Adaptivemobile activity recognition system with evolving data streams
Zahraa Said Abdallah ,MohamedMedhatGaber
Bala Srinivasan , Shonali Krishnaswamy a Faculty of Information Technology,MonashUniversity,Melbourne, Australia b School of Computing Science andDigitalMedia, RobertGordonUniversity,UK c Institute for Infocomm Research (I2R), Singapore a r t i c l e i n f o a b s t r a c t Article history: Mobile activity recognition focuses on inferring current user activities by leveraging sensory data Received 18 February 2014 vailableon today's sensor richmobilephones. Supervised learningwith staticmodelshasbeen applied Received in revised form pervasively formobile activity recognition. In this paper, we propose a novel phone-based dynamic 10 September 2014 recognition framework with evolving data streams for activity recognition. The novel framework Accepted 13 September 2014 Available online incorporates incremental and active learning for real-time recognition and adaptation in streaming settings.While stream evolves, we reﬁne, enhance and personalise the learning model in order to Keywords: accommodate thenaturaldrift in a given data stream. Extensive experimental resultsusing real activity Ubiquitous computing recognition data have evidenced that the novel dynamic approach shows improved performance of Mobile application recognising activities especially across different users.Activity recognition & Streammining 2014 Elsevier B.V. All rights reserved. Incremental learning Active learning
activity recognition researchhas focused on traditional classiﬁcatory learning techniques. First,data is collected and annotatedbydomain Data stream mining has unique characteristics that make it experts.
Then, labelled data is deployed to build and train the more challenging than static data mining. In typical streaming classiﬁer learningmodel.When themodel is ready, the recognition settings, data has an inﬁnite length; therefore, traditionalmining system is used to predict activities from the sensory data. Awide techniques that require severalpassesondata cannotbeapplied in rangeof classiﬁcationmodelshasbeendeployed foractivity recogni- a streaming environment. Concept-drift nature of streaming data tion such as Decision Trees, Naive Bayes and Support Vector makes ithard topredict and classifynew incomingdata. Theprior Machines. The premise underlying machine learning in activity knowledge of data eventually becomes outdated while stream recognition is that new activities can be recognised using prior evolves. Thus, the classiﬁcatory model has to be continuously knowledge of previously collected data representing different activ- updated and reﬁned to copewith changes occurring naturally in ities. State-of-the-art activity recognition systems rely strongly on data stream.
There aremany emerging applications inwhich data priorknowledge, ignoringpostdeployment essential adaptation and streamsplayan important andnatural role,oneof these is activity reﬁnement that are naturally occurring in a dynamic streaming recognition. Activity recognition aims to provide accurate and environment.Moreover,personalisation ofmodel to suit aparticular opportune informationonpeople's activities andbehaviour.Activ- user had little focus in the research area. Typically,walking for one ity recognition has become an emerging ﬁeld in the areas of usermaywell be running for another, therefore tuning the general pervasive sensory data processing and ubiquitous computing model to recognise agivenuser'spersonalised activities is crucial for because of its important proactive and preventing applications. building a robust activity recognition system.Thus, itbecame crucial The increased interest in the ﬁeld of human activity recognition to handle the emerging change of activities resulted from the contributes to numerous domains, such as health care [18,32] modiﬁcation of users' activities patterns or personalisation of user's surveillance and security [26,11], personal-informatics [32,5] and activitieswhile stream evolves. just-in-time systems [13,25]. In this paper, we aim to build a personalised and adaptive Activity recognition (AR) has beenwidely studiedwith different framework for activity recognition that incrementally learns from approaches and from different perspectives. The state of the art in evolving data stream. The developed framework dealswith high speed,multi-dimensional streaming data to learn,model, recog- nisepersonaliseduser's activities. Thenovel approach extends then Corresponding author. http://dx.doi.org/10.1016/j.neucom.
