Winning Approach of the Tencent Advertising Algorithm Competition: Feature Engineering, Model Selection, and Future Work

Hello, we are the Ge Wenqiang team, from Jilin University, Harbin Institute of Technology, and Beijing University of Posts and Telecommunications, and we are fortunate to have won first place in the preliminary round. We will introduce our method from three aspects.

1. Feature Engineering : In the preliminary round we found that one‑hot features differed greatly from our statistical features, so we added one‑hot features to our nearly hundred‑dimensional statistical features, gaining about 7k improvement. Our statistical features were extracted using 5‑fold cross‑validation to avoid leakage. In the final round, due to the massive increase in data volume, one‑hot features could no longer be used, causing our LightGBM model to drop about 8k.

2. Model Selection : In the preliminary round we trained three models—LightGBM, DeepFM, and DeepFFM. LightGBM achieved the highest score, and a simple weighted average gave good results. In the final round, DeepFFM consumed too much GPU memory, leading to OOM; reducing its embedding dimension hurt performance, so we abandoned it. Combining the LightGBM model without one‑hot features and DeepFM yielded a final score around 757, about 8k lower than the preliminary round.

3. Later Work : In the final round we faced memory and GPU constraints, partly due to hardware and partly because our model design was unsuitable for the large‑scale data. We believe neural network models have great advantages in such data: they train quickly and benefit from the data volume. When traditional feature‑based tree models are insufficient, we must optimize neural network models.

Finally, we wish everyone an enjoyable competition and good results.