Human-robot teaming largely relies on the ability of machines to respond and relate to human social signals. Prior work in Social Signal Processing has drawn a distinction between social cues (discrete, observable features) and social signals (underlying meaning). For machines to attribute meaning to behavior, they must first understand some probabilistic relationship between the cues presented and the signal conveyed. Using data derived from a study in which participants identified a set of salient social signals in a simulated scenario and indicated the cues related to the perceived signals, we detail a learning algorithm, which clusters social cue observations and defines an "N-Most Likely States" set for each cluster. Since multiple signals may be co-present in a given simulation and a set of social cues often maps to multiple social signals, the "N-Most Likely States" approach provides a dramatic improvement over typical linear classifiers. We find that the target social signal appears in a "3 most-likely signals" set with up to 85% probability. This results in increased speed and accuracy on large amounts of data, which is critical for modeling social cognition mechanisms in robots to facilitate more natural human-robot interaction. These results also demonstrate the utility of such an approach in deployed scenarios where robots need to communicate with human teammates quickly and efficiently. In this paper, we detail our algorithm, comparative results, and offer potential applications for robot social signal detection and machine-aided human social signal detection.