Instagram Business Experiences

Help improve Instagram feed ads rendering experience, enhancing user consideration and shorten conversion journey.

lead workstream including Call-to-Action system, tappability and others.

Offline Web Content Integration --- Thumbnail System

1. Developed a thumbnail subsystem for extracting thumbnail information from web pages for different vertical searches.

2. Reduced 40% database storage and 20% crawler resources by implementing a URL deduping algorithm. Speeded the algorithm 20 times by filtering top duplicate hosts and implementing Pyspark.

3. Completed the clean and refresh process of the system, using sm-db as database and sm-stream for data streaming; enabled automatic alerts to detect the inconsistency between DB and OSS cloud during the clean process. Webpage finally updates thumbnail sub-contents at a delay of fewer than 2 hours.

4. Implemented multi-business control, logging, and data statistics on the selection process of the thumbnail subsystem.

5. Built a dashboard on metrics of crawling delays and business statistics by MaxCompute to monitor the health of the system.

Offline Web Content Integration – Quark Medical Data Process

1. Migrated Quark Medical enter-database process from local servers to the cloud.

2. Used Blink for stream computing, Lindorm as database, and DataX for building ETL pipeline. Wrote Blink plugins by Java to extract medical atlas entity and relation from session data between patients and Quark AI doctor.

3. Utilized tricks such as Session Window and Streaming Joins to optimize the enter-database system, which finally can process data up to 1000qps, improved by 50% compared to the previous system.

Political Sensitivity Analysis

1. Conducted model integration of Fasttext and Bert for politically sensitive webpage classification. Increased the recall by 20%; wrote an online C++ plugin based on the model and published it to the online process, proved to be successful.

Blockchain News Sentiment Analysis

1. Crawled nearly 70,000 Blockchain news corpora from websites using Scrapy and multi-processing in Python.

2. Encoded sentence by BERT and fed encoding results into Bi-LSTM network for document-level sentiment classification (Pytorch); got a 0.93 F1 score and 6% improvement in accuracy compared to CNN network.

Data Visualization System

3. Developed a simple event detection algorithm using the sentiment index as signals and Bollinger Band as event filters, extracted two main events successfully: 1.ETH being hacked (2019.5) 2. The release of Libra (Facebook).

4. Integrated visual analytics, such as T-SNE and LDA, into a visualization system (JavaScript).

1. Conducted feature engineering, hyperparameter tuning by Hyperopt in Lightgbm model for retail sales prediction; improved the model’s MAE loss by 0.05 compared to the baseline.

2. Modeled customers’ requests as a linear programming model on commodity transportation problems; used Gorubi to solve the LP problem; implemented a greedy algorithm for post-processing to satisfy discrete pallet constraints.

3. Helped the project won the POC battle against other rivals.

4. Helped maintain Lightgbm baseline model for sales forecasting; modified the feature engineering process monthly; conducted a post-process KA model by Xgboost to align the baseline model.