Predicting Structured Outputs k-Nearest Neighbours Method
In this work, we address several tasks of structured prediction and propose a new method for handling such tasks. Structured prediction is becoming important as data mining is dealing with increasingly complex data (images, videos, sound, graphs, text,...). Our method, k-NN for structured prediction (kNN-SP), is an extension of the well known k-nearest neighbours method and can handle three different structured prediction problems: multi-target prediction, hierarchical multi-label classification, and prediction of short time-series. We evaluate the performance of kNN-SP on several datasets for each task and compare it to the performance of other structured prediction methods (predictive clustering trees and rules). We show that, despite it’s simplicity, the kNN-SP method performs satisfactory on all tested problems.
KeywordsNear Neighbor Dynamic Time Warping Structure Output Descriptive Attribute Short Time Series
Unable to display preview. Download preview PDF.
- 1.Bakir, G.H., Hofmann, T., Schölkopf, B., Smola, A.J., Taskar, B., Vishwanathan, S.V.N.: Predicting Structured Data. The MIT Press, Cambridge (2007)Google Scholar
- 2.Blockeel, H., De Raedt, L., Ramon, J.: Top-down induction of clustering trees. In: Proc. 15th International Conference on Machine Learning, pp. 55–63. Morgan Kaufmann, San Francisco (1998)Google Scholar
- 5.Dasarathy, B.V. (ed.): Nearest Neighbor (NN) Norms: NN Pattern Classification Techniques. IEEE Computer Society, Los Alamitos (1990)Google Scholar
- 8.Kocev, D.: Ensembles for Predicting Structured Outputs. PhD Thesis, Jozef Stefan International Postgraduate School, Ljubljana, Slovenia (2011)Google Scholar
- 13.Yianilos, P.N.: Data structures and algorithms for nearest neighbor search in general metric spaces. In: Proc. 4th ACM-SIAM Symposium on Discrete Algorithms, pp. 311–321. SIAM, Philadelphia (1993)Google Scholar