Time series analysis plays an essential role in today’s society due to the ease of access to information. This analysis is present in the majority of applications that involve sensors, but in recent years thanks to technological advancement, this approach has been directed towards the treatment of complex signals that lack periodicity and even that present non-stationary dynamics such as signals of brain activity or magnetic and satellite resonance images. The main challenges at the time of time series analysis are focused on the representation of the same, for which methodologies based on similarity measures have been proposed. However, these approaches are oriented to the measurement of local patterns point-to-point in the signals using metrics based on the form. Besides, the selection of relevant information from the representations is of high importance, in order to eliminate noise and train classifiers with discriminant information for the analysis tasks, however, this selection is usually made at the level of characteristics, leaving aside the Global signal information. In the same way, lately, there have been applications in which it is necessary to analyze time series from different sources of information or multimodal, for which there are methods that generate acceptable performance but lack interpretability. In this regard, we propose a framework based on representations of similarity and multiple-instance learning that allows selecting relevant information for classification tasks in order to improve the performance and interpretability of the models