Multi-domain Modeling and AutoML Techniques from Kaggle/KDD Cup Championships

This article shares the author’s experience from seven championship wins in Kaggle and KDD Cup competitions, focusing on multi‑domain modeling optimization, an AutoML framework, and general modeling methodologies.

Background and Introduction : Highlights the significance of algorithm competitions for advancing techniques, citing examples such as Field‑aware Factorization Machines and ResNet.

Multi‑domain Modeling Optimization : Discusses three problem types—recommendation systems, time‑series forecasting, and automated machine learning—detailing representative competitions (Outbrain Click Prediction, KDD Cup 2020 Debiasing, KDD Cup 2018 Air Quality, etc.) and the proposed multi‑level, multi‑factor model‑fusion strategies.

AutoML Technical Framework : Describes a three‑module AutoML pipeline (data preprocessing, automated feature engineering, automated model optimization), including feature operators, fast feature selection, high‑order feature generation, importance‑driven grid search, and model‑fusion techniques.

General Modeling Methodology : Presents a three‑stage workflow—exploratory modeling, key‑issue modeling, and automated modeling—emphasizing consistent evaluation, bias‑variance trade‑off, and robustness.

Conclusion : Summarizes the shared insights and invites readers to apply these methods in competitions and industrial projects, also providing recruitment information.