2.1 Information Sharing: The Paradigm Shift

Humans are indeed the most versatile sensors. They can make judgments, where traditional sensing technology falls short. They are already embedded in spaces that are interesting to measure. They are good at understanding context, detecting anomalies, interpreting complex scenes, and prioritizing information transfer. Hence, leveraging their help can lead to a better understanding of many human-centric eco-systems, such as urban spaces, transportation systems, agricultural processes, human-operated supply chains, and residential energy consumption, to name a few. Several surveys appeared in recent literature that offer different visions for the social sensing landscape [51, 52]. There are many ways humans can be involved in sensing systems. For example, they may serve as:

• Information sources: The versatility of humans as sophisticated observers of their surroundings has led to their exploitation as information sources since the early days of civilization. Military intelligence, for example, has relied on human sources long before the introduction of physical sensors. The recent advent of social networks reaffirms the importance of humans as an information source. Coupled with technological advances that enable instantaneous global information transfer, media that collect grassroots human observations are supplanting the news in delivery of timely information. The author has recently been in a conversation with a friend from California, who confessed that when experiencing what feels like an Earthquake, he first turns to Twitter. Chances are, if this is not a false alarm, the social network will already be teaming with real-time tweets announcing the event.

• Sensor operators: Humans can alternatively operate sensors such as cameras or geotagging apps. In this case, we still leverage the versatility of humans in identifying what sensory observations to report. However, the act of sensing itself is carried out by another device, under human control. Early participatory sensing applications adopt this model. A common application in that category is geotagging. Humans would participate in geotagging campaigns, where the goal is generally to record locations of observations of interest. For example, a campaign might record locations of invasive species in a park, or locations of offensive graffiti on city walls. The human would have to observe the item of interest, then take a picture or push a button on their phone to record the current location.

• Sensor conveyors: Humans increasingly serve as sensor conveyances. For example, they carry smartphones that now come equipped with an impressive array of sensors, they drive cars that have many sensors of their own, and they wear smart pieces such as watches and vision devices (e.g., Glass), that can record events on the move. The exploitation of human mobility for ferrying sensors is natural in applications that aim at measurements of social spaces. It offers coverage of the space and has a natural tendency to deliver more measurements from more heavily utilized (and hence, typically more important) areas.

• Sensor data processors: It is common for humans to process sensory data as well. For example, tagging your friend in a picture on Facebook is an act of sensory data processing. In this act, the potentially complex scene in the original picture (which itself is a sensor output) is processed by you. The processing results in the generation of additional metadata (the tags) that can, in principle, be exploited by collection applications downstream. Individuals post comments on photos, which also constitute sensory data processing for the same reason. The emergence of crowd-sourcing platforms, such as the Mechanical Turk [53], enables large-scale application of human capabilities to sensor data processing tasks. Humans can also perform data fusion and analysis, producing their own inferences and opinions regarding the underlying observations, and sharing information of higher semantic content. For example, they may recognize a video to indicate that “police attacks demonstrators” or that “demonstrators attack police,” a distinction that is harder to do in an automated fashion by a machine.

In addition to the role humans play in sensing, it is often interesting to distinguish the manner in which they participate. For example, they may explicitly and intentionally join a sensing campaign or may simply be sharing data on social media with no awareness of the sensing application. They may share data on their own, out of conviction to do a public good, or may be incentivized to do it in exchange for a benefit. As mentioned earlier, a confluence of three social and technology trends explains the increasing ubiquity of social sensing. Those trends are:

• Sensing device proliferation: The first trend that fuels social sensing applications is the commercial proliferation of sensing systems that are commonly accessible to large consumer populations. Active RFIDs, smart residential power meters (with a wireless interface), camera cell-phones, in-vehicle GPS devices, accelerometer-enhanced entertainment platforms (e.g., Wii-fit), and activity monitoring sportsware (e.g., the Nike+iPod system) have all reached mature market penetration, offering unprecedented opportunities for data collection.

