Learning humans' behavior from activity logs requires choos- ing an adequate machine learning technique regarding the situation at hand. This choice impacts signi cantly results reliability. In this paper, Hidden Markov Models (HMMs) are used to build intentional process models (Maps) from activity logs. Since HMMs parameters require to be learned, the main contribution of this paper is to compare supervised and unsupervised learning approaches of HMMs. After a theoretical compari- son of both approaches, they are applied on two controlled experiments to compare the Maps thereby obtained. The results demonstrate using su- pervised learning leads to a poor performance because it imposes binding conditions in terms of data labeling, introduces inherent humans' biases, provides unreliable results in the absence of ground truth, etc. Instead, unsupervised learning obtains e cient Maps with a higher performance and lower humans' effort.