Weekly Champion nju_newbiew Shares Competition Experience and Technical Insights

We are the nju_newbiew team from Nanjing University, consisting of yy, cold rain, and sf, and we proudly share our experience after winning the weekly champion in the Tencent Social Ads University Algorithm Competition.

Our A‑list best score of 0.101717 was achieved by early model fusion, which gave us an advantage despite individual models not outperforming other top teams.

1. Data Processing After entering the finals the data volume grew significantly, causing slow file I/O and memory issues. We converted data to int8 where possible and stored files in H5 format for fast read/write. We trained on the full dataset after removing obvious outliers.

2. Offline Validation Set To avoid the common online‑downward trend after offline improvements, we constructed validation sets strictly following chronological order (e.g., using data from the 27th, 28th, and 29th as test sets) to prevent information leakage and maintain consistency between offline and online performance.

3. Feature Engineering Our features are mainly basic statistical metrics, with the best single‑model feature set containing about 40 dimensions. We experimented with conversion rate statistics, windowed and smoothed calculations, and even tried Word2Vec/Doc2Vec embeddings, but found simple conversion‑rate features to be fast and effective for our models.

4. Model Similar to last week’s champion, we used GBDT and FFM in the preliminary stage, noting GBDT performed better online while FFM excelled in the finals. Our best single model is an FFM, and a wide‑and‑deep model also achieved good results. Recent work focuses on improving training efficiency for stacking, and we plan to concentrate on better model fusion in the final week.

We encourage everyone to treat the competition as fun rather than a grind, to read others' experiences and related papers, and to keep learning.

For more details, visit the official competition website: http://algo.tpai.qq.com. Follow the official WeChat account TSA-Contest for additional resources and gifts.