Tencent Advertising Algorithm Competition: FFM Approach and Feature Engineering by the Wenqiang Ge Team

The Wenqiang Ge team from Jilin University, Harbin Institute of Technology, and Beijing University of Posts and Telecommunications won the first week of the Tencent Advertising Algorithm Competition rematch.

They built upon a FFM baseline shared by Zhihu user “追溯星霜”, initially using only raw ID features achieving an online score of about 0.745.

To improve performance, they discretized continuous features into ten equal‑width bins and fed low‑correlation LGBM continuous features into FFM, raising the score to roughly 750 on the initial A榜.

Recognizing the importance of ad‑user cross features, they further discretized LGBM cross‑conversion‑rate features and fed them into the model, attaining an A榜 score of 0.7557.

The team noted that FFM predictions differ substantially from those of LGBM and neural networks, leading to significant gains when ensembled.

For handling large intermediate data in the rematch, they recommend the Feather package for fast, reliable storage of DataFrames.

They used the open‑source xlearn library from Peking University, which offers faster prediction than libffm, supports AUC as an evaluation metric, and can be called directly from Python.

Key FFM hyperparameters mentioned are the latent vector dimension k, learning rate lr, and regularization coefficient λ, which can be tuned based on past Tencent competition shares and the original FFM paper.

Finally, they wish all participants enjoyment and success in the competition.