Insights from the First Tencent Social Advertising University Algorithm Competition: Teams’ Strategies and Experiences

On July 6, the first Tencent Social Advertising University Algorithm Competition concluded, with the Nanjing University team "nju_newbie" winning the championship and a prize of 300,000 CNY, while "Raymone" and "我很难受" secured second and third places respectively.

The competition task was to predict the conversion rate (pCVR) of mobile app ads, i.e., the probability that a user who clicks an ad will later activate the app, a more challenging metric than click‑through rate.

Interviews with the three finalist teams reveal their backgrounds, motivations, and technical solutions. "nju_newbie" combined deep‑learning models (wide&deep, PNN, NFFM) with LightGBM, emphasizing data preprocessing, feature extraction (conversion, click, install, time features), and model stacking, which boosted performance by about 2.5 ten‑thousandths.

"Raymone" focused on extensive data cleaning, five‑fold cross‑validation, and a rich set of features including user‑app interaction history, position‑related attributes, and probability‑estimated features, using LightGBM, FFM, LR, GBDT, and ET models before stacking.

The "我很难受" team also treated the problem as binary classification, employing XGBoost, LightGBM, and FFM, with stacking and careful feature correlation analysis, while addressing memory errors through code optimization.

All teams reported mental and technical challenges such as intense competition pressure, large data volume, and model bottlenecks, which they overcame through teamwork, iterative feature engineering, and consulting literature and mentors.

They advise future participants to persist through difficulties, innovate beyond existing frameworks, leverage real‑world data, and continuously document insights, emphasizing that practical competition experience is essential for growth in machine‑learning and data‑analysis fields.