In the current world context, with a globalized production system and highly interconnected farms, understanding and controlling livestock diseases and their spreading are critical issues both for food industry and for public decision-makers. In order to prevent outbreaks, it appears essential to characterize and predict animal trade movements, as major pathogen-inducing pathways between farms. A convenient way to grasp and explore the general structure of this complex system is to use a dynamic graph representation. Indeed, the vertices of this network are the farms (and also the commercial operators such as markets and assembly centers), while the edges represent animal exchanges between farms. This graph is directed (sellers send cattle to buyers), weighted (each edge is labelled with the number of exchanged animals), and time-varying (a transaction occurs at a given time-instant). Furthermore, additional information on the vertices, such as geographical situation, type of farm, ... is usually available. With this representation, mathematical models integrating the dynamics of the network, but also the epidemics spreading on the temporal network and the farmers’ behaviour with respect to trading may be developed and investigated. However, one of the limits of these models is their scalability, whereas cattle-trading networks are usually of substantial size. For instance, the data at our disposal, concerning the exchanges involving French farms from 2005 until 2009 with a daily resolution level, contains millions of edges and hundreds of thousands of nodes. In this context, it is important to build a reduced version of this network, by identifying groups of vertices with common features (in a broad sense). In this paper, we propose an exploratory study of the French cattle-trading network using selforganizing maps and, more particularly, the recently developed R-package SOMbrero, which implements both numerical and relational versions of the algorithm. The output consists in a clustering of the vertices and a nonlinear mapping of the graph, which is a reduced version of the original network. The SOM algorithm may be trained either on vector data, which may be numerical features extracted from the dynamical graph or from snapshots of it, or on relational data, which may be any kernel or dissimilarity computed on the vertices or on the edges, such as, for example, the dynamic shortest-path distance. We will illustrate both versions of the algorithm by providing some of the outputs (see Figures 1 and 2 for an example of the algorithm trained on numerical features extracted from one yearly snapshot of the network) and explaining the benefit of each of them in the process of understanding the inner mechanisms of the network