Practical Guide to Machine Learning at Meituan: From Problem Modeling to Model Optimization

This article presents an end‑to‑end practical guide on applying machine learning to solve real‑world problems at Meituan, focusing on the offline training pipeline.

It begins with a brief overview of machine learning, defining it as the discipline that builds algorithms capable of learning from data, and distinguishes supervised from unsupervised learning, emphasizing that supervised methods dominate industrial applications.

The first technical step is problem modeling. Using the example of estimating the revenue of a DEAL (group‑buying order), the article explains how to collect domain knowledge, decompose the business goal into machine‑learnable sub‑tasks, and decide between a single‑model or multi‑model approach.

Data preparation follows. The guide stresses the importance of consistent data distribution between training, testing, and online environments, low label noise, and careful sampling only when necessary.

Feature extraction is then discussed. High‑level (generic) and low‑level (specific) features are introduced, with examples of POI and per‑capita consumption. The article notes that linear models require richer feature engineering, while non‑linear models can rely more on high‑level features.

Feature normalization techniques such as rescaling, standardization, and scaling to unit length are described, along with feature selection methods (filter, wrapper, embedded) to avoid over‑parameterization.

Model selection considerations include alignment with business objectives, compatibility with data and features, and the prevalence of the model in industry. Typical choices mentioned are Logistic Regression, GBDT, Random Forest, Linear SVM, and Deep Neural Networks.

The training section details the formulation of a hypothesis function and loss function (e.g., maximum likelihood for logistic regression) and introduces several optimization algorithms: gradient descent (stochastic and batch), Newton’s method, quasi‑Newton methods (BFGS, L‑BFGS, OWL‑QN), and coordinate descent.

After training, the guide addresses model evaluation and common pitfalls such as under‑fitting and over‑fitting, providing a diagnostic table and practical remedies.

Finally, the article summarizes key take‑aways: understand the business, ensure high‑quality data, perform thoughtful feature engineering, choose models that match the task, and iteratively diagnose and improve the model pipeline.