Z.S. Abdallah et al. /Neurocomputing
7a complete retraining of the model. However, deploying this model replacesoldone for futurepredictionwith streamevolution model in streaming settings requires real rime adaptation with (line 8). limited availability of labelled data. Alternatively, our proposed approachuses adynamicmodel that couldbe reﬁned and adapted Algorithm 1. STAR top level algorithm. automatically while stream evolves. Themodel is also persona- Input: TrainigData: Annotated data for building LM lised to a particular userwithno need to retrain.Our technique is Stream: Sensory data evolving from sensors deployed on the mobile streaming environment for real time Output: Predcticed : activities predicted labels recognition and adaptation. Finally, our technique addresses the
1: LM’ModellingComponent (TrainingData) limited availability of labelled data by integrating batch active
2: while Stream not empty do learning in data streammining for activity recognition to select
3: WindowDatat’Stream data of the slidingwindow at only themost proﬁtable data to label in batches. time t
4: WindowClusterst’OnlClus (WindowDatat)
5: for all Clusteri WindowClusterst do 3. STAR top level description A
6: Predictedti’PredictActivityLabel (Clusteri, LM) new’ In this section, we introduce our novel STream learning for
7: LM UpdateFramework (Predictedti, Clusteri) ’ new mobileActivityRecognition framework- termed STAR– alongwith
8: LM LM its phases and components. In terms of the learning paradigm,
9: end for STAR is divided across two phases: ofﬂinemodelling phase and
10: endwhile online recognition and adaptation phase. In the ofﬂinemodelling phase, we build a learningmodel (LM) from a set of annotated sensory data that represents different activities. The output of the Ofﬂine phase: The aim in this phase is to build a ﬂexible, ofﬂine modelling phase is a ﬁne grained learning model that expandable, ﬁne grained and light-weight learning model for represents activities existing in the training examples. In the activity recognition. The novel dynamic model consists of a set online phase, we introduce a dynamic recognition technique of of clusters, each contains its corresponding sub-clusters. Sub- unlabelled streaming datawith different activities. The key chal- clusters represent different patterns inside a cluster. Themodel- lenge in this phase is to enable incremental and continuous ling component processes annotated sensory data for training. learning so that the recognition system can copewith and reﬂect Only summarised characteristics of the data remain intomemory the expected changes in single user's activities or across different to represent various activities for the following phase.
The light- users. Fig
1 shows the top level explanation of STAR with its weight and ﬂexibility of the created learningmodel allow future phases and components. necessarymodel tailoring to best ﬁt drifting data stream. Theoverallprocess forourapproach isdepicted inAlgorithm
1. Online phase:
The onlinephase consists of both the recognition Subsets of the algorithm are descried in subsequent sections. and adaptation components. The two components are implemen- Modelling component builds the initial learning model from ted on board the mobile phone for real-time recognition and annotated data.We split data stream into equal size chunks of adaptation. Firstly,wedeploy the light-weight learningmodel and unlabelleddata (line3). Inorder todetect concurrentactivities,we integrate itwith the recognition component forpredicting incom- apply an online clustering for each chunk of data (line 4).
