Insights from the Tencent Social Advertising University Algorithm Competition – Week 3 Champion Team “Daodi Dui Bu Dui”

The third‑week champion of the Tencent Social Advertising University Algorithm Competition is the team “Daodi Dui Bu Dui”, composed of two female students and one male student from Peking University.

In an interview they explained that they had no prior competition experience, making the victory especially rewarding.

Key takeaways from their competition experience include:

1. Data handling: They spent considerable time constructing the dataset, applying data‑cleaning strategies to remove unreliable records, and found that using the full dataset performed better than a subset. For validation they randomly selected 10% of the data, experimenting with different random seeds to minimize the gap between offline and online log‑loss.

2. Feature engineering: Initially they added all raw and potential features, but later refined the set by removing highly correlated or useless features through ablation tests, consulting competition blogs, papers, and past Kaggle solutions, and performing statistical analysis to identify truly informative features.

3. Model training: To address high dimensionality from one‑hot encoding they used sparse encoding and optimized XGBoost’s parallel configuration for faster iteration, emphasizing the importance of an efficient training pipeline for large‑scale data.

They conclude by wishing all participants good results and encourage others to learn from their approach.

For more details and registration, visit the official competition website: http://algo.tpai.qq.com.