A Calculation Method of the Similarity Between Trained Model and New Sample by using Gaussian Distribution

In many Human Robot Interaction scenarios, social robots are expected to communicate with human naturally. Especially, the skill of predicting human internal state is a field of great interest in many applications, including video surveillance, behavior analysis, human robot interaction, life-logging. While accurately predicting these technologies (e.g. emotion, confident for talking a topic) could have benefits for many fields, generic machine learning systems still yield low performance in some situation. We hypothesize that these sophisticated models suffer from individual differences of human’s personality. Therefore, we proposed a multi characteristic model architecture which combines the personalized machine learning models and utilize each model’s prediction score in the inference. This architecture formed with reference to ensemble machine learning architecture. In this research, we focus on a similarity between new user and trained user model by using the idea of applicability domain of machine learning models. In the empirical result, we confirmed that data distribution (one way of checking applicability domain) of each user model correspond to the performance of models and we estimated confidence during communication as a human internal state.