• Mobile connectivity: The second trend lies in ubiquitous mobile Internet access available to sensing platforms on the move. This untethered connectivity allows events to be measured and reported in real time, anytime, anywhere. A clear example is the case of GPS measurements and pictures taken by cell-phones. Besides GPS and cameras, modern smartphones currently host myriads of other sensors as well, such as accelerometers, magnetometers, and gyroscopes, and offer 3G/4G, WiFi, and Bluetooth network access, which enables sharing their data. Vehicular Internet access, is also becoming available, for example, Chrysler and BMW were some of the earliest car manufacturers to enable applications that exploit network connectivity. The vehicular OBD-II interface is already being used by services such as OnStar for remote diagnostics. Other applications may perform traffic statistics, alert to nearby accidents, or detect emergency conditions. In medical spaces, significant investments have been made in sensor technology for longitudinal monitoring. Microsoft HealthVault was one early example of an initiative to automate collection of and access to medical information. A significant number of vendors announced wearable health and biometric monitoring sensors that automatically upload user data to HealthVault. WiThings, a company that offers a variety of WiFi enabled health and fitness monitoring devices is another example of sensors in social spaces. The proliferation of sensing devices with Internet connectivity that collect data in social spaces makes it feasible to build human-centric data collection and sharing applications that augment human capabilities and improve situation awareness.

• Social networks: The existence of sensors and Internet connectivity, however, would not have been sufficient, by itself, to support social sensing in the mainstream. The third key trend that fueled social sensing was the increased popularity of mass information dissemination channels, afforded by social networks. Twitter, Flickr, Twitpic, YouTube, and other networks allow individuals to globally broadcast their observations. It is this global dissemination opportunity that makes it easy to build large-scale applications that utilize commonly-available sensors, upload data in real-time, and share the observations at scale.

The above opportunities have generated significant interest in the research community in building application prototypes that rely on observations made by humans or by sensors in their vicinity. Below, we present an application taxonomy then outline recent work related to social sensing, and conclude with a note on privacy.

2.2 An Application Taxonomy

One can generally divide social sensing applications into three types, depending on their functionality. The three types differ in the level of complexity of data processing done to the measurements.

• Data-point-centric applications: The first and simplest social sensing application is one where individuals share single data points (the observations) that are then made available to clients or decision-makers. An example is geo-tagging applications, where individuals share pictures (tagged by location) of entities of relevance to the application. For example, a sensing campaign might ask participants to document locations of invasive species in a park, or locations of garbage on a beach. These observations (and pictures) can then be displayed on a map, or offered to municipal decision-makers for appropriate action.

• Statistics-centric applications: The second type of social sensing applications is one where statistics are computed from the data. An example might be a traffic speed monitoring or a pollution monitoring application where the speed or pollution levels measured by different individuals are used to compute statistics such as averages and probability distributions. Many early examples of social sensing belong in that category. For example, traffic patterns were monitored in a city to help drivers avoid congestion areas [54], bike route data were collected by biking enthusiasts to help them pick better routes [55], and hiker encounters were recorded on mountain trails to help locate missing hikers [56]. These applications offer useful statistics about a given locale that are of interest to individuals in that locale.

• Model-centric applications: A third and most general type of social sensing applications has been described in literature, where generalizable models are learned from sensory data collection, that can be used to affect human decision making outside of the collection locale. For example, sharing data collected by smart energy meters installed in some households, together with relevant context, can lead to a better understanding of energy consumption in contemporary homes and best practices that increase energy efficiency elsewhere around the nation. Similarly, sharing data collected by activity sensors among fitness enthusiasts can lead to lifestyle recipes that promote healthier behaviors for multitudes of others. Also, sharing data on environmental pollutants and personal well-being (e.g., locations and incidents of asthma attacks) can establish links between likelihood of attacks and exposure to specific contaminants, which may help individuals reduce their exposure to those contaminants. In a recent study of vehicular fuel-efficiency, a model predicts the total fuel consumption for a vehicle on a road segment as a function of several variables such as road speed, degree of congestion, and vehicle parameters. Once the model is known, it is possible to optimize human decisions by offering better (GPS) navigation advice for any vehicle on any street.

