Runner‑up Team’s Experience and Practical Tips from the First Tencent Social Advertising University Algorithm Competition

The first Tencent Social Advertising university algorithm competition focused on mobile‑app ad conversion‑rate prediction and provided participants with real‑world advertising data, aiming to solve the core problem of conversion estimation.

The team “Raymone”, composed of Li Miao, Li Qiang from Dalian University of Technology and Li Da from Tsinghua University, won the runner‑up prize.

Li Qiang, now a prospective Tencent employee, shares his experience: the competition broadened his knowledge, sharpened existing skills, and offered access to authentic industry data that is rarely available in academia, as well as internship and recruitment opportunities.

1. Data Set Splitting – Split the provided training set into a local training set and a local validation set. Common strategies are random split (enabling cross‑validation) and time‑based split when the data has a temporal order, to avoid data leakage.

2. Feature Engineering – Features determine the model ceiling. Build features from the business perspective (e.g., user demand for the advertised app) rather than blindly engineering them; consult existing competition blogs for ideas.

3. Data Scale – If the dataset is too large for your machine, down‑sample for quick feature validation, then retrain on the full data once a feature proves useful. Save extracted features to disk to avoid repeated computation.

4. Single Model Choice – Popular models include XGBoost, LightGBM, GBDT, and FFM. LightGBM often matches XGBoost’s accuracy while being several times faster, making it a good default.

5. Model Ensemble – Ensemble can boost scores. Focus on feature engineering early, then ensemble later. Train diverse models, each suited to specific feature sets, to achieve larger gains.

Final advice: seek guidance from experienced participants (“senior drivers”), even if they are not on your team, to accelerate learning.

References: [1] Kaggle Data Mining Competition Experience [2] Stacking Techniques for Competition Late‑Stage