Streaming PTAS for Constrained k-Means

TL;DR

This paper introduces a streaming PTAS for constrained k-means clustering, leveraging list-based solutions and stability assumptions to improve efficiency and applicability in large-scale and constrained scenarios.

Contribution

It generalizes list-k-means algorithms to streaming settings and enhances PTAS efficiency under stability assumptions, enabling practical constrained k-means clustering.

Findings

Developed a 2-pass streaming algorithm for list-k-means.

Converted the streaming algorithm into a 4-pass logspace PTAS for constrained k-means.

Significantly improved the running time of stable k-means algorithms under certain conditions.

Abstract

We generalise the results of Bhattacharya et al. (Journal of Computing Systems, 62(1):93-115, 2018) for the list-$k$-means problem defined as -- for a (unknown) partition $X_1, ..., X_k$ of the dataset $X \subseteq \mathbb{R}^d$, find a list of $k$-center sets (each element in the list is a set of $k$ centers) such that at least one of $k$-center sets $\{c_1, ..., c_k\}$ in the list gives an $(1+\varepsilon)$-approximation with respect to the cost function $\min_{\textrm{permutation } \pi} \left[ \sum_{i=1}^{k} \sum_{x \in X_i} ||x - c_{\pi(i)}||^2 \right]$. The list-$k$-means problem is important for the constrained $k$-means problem since algorithms for the former can be converted to PTAS for various versions of the latter. Following are the consequences of our generalisations: - Streaming algorithm: Our $D^2$-sampling based algorithm running in a single iteration allows us to design a 2-pass, logspace streaming algorithm for the list-$k$-means problem. This can be converted to a 4-pass, logspace streaming PTAS for various constrained versions of the $k$-means problem. - Faster PTAS under stability: Our generalisation is also useful in $k$-means clustering scenarios where finding good centers becomes easy once good centers for a few "bad" clusters have been chosen. One such scenario is clustering under stability where the number of such bad clusters is a constant. Using the above idea, we significantly improve the running time of the known algorithm from $O(dn^3) (k \log{n})^{poly(\frac{1}{\beta}, \frac{1}{\varepsilon})}$ to $O \left(dn^3 k^{\tilde{O}_{\beta \varepsilon}(\frac{1}{\beta \varepsilon})} \right)$.