Insights and Strategies from Winning the Tencent Advertising Algorithm Competition

Hello, I am Guo Daye, an undergraduate at Sun Yat‑sen University. I was fortunate to win the weekly champion again, and I would like to share my competition experience and insights.

1. Data Processing The competition involved a large amount of data, which can be challenging for machines with limited memory. Two approaches work well: (1) use a streaming‑training model such as FFM, which trains quickly and reads data batch‑by‑batch; (2) split the data into five parts, train five separate models, and average their predictions, a method that performed well in last year’s finals and allows lower‑performance machines to run the task.

2. Feature Engineering Feature engineering is crucial for ranking. We recommend (1) using a small subset (1% of the data) to discover strong features, and (2) understanding the business context so that features reflect real‑world behavior. For example, creating ID features like aid_age captures how ad characteristics affect different age groups. Cross features between ads and users proved very useful, and adding statistical cross‑features helped us reach a score of 0.755.

3. Model Selection We employed both GBDT and deepFFM. GBDT offers excellent performance but cannot be trained in a streaming fashion, making data loading slow. deepFFM trains quickly but requires preprocessing. By using GBDT on a small dataset to discover features and then feeding those features to a custom deepFFM (enhanced with a dropout layer and a modified loss function), we achieved a single‑model score of 0.753.

Finally, I wish everyone enjoys the competition process and hopes you achieve great results.

▶▶ Click Read Original to register for the competition and remember to complete real‑name authentication after logging in.