The ing sensory streaming data. Then, we apply our novel stream predictionand adaptation techniques aredeployedoneach cluster mining ensemble prediction technique for recognising different (lines 6 and 7). Themodel is reﬁned with recent data. Updated actenvironment, our recognition approach deals with activities as learningmodel is described in the following equation: clusters rather than responding toeach single instance in the streamsc21;sc22…sc2P2 ;…Cn scn1;scn2…scnPn Thus, the predicted activitywith the cluster-based approach corre-sponds to themajor activity performed in the target timewindow. When a new cluster Cnew emerges, the ensemble classiﬁer is Another advantage of using the cluster-based approach is the applied to choose the best candidate cluster among n clusters preservationof the limited resourceson themobiledevice especially withm sub-clusters in the LM.Wherem is the total number of n when dealingwith unbounded very fast sensory data stream. sub-clusters ∑i 1Pi. Number of sub-clusters varies from one ¼ ¼ Capturing concurrent and interleaved activities: One of the chal- cluster to another based on the diversity of patterns within the lengeswith
thewindowing, cluster-based approach for recognition is cluster. The four assessingmeasures are the following: how to handle concurrent and interleaved activities. Typically, some activitiesarehighlycorrelated suchaswalkingand standing.Even ina Distance: This basicmeasure focuses on the closeness and the small sizewindow one personwould have pauses of “stand” activity, separation of incoming data from LM clusters. Cj is the best for instance,while “walking”. Inour cluster-basedapproachweaim to candidate from distance perspective if the Euclidean distance detect the major activity performed in a single window. In this between CentroidCnew and Centroidsci is the shortest among all n example, only themajor activity, namelywalking,will be detected. clusterswithm sub-clusterswhere sciACj. However, capturing concurrent and highly related activities improves Gravity: Thismeasure concerns not only the closeness but also theoverallefﬁciencyof the system.Therefore,wecombine thecluster- theweightof the clusters.Each clusterhas itsowngravitational based approachwith a concurrent activity detection technique to be force generated from itsweight. The gravitymeasure focuses able to detect concurrent and interleaved activities in eachwindow/ on the attraction among clusters caused by their gravitational cluster. force as per Eq. (4.11). The bigger theweight of the cluster the In order to capture concurrent activities in a singlewindow,we stronger the gravitational force produced around it. Therefore, propose an online clusteringmethod (OnlClus) for eachwindow – as the probability it could attractmore data is increased.When explained in Fig. 3. OnlClus aims to detect and separate concurrent the gravitational force between Cnew and sci is the biggest activities thatoccur at the samewindow.At this step,we applywell- among alln clusterswithm sub-clusters,where sciACj, then Cj known clustering techniques suchasK-meansorEMon the small size is the best candidate from the gravity perspective. window data. The formed clusters of eachwindow are assessedwith Density: Among the choices ofmeasures is the densitywhich an ensemble classiﬁer for recognising their corresponding activities. concerns data cohesiveness and dispersion for both the cluster level and collective levelof all clusters. Thedensityperspective 4.2.1.
Ensemble classiﬁcation technique studies the impact on
LM cluster/sub-cluster density if new Theﬂowof the recognition component continuewith applying data joins this particular cluster. In order to choose the best an ensemble classiﬁcation technique on clusters of eachwindow. candidate from density perspective, we compute the virtual The ensemble classiﬁer is a light-weight algorithm based on a densitydifference (VDD) forboth cases of gainor loss. TheVDD hybrid similarity measures approach for prediction. Thus, the computes the virtual value of gain/loss when the Cnew is classiﬁer deploys an ensemble of four measures to assess the mergedwith existing LM sub-cluster compared to sub-cluster similarity of each new cluster (of the current window)with LM original density Densitysc. Cj is the best candidate from the clusters. State of the art classiﬁcation methods adopt a single density perspectivewhen the VDD of Cnew ifmergedwith sci is measure for prediction, combing variousmeasures beneﬁts from either of the maximum gain or minimum loss among all n the strengthof each andprovides a comprehensiveunderstanding clusterswithm sub-clusters,where sciACj,. of data from various perspectives. Eachmeasure votes for its own
Deviation:Unlike density, deviation focuses on cluster's internal “best candidate” cluster from themeasure'sperspective. Then, the cohesiveness around centroid.Whennew cluster Cnew emerged, classiﬁer decides upon the predicted label as the one with the we test the impact onWISCSD for each sub-cluster. The best majority votes among allmeasures. The deployedmeasures are candidate from the deviation prospective is the onewith least distance, density, deviation and gravity. The deployment of the affect on itsWISCSDwhenmergedwith the new data. ensemble classiﬁcation for clustering, classiﬁcation and activityivities onlinewith continuouswindowing of the data stream.
marketing assignment, marketing management assignment, marketing management assignment help, digital marketing assignment help, marketing assignment help online, marketing homework help, marketing assignment writing, online marketing assignment help, strategic marketing assignment help, online marketing assignment, marketing assignment writing service, international marketing assignment help, marketing mix assignment help, marketing homework assignment