Database Reference
In-Depth Information
and the second includes the sites within 140 miles. For electroencephalo-
grams, the first neighborhood is the 3
×
3 rectangle in the grid of electrodes,
and the second is the 5
5 rectangle.
For each series, we have found the five closest matches, and then
determined the average number of matches that belong to the same
neighborhood. In Table 4, we give the results and compare them with the
perfect selection and random selection; larger numbers correspond to better
selections.
The results show that the peak similarity is usually more effective than
the other three similarities. If we use the 90% compression, the peak simi-
larity gives better selection than the other similarity measures for the stock
prices, air and sea temperatures, and
×
; however, it gives worse results
for the wind speeds. If we use the 95% compression, the peak similarity out-
performs the other measures on the stocks, temperatures,
eeg
,andlarge-
neighborhood selection for the wind speeds; however, it loses to the mean
similarity and correlation coecient on the small-neighborhood selection
for the wind speeds.
We have also checked how well the peak similarity of original series
correlates with the peak similarity of their compressed versions (Figure 6).
We have observed a good linear correlation, which gracefully degrades with
an increase of compression rate.
eeg
5. Pattern Retrieval
We give an algorithm that inputs a pattern series and retrieves similar series
from a database. It includes three steps: identifying a “prominent feature”
of the pattern, finding similar features in the database, and comparing the
pattern with each series that has a similar feature.
We begin by defining a leg of a series, which is the segment between
two consecutive important points. For each leg, we store the values listed
in Figure 7, denoted vl, vr, il, ir, ratio ,and length; we give an example of
these values in Figure 8. The prominent leg of a pattern is the leg with the
greatest endpoint ratio.
The retrieval procedure inputs a pattern and searches for similar seg-
ments in a database (Figure 9). First, it finds the pattern leg with the
greatest endpoint ratio, denoted
ratio p ,
and determines the length of this
leg,
length p . Next, it identifies all legs in the database that have a similar
endpoint ratio and length. A leg is considered similar to the pattern leg
if its ratio is between
ratio p /C
and
ratio p · C,
and its length is between
Search WWH ::




Custom Search