Conclusions - Ranking Queries on Uncertain Data - page 212

9.2.3.2 Uncertain data streams with correlations
Correlations may exist among uncertain data streams. For example, in the speed
monitoring application, if two speed sensors are deployed at the same location, then
the readings of the two sensors are mutually exclusive. That is, only the reading of
one sensor can exist in a possible world. More generally, complex correlation among
two or more uncertain data streams can be represented by the joint distribution of
their readings. How to continuously monitor the top- k uncertain data streams in such
cases is highly interesting.
9.2.3.3 Continuously monitoring probabilistic threshold top- k queries with
different parameter values
There are two parameters in probabilistic threshold top- k queries, the query param-
eter k and the probability threshold p . In some applications, we may be interested in
how query results change as parameters vary. To support the interactive analysis, it
is highly desirable to monitor probabilistic threshold top- k queries on uncertain data
streams with different parameters. To achieve this goal, it is interesting to study if
we can extend the PRist+ index developed in Section 5.5 to uncertain data streams.
9.2.3.4 Continuously monitoring probabilistic threshold top- k aggregate
queries
In this chapter, we focus on probabilistic threshold top- k selection queries, where the
ranking function is applied on a single instance. Another category of top- k queries
is top- k aggregate queries [29], where the ranking function is applied on a group of
instances. The top- k groups with highest scores are returned as results. It is interest-
ing to investigate how to extend the techniques discussed in this chapter to handle
top- k aggregate queries on uncertain data streams.
9.2.4 Probabilistic path queries on road networks
An important future direction of probabilistic path queries on uncertain road net-
works is to explore the temporal uncertainty and correlations of travel time along
road segments.
The HP-tree can be maintained incrementally as edge weights change. For each
node in an HP-tree that contains a set of edges, an optimal weight stochastically
dominating all edge weights is stored. The optimal weight can be constructed using
a set of (value, probability) pairs that are the skyline points among the (value, prob-
ability) pairs of all edges. Therefore, the optimal weight in a sliding window can