Conclusions - Ranking Queries on Uncertain Data - page 213

be maintained using any efficient skyline maintenance algorithms for streams, such
as [196].
9.3 Future Directions: Enriching Data Types and Queries
It is highly interesting to study more general types of ranking queries on complex
uncertain data. We give some examples here.
9.3.1 Handling Complex Uncertain Data and Data Correlations
In some applications, data tends to be heterogeneous, semi-structured or unstruc-
tured. For example, in medical applications, information collected about patients
may involve tabular data, text, images, and so on. Uncertainty may widely exist in
those applications due to factors like equipment limitations and ambiguity in natu-
ral language presentation. Such uncertain data types pose grand challenges in data
analytics.
First, it is difficult to model uncertainty in complex data types. For example,
documents extracted by hand-writing recognition may be prone to mistakes. How
can we represent the uncertainty in each word or prase in documents? Neither the
probabilistic database model nor the uncertain object model can be directly adopted.
Second, ranking queries on complex uncertain data may have different forms
from the ranking queries on simple uncertain data. In this topic, we extend top- k
selection queries on uncertain data, where a score is computed for each data record
on a set of the attributes. In many applications, complex uncertain data objects may
be ranked using different methods. For example, a set of documents or images are
often ranked according to their relevance scores to a set of key words. Answering
such complicated ranking queries on complex uncertain data is non-trivial.
Last, complex data correlations may exist among uncertain data objects. It is
difficult to model the correlations, not to mention answering queries on uncertain
data with complex correlations.
9.3.2 Answering More Types of Ranking and Preference Queries
on Uncertain Data
In this topic, we focus on top- k selection queries on uncertain data, where a total
order on all objects or instances is available. More generally, given a set of partial
orders on tuples as user preferences, a preference query [197] finds the tuples that
best match the preferences. For example, when searching for used cars, a user may
specifies his/her preferences as “I like Toyota better than Ford” and “I prefer low