Information Technology Reference
In-Depth Information
number of time-series data sets. The results of our experiments provide direct indi-
cations for the application of hubness-aware classifiers for real-world time-series
classification tasks. In particular, the HIKNN approach seems to have the best over-
all performance for time-series data.
Furthermore, we pointed out that instance selection can substantially speed-
up time-series classification and the recently introduced hubness-aware instance
selection approach, INSIGHT, outperforms the previous state-of-the-art instance
selection approach, FastAWARD, which did not take the presence of hubs explicitly
into account. Finally, we showed that the selected instances can be used to construct
features for instances of time-series data sets. While mapping time-series into a vec-
tor space by this feature construction approach is intuitive and leads to acceptable
overall classification accuracy, the particular instance selection approach does not
seem to play a major role in the procedure.
Future work may target the implications of hubness for feature construction
approaches and how these features suit conventional classifiers. One would for exam-
ple expect that monotone classifiers [ 2 , 13 , 23 ], benefit from hubness-based feature
construction: the closer an instance is to a good hub, the more likely it belongs to
the same class. Furthermore, regression methods may also benefit from taking the
presence of hubs into account: e.g. hw- k NN may simply be adapted for the case
of nearest neighbor regression where the weighted average of the neighbors' class
labels is taken instead of their weighted vote. Last but not least, due to the novelty
of hubness-aware classifiers, there are still many applications in context of which
hubness-aware classifiers have not been exploited yet, see e.g. [ 47 ] for recognition
tasks related to literary texts. Also the classification of medical data, such as diagno-
sis of cancer subtypes based on gene expression levels [ 31 ], could potentially benefit
from hubness-aware classification, especially classifiers taking class-imbalance into
account [ 51 ].
Acknowledgments Research partially performed within the framework of the grant of the Hungar-
ian Scientific Research Fund (grant No. OTKA 108947). The position of Krisztian Buza is funded
by the Warsaw Center of Mathematics and Computer Science (WCMCS).
References
1. Aha, D., Kibler, D., Albert, M.: Instance-based learning algorithms. Mach. Learn. 6 (1), 37-66
(1991)
2. Altendorf, E., Restificar, A., Dietterich, T.: Learning from sparse data by exploiting monotonic-
ity constraints. In: Proceedings of the 21st Annual Conference on Uncertainty in Artificial
Intelligence, pp. 18-26. AUAI Press, Arlington, Virginia (2005)
3. Barabási, A.: Linked: How Everything Is Connected to Everything Else and What It Means for
Business, Science, and Everyday Life. Plume, New York (2003)
4. Bellman, R.E.: Adaptive Control Processes—A Guided Tour. Princeton University Press,
Princeton (1961)
5. Botsch, M.: Machine Learning Techniques for Time Series Classification. Cuvillier, Munchen
(2009)
 
Search WWH ::




Custom Search