PP: Extracting statisticla graph features for accurate and efficient time series classification

Problem: TSC, time series classification;

Traditional TSC: find global similarities or local patterns/subsequence(shapelet).

We extract statistical features from VG to facilitate TSC

Introduction:

Global similarity:

the difference between TSC and other classification: deal with sequentiality property.

traditional methods: K-NN algorithm + DTW, one intrinsic issue with DTW, is that it focuses on finding global similarities. 在我看来这句话，简直是boo shit，一个距离测量只关注与全局的相似度？它应该是全部的距离都包含。

Local features:

Bag-of-patterns; SAX-VSM; shapelets-based algorithms.

Suffering:

1. high computation complexity
2. suboptimal classification accuracy

Time series --------> VG --------> graph features

graph features: Motif distribution, density;

Q:

1. why it's called multiscale  VG
2. the statistical graph features: probability distributions of small motifs, assortativity and degree statistics.

much faster than Learning Shapelets and Fast Shapelet.

Future work:

1. Other useful and efficient graph features: degree distribution entropy, centrality, bipartivity, etc.

2. adopt MVG for multivariate TSC.