Interview with ‘拔萝卜’: Lessons Learned from the Tencent Social Ads Algorithm Competition

During the second week of the Tencent Social Ads University Algorithm Competition finals, the team that made the most progress was a solo female contestant named “拔萝卜”. The article congratulates her, shows a photo, and presents a brief interview.

The interviewee, Feng Qianyu from Shanghai Jiao Tong University, joined the contest with a learning mindset and was delighted to receive the progress award. As a beginner she was unfamiliar with many basic operations and spent most of the competition battling data volume, memory constraints, and feature extraction, leaving little time to practice advanced techniques.

She reflects that senior participants already master sophisticated feature‑model usage, and she shares her own thoughts and takeaways.

Her initial workflow was to extract features, filter them, train several basic models, and then consider ensemble methods. However, data size and memory issues caused frequent mysterious errors, making debugging very time‑consuming. She has not yet performed feature fusion and has only tuned an XGBoost model, believing there is still room for improvement.

1. Efficiency : Besides focusing on model performance, she emphasizes the importance of time and resource costs, advocating for good code architecture, modularity, and caching frequently used datasets to avoid repeated costly reads.

2. Data Leakage : She warns against using features derived from the entire dataset, which can cause severe leakage and online failures. She currently uses a two‑day training set that balances data size and runtime efficiency.

3. Features : In ad recommendation tasks, features are crucial. She employed basic CVR statistics, cross‑features with Bayesian smoothing, and a leakage feature that provided the biggest performance boost—while stressing the need to avoid leakage. She also extracted category‑specific features through interaction and temporal signals, and analyzed two install tables for users and apps, though the gains were modest.

4. Models : XGBoost and LightGBM remain the primary models; understanding their algorithms and parameter meanings enables more effective coarse‑to‑fine tuning. She also expresses interest in trying Field‑aware Factorization Machines (FFM) if time permits.

She concludes by hoping everyone gains knowledge from the competition, encourages continuous learning, and invites further discussion.

For more details, visit the official competition website and follow the official WeChat account “TSA‑Contest”.