PP: Learning representations for time series clustering

Problem: time series clustering TSC - unsupervised learning/ category information is not available.

time-series clustering for anomaly detection/ pattern detection. 

Feature-based time series clustering methods typically rely on domain knowledge to manually construct high-quality features. 

Deep temporal clustering representation DTCR: add temporal reconstruction and k-means into the seq2seq model. 

Introduction:

time-series clustering ----- data mining technology: from data to knowledge/ extract valuable information;

Feature-based methods: extracts features and then clusters. This kind of methods is robust to noise and outliers. It can conduct dimension reduction to improve efficiency. 

However, most existing methods require domain knowledge to construct high-quality features manually. 

discriminative features. 

Seq2seq Model: it can learn general representations from sequential data. 

We aim to learn a non-linear temporal representation for TSC using seq2seq model. 

当使用seq2seq模型时，由于缺少labels，无法进行学习，guide the learning process to generate cluster-specific representations. 所以该论文如何解决这个问题？

generate cluster-specific temporal representations. 

DTCR = temporal reconstruction + k-means + seq2seq model

个人观点： 这不会是主流方法，而且结果图有分类比较和聚类比较，结果感觉不真实。

对于TSC，本不应该应用方法的拼拼凑凑，而且他说有辅助分类，在真实世界中，不可能有辅助分类帮助你进行聚类。

Supplementary knowledge:

1. other TSC methods:

• encode time series into images, and then use CNNs etc, like recurrence plots/  Gramian angular summation/ Gramian angular difference fields/ Markov transition fields. 

2. PP: Motif difference field (MDF): A simple and effective image representation of time series for classification. 

encode time series into MDF images.

This paper tries to include temporal information while encoding.