The research in IoT and Smart Homes elds is rapidly growing, leading to the emergence of new services to improve the health and lifestyle of people based on the analysis of data that they produce performing their daily activities. However, researchers report a lack of high-quality publicly-available datasets: conducting experiments gathering such data is long and expensive, especially if the annotation of meaningful information (environment, person's activity, health status) is required. Moreover, there are even more speci c settings (e.g., dementia detection) where data must be related to a change in inhabitants' behavior. We present a collection of new publicly-available datasets generated with the SHARON simulator. Thanks to this software, researchers can obtain synthetic data suiting their speci c requirements. Two classes of datasets are described: one extends existing datasets preserving the original statistical properties, the other is composed of simulations of virtual inhabitant-environment systems. Moreover, we induced behavioral drifts compatible with dementia symptoms, generating further datasets. We believe that these Copyright © CIKM 2018 for the individual papers by the papers' authors. Copyright © CIKM 2018 for the volume as a collection by its editors. This volume and its papers are published under the Creative Commons License Attribution 4.0 International (CC BY 4.0). resources may help the progress of research, as long as new real-life high-quality datasets are not available. 1
The possibility of gathering large amounts of data from Smart Home environments is a valuable opportunity for the development of numerous applications, like, e.g., security, home automation, remote monitoring, etc.
Data are collected by using di erent types of sensors, connected to a home (usually wireless) network and stored in a central database. Localization of the inhabitants, state of the house such as brightness, temperature, humidity, doors and windows opening, as well as the activation of household appliances can be a source of knowledge for advanced analytics.
Moreover, in addition to the mentioned data, there is often a need for extensive descriptions of the context in which the data were collected: the so-called \ground-truth". For example, much attention has been dedicated to the research in the Activity Recognition (AR) eld { that is the task of identifying the ongoing Activity of Daily Living (ADL) from sensors data. As highlighted by Sprint et al. [SCFSE16], in order to access the Health Events related to a person living in a Smart Environment, supervised machinelearning algorithms are commonly used. Usually, AR requires a set of labels related to the performed ADLs: these data are provided by external annotators (often called oracles) which look at them and utilize extra information (such as videos, the house oor-plan, the resident pro le, etc.) to generate corresponding ground-truth labels.
s e s u o
H
It comes clear that creating Smart Home datasets with ground-truth information related to the inhabitant's activities and well-being status is a long and costly operation that often slows down the progress of research and advanced applications. In the next section we provide an overview of the currently publiclyavailable datasets, highlighting the strengths and weaknesses of the various resources. In Section 3 we describe a new set of resources, their peculiarities and how have they been generated. Finally, in Section 4 we conclude this work discussing future challenges about the dissemination of Smart Home datasets. 2
In recent years, many papers have been discussing the importance of the continuous monitoring of the person's behavior as a source of information concerning his/her well-being [RBC+15, PKL+05, PLJ+15]. According to Saives et al. [SPF15], improving the life of the inhabitant with new technological services makes a house \Smart"; those applications cover several interesting research elds, all of them sharing the same need to collect Home Automation datasets. A literature review by Rashidi et al. from 2013 reports 18 noticeable projects in Ambient Assisted Living, and con rms that \one of the most important component is Human Activity Recognition" [RM13]. Despite the great interest in research concerning Activity Recognition (AR) and Behavioral Drift Detection (BDD), the amount of publicly-available high-quality datasets is particularly small. Indeed, the collection of Home Automation data in controlled settings, with good annotation, is a hard and resource demanding task.
Table 2 summarizes the features of the most widely used datasets in the literature to evaluate AR and BDD research; as reported by Benmansour et al. [BBF16], AR and BDD with multiple residents introduce complexity in identifying the dwellers and disassociating data and activities.
ARAS (Activity Recognition with Ambient Sensing) is a project developed aiming at ADL recognition [AEIE13]. The authors have published their datasets, that comprise data collected from two houses with two inhabitants, for a duration of one month each. The deployed sensors set was composed of 20 boolean sensors, and data were annotated with 27 di erent ADLs. The dataset however reports erratic routine of the inhabitants (unusual meal times, unexpected behavior during the ADL, etc.), speci es only one activity at a time (even when two happens concurrently), and reports ADLs which cannot be identi ed due to sensor lack (e.g., no sensor to detect \using internet" and \reading" activities were present).
CASAS (Center for Advanced Studies in Adaptive Systems) is a research project and a department of the Washington State University very active in AR studies. Their focus is to design a smart home \small in form, lightweight in infrastructure, extendable with minimal e ort, and ready to perform key capabilities out of the box", through their Smart Home in a Box project [CCTK13]. The success of this project enabled the collection and publication of several datasets, so that many AR research studies worked using CASAS data [CSEC+09]. Nonetheless, the annotation of the datasets is restricted to a reduced subset of the freely available data, and in most of these cases it was obtained thanks to an automatic labeling method rather than using a personal diary or an oracle. Finally, the variety of the installed sensors is often restricted to two di erent types (motion and temperature), reducing the possibility of advanced data analysis.
