特征工程

Using intuition to design new features, 基于业务理解 + 基于常用套路 + 基于EDA

我个人喜欢基于类似OOD设计的ER图进行思考

• numerical feature

• categorical feature

• sparse feature

• dense feature

• time feature

• sequence feature

• graph feature

特征生成

• 类型角度: 比如数值类和类别类

• domain角度: 比如推荐常用的用户，物品，context, 交叉特征

特征选择

• correlation filtering

Feature store

数据经过feature engineering pipeline之后进入feature store，feature store负责特征的存储与serving

• Ingestion: Batch (ETL, Spark) and Stream (Flink, Kafka)

• Store: Offline (parquet, BigQuery) and Online (Redis)

• Service: offline and online access

问答

• high cardinality

• feature backfill: backfill offline first → experiment → promote only if useful.

  • feature lifestyle: draft → backfill → validate → promote → deprecate

  • How Pinterest Accelerates ML Feature Iterations via Effective Backfill

参考

精读

• 特征工程到底是什么？ - 砍手豪的回答 - 知乎

• Feature Engineering

扩展

• 刀功：谈推荐系统特征工程中的几个高级技巧 - 石塔西的文章 - 知乎

• Feature Store 101

• 有哪些精彩的特征工程案例？ - 冯国添的回答 - 知乎

• A Closer Look at Our Machine Learning Feature Store

• Lessons from Building a Feature Store on Flink

代码

• https://github.com/senkin13/kaggle

• https://www.kaggle.com/code/fabiendaniel/elo-world