SVM-Based Clustering Algorithm
Author Information
Author(s): Stephen Winters-Hilt, Sam Merat
Primary Institution: University of New Orleans
Hypothesis
Can SVMs be effectively used for clustering without prior knowledge of input classes?
Conclusion
Non-parametric SVM-based clustering methods may allow for much improved performance over parametric approaches.
Supporting Evidence
- The algorithm improves accuracy by iteratively re-labeling misclassified vectors.
- Clustering performance is evaluated using measures like Purity and Entropy.
- The SVM-External approach is not biased towards the shape of the clusters.
Takeaway
This study shows that a special computer program can group things together without knowing what they are beforehand, making it better than older methods.
Methodology
An SVM-based clustering algorithm was introduced that clusters data with no prior knowledge of input classes, using iterative re-labeling and SVM re-training.
Potential Biases
The algorithm may be biased towards the shape of the clusters in feature space.
Limitations
Convergence to a global optimum is not guaranteed.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website