Alexander Karpekov

PhD in Computer Science

Current research areas focus on Interactive Computing, Explainable AI, and Ubiquitous Computing.

Advisors: Sonia Chernova and Thomas Plötz .

MSc in Computer Science | GPA: 3.9/4.0

Completed 2nd Master’s Degree remotely while working full time at Google. Focused on Machine Learning and Artificial Intelligence.

MA in Economics | GPA: 3.8/4.0

Worked as a Teaching Assistant for 3 graduate-level classes in Statistics and Econometrics, leading sessions for 120+ students. Received the best TA award. Regional focus was on China. Studied Mandarin Chinese.

BA in Political Science | GPA: 92/100

Studied Comparative Politics. Languages: English, French. Thesis on History of Migration in the United Kingdom.

Senior Data Scientist (L5)

Worked as a Data Scientist in Google Search and YouTube Music, with the main focus on statistical data analysis and A/B experiment design and evaluation to improve search quality and music recommendations. Was promoted twice to L5. Presented my work and findings at regular director, VP, and executive level meetings, including YouTube CEO Susan Wojcicki. A few notable projects:

• Developed a pathfinding algorithm in song embedding space, improving music recommendations that led to 3% boost in user engagement and music discovery rates. This work was presented at Google-level Data Science Conference in 2023.
• Implemented a new methodology to cluster YouTube multi-billion music corpus using text, sound, search, and co-watch embeddings, which led to a 30% reduction in harmful watchtime and a 0.5% increase in music revenue ($100s millions).
• Created a new counterfactual causal impact methodology to evaluate the impact of the new feature launch on user engagement and conversion that helped establish no statistically significant long-term effects on key business metrics. The analysis was instrumental to halt the global rollout at Engineering and Product VP-level.

Data Analyst

Worked as a Data Analyst in the Data Science team, focusing on Twitter data analysis and news discovery algorithms.

• Built statistical models to automatically classify Twitter user handles.
• Conducted Twitter user clustering and unsupervised learning using networks analysis methodologies to improve news discovery algorithms.
• Led company-wide effort for reporting automation using Python instead of Excel.

Alexander Karpekov, Sonia Chernova, Thomas Plötz

In this paper we introduce DISCOVER, an active-learning method to identify fine-grained human activities from unlabeled smart home sensor data. DISCOVER combines self-supervised feature extraction and embedding clustering with a custom built visualization tool, which allows researchers to identify, label, and track human activities and changes over time.

Aeree Cho, Grace Kim, Alexander Karpekov, et al.

An interactive visualization tool that helps users understand how transformer models work through hands-on experimentation and real-time feedback.

• Best Poster Award at IEEE Viz 2024
• Went Viral: 150K+ visitors in the first 3 months

Alexander Karpekov, Sidney Miller

This project investigates the use of Transformer attention weights for deriving feature importance in NLP tasks, demonstrating that combining attention weights with gradient information improves explainability and providing an open-source GitHub tool for applying this method to any Transformer model.

Alexander Karpekov

This project examines Guangdong's shifting economic growth using Night Lights data from satelites, focusing on development beyond the Pearl River Delta and the impact of 2008 government policies.