Graph-based fine-grained model selection for multi-source domain

The prosperity of datasets and model architectures has led to the development of pretrained source models, which simplified the learning process in multi-domain transfer learning. However, challenges such as data complexity, domain shifts, and performance limitations make it difficult to determine which source model to transfer. To meet these challenges, source model selection has emerged as a promising approach for choosing the best model for a given target domain. Most literature utilizes transferability estimation combined with statistical methods to deduce the model selection probability, which is a coarse-grained method that selects a single model with limited accuracy and applicability in multi-source domains. To break through this limitation, we propose a graph-based fine-grained multi-source model selection method (GFMS) that aims to adaptively select the best source model for any single target domain data. Specifically, our proposed method comprises three main components: building a source model library through cross-training; generating the selection strategy by exploring the similarities among the data features, the associations between the features and models based on graph neural networks; blending the selected models using a weighted approach to obtain the best model adaptively. Experimental results demonstrate that the proposed adaptive method achieves higher accuracy in both model selection and image classification than the current state-of-the-art methods on compared datasets.