Subspace Clustering via Integrating Sparse Representation and Adaptive Graph Learning

Abstract

Sparse representation is a powerful tool for subspace clustering, but most existing methods for this issue ignore the local manifold information in learning procedure. To this end, in this paper we propose a novel model, dubbed Sparse Representation with Adaptive Graph (SRAG), which integrates adaptive graph learning and sparse representation into a unified framework. Specifically, the former can preserve the local manifold structure of data, while the latter is useful for digging global information. For the objective function of SRAG has multiple intractable terms, an ADMM method is developed to solve it. Numerous experimental results demonstrate that our proposed method consistently outperforms several representative clustering algorithms by significant margins.

FedGS: Federated Graph-Based Sampling with Arbitrary Client Availability

“…We first review the client availability settings discussed in existing FL literature (Ribero, We observe that the client's active probability may depend on data distribution or time (Ribero, Vikalo, and De Veciana 2022;Gu et al 2021). We mainly conclude the existing client availability modes and propose a comprehensive set of seven client availability modes in Table 1 to conduct experiments under arbitrary availability.…”

“…In a FL system, there is a server that broadcasts a global model to clients and then aggregates the local models from them to update the global model. Such a distributed optimization may cause prohibitive communication costs due to the unavailability of clients (Gu et al 2021).…”

“…(Li et al 2020) sample clients in proportion to their data sizes with replacement to obtain an unbiased estimator of update. More recently, some works take the client availability into account when sampling clients (Yan et al 2020;Gu et al 2021;Balakrishnan et al 2021;Cho, Wang, and Joshi 2020; Figure 1: A motivating example: there are significant challenges to achieving both stable model updates and bias-free training with the intermittent client availability in FL. Huang et al 2020).…”

“…They show that selecting clients without considering whether the clients are active will lead to unbounded waiting times and poor response rates. As a result, they sample only the active clients to guarantee immediate client availability (Gu et al 2021;Cho, Wang, and Joshi 2020). For example, in Fig.…”

“…Nevertheless, these works ignored the long-term bias introduced by arbitrary client availability, which will result in the model overfitting on a particular data subset. Recent works(Ribero, Vikalo, and De Veciana 2022;Gu et al 2021;Huang et al 2020) are aware of such longterm bias from the arbitrary availability of clients. However, these two competitive issues (e.g.…”