(IGTS) Temporal Segmentation for multivariate time series

Data collected in this repository contains resources used and described in the paper. The repository is structured as follows:

paper/: Formal description of the algorithm and evaluation results
code/: The Matlab project code
presentation/: The presentation slides

Code
The codes are available in code/.

"GetTS_Syn.m": The function generates synthetic data
"main": Run this file for the implementation of the algorithm. All the variables are defined in this file.
"best_k": It finds the best candidate for the number of the segments based on the given information gain.
"Clean_TS": This function normalizes the time series and doubles the number of the time series to address the hetergenousity in the time series.
"DP_IG": Implementation of the dynamic programming for IGTS
"FullSearch": Implementation of the full search for IGTS
"GetTS_USC": It preprocesses the USC data for the algorithm
"Hierarchical": Implementation of the TOP-Down algorithm for IGTS
"IG_Cal": It calculates the information gain
"ksegmentation": Implementation of k-segmenatation algorithm
"movingvar": It is an optional function that adds variance of the time series
"Sh_Entropy": It calculates the entropy

The input is m time series with the length of n that should be stored in an m*n matrix.

Paper Abstract

This paper aims to observe and recognize transition times, when human activities change. No generic method has been proposed for extracting transition times at different levels of activity granularity. Existing work in human behavior analysis and activity recognition has mainly used predefined sliding windows or fixed segments, either at low-level, such as standing or walking, or high-level, such as dining or commuting to work. We present an Information Gain-based Temporal Segmentation method (IGTS), an unsupervised segmentation technique, to find the transition times in human activities and daily routines, from heterogeneous sensor data. The proposed IGTS method is applicable for low-level activities, where each segment captures a single activity, such as walking, that is going to be recognized or predicted, and also for high-level activities. The heterogeneity of sensor data is dealt with a data transformation stage. The generic method has been thoroughly evaluated on a variety of labeled and unlabeled activity recognition and routine datasets from smartphones and device-free infrastructures. The experiment results demonstrate the robustness of the method, as all segments of low- and high-level activities can be captured from different datasets with minimum error and high computational efficiency.

本仓库所收录的数据包含论文中使用并阐述的相关资源，仓库结构如下：

paper/：存放算法的正式阐述内容与评估结果
code/：Matlab 项目代码
presentation/：演示文稿幻灯片

### 代码说明
相关代码均存放于code/目录下：
"GetTS_Syn.m"：用于生成合成数据的函数
"main"：运行该文件以实现算法，所有变量均在此文件中定义
"best_k"：基于给定的信息增益，选取最优的分段数目候选值
"Clean_TS"：该函数对时间序列进行归一化处理，并将时间序列的数量翻倍，以解决时间序列中的异质性问题
"DP_IG"：针对IGTS的动态规划实现
"FullSearch"：针对IGTS的全搜索实现
"GetTS_USC"：为算法预处理USC数据
"Hierarchical"：针对IGTS的自上而下（TOP-Down）算法实现
"IG_Cal"：用于计算信息增益
"ksegmentation"：k分割算法的实现
"movingvar"：可选函数，用于添加时间序列的方差特征
"Sh_Entropy"：用于计算熵值

输入为m条长度为n的时间序列，需以m×n矩阵的形式存储。

### 论文摘要
本文旨在观测并识别人类活动发生转变时的过渡时刻。目前尚未有通用方法可用于提取不同活动粒度层级下的过渡时刻。现有人类行为分析与活动识别研究多采用预定义滑动窗口或固定分段策略，且仅适用于低层级活动（如站立、行走）或高层级活动（如用餐、通勤上班）。本文提出一种基于信息增益的时间分段方法（Information Gain-based Temporal Segmentation, IGTS），这是一种无监督分段技术，可从异质传感器数据中识别人类活动与日常作息的过渡时刻。所提出的IGTS方法既可应用于低层级活动场景——此时每个分段对应单一待识别或预测的活动（如行走），也可适用于高层级活动场景。针对传感器数据的异质性问题，本文通过数据预处理阶段予以解决。该通用方法已在来自智能手机与无设备感知基础设施的各类标注及未标注活动识别与日常作息数据集上进行了全面评估。实验结果表明，该方法具备良好鲁棒性，可从不同数据集以最小误差与高计算效率捕获高低层级活动的所有分段。