ABSTRACTUsing the data set of The Three Hundred project, i.e. 324 hydrodynamical resimulations of cluster-sized haloes and the regions of radius 15 ${{h^{-1}\, {\rm Mpc}}}$ around them, we study galaxy pairs in high-density environments. By projecting the galaxies’ 3D coordinates onto a 2D plane, we apply observational techniques to find galaxy pairs. Based on a previous theoretical study on galaxy groups in the same simulations, we are able to classify the observed pairs into ‘true’ or ‘false’, depending on whether they are gravitationally bound or not. We find that the fraction of true pairs (purity) crucially depends on the specific thresholds used to find the pairs, ranging from around 30 to more than 80 per cent in the most restrictive case. Nevertheless, in these very restrictive cases, we see that the completeness of the sample is low, failing to find a significant number of true pairs. Therefore, we train a machine learning algorithm to help us identify these true pairs based on the properties of the galaxies that constitute them. With the aid of the machine learning model trained with a set of properties of all the objects, we show that purity and completeness can be boosted significantly using the default observational thresholds. Furthermore, this machine learning model also reveals the properties that are most important when distinguishing true pairs, mainly the size and mass of the galaxies, their spin parameter, gas content, and shape of their stellar components.