Tapia et al. [TIL04] presented two datasets related to two houses with a single resident each collected by MIT. They comprise data collected from many Boolean sensors (up to 85) for two weeks each. Activity annotation was achieved asking the inhabitant to use a Person Digital Assistant (PDA). Every 15 minutes candidates were reminded by the PDA to record the performed activities. Even if this methodology is less intrusive and less demanding than spontaneous annotation, it resulted to be less accurate probably because it is not spontaneous. Moreover, the reduced duration makes it less relevant for traditional machine learning methods.
T.Van Kasteren [VKNEK08], working at an Activity Recognition project at University of Amsterdam, has collected a dataset concerning two houses with single inhabitant. The volunteers houses were instrumented with 20 boolean sensors, collecting data for 28 days. The annotation was done directly by the inhabitant, but it reports some inaccurate entries, as well as some unexpected data (e.g., sensors always on/o ).
Referring to the reported projects, we can subsume the weak points of publicly-available datasets as follows:
Limited Sensor Variety: many projects use few sensors or a limited variety in sensed quantity. Limited Extension: projects involving several volunteers, present short duration per-person; conversely, long lasting collections refer only to few participants; Limited Annotation Reliability: inhabitants and automatic methods could lead to insu cient results in terms of accuracy and, in some cases, the single activity annotation is not su cient to describe properly the experimental settings; Heterogeneity: every project de nes its set of activities, sensors, standards, and protocols, resulting in non-comparable datasets; Speci city and Applicability: most of the projects report data collected with a speci c intent, not necessarily matching the aim of other research groups; dually, if a dataset is collected in generic settings, it might not contain some speci c situations required by other research groups.
Moreover, we would like to emphasize the lack of attention devoted to the behavioral change annotation. Indeed, all the mentioned datasets have a too short time duration and/or have no annotation concerning such modi cations in the inhabitant behavior.
Alternative approaches for the dataset collection phase consist in substituting the real world system with a simulation software producing synthetic data [Mas, MN06, AR07]. In this paper we present a collection of datasets generated with SHARON simulator, which can be tuned to produce highly customized synthetic home automation data for advanced applications. 3
The datasets we present have been obtained exploiting SHARON's sensor data generation algorithms, with di erent environments and inhabitant behaviors.
The resources are accessible at the persistent URL http://www.purl.org/synthetic sh dataset, and are available under the Creative Commons Attribution 3.0 CC-BY License; when exploiting the hereby included data, please cite the work of Veronese et al. [VMT+16]. The resources and the software to generate further data are also available at the institutional website of the Assistive Technology Group ATG [b1115]. 3.1
SHARON is a tool developed in the BRIDGE project [MSV+15] to face the lack of data for advanced Smart Home applications such as Activity Recognition. The simulator is structured in two main layers: the top layer generates the daily activity schedule, the bottom layer translates them into sensor activations. The software can be tuned designing the dwelling characteristics, the virtual sensors models, and a set of parameters representing the inhabitant response to needs (e.g., hunger, tiredness, boredom, stress, etc.). The activity schedule attempts to satisfy the person needs in relation to the time of the day, the weekly cycle, the weather conditions and other non-deterministic components. The bottom level relies either on a virtual agent, performing the scheduled action in the environment and activating the sensors following a set of alternative patterns, or on a statistical module, reproducing the activations of sensors given an activity as performed in an available training dataset.Finally it is possible to program a change in the simulation parameters so that the inhabitant behavior is a ected accordingly.
All the details about the data generation model implemented in SHARON to produce Synthetic Smart Home Data are available in the work of Veronese et al. [VPC+14]. The evaluation of the simulator has already been performed through a cross-validation process applied on a real world dataset (ARAS [AEIE13]); the work in Veronese et al. [VMT+16] reports the results for both the layers of the SHARON simulator: Top layer validation (ADL scheduling): three di erent validation metrics (Bhattacharyya distance [Bha46], Earth Mover Distance [Hit41] and Kullback-Leibler divergence [KL51]) have been used to evaluate the di erence between activity distributions in the generated dataset with respect to a test set extracted from the original dataset. The same distance has been computed between a training set of the above mentioned real world dataset (original dataset ) and the test set; Figure 1 shows that the ADL scheduling generated by the SHARON simulator is compatible with the schedule of the original dataset.
