In our data collection, Nokia N82 mobile phones were embedded with .... [14] X. L. Wang, High accuracy distributed target detection and classification in sensor ...
Online Sequential ELM based Transfer Learning for Transportation Mode Recognition Zhenyu Chen1,2,3,4 , Shuangquan Wang1,2 , Zhiqi Shen5 , Yiqiang Chen1,2 and Zhongtang Zhao1,2,4 1 2
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China Beijing Key Laboratory of Mobile Computing and Pervasive Device, Beijing, China 3 Computer Science Department, Dartmouth College, Hanover, NH, USA 4 University of Chinese Academy of Sciences, Beijing, China 5 School of Computer Engineering, Nanyang Technological University, Singapore
Abstract—Transportation mode recognition plays an important role in discovering life patterns from people’s physical behavior. Learning knowledge from mobile sensing data enables transportation mode recognition on mobile phone. However, existing transportation mode recognition methods are mostly based on fixed recognition models, which do not consider the diversities in different users and their transportation context. In this paper, an online sequential extreme learning machine based transfer learning method (TransELM) is proposed to recognize various transportation modes. TransELM is mainly comprised of three steps: firstly, an initial ELM classifier is trained on the labeled training data from the source domain; secondly, the mean and standard deviation are calculated as multi-class trustable intervals in source domain, and then the partially trustable samples are effectively extracted from the target domain; thirdly, the trustable samples are integrated, where an incremental OSELM method is employed to update the original ELM classifier. Experimental results show that TransELM obtains higher accuracy than the traditional ELM classifier in real world transportation mode recognition problems.
I.
I NTRODUCTION
With the advances of wireless communications, micro electrical mechanical systems (MEMS) and sensor technology, massive sensor-embedded mobile phones are able to continuously collect sensor data at a large scale to make reliable inferences about people’s life patterns [1], [2], [3]. Transportation mode recognition plays an important role in mining the life patterns from people’s physical behavior, and machine learning methods applied on mobile sensing data enable transportation mode (e.g. walking, driving, bicycling, taking light-rail, etc.) recognition on mobile phones. Transportation mode recognition can also be applied in various fields such as mobile recommendation, route planning, mHealth, social networks, behavioral and cognitive sciences. Traditional transportation mode recognition methods are commonly based on the hypothesis that the labeled training data and the unlabeled testing data have the same distribution in feature spaces, which means that these fixed recognition models are only established on the labeled training data but not the unlabeled testing. However, in real world applications, the training and testing datasets usually have different distributions and significant differences in the structure of feature spaces between the source domain and the target domain, which results in the significant decrease of recognition accuracy using fixed recognition models. Hence, we need to transfer the valuable knowledge learned from the original domain
(source domain) to the new domain (target domain), where such valuable knowledge comes from the common features and trustable samples shared by the two domains. To address the transfer learning problem for transportation mode recognition, this paper proposes an online sequential extreme learning machine (OSELM) based transfer learning method called TransELM. This method can utilize valuable features and trustable samples to effectively transfer common knowledge across labeled source domain and unlabeled target domain. TransELM mainly includes three steps: Firstly, an initial ELM classifier is trained on the labeled training dataset from the source domain. Secondly, relevant mean and standard deviation values are separately computed as trustable intervals for each class of transportation modes. The unlabeled dataset of target domain is classified with the initial ELM model and trustable samples whose output values belong to corresponding trustable intervals are effectively extracted. Thirdly, for integrating these trustable samples, an incremental OSELM method is employed to incrementally update the original ELM classifier. We present experimental results from our user study in this paper, including five people with six typical transportation modes in the daily life. Experimental results show that TransELM has following advantages in transportation mode recognition: 1) combining accelerometer and GPS data generated on mobile phones, our TransELM method can effectively adopt distinctive characteristics for transportation mode recognition without additional sensing device or equipment required; 2) mobile phone does not require fixed device orientation and displacement position as is in other wearable devices; 3) TransELM does not have iterative or retraining step, which means the computation and storage costs are comparative small; 4) TransELM is promising to be applied in large-scale application in the real world. The rest of the paper is organized as follows. In Section 2, related work for transportation mode recognition is introduced, and Section 3 describes the methodology of TransELM in detail, then experimental results are presented and discussed in Section 4, finally Section 5 concludes this paper and gives the future work. II.
