EPLS: A novel feature extraction method for migration data clustering

Highlights

• A numerical feature extraction approach EPLS is proposed.

• EPLS attempts to preserve the most valuable features which are adaptive to different distance measures and different clustering approaches.

• EPLS-based clustering algorithm can scale to large-volumes of data for its dimensionality reduction characteristic.

• EPLS can be efficiently suitable for migration data clustering.

Abstract

Nowadays human activity data such as migration data can be easily accumulated by personal devices thanks for GPS. Analysis on migration data is very useful for society decision. Migration data as non-line time series have the properties of higher noise and outliers. Traditional feature extraction methods cannot address this issue very well because of inherent characteristics. Aiming at this problem, a novel numerical feature extraction approach EPLS is proposed. It is an integration of the Ensemble Empirical Mode (EEMD), Principal Component Analysis (PCA) and Least Square (LS) method. The EPLS model includes (1) Mode Decomposition in which EEMD algorithm is applied to the aggregation dataset; (2) Dimension Reduction is carried out for a more significant set of vectors; (3) Least Squares Projection in which all testing data are projected to the obtained vectors. Experimental results show that EPLS can overcome the higher noise and outliers based on migration data clustering. Meanwhile, EPLS feature extraction method can achieve high performance compared with several different clustering methods and distance measures.

Introduction

Human activities data can be easily collected by personal devices in recent years. An increasing number of human activity data is accumulated for the decision of society  [33]. Through human activities data analysis  [16], researches can find some interesting patterns. For example, some patterns can be mined from twitter data which are useful to forecast stock price  [33]. Air quality index can be used to find PM2.5 patterns  [5].

Migration data from Baidu system are used to track human migration activities; for example, during Chinese New Year of 2015, 3.6 billion passenger trips data have been recorded. Baidu map gathers an amount of migration activities dataset from smartphones which own Baidu Maps or other apps using its location-based platform. This dataset is human activity related, which reveals human movement pattern. Analysis on migration is an important aspect of many fields such as economy, traffic, and culture  [18].

How to analyze and apply these data is a challenge because the database is time related and has a higher noise and nonlinear level. Usually these data are treated as chaotic time series. A chaotic time series naturally has the properties of high dimension and large data size  [23], [25]. Usually, clustering is used for exploratory data analysis and act as a major processing step for other tasks [21], [24], [28]. It can be concluded that clustering is classified into three main branches: (1) whole time series clustering; (2) subsequence clustering; (3) time point clustering  [12]. As for whole time series clustering, there are three different categories, namely shape-based approach  [23], [46], [27], feature-based  [2], [19] and model-based  [44], [35]. Feature-based clustering is taken into consideration in this study.

Traditional feature extraction methods usually have some limitation to deal with time series clustering  [2], [19] because of the nature properties of chaotic time series such as nonlinear, high level of noise and outlier, and non-stationary. So a new representative method is needed to address these problems. It is suggested that not all of these clustering methods and similarity measures are appropriate for every time series databases  [41]. Usually Euclidean Distance (ED) can lead to good clustering results as a useful method  [11]. But ED measure is not a general method; for some databases, Elastic measure including Dynamic Time Warping (DTW) and Edit Distance can achieve higher performance  [6]. Finding a suitable distance measure or a clustering method in specific dataset with best result is difficult. However, a relatively general feature extraction approach for fixed database can deal with this problem to some extent.

In this paper, EPLS as hybrid model-based approach is proposed to extract numerical features for migration data which is gathered from Baidu map engine. EPLS attempts to preserve the most valuable features which are adaptive to different distance measures and different clustering approaches. EPLS approach is based on Ensemble Empirical Mode Decomposition (EEMD)  [9], Principal Component Analysis (PCA)  [32] and Least Square (LS) method  [3] which transfers the original time series into feature space. The proposed method is immune to dataset with higher noise and outliers. Meanwhile, the extracted feature from EPLS has relatively low dimension which shows that it can be adapted to different distance measures and clustering methods.

This paper is organized as follows. In Section  2, the related work is discussed. In Section  3, the details of EPLS algorithm, as well as EEMD, PCA and LS, are shown. Then, some experiment settings and database description are given in Section  4. In Section  5, the EPLS model based approach is applied to Migration data from Baidu system. Section  6 provides a summary of the results and concludes the whole paper.