Bottom layer validation (Sensor activations): the Bhattacharyya distance have been computed to compare the sensor activation distributions in the ARAS dataset with respect to 500 450 400 350 taa300 d ing250 n i raT200 150 50 100 Dinner
Toilet OtherReading
WaGtcohininggOTuVt Shower Internet
Lunch
Sleeping 0 Breakfast 0 100
Conversation
Snack 200 300 Simulated data 400 500 the generated datasets (using both the agent module and the statistical module). Table 2 reports the results for three relevant activities: Cleaning (where the sequence of sensor activations is almost random), Lunch (where several executions are di erent, but keeping an overall procedure), and Having Shower (where the procedural connotation is strong).
Music Napping
Laundry
Cleaning The generated dataset has been obtained using the SHARON simulator; every day of simulation is represented by two text les: one describing the ADL scheduling and one describing sensor activations. The former contains all the performed activities - one activity per line - with the starting time, the activity identi er, and the activity name in a comma separated format. The latter contains 86400 lines - one for every second of the day - reporting the boolean status of every sensor of the house separated by blank characters.
The proposed datasets refer to three di erent house models. Each dataset comprises 90 days of the virtual inhabitant life, and has an alternative version comprising an injected behavioral drift compatible with dementia symptoms, that can be used for comparison. In the following, the characteristics of di erent classes of datasets are described; they are summarized in Tables 3 and 4. This rst group of datasets comprises four synthetic home automation datasets (their names start with A-* ) based on a virtual reproduction of the ARAS project test environment [AEIE13]. Two of them (Aext-* ) have been obtained by training SHARON over the behavior of one of the original ARAS project inhabitants, resulting in an extension of the original data. The other two (A-agn-* ) have been obtained using the same ADL scheduling but with an agentbased simulation. Two variants with behavioral drift due to dementia (*-dem) are also available.
The house environment exploited for simulation comprises 20 binary home automation sensors. The location is a simple apartment with four main spaces: bedroom, bathroom, and an openspace with kitchen and living room. Most common sensors are motion detectors, but in this environment there are also tap, toilet, and shower sensors, pressure detectors on chairs, sofa and bed.
The inhabitant routine comprises two di erent patterns for weekdays and weekends. During the weekdays the inhabitant spends a daily amount of time outside the dwelling (for working activities), while during the weekend leisure is the main occupation (relax, reading, internet, etc.). There are 13 performed activities, as described in Table 4, plus an unquali ed activity \Other". 3.2.2
The second dataset group (K-* ) is related to the research project home by Van Kasteren et al. [VKNEK08]. In this case the virtual environment reproduces the experimental house, as well as the sensor activations, which are produced after a training on the original data. The results are two datasets: one with the extension of the real dataset (K-ext-norm), the other with the superimposed behavioral drift (Kext-dem).
The house environment exploited for simulation comprises 21 binary home automation sensors. The location is a two-storey apartment: on the rst oor there are a bathroom and an open-space with kitchen and livingroom; the second oor is composed by two bedrooms, a bathroom, and a study room. Installed sensors include motion sensors to detect doors, drawers and cupboards openings, tap and shower sensors, sensors to detect appliances uses, pressure detectors on chairs, sofa and bed.
The dataset describes 12 activities, the same of the ARAS datasets, except for \Working" and \Internet" activities that are missing (Table 4). The inhabitant routine comprises two di erent patterns for weekdays and weekends, mainly by di erentiating the time and duration of meals. 3.2.3
This last group of datasets are fully virtual (V-*). The authors designed a simple four room house, and programmed an easy routine for a virtual inhabitant. The obtained datasets are based on an agent based sensor activation simulation, one with plain routine (V-agnnorm), the other with the injected drift (V-agn-dem).
The virtual designed environment includes 11 binary sensors. The house is designed with four main rooms: kitchen, bedroom, bathroom, and livingroom. Most devices are movement sensors, with open-close detectors on main door and bathroom cabinet.
