Unsupervised Similarity Learning for Spectral Clustering
Abstract: Spectral clustering has been popularized due to its ability to identify non-convex boundaries between individual clusters.
However, it requires defining a similarity metric to construct the Laplacian matrix.
Instead of predefining this metric upfront, we propose to learn it by finding the optimal parameters of a kernel function.
This learning approach parameterizes the data topology by optimizing a similarity function that assigns high similarity values to a pair of data that share discriminative features and vice versa.
While some existing approaches also learn the similarity values, they rely on hyperparameters to do so.
However, these hyperparameters cannot be validated in an unsupervised setting.
As a result, suboptimal hyperparameter values can lead to detrimental performance.
To circumvent this drawback, we propose a method that eliminates the need for hyperparameters by learning the optimal parameter for a similarity metric used in spectral clustering.
This enables unsupervised learning of the similarity metric while performing spectral clustering.
The method's capability is verified on several benchmark datasets with a large scale of non-convexity.
Our method outperforms SOTA approaches on accuracy and normalized mutual information measures up to 10$\%$ when applied to popular image and text datasets.
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=FLloecOFZn&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DTMLR%2FAuthors%23your-submissions)
Changes Since Last Submission: Updated the motivation in the text.
Assigned Action Editor: ~Brian_Kulis1
Submission Number: 2458
Loading