DTCR(Learning Representations for Time Series Clustering)论文理解
DTCR模型结构:
一.模型思路来源:
重构损失能够包含以前时间序列的相关特性,但学到的特征对分类任务并不一定适用。为使学到的特征形成类结构,从而获得特定于类的表示,所以模型中引入了k-means。另外,在序列到序列的模型中,encoder的能力越好,学到的潜在特征越好。因此为了提升encoder的能力,在模型中加入了假样本生成和辅助分类机制
二.损失函数解析
1.重构损失
Lreconstruction=1n∑i=1n∣∣xi−xi^∣∣22
L_{reconstruction}=\frac{1}{n}\sum^n_{i=1}||x_i-\hat{x_i}||^2_2
Lreconstruction=n1i=1∑n∣∣xi−xi^∣∣22
2.k-means损失
(1)原理:
给定Hm×NH^{m\times N}Hm×N,k-means的最小化可以被重构为和Gram矩阵HTHH^THHTH相联系的最大化问题。谱松弛将k-means的损失函数转换为如下形式:其中F是类指示矩阵
Lk−means=Tr(HTH)−Tr(FTHTHF)
L_{k-means}=Tr(H^TH)-Tr(F^TH^THF)
Lk−means=Tr(HTH)−Tr(FTHTHF)
由于H是固定的,所以上述损失函数可以转化为max Tr(FTHTHF),s.t. FTF=Imax \space Tr(F^TH^THF),s.t.\space F^TF=Imax Tr(FTHTHF),s.t. FTF=I
(2)本模型中的应用:
在DTCR中,H是动态的,Tr(HTH)Tr(H^TH)Tr(HTH)可以作为训练H的一个正则项,所以损失函数为:minH,F J(H)+λ2[Tr(HTH)−Tr(FTHTHF)],s.t. FTF=I:min_{H,F}\space J(H)+\frac{\lambda}{2}[Tr(H^TH)-Tr(F^TH^THF)],s.t.\space F^TF=I:minH,F J(H)+2λ[Tr(HTH)−Tr(FTHTHF)],s.t. FTF=I,其中J(H)是重构损失+分类损失。
在训练过程中,迭代更新F和H:
①固定F,用SGD更新H
②固定H,用max Tr(FTHTHF),s.t. FTF=Imax\space Tr(F^TH^THF),s.t.\space F^TF=Imax Tr(FTHTHF),s.t. FTF=I来更新F(每10步更新一次)
3.分类损失
随机打乱一些时间步来生成对应输入的假样本,送入encoder,再对这些数据进行分类,判断是real还是fake,分类结果用二维one-hot向量yi^\hat{\textbf {y}_i}yi^表示,这里y^i=Wfc2(Wfc1hi)\hat{y}_i=\textbf{W}_{fc2}(\textbf{W}_{fc1} \textbf{h}_i)y^i=Wfc2(Wfc1hi),其中,Wfc1∈Rm×d,Wfc2∈Rd×2\textbf{W}_{fc1}\in R^{m\times d},\textbf{W}_{fc2}\in R^{d\times 2}Wfc1∈Rm×d,Wfc2∈Rd×2是全连接层的参数。
损失函数为:
Lclassification=−12N∑i=12N∑j=121{yi,j=1}logexp y^i,j∑j=12exp y^i,j
L_{classification}=-\frac{1}{2N}\sum^{2N}_{i=1}\sum^2_{j=1}1\left\{y_{i,j}=1\right\}log\frac{exp\space \hat{y}_{i,j}}{\sum^2_{j=1}exp \space\hat{y}_{i,j}}
Lclassification=−2N1i=1∑2Nj=1∑21{yi,j=1}log∑j=12exp y^i,jexp y^i,j
总结损失函数:LDTCR=Lreconstruction+Lclassification+λLK−meansL_{DTCR}=L_{reconstruction}+L_{classification}+\lambda L_{K-means}LDTCR=Lreconstruction+Lclassification+λLK−means
三.DTCR具体训练算法
- Salient Subsequence Learning for Time Series Clustering论文笔记
- 论文笔记之Learning Deep Representations for Graph Clustering
- 【Paper Note】Convolutional Clustering for Unsupervised Learning 论文理解
- K-means hashing: An affinity-preserving quantization method for learning binary compact codes论文理解
- 论文研读--Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification
- Multi-style Generative Network for Real-time Transfer论文理解
- 论文阅读:Joint Learning of Single-image and Cross-image Representations for Person Re-identification
- 论文笔记之:Learning Cross-Modal Deep Representations for Robust Pedestrian Detection
- Anomaly Detection for Time Series Data with Deep Learning——本质分类正常和异常的行为,对于检测异常行为,采用预测正常行为方式来做
- A Simple Pooling-Based Design for Real-Time Salient Object Detection------论文理解
- 个人笔记An intelligent hybridization of ARIMA with machine learning models for time series forecasting
- 论文笔记:Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation
- 论文阅读——Generic and Scalable Framework for Automated Time-series Anomaly Detection
- DeepSense: a Unified Learning Framework for Time-Series Mobile Sensing Data Processing 论文笔记
- 【Paper Note】Convolutional Clustering for Unsupervised Learning 论文翻译
- 论文阅读《Joint Learning of Single-image Cross-image Representations for Person Re-identification》
- 论文阅读理解 - Learning Feature Pyramids for Human Pose Estimation
- 【Hybrid Neural Networks for Learning the Trend in Time Series】 阅读笔记
- Deep Neural Networks for Learning Graph Representations论文笔记
- 【时序模型】《Deep and Confident Prediction for Time Series at Uber》论文阅读