Applying Machine Learning in Shumei's Business: Supervised, Unsupervised, and Reinforcement Learning Cases

This article, based on Li Tian's talk at DataFun AI+ Talk, summarizes the practical deployment of machine learning at Shumei Technology, covering the three major ML domains—supervised, unsupervised, and reinforcement learning—and their specific applications in the company's products.

In the supervised learning segment, Shumei uses classification and scoring models such as logistic regression, decision trees, XGBoost, DNN, CNN, and RNN for credit scoring (TianXin), customer scoring/advertising (TianWang), and content moderation (TianJing). Feature engineering follows a traditional pipeline of feature extraction, selection, and model training, often employing sequence‑based modeling to reduce manual feature work.

Unsupervised learning is applied to fraud and anti‑cheat scenarios, where clustering algorithms like K‑means, DBSCAN, and isolation forest (iForest) detect anomalous behavior patterns that differ from normal user activity.

Reinforcement learning is discussed as a future direction; algorithms such as Q‑learning, Monte‑Carlo Tree Search, and rolling genetic algorithms are mentioned, though their current applicability to fraud detection is limited due to the need for well‑defined play‑review environments.

The article then details how these techniques are integrated into specific Shumei services:

TianXin: Multi‑source credit scoring using structured data, feature selection, and model deployment; challenges include data source bias and the need for per‑source model fusion.

Model Fusion: Stacking and other ensemble methods combine multiple models to improve coverage and robustness, handling both multi‑model and multi‑source scenarios.

TianWang: Real‑time event‑based scoring for advertising behavior, emphasizing low computational overhead, minimal storage, and simple deployment via RNN cells.

TianJing: Audio content moderation using features like MFCC, Pitch, and iVector, with classifiers such as HMM/GMM, DNN, and RNN.

Finally, the article outlines future exploration of reinforcement learning for fraud detection, noting the need for extensive reward design, play‑review definition, and large computational resources.

The author, Li Tian, is a machine‑learning and deep‑learning engineer at Shumei with a master's degree in Big Data and Text Analytics from the University of Essex.