Weekly Champion Insights from the Tencent Social Ads Algorithm Competition – The ThreeIdiots Team

We are the ThreeIdiots team, consisting of Charles and Wepon from Peking University and wsss from USTC. We clarify that we are not the Kaggle 3Idiots team; our name was chosen out of respect and humor, and we apologize for any confusion.

Our success this week stems from a feature‑centric approach: we believe that features set the upper bound of model performance. Understanding the user’s context, conversion drivers (primary demand, ad presentation, and constraints such as network speed), and extracting relevant features from real business logic are crucial.

For data splitting we adopted a time‑based method rather than random splitting to avoid leakage, ensuring that online and offline data remain synchronized, especially given the strong temporal characteristics of the dataset.

Feature engineering consumed most of our effort. We recommend focusing on robust feature extraction rather than excessive parameter tuning, particularly for tree‑based models where parameter impact is limited. Besides common ID conversion‑rate features, we experimented with word embeddings from the user_installapp table and generated dense features using a Wide & Deep model.

Handling the large data scale presented memory and speed challenges. We suggest down‑sampling, applying feature selection (Filter, Wrapper, Embedded methods), and using models like ensemble trees or Lasso that inherently perform selection. Incremental feature extraction—saving intermediate features to disk and concatenating later—helps manage resources.

In the preliminary round we used XGBoost, which outperformed LightGBM, but for the final round we switched to LightGBM and FM due to faster training on the larger dataset. We have not yet performed model fusion but plan to explore it later; however, we advise prioritizing strong features before attempting complex ensembles.

In summary, observe the data carefully, invest time in thoughtful feature design, minimize unnecessary coding and parameter tweaking, and iterate quickly. We wish everyone success in the competition.