drop-out / Rnn Active User Forecast
Licence: mit
1st place solution for the Kuaishou Active-user Forecast competition
Stars: ✭ 179
Labels
Projects that are alternatives of or similar to Rnn Active User Forecast
Ethereum demo
This is the code for "Ethereum Explained" by Siraj Raval on Youtube
Stars: ✭ 177 (-1.12%)
Mutual labels: jupyter-notebook
Autofeat
Linear Prediction Model with Automated Feature Engineering and Selection Capabilities
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
Chess Surprise Analysis
Find surprising moves in chess games
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
Unet In Tensorflow
U-Net implementation in Tensorflow
Stars: ✭ 177 (-1.12%)
Mutual labels: jupyter-notebook
Lede Algorithms
Algorithms course materials for the Lede program at Columbia Journalism School
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
Super resolution with cnns and gans
Image Super-Resolution Using SRCNN, DRRN, SRGAN, CGAN in Pytorch
Stars: ✭ 176 (-1.68%)
Mutual labels: jupyter-notebook
Papergraph
AI/ML citation graph with postgres + graphql
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
Deeplearninglifesciences
Example code from the book "Deep Learning for the Life Sciences"
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
Python data science and machine learning bootcamp
Jupyter notebook for Udemy course: Python data science and machine learning bootcamp
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
Tensorflow2 Docs Zh
TF2.0 / TensorFlow 2.0 / TensorFlow2.0 官方文档中文版
Stars: ✭ 177 (-1.12%)
Mutual labels: jupyter-notebook
Ocaml Jupyter
An OCaml kernel for Jupyter (IPython) notebook
Stars: ✭ 177 (-1.12%)
Mutual labels: jupyter-notebook
2017 Ccf Bdci Aijudge
2017-CCF-BDCI-让AI当法官(初赛):7th/415 (Top 1.68%)
Stars: ✭ 177 (-1.12%)
Mutual labels: jupyter-notebook
Code Of Learn Deep Learning With Pytorch
This is code of book "Learn Deep Learning with PyTorch"
Stars: ✭ 2,262 (+1163.69%)
Mutual labels: jupyter-notebook
Lstm anomaly thesis
Anomaly detection for temporal data using LSTMs
Stars: ✭ 178 (-0.56%)
Mutual labels: jupyter-notebook
作者按:由于比赛时间仓促,这份代码中有些地方写的并不规范。更规范的tensorflow RNN构建,可以参考作者的另外一个项目tenosrflow-RNN-toolkit,该项目使用更高程度抽象的building block构建RNN,同时不失灵活性。
赛题回顾
这是一个活跃用户预测问题。给定快手用户注册、登陆、视频观看与发布、互动的记录,预测未来7天活跃用户。
详情可参见比赛页面。
RNN: Many2One vs Many2Many
使用RNN,一般地会想到如下解决方案:以几天内的用户行为序列为输入,以未来七天该用户是否活跃为标签,标注该序列。这是一种Many2One的解决方案。
为了充分利用数据,需要对训练数据做大量的滑窗,以实现数据增广,计算成本高。另外,每个序列只有一个标签,梯度难以传导,导致训练困难。相反的,我们可以考虑Many2Many结构,即每个输入都对应输出之后7天是否活跃,充分利用监督信息,减轻梯度传到负担,使训练更加容易。
Many2One和Many2Many结构的简单对比如下。
Many2One | Many2Many | |
---|---|---|
无需滑窗 | √ | |
充分利用监督信息 | √ | |
变长序列 | √ |
输入序列
相比xgboost的历史统计量为特征的解决方案,RNN无需对输入序列做过多处理,对各类行为序列直接输入即可。简单列表如下:
- 当天是否登陆(0/1)
- 当天观看次数(加1取对数)
- 分action_type行为记录数(加1取对数)
- 分page行为记录数(加1取对数)
Intercept
另外,在输出层直接做一个intercept拼接,将日期、device_type、register_type one-hot后输入。低频类别可归为一类。
Variable Length
因为序列是变长的,采用dynamic-RNN,每个batch中取相同长度的序列,不同batch长度不同,每次随机取某一长度的batch。
余弦退火快照集成
采用余弦退火快照集成,可以以极低的成本获得大量有差异的局部最优,最后再进行融合,能获得显著的提升。
Note that the project description data, including the texts, logos, images, and/or trademarks,
for each open source project belongs to its rightful owner.
If you wish to add or remove any projects, please contact us at [email protected].