The inhabitant routine represents a remote-worker, working 8 hours at home in weekdays, and relaxing in the weekends. The activities are 14, with the addition of an unquali ed \Other". 3.3
Alzheimer's Disease (AD) is becoming widespread as reported by AD International [WJB+13]: there will be up to 65.7 million people living with dementia worldwide by 2030 and up to 115.4 million by 2050. The typical symptoms of AD involve the daily routine, concerning: forgetfulness, di culty performing ADL, incontinence, speech problems, wandering and getting lost, depression, sleep disorders. In the provided dataset (*-dem) this condition is simulated by replicating part of the symptoms. The time taken to perform complex tasks such as \Take a shower" is increased by 20%, its rate is decreased by 15%. The duration of nighttime sleep passes from an average of 8 uninterrupted hours to 4.5 hours fragmented up to 5 times, while naps appear during the day. The frequency of activities such as \Dinner" and \Going out" slightly decreases. 4
The presented datasets, generated with SHARON, are a support resource for research groups working on smart home data processing for advanced applications. Even if with some limitations, the proposed data are a resource to foster such research, avoiding the costs of creating a real world testbed. Moreover, the software SHARON is publicly-available, enabling to generate further di erent data with particular conditions and behavioral drifts, and overcoming the lack of highquality real world datasets. The quality of the data generated by the simulator has been discussed in the work of Veronese et al. [VMT+16], which has already attracted the attention of the scienti c community that has expressed a willingness to access data. We believe that this could be used as a tool to provide early tests for new methods development (e.g., Activity Recognition and Behavioral Drift Detection), before allocating time and nancial resources. The provided datasets are only a possible application of the simulation software, whose next releases will include further features and a user friendly web interface to allow the generation of high quality synthetic datasets. [AEIE13] [CSEC+09] Diane Cook, M Schmitter-Edgecombe, Aaron Crandall, Chad Sanders, and Brian Thomas. Collecting and disseminating smart home sensor data in the [Hit41] [KL51] [Mas] [MN06] [MSV+15] [PKL+05] [PLJ+15] [RBC+15] casas project. In Proceedings of the CHI Workshop on Developing Shared Home Behavior Datasets to Advance HCI and Ubiquitous Computing Research, 2009.
Frank L Hitchcock. The distribution of a product from several sources to numerous localities. J. Math. Phys, 20(2):224{ 230, 1941.
Solomon Kullback and Richard A Leibler. On information and su ciency.
The annals of mathematical statistics, pages 79{86, 1951.
Mason project website. http://cs.gmu.edu/eclab/projects/mason.
Accessed: 2015.
Miquel Martin and Petteri Nurmi. A generic large scale simulator for ubiquitous computing. In 2006 Third Annual International Conference on Mobile and Ubiquitous Systems: Networking & Services, pages 1{3. IEEE, 2006.
Simone Mangano, Hassan Saidinejad, Fabio Veronese, Sara Comai, Matteo Matteucci, and Fabio Salice. Bridge: Mutual reassurance for autonomous and independent living. Intelligent Systems, IEEE, 30(4):31{38, 2015.
Paula Paavilainen, Ilkka Korhonen, Jyrji Lotjonen, Luc Cluitmans, Marja Jylha, Antti Sarela, and Markku Partinen.
Circadian activity rhythm in demented and non-demented nursing-home residents measured by telemetric actigraphy.
Journal of sleep research, 14(1):61{68, 2005.
Kirsten KB Peetoom, Monique AS Lexis, Manuela Joore, Carmen D Dirksen, and Luc P De Witte. Literature review on monitoring technologies and their outcomes in independently living elderly people. Disability and Rehabilitation: Assistive Technology, 10(4):271{ 294, 2015.
Daniele Riboni, Claudio Bettini, Gabriele Civitarese, Za ar Haider Janjua, and Viola Bulgari. From lab to life: Fine-grained behavior monitoring in the elderly's home. In Pervasive Computing and Communication Workshops (PerCom Workshops), 2015 IEEE [RM13] [SCFSE16] [SPF15] [TIL04] on, pages Parisa Rashidi and Alex Mihailidis. A survey on ambient-assisted living tools for older adults. IEEE journal of biomedical and health informatics, 17(3):579{ 590, 2013.
Gina Sprint, Diane Cook, Roschelle Fritz, and Maureen SchmitterEdgecombe. Detecting health and behavior change by analyzing smart home sensor data. In Smart Computing (SMARTCOMP), 2016 IEEE International Conference on, pages 1{3. IEEE, 2016.
Jeremie Saives, Clement Pianon, and Gregory Faraut. Activity discovery and detection of behavioral deviations of an inhabitant from binary sensors. IEEE Transactions on Automation Science and Engineering, 12(4):1211{1224, 2015.
Emmanuel Munguia Tapia, Stephen S Intille, and Kent Larson. Activity recognition in the home using simple and ubiquitous sensors. Springer, 2004. [VMT+16] [VPC+14] [WJB+13]
Fabio Veronese, Andrea Masciadri, Anna A Tro mova, Matteo Matteucci, and Fabio Salice. Realistic human behaviour simulation for quantitative ambient intelligence studies. Technology and Disability, 28(4):159{177, 2016.
F Veronese, D Proserpio, S Comai, M Matteucci, and F Salice. Sharon: a simulator of human activities, routines and needs. Studies in health technology and informatics, 217:560{566, 2014.
Anders Wimo, Linus Jonsson, John Bond, Martin Prince, Bengt Winblad, and Alzheimer Disease International.
The worldwide economic impact of dementia 2010. Alzheimer's & Dementia, 9(1):1{11, 2013.