2.3 Early Research

An early overview of social sensing applications is presented in [57]. Some early applications include CenWits [56], a participatory sensor network to search and rescue hikers in emergency situations, CarTel [54], a vehicular sensor network for traffic monitoring and mitigation, CabSense [58], an application that analyzes GPS data from NYC taxis and helps you find the best corner to catch a cab [59] and BikeNet [55], a bikers sensor network for sharing cycling related data and mapping the cyclist experience. More recently, social sensing applications in healthcare have become popular. Numerous medical devices have been built with embedded sensors that can be used to monitor the personal health of patients, or send alters to the clinic or through the patient’s social network when something unexpected happens. Such social sensing can be used for activity recognition for emergency response [60], long term prediction of diseases [61–63], and life-style changes that affect health [64, 65].

Early work in social sensing focused on challenges such as preserving privacy of participants [66, 67], improving energy efficiency of sensing devices [68, 69] and building general models in sparse and multi-dimensional social sensing spaces [70, 71]. Examples include privacy-aware regression modeling, a data fusion technique that produces the same model as that computed from raw data by properly computing non-invertible aggregates of samples [66]. Authors in [67] gave special attention to preserving privacy over time-series data based on the observation that a sensor data stream typically comprises a correlated series of sampled data from some continuous physical phenomena. Acquisitional Context Engine (ACE) is a middleware that infers unknown human activity attributes from known ones by exploiting the observation that the values of various human context attributes are limited by physical constraints and hence are highly correlated [68]. E-Gesture is an energy efficient gesture recognition architecture that significantly reduces the energy consumption of mobile sensing devices while keeping the recognition accuracy acceptable [69]. The sparse regression cube is a modeling technique that combines estimation theory and data mining techniques to enable reliable modeling at multiple degrees of abstraction of sparse social sensing data [70]. A further improved model to consider the data collection cost was proposed in [71].

The concept of sensing campaigns have been introduced in literature, where participants are recruited to contribute their personal measurements as part of a large-scale effort to collect data about a population or a geographical area. Examples include documenting the quality of roads [72], measuring the level of pollution in a city [73], or reporting locations of garbage cans on campus [74]. In addition, social sensing covers scenarios where human sources spontaneously report data without prior coordination, such as data describing important events. Examples include large volumes of reported observations of political unrest, riots, and natural disasters on Twitter. Recent research attempts to understand the fundamental factors that affect the behavior of these emerging social sensing applications, such as analysis of characteristics of social networks [75], information propagation [76], and tipping points [77].

A critical question about trustworthiness arises when the data in social sensing applications are collected by humans whose “reliability” is not known. In social sensing, anyone can contribute data. Such openness greatly increases the availability of the information and the diversity of its sources. On the other hand, it introduces the problem of understanding the reliability of the contributing sources and ensuring the quality of the information collected. Trusted Platform Module (TPM), commonly used in commodity PCs, can be used to provide a certain level of assurance at the expense of additional hardware [78]. YouProve is a recent technique that relies on trust analysis of derived data to allow untrusted client applications to verify that the meaning of source data is preserved [79]. Trust analysis can also be performed at the server side by building a likelihood function for sensed data to provide a quantifiable estimate of both source reliability and the correctness of observations. A rich set of work that are referred to as fact-finders have been developed to perform trust analysis in information networks by jointly assessing the reliability of sources and the credibility of facts reported by them. A detailed review of literature on fact-finders is provided in Chapter 4.

