Heather Song

Ads Quality Team • Launched multi‑label multi‑task (MTML) DNN models for Ads ranking, achieving a 0.85% revenue gain and addressing online‑offline AUC gaps. • Developed a Grocery Ads LightGBM model, later upgraded to an MTML DNN model, resulting in 2.52% and 2.51% revenue gains, respectively. • Built auto‑calibration pipelines (Logistic Regression, Platt Scaling, Polynomial Regression) to enhance auction fairness across all Ads models. • First integrated Homepage Banner rankers into a multi‑armed bandit (MAB) platform, improving CTR by 4% and addressing attribution gaps. • Optimized the Order Value Prediction model for auto‑bidding, reducing Cost Per Action (CPA) by 4.63% and prediction error by 7.8%. • Enhanced Search Ads models by incorporating context and price‑sensitivity features, increasing restaurant search ads revenue by 1.6%. • Identified year‑long broken Search Ads feature pipelines during on‑call, implementing fixes that led to +0.50% CVR and +1.55% revenue gains. • Mentored junior Machine Learning Engineers on the Ads team through 1:1 supervision for feature engineering and model development.

• Expanded coverage of retail product taxonomy classification models (BERT, LSTM, FastText) from 50% to 97% with 85% precision, improving downstream substitution recommendation approval rates by +0.5% and reducing item‑not‑found rate by 2%. • Deployed a document understanding model for food type classification using Horovod + Petastorm, increasing coverage from 89% to 97%. • Launched the first Alcohol Compliance model (alcohol vs. non‑alcohol) with 94% Precision and 99% Recall, reducing annotation costs.

• Developed the first query understanding model (LSTM) for classifying food vs. non‑food queries using Databricks, Snowflake, and PyTorch. • Led query expansion efforts (mapping queries to synonyms) for hybrid restaurant and grocery research in collaboration with PMs and engineers.

1. Established pipelines to process Android apk files(AndroZoo) to extract features using multi-thread programming (Python) . 2. Extracted sensor features (JavaScript) and visualized data of sensors by R and Python. 3. Proposed machine learning models and neural network combined sensor features to classify repackaged malicious software. 4. Submitted an IEEE 2020 Poster named “Sensor-Based Repackage Malware Detection” and presented it with 100+ schoolmates.

1. Performed feature engineering to build a model, which includes city marketing profile and geographic supplement profiles. 2. Utilized PageRank and other algorithms target outlier uers which helped to improve the precision and recall of other models. 3. Designed machine learning model to predict user's next day exposure pages ( Python and Spark ) Scheduled training and inference tasks to predict labels of all users ( 400 millions ) (HIVE, Spark, SQL) offline every day, which brought a 2-3% absolute improvement in CVR, and 2% uplifts in total revenue and 3% uplifts in total trips. 4. Designed deep learning models using user's behavior sequence to do purchasing plan recommendation. Models including Youtube DNN, Bert, LSTM. 5. Used graph embedding to generate geohash embeddings and tested its effectiveness in another CVR model. 6. Used dynamic programming and decision tree (Scikit-Learn, Python, SQL, PySpark) to analyze the market potentials of a new purchasing card plan. Formed simple rules trying to lower the influence on the current MIP. Designed A/B test to test the effectiveness of rules.

Coded a semi-supervised learning method based on Optical Flow for obstacle detection and tracking function of the ROS equipment by Python ● Designed an ENN to get obstacle avoidance strategies in virtual environment Gazebo ● Obtained a 89% accuracy in simulator test, and the velocity of ROS vehicle is faster than non-ENN method, and the accuracy is also 3% higher than non-ENN method

I work as a Robot Operation System (ROS) intern and develop machine learning methods. 1. De-nosed and balanced color of warehouse video data for preprocessing, and real-time obstacle classification by CNN model. 2.Tracked obstacles by Optical Flow and EM algorithm and predicted the next possible position of obstacles by ARIMA. 3.Tested former methods on a real ROS car and reaching an accuracy of about 83%. Learned obstacle avoidance strategies using a genetic algorithm achieving 79% avoidance accuracy in simulation.

Using RGB-D camera collecting real-time data, programmed for the segmentation and classification of parking plot based on CNN.

● Proposed a mathematical model to quickly predict road smoothness ● Built models to analyze the smoothness of road and fuel consumption using Python ● Enlisted as a Shanghai College Student Innovation Project ● Wrote a paper that won Grand Prize of National College Student Surveying Paper Competition