AgentSkillsCN

Data Science

从探索阶段到特征工程，全面规范数据科学工作流的标准。

SKILL.md
--- frontmatter
name: Data Science
description: Standards for Data Science workflows, from exploration to feature engineering.

Data Science Standards

1. Exploration vs. Production

  • •Notebooks: Use .ipynb ONLY for initial EDA (Exploratory Data Analysis) and visualization.
  • •Scripts: All logic destined for the pipeline must be refactored into .py modules (e.g., src/features/).
  • •Rule: Never run a notebook in a production DAG.

2. Feature Engineering Principles

  • •Point-in-Time Correctness: Features for a given row (Player-Week) must ONLY use data available prior to that game.
    • •Example: Calculating "Next Week's Salary" valid? NO.
    • •Example: Calculating "Avg Score Last 3 Weeks"? YES.
  • •Leakage Prevention: Strictly split training/validation by TIME (e.g., Train 2011-2022, Valid 2023), never random shuffle, due to time-series correlations.

3. Metrics that Matter

  • •Business Proxy: Always prioritize metrics that map to business value (e.g., "Financial Lift", "Dead Cap Saved") over raw technical metrics (MSE, Accuracy) unless debugging.
  • •Baselines: Always compare model performance against a naive baseline (e.g., "Predict last year's performance").