Database Reference
In-Depth Information
individual gid may contain a very large number of items which have
traveled together. In order to materialize the measures such as counts
the algorithm does not need to access the counts of the individual EPCs.
It has been observed in [30] that the movement trails of RFID data
form gigantic commodity flowgraphs, which represent the locations and
durations of the path stages traversed by each item. The work in [30]
proposes a method to construct a warehouse of commodity flows, which
is also referred to as a
FlowCube
. Similar to OLAP, the model comprises
cuboids, which aggregate the item flows at a given abstraction level. In
this case, the measure of each cell is a commodity flowgraph, which cap-
tures the major movements and trends, as well as significant deviations
from the trend in each cell. The flowgraph can also be viewed at multi-
ple levels by changing the abstraction levels of path stages. The latter is
achieved by performing simultaneous aggregation of paths to all inter-
esting abstraction levels, In addition, path segments with low frequency
and rarely occurring cells are removed from the representation.
It has been observed in [28] that a clustered path database, which is
natural to RFID applications, can be naturally modeled as a
compressed
probabilistic workflow
. Each location corresponds to an activity, and lo-
cations are linked according to their order of occurrence. A link between
two activities has a probability, which represents the percentage of time
that one of the activities occurred immediately after the other. A prob-
abilistic representation of the workflow can also be used in the context
of the FlowCube. The details of such a concrete probabilistic workflow
are provided in [28].
4. Semantic Event Extraction from RFID Data
Streams
The discussion so far has focussed on low level cleaning, event extrac-
tion and data management of RFID. However, in many applications,
the events to be discovered are
high level semantic events
,asopposedto
the primitive event of an object moving from one location to another.
Such events are also referred to as
complex
events. The problem of event
mining in RFID processing is related to previous research on complex
event detection in active databases and high fan in sensor systems [1, 7,
13, 14, 18, 26, 62, 76, 70, 79]. In particular, the work in [62] discusses a
high fan-in architecture for a sensor network, and shows how it can be
used in order to process complex events by combining RFID data with
other kinds of sensor readings and stored data.
An example of such a high-level semantic event discussed in [78] is
the shoplifting example, in which the event corresponds to an item be-