2.4 The Present Time

Information distillation (or reduction of large amounts of data into smaller amounts of actionable information) is an increasingly important interaction modality between humans and data. Information distillation services are made popular by a shift in the digital information landscape. This shift is from a web of slowly updated cross-linked objects (e.g., Web pages) to streams of continually generated real-time data emitted by humans and sensors. The availability of real-time data offers both new opportunities and new challenges. There are unprecedented opportunities for building real-time situation awareness applications such as disaster-response services that help first-responders assess current damage, transportation advisories that help individuals avoid traffic bottlenecks, and citizen-science tools that collect and process data from speciality sensors (such as rain gauges or pollution sensors) owned by interested individuals. Reliable information must be distilled from unreliable data. Applications that use the data must address modeling challenges, in order to offer predictive services that learn from past observations. Several categories of sensing applications were discussed ranging from those where humans collect data points that are individually significant to those where models or statistical properties of aggregates are sought.

Many other challenges remain topics of current research. Those include front-end challenges (e.g., energy consumption), coordination challenges (e.g., participatory sensing campaign recruitment), back-end challenges (e.g., modeling and prediction), and challenges in the overall understanding of the emergent behavior of social sensing systems at large. While a significant amount of research has already been undertaken along those fronts, much remains unsolved. New interdisciplinary research is needed to bring about better solutions for a better theoretical understanding of emerging social sensing systems in a future sensor- and media-rich world. In the rest of this book, we focus on one such challenge: namely, the challenge of ascertaining data reliability.

2.5 A Note on Privacy

A discussion of social sensing would not be complete without commenting on a topic that lies at the crux of information sharing; namely, that of user privacy. What do advances in social sensing mean in terms of privacy expectations of individuals? There has been significant advances in research on user privacy, but the problem is by no means solved. It is the authors’ belief, however, that privacy will be solved through legal protection and not by technical means.

It is interesting to observe that modern society has already given up on privacy in very significant ways. For example, while most would consider a credit card number to be private data, they share that number voluntarily with complete strangers on regular basis; often multiple times a day! Every time we use a credit card at a restaurant, a waiter we might have never seen before disappears with our card but we prefer to use our cards instead of cash. The reason is simple: we trust the legal system to be efficient at catching credit card fraud. Our liability due to fraud is limited, and credit cards are far more convenient than cash, which can attract even more crime.

Our phone’s location services offer another example of loss of privacy. Most of us have gotten used to turning on navigation on our smartphones when driving in an unfamiliar locale. Clearly, turning on a location service reveals our location, but many people are accustomed to that and do not mind it in exchange for the perceived benefit of not having to use paper maps. In fact, some car manufacturers advertise loss of privacy as a feature. Their cars, they say, will transmit their location to get help should any problem or accident occur. The reason we have grown to accept some loss of privacy is in part the perceived benefit and in part the protection associated with the service. Should private information be somehow misused, the user has a legal recourse. Hence, good business practice suggests that privacy be respected.

Storage of our private email is another example of voluntary loss of privacy. Many of us use providers, such as Microsoft and Google, to handle email. Such email becomes observable to various analytics engines that can mine the text of that email. Nevertheless most people are comfortable using these services because we trust the providers.

Our trust in providers, good business practices, and the legal system is not a new development. Several hundred years ago, much of human society lived in small towns and rural areas. In a small town everyone knows everyone else’s business. It is very hard to hide. Residents of little towns did not have privacy expectations. The reason was simple. While it is possible for others to know much about you, it is hard for them to get away with crimes that exploit such data. After all, everyone does know everyone else’s business, so perpetrators will be caught.

Over the centuries, little towns grew into bigger cities, where it became possible for perpetrators to hide. Privacy became an issue. We conjecture that the advent of communication technology, sensing, and data analytics will eventually produce a world, where distances are drastically reduced, and the city turns again into a village—a global village, if you will, where data may be shared but crimes are not committed because it is hard to hide from the ubiquitous eye of data-centric crime-detection and investigation systems, backed by legal protection of the individuals’ right to share data without fear of misuse. Hence, privacy, as a technical problem, is not discussed in the rest of this book.