INDEXING TIME-SERIES UNDER CONDITIONS OF NOISE - Data Mining in Time Series Databases - page 95

Fig. 19. Noisy ASL data: The correct clusterings of the LCSS method using complete

linkage.

The LCSS proves to be more robust than the Euclidean and the DTW
under noisy conditions (Table 4, Figures 17 and 18). The Euclidean again
performed poorly, recognizing only 1 cluster, the DTW recognized 2 and
the LCSS up to 12 clusters (consistently recognizing more than 6 clusters).
These results are a strong indication that the models based on LCSS are
generally more robust to noise.

6.3. Evaluating the Quality and E ciency of the Indexing
Technique

In this part of our experiments we evaluated the eciency and effective-
ness of the proposed indexing scheme. We performed tests over datasets of
different sizes and different number of clusters. To generate large realistic
datasets, we used real sequences (from the SEALS and ASL datasets) as
“seeds” to create larger datasets that follow the same patterns. To perform
tests, we used queries that do not have exact matches in the database, but
on the other hand are similar to some of the existing sequences. For each
experiment we run 100 different queries and we report the averaged results.
We have tested the index performance for different number of clusters
in a dataset consisting of a total of 2000 sequences. We executed a set of
K 1, 5, 10, 15 and 20 and we plot

the fraction of the dataset that has to be examined in order to guarantee

-Nearest Neighbor (K-NN) queries for K