image.png
模型
本文主要介绍一个学习路线,后续详细介绍各部分内容。常用的模型,以下基本可以涵盖主流思想:
- 传统时序模型:ARIMA,Prophet,EMD
- 构造时序特征的统计学习方法:LR,GBDT(xgboost\lightgbm)
- 深度学习方法:seq2seq,wavenet,transormer
企业研究
来自工业届的研究经验与创新模型,值得一看的文章和模型:
- Uber:Time-series Extreme Event Forecasting with Neural Networks at Uber
- Uber:How Uber Manages Uncertainty in Time-Series Prediction Models
- Google:Temporal Fusion Transformers for Multi-horizon Time Series Forecasting
- Google:Using AutoML for Time Series Forecasting
- Amazon:DeepAR
- Amazon:DeepGLO
- Facebook:Peophet
- Facebook:AR-Net research that combines autoregressive models and neural networks
Python库
发挥python的优势,多用已经造好的轮子。tsfresh可以自动构造时序特征,sktime是类似sklearn写法的预测库,pytorch和tensorflow实现了常见的深度学习模型。前面提到的模型ARIMA和prophet也有可调用的stats,pyprophet。
- tsfresh:https://github.com/blue-yonder/tsfresh
- sktime:https://github.com/alan-turing-institute/sktime
- pytorch prediction:https://github.com/jdb78/pytorch-forecasting
- tensorflow pediction:https://github.com/LongxingTan/Time-series-prediction
kaggle比赛
一些经典的kaggle比赛,Top方案中既有构造时序特征的gbdt方法,也有深度学习方法。优秀的方案里既有对模型的深刻认识,更是对数据和任务的精细解读。
- Corporación Favorita Grocery Sales Forecasting
- Web Traffic Time Series Forecasting
- Recruit Restaurant Visitor Forecasting
- M5 Forecasting - Accuracy
联系方式
YueTan,欢迎关注