摘要
clustering的quality取决于两个problem-dependent factors
- the chosen similarity metric and
- the data representation.
本文贡献：
propose a deep supervised clustering metric learning method that formulates a novel loss function.
作者derive a closed-form expression for the gradient that is efficient to compute
intro
主要是学习representation，如下图
learn之后能控制close的程度，也就是distance metric。
the representations of training examples are closer to the representative vector of their category than to the representative vector of any other category.
本文是supervised clustering？
方法
用Bregman divergences将the problem of partitioning a dataset 给relax一下
这个变化后的问题（公式2）是NP-hard的，因此需要approximate一下，即relax，作者present a spectral relaxation of this problem。
以后再细看。