.png){kind=icon}
.png){kind=icon}
Return
A distributed inference algorithm for Dirichlet process mixture models with exponential family components
.png){kind=icon}
OA
AI
DOI：10.1016/j.neucom.2025.131119.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• In the federated context, we propose a novel distributed inference framework for Dirichlet Process Mixture Models (DPMMs) based on a Master/Worker architecture. Data is evenly partitioned among workers to ensure a balanced workload. Workers operate independently without sharing information with each other, communicating only with the master by exchanging minimal necessary statistics. • Each worker executes a local collapsed Gibbs sampler to discover local clusters and infer a local DPMM. The sufficient statistics associated with these local clusters are then transmitted to the master node. • At the master level, the global DPMM and clustering structure are estimated solely using sufficient statistics, without accessing the raw data from each cluster. • We demonstrate the effectiveness of our approach on both continuous and discrete data. For continuous data, we focus on a multivariate Gaussian mixture model and achieve a runtime of just 3 min for 100 iterations on a dataset with 100 K points, compared to 12 h required by the centralized collapsed Gibbs sampler. For discrete data, we introduce a multinomial DPMM with an application for text clustering, showing the versatility of our framework. • Furthermore, we demonstrate how our approach can be generalized and still work with the exponential family of distributions. The computational details are provided. The overall workflow of our model is illustrated in Fig. 1.
Keywords:
Horizontal federated learning
Distributed computing
Dirichlet process mixture models
Markov chain Monte Carlo
Bayesian non-parametric modeling
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
View more organizations