Practical Guide to Scoring 85+ in the Tencent Advertising Algorithm Competition

As a newcomer who first encountered a big‑data competition at the second Tencent contest and later managed to score around 85 points within ten days, the author summarizes personal experiences aimed at participants who have not yet entered the top 200 rankings and are seeking direction.

The article begins by recommending ways to stay informed about upcoming competitions, such as following official websites, competition handbooks, FAQs, forums, groups, and public accounts. After registration, participants should download the dataset and review the problem description, but they need not explore every data dimension due to limited time.

Three main channels for understanding the competition are highlighted: the official "Problem Description" and "Participant Manual", the FAQ released after the competition starts, and community resources like forums, chat groups, and public accounts that provide deeper analyses.

Once the problem is grasped, the author suggests a practical workflow: first perform a quick data inspection, then create a simple rule‑based baseline (e.g., using bid divided by a constant) to obtain an initial submission. This baseline is treated as a simple model and can already achieve modest scores.

For model development, Python is the recommended language, with tree‑based models (XGBoost, LightGBM) and neural networks being common choices. LightGBM is suggested for beginners. Feature engineering should start with a solid baseline, then incrementally add features, paying attention to potential leakage. Simple imputation and smoothing can quickly raise scores to the low 80s.

Further optimization involves tuning hyper‑parameters (using publicly shared parameter sets) and refining features (handling outliers, missing values, creating statistical and interaction features). The author also shares several practical tricks, such as focusing on CPC ads, labeling non‑CPC ads, ensuring model generalization for unseen ad IDs, and experimenting with bid dropping.

Finally, the article reminds participants of the submission deadline (May 23, 12:00 PM Beijing time) and notes that the top 20% of scores will advance to the semifinals.