R ELATED W ORK
In recent years, encouraging progress has been made in transportation mode recognition, which can monitor people’s travel behavior and infer their life patterns. At present, existing
transportation mode recognition methods mainly utilize popular sensors such as Global System for Mobile communication (GSM), accelerometer, Global Positioning System (GPS), etc. [4] and [5] detect received signal strength of GSM signals and the number of different GSM cell towers to recognize mobility modes such as stationary, walking and driving. [6] combines two sensor modules of GSM and WiFi to train a Decision Tree (DT) classifier based on extracted temporal and spatial features. Reddy et al. [7] propose a two-stage classification method, the first stage is also a DT classification stage and the second stage is using a first-order Hidden Markov Model (HMM) method, which can eliminate unlikely transitions between adjacent two modes like transiting from bicycling mode to driving mode. Troped et al. [8] explore adding GPS data to supply accelerometer data in the recognition of activity modes, which indicates some extent of improving the classification accuracy. Wang et al. [9] derives an acceleration synthesization based feature extraction method to recognize six representative kinds of transportation modes using only one tri-axial accelerometer sensor embedded on mobile phone, which can prove that the proposed acceleration synthesization method is good at dealing with the variations of wearable position and placement orientation of mobile phone devices. [10] firstly applies a supervised learning based approach to obtain the recognition result of transportation mode from raw GPS data, then consider the transition probability between different transportation modes and the recognition result is post-processed accordingly, which can improve the accuracy of final recognition result. [11] and [12] use the map matching to enhance recognition performance by fused GPS and GIS information. However, these traditional recognition methods are generally considered as fixed models, which cannot take transportation mode diversities into account, such as the differentiation of user personalization in phone usage, the change of traffic situation in various contexts and so forth.
Considering most practical transportation mode applications, there actually exist some extents of differentiations and inconsistencies, which probably come from 1) different people have personalized behavior modes in the daily life; 2) transportation situation occurs obviously over various contexts of traffic districts and time periods such as the workload of road condition, the extent of traffic congestion, etc.; 3) mobile phones are usually carried with varying manners by personalized users like randomly wearable displacement position, device orientation and so on. Thus, we need a transfer method of transportation mode recognition, while fixed recognition methods cannot deal with diversities of transportation modes. To achieve good recognition accuracy and generalization capability, this paper proposes a novel transportation mode recognition method on mobile phones using OSELM based transfer learning. The initial ELM classification model is trained over accelerometer and GPS fused features in the labeled training dataset. After trustable samples extraction, the OSELM based incremental method is increasingly updated, which combines parameters of the initial ELM classifier and features of newly extracted trustable samples in the unlabeled testing dataset, then completes the update process to finally generate classification results for transportation mode recognition.
Step 3. The TransELM model generation
Step 2. New unlabeled instances selection
Step 1. The initial ELM model training
Source domain (labeled data)
Fig. 1.
Target domain (unlabeled data)
The framework of proposed TransELM method.
III.
OVERVIEW OF T RANS ELM
The framework of proposed TransELM method is shown in Figure 1, which mainly consists of three steps. Step 1: The common features of accelerometer data based on the acceleration magnitude series and GPS data based on the the GPS velocity series are separately extracted; and then by combining accelerometer features and GPS features, an initial ELM classifier is trained on the labeled training dataset of source domain. The parameters of this initial ELM model, such as randomly selected input weight vector a, bias vector b, activation function G(a, b, x) and the number of hidden nodes ˜ the hidden layer output matrix H and the output weight β, N, are reserved when the initial ELM classifier is generated. Step 2: on the one hand, from all output values of each class of transportation modes in source domain, corresponding mean and standard deviation values are calculated as multi-class trustable intervals; on the other hand, the unlabeled testing data in target domain is classified by the initial ELM classifier, then the classification output values are estimated according to defined trustable intervals, and a portion of samples whose output values belong to corresponding classes of trustable intervals are extracted, which are considered as trustable samples in the new unlabeled instances selection. Step 3: integrating these trustable samples, an incremental OSELM method is employed to increasingly update the current ELM classifier. With the unlabeled samples coming through the data stream one by one or chunk by chunk, the incremental classifier is continuously repeated from Step 2 to Step 3 until no new sample arrives. A. Feature Extraction 1) Accelerometer Feature Extraction: As the accelerometer detects and transforms changes in capacitance into an analog output voltage, acceleration features are extracted from sensor readings of triaxial accelerometer. These output voltages can be mapped into acceleration along three axes, ax , ay , az , which are the orthogonal decompositions of real acceleration, thus the magnitude of synthesized acceleration can be expressed as:
a=
q
a2x + a2y + a2z
(1)
Based on the acceleration magnitude series, 14 acceleration features [13] are extracted from a sliding window of 8 seconds without overlapping between consecutive windows. They are maximum, minimum, mean, standard deviation, energy, zero crossing rate, four amplitude statistics features and four shape statistics features of the power spectral density (PSD) [14].
Figure 2 illustrates the network structure of ELM. Given a training set ℵ = {(xi , ti )|xi ∈