Aug 2018 · 3 mos · Beijing City, China

• Worked for toutiao.com car channel and toutiao App. Optimized on-line and off-line recommendation model for advertisements, improved advertisement push, banners, and SMS performance.
• Wrote HiveQL and SQL to gather original data, preprocessed data set by data cleaning, categorical feature transformation and normalization.
• Enriched user-profile features and optimized user-profile update strategies to supply reliable features for machine learning models .
• Trained XGboost and LR model, increased AUC from 0.67 to 0.8. Did data resampling and undersampleing for imbalanced data, imputed missing values, evaluated model by recall and precision.
• Built MapReduce ETL pipeline on Hadoop, which could automatically fetch original data in HDFS, generate instances, calculate feature and data coverage rate, trained machine learning models, calculated model evaluation indexes, recorded debug log.
• Designed A/B test to evaluate model performance. Increased push service CTR from 3% to 5.5%, ROI for SMS from 8 to 12.

Mar 2016 – May 2016 · 2 mos · Beijing, China

• Conducted research on machine learning, recommendation system at CSLT lab.