Invariant Learning Beyond Explicit Environment Modeling

View the PDF file from the paper entitled Rustle of the tape in the generalization of the graph: fixed learning beyond the modeling of the explicit environment, by Shu Shen and 7 other authors

PDF HTML (experimental) view

a summary:The generalization outside the distribution (OOD) has emerged as a decisive challenge in learning the graph, as the real -world graph data often appear various environments that fail traditional models in circulating. The promising solution to address this problem is the graph of learning (GIL), which aims to learn fixed representations by dismantling the unrelated subcategrams to naming the sub -spaces of the environment. However, the current GIL methods face two main challenges: (1) The difficulty of capturing and modeling various environments in the graph data, and (2) semantic shelf, where it is difficult to distinguish between fixed sub -hoses of different categories, which leads to the poor separation of the layer and the increase in the wrong classification. To deal with these challenges, we suggest a new method called fixed learning (MPHil), which offers major innovations: (1) Extracting unexpected extraction, providing superior classrooms, and multi -category classification, which is a vary in a overlapping category. Get rid of the need for an explicit environmental modeling in Gil and alleviate the issue of semantic cliff. Discovered from the theoretical framework of GIL, we offer two new objective functions: the loss of matching the initial model to ensure the matching samples with the initial models of the right layer, and the loss of the initial model for increasing the distinction between the initial models of the different groups in the high -speed space. Intensive experiments on 11 ooud circulation groups show that MPHil achieves modern performance, and greatly outperforms the performance of current methods through graph data from various fields and with different distribution transformations.