The emerging technology of radio-frequency identification (RFID) opens entirely new ways to record various aspects of human social interactions in a broad set of settings. First of all, this technology allows to log the temporal dynamics of face-to-face interactions by detecting the physical proximity of participants. The power of this method has been demonstrated earlier in several studies addressing the temporal and spatial evolution of social interactions with high resolution [1, 2]. As examples, such experiments brought us understanding about the structure and evolution of networks formed by face-to-face interactions [3, 4] or how epidemics are spreading in specific environments like in schools or hospitals [2, 5, 6]. Building on these advancements we have taken an uncharted direction by designing RFID tags, which capture not only proximity but also the verbal communication of participants [7]. We use our technology for a large-scale longitudinal sociolinguistic experiment, where we record the oral and face-to-face interactions of children between age 3 to 6 in a preschool setting for one week per month over three years. Our primary goal is to understand how the acquisition of linguistic patterns of children, coming from families of various socioeconomic, cultural, and linguistic background, co-evolve with their social network in time towards a consensus matching the linguistic standards. There are several challenges in such an experiment starting with the reliable reconstruction of social interactions from the recorded RFID signals. Although this problematic has appeared earlier in other RFID studies, no systematic methods have been proposed so far, despite its importance to determine the outcome of any proximity study based on this technology. Packages exchanged between RFID sensors are sometimes lost due to noise or environmental conditions, which easily cause falsely interrupted or missed interactions or lead to the identification of fake ones. These can considerably bias the observations of the temporal social network and any dy-namical process, like language or epidemics, evolving on the top of it. In this contribution our aim is to close this gap by exploring several signal reconstruction strategies using supervised learning methods trained on recorded and annotated data obtained from controlled RFID experiments. Our aim is to find the best method, which provides the highest reconstruction score of temporal interactions for the best approximation of the original social network. More precisely, we consider the communication between pairs of RFID badges, which may receive a package from each other every five seconds with signal strength depending on the distance and orientation of the hosting people. First, using regression methods, we determine a baseline threshold of signal strength corresponding to a distance-range of 1 − 2 meters of realistic verbal face-to-face interactions. Using this threshold we convert our raw sequence to a sequence of pairs of binary signals indicating mutual simultaneous packages observed between sensors (see pairs of green arrows in Fig. 1b) if they capture a meaningful social interaction, or missing packages otherwise (red arrows in Fig. 1b). Reconstruction from a perfect signal would give us periods of interactions built from consecutive mutually observed packages, separated by 1