D1.1 Technology survey: Prospective and challenges - Revised version (2018)
5 Participatory / citizen science for water management
5.4 Reputation models
For encouraging participation, various reputation models have been proposed and used for participatory sensing. The simplest reputation models are ones that are summation and average based. They use an aggregation of ratings (i.e., by summing, as in case of eBay, or averaging, as in case of Amazon), to create an overall single reputation score [Schlosser, 2004]. An alternative scheme to having reputations being a numerical value is to use discrete labels. For example, the Slashdot web site aggregates ratings on actions, such as story submissions, postings, moderation activities, into tiers for participants that include terrible, bad, neutral, positive, good, and excellent [Reddy, 2010].
Reputation models based on Bayesian frameworks have also been popular for sometimes [Ganeriwal, 2008]. Particularly, such models rely on ratings, either positive or negative, and use probability distributions, such as the Beta distribution, to come up with reputation scores. Reputation is determined using the expectation of the distribution, and the confidence in this reputation score is captured by analyzing the probability that the expectation lies within an acceptable level of error. Additional features are easily enabled, such as aging out old ratings by using a weight factor when updating reputation and dealing with continuous ratings by employing an extension involving the Dirichlet process [Ganeriwal, 2008].
Another challenge for participatory sensing comes from the dynamic conditions of the set of mobile devices and the need for data reuse across different applications. Unlike traditional sensor networks, applications of participatory sensing rely on population of mobile devices, on the type of sensor data each can produce. Data quality in terms of accuracy, latency, and confidence can change all the time due to device mobility, variations in their energy levels and communication channels, and device owners’ preferences. Identifying the right set of devices to produce the desired data and instructing them to sense with proper parameters to ensure the desired quality is a complex problem. In traditional sensor networks, the population and the data they can produce are mostly known a priori; thus, controlling the data quality is much easier. The same sensor data have been used for different purposes in many existing participatory applications. For example, accelerometer readings have found usage in both transportation mode identification, and human activity pattern extraction.
Related to reputation is the need to understand human behaviour, as people are the carrier of the sensing devices, and their recruitment depends on their capability to correctly collect sending data [Mascolo, 2016]. A variety of data mining and statistical tools can be used to distill information from the data collected by mobile phones and calculate summary statistics related to human behaviour recognition. Still, recognizing human behaviour is still a somewhat unsolved research direction, which is why it is generally mention as potential enabler for participatory campaigns, but thus far not many frameworks managed to successfully incorporate this aspect in the recruitment decisions [Lane, 2010].