Tencent Advertising Algorithm Competition: Experience and Tips from the Runner‑Up

Hello everyone, I am Xu An from the University of Electronic Science and Technology of China. I am honored to share my experience in the Tencent Advertising Algorithm Competition where our team, Three Little Bees, won the runner‑up prize.

Feature engineering greatly influences competition results; I focused on two lines of thought: deriving features from the problem itself (considering user intent, product‑user match) and from feature types (cross‑features, strange distributions, multi‑value categorical features, consistent scales). Features must be consistent across train, validation and test sets to avoid model side‑effects.

Recent machine learning competitions mainly use XGBoost and LightGBM; FFM works well for many categorical features, while deep learning shines on large data. It is better to master one model family first before trying others, or let teammates specialize and combine strengths.

With large datasets, training a single model can take hours; speeding up via multiprocessing or separating feature scripts helps. Caching features to files, and including data length in filenames prevents loading wrong caches when switching datasets.

To avoid forgetting ideas, I jot down sudden insights in my phone memo and review them while coding.

The competition lasts over two months like a marathon; allocating fixed daily hours, not obsessing over daily leaderboards, and staying relaxed often yields inspiration (e.g., while playing games).

Study past winners’ code on Kaggle or similar domestic contests, read line by line, ask what each line does, how you would write it, and what differences might appear; this accelerates entry into ML contests.

Finally, I wish everyone good results in the competition.