IJCAI 2018 International Advertising Algorithm Competition Champion Uses Transfer Learning and LightGBM for Ad Conversion Prediction

On June 5, 2018, the IJCAI 2018 International Advertising Algorithm Competition concluded in Hangzhou, where Hua Zhixiang, a solo participant from JD.com’s middle‑platform trading platform, outperformed more than 6,000 contestants to claim the championship and a $10,000 prize.

IJCAI is a premier AI conference covering machine learning, computer vision, speech, and video technologies. The 2018 algorithm contest aimed to stimulate algorithmic talent by challenging participants to predict advertising conversion rates using real e‑commerce transaction data.

Hua, also a Kaggle Grandmaster, entered the competition under the team name “DOG” and applied transfer learning to the ad conversion prediction task, achieving outstanding results.

The competition used a massive real‑world advertising dataset to build a model that forecasts users' purchase intent, i.e., the ad conversion rate. Judges praised the DOG team’s solution for its overall simplicity and clear design thinking.

Technically, the solution consists of two LightGBM models (Level‑1 and Level‑2). Level‑1 is trained on all pre‑heat period data to model shopping behavior, and its predictions for the shopping‑festival day become inputs for Level‑2, which also uses the day’s data, thereby mitigating the distribution shift between the pre‑heat period and the festival.

Feature engineering focused on four groups: statistical features (click counts, page views, hourly search averages), temporal‑difference features (time gaps between user and item interactions), ranking features (interaction frequencies between user and item), and representation features (user interests in item attributes).

The entire codebase fits on a single page, highlighting its practicality and ease of deployment. Hua’s success also reflects JD.com’s strong emphasis on algorithmic talent, as he previously won the first JD JDATA algorithm competition and was recruited as an outstanding algorithmic talent.

During JD’s 6.18 shopping festival, similar algorithmic models were widely applied to improve the shopping experience, with Hua playing a key role in those efforts.