Database Reference
In-Depth Information
might differ depending on other parameters. For example, it could be
that the key predictors of fuel eciency for hybrid cars and gas-fueled
cars are different. It is the responsibility of the model construction ser-
vices to offer not only a general mechanism for applications to build
good models quickly from the data collected, but also a mechanism for
identifying the scope within which different predictors are dominant. A
single “one-size-fits-all” prediction model, computed from all available
data, is not going to be accurate. Similarly computing a model for each
special case (e.g., a model for each type of car) is not going to be useful
because, as stated above, the sampling is sparse. Hence, it is key to be
able to generalize from experiences of some types of vehicles to predic-
tions of others. Recent work combined data mining techniques based on
regression cubes and sampling cubes to address the model generalization
problem for sparse, high-dimensional data [64].
3.3 Real-time Decision Services
Ultimately, a generalized model, such as that described above, may
be used as an input to an application-specific optimization algorithm
that outputs some
decisions
for users in response to user queries. For
example, estimates of fuel consumption on different roads on a map can
be input to Dijkstra's algorithm to find the minimum fuel route between
points specified by the user. This route constitutes a decision output.
Hence, support for real-time stream processing and decision updates
must be provided as part of the social sensing architecture.
A key property of real-time decision services is the involvement of
humans in the loop. A significant challenge is therefore to design appro-
priate user interfaces. End-user devices will act as data custodians who
collect, store, and share user data. The level at which these custodians
interact with the user, as well as the nature of interactions, pose signifi-
cant research problems with respect to minimizing inconvenience to the
user while engaging the user appropriately. Context sensing, collabora-
tive learning, persuasion, and modeling of socio-sensing systems (with
humans in the loop) become important problems. Participation incen-
tives, role assignment, and engagement of users in modeling and network
learning become important application design criteria that motivate fun-
damental research on game theoretic, statistical, machine learning, and
economic paradigms for application design.
3.4 Recruitment Issues
The quality of the social experience gained from a sensor-based frame-
work is dependent on the ability to recruit high quality participants for
