Choosing Mainstream CTR Models: LightGBM, FFM, and Deep Learning Approaches

Greetings, I am Ge Yunpeng, a graduate student from Harbin Institute of Technology Shenzhen, the weekly champion of the Tencent advertising algorithm competition, and I would like to share my insights on choosing mainstream CTR models for this contest.

1. LightGBM (statistical feature model) – This model is robust for numeric features and performs well on conversion‑rate features due to the look‑alike audience expansion being framed as a CTR problem. It is important to mitigate data leakage (e.g., via Bayesian smoothing) and to handle many categorical features; open‑source features from the ByRAN community can add about 0.01 AUC.

2. Field‑aware Factorization Machines (FFM) – A classic CTR model that can be used via the command‑line ../../libffm/ffm-train -l 0.00002 -k 8 -r 0.05 -s 30 -t 25 -p valid.ffm train1.ffm v1_model . The key step is converting categorical features into field+feature_index+value format. Simple cross‑encoding of IDs and occasional use of numeric‑feature discretization can help; set a slightly larger learning rate, train 50‑100 rounds, keep a validation set, and disable auto‑stop to avoid over‑fitting.

3. Deep Learning models – Given the strong sequential information in the feature set, deep models (wide‑and‑deep, DeepFM, PNN, etc.) work well. Recent community contributions such as DeepFFM are worth trying, and an NN model can also be built from an NLP perspective.

Finally, I hope everyone gains more machine‑learning knowledge from the competition, enjoys the process, and competes honestly without cheating.

Small suggestion: Because the provided dataset is large, try to limit the use of memory‑heavy libraries like NumPy and pandas; using native Python lists can greatly reduce memory consumption.