Tencent Social Ads Data Mining Expert Q&A: Feature Engineering, Modeling, and Competition Insights

Today, a data mining expert from Tencent's Social Ads department held a Q&A session for competition participants in their study group (ID: 150522270), aiming to share knowledge and experience.

Q1: Why does data from the 16th appear with only a few rows? A: The log data was delayed and landed in the 17th day's directory; you can choose whether to use it.

Q2: Why does calculating historical conversion rates on the full set and using a sliding window both show effects, and isn’t full‑set calculation a data leakage? A: Full‑set statistics use past and future data, which are correlated, so both methods can show positive effects; however, using future data causes leakage and overestimates results.

Q3: Is the competition dataset completely real or sampled by user ID? A: It is real data, sampled randomly by appid and userid .

Q4: For Bayesian smoothing of conversion rates, should we use windowed processing or full‑set processing? Does full‑set cause data leakage? A: Consider whether the test set contains that information; using full‑set statistics from the training set is acceptable if the same feature extraction logic is applied locally.

Q5: Highly correlated features like appid and adid have almost identical statistics; should both be used or handled specially? A: This is a normal feature selection task; rely on local performance.

Q6: How to handle missing values for combined statistical features, e.g., using the mean? A: Refer to previous Kaggle competition experience; choose imputation methods based on feature distribution, and treat missing values as a special category when using models like XGBoost or Random Forest.

Q7: For large‑scale training sets with long training times, what are efficient ways to build a smaller local training/validation set? A: See the public article "Data Mining Competition Big Data Processing and Modeling Experience" for guidance.

Q8: If predicting day 31 data, is full‑set statistics feasible, or will it be rejected during defense for using unavailable data? A: This method does not trigger the "unavailable data" issue because day 31 data is not used during prediction; it may still cause potential leakage during training but can be acceptable if results are satisfactory.

Q9: How to discover effective feature cross combinations without exhaustive groupby? A: Use common statistical metrics or visual tools like boxplots.

Q10: Can Factorization Machines (FM) replace manual feature crossing and achieve the same effect as Logistic Regression with many crossed features? A: Simple cross features (e.g., x1 * x2) need not be handcrafted, but complex features (e.g., userID statistics under specific appID & positionID) may still require manual extraction.

Q11: How to select features effectively when low‑importance features may help and high‑importance ones may be redundant? A: Feature importance is hard to measure; consider wrapper or embedded methods that select features based on model performance.

Reference link

Q12: Is prediction typically real‑time or offline? If real‑time, can we use day‑31 data before the event? A: Prediction is real‑time; the model for day 31 is trained before day 31 arrives, so day‑31 data is unavailable during training, matching the competition setup.

Q13: Resources for Vowpal Wabbit supporting LR + high‑order feature interactions and online learning? A: Vowpal Wabbit GitHub and a third‑place solution using VW with high‑order features: GitHub repo .

Q14: References for model ensembling methods and parameter selection? A: Kaggle experience article link and a comprehensive survey Kaggle Ensembling Guide .

Q15: Does stacking work if some models perform poorly? A: Successful stacking requires diverse models with relatively balanced performance; large performance gaps can reduce ensemble benefit.

Q16: Does Tencent currently use online training? A: Yes, online mini‑batch training is used.

Q17: Why do some records show two conversions for the same event, and how to handle it? A: The train.csv was inspected and no such case was found; the issue may stem from data merging problems.

Q18: Why does using the full dataset for baseline in the final round perform worse than using only a few days, unlike the preliminary round? A: The effect is not directly tied to data volume; recent data often carries more importance, highlighting the advantage of online learning that continuously adapts to new distributions.

For more details, visit the official competition website: http://algo.tpai.qq.com . Official WeChat for contest updates and gifts: TSA-Contest .

Reference material: Kaggle experience sharing (see images above).