This thesis is built on the idea of modeling episodes of multiple time series which can be briefly defined as multivariate time series whose individual dimensions vary in time and nature. This kind of data arises naturally when we observe repeatedly scenarios where collections of individual elements that may or may not take part in the collective observed behaviour. We illustrate the ideas constructed around this kind of data making use of datasets related to crowdfunding and video-on-demand. These datasets are prolonged periods of observation of these scenarios and provide natural examples to the ideas we develop. How to relate seemingly disconnected individual episodes and how to incorporate information from them into the general view of the multiple episodes is the main goal of this thesis. We focus on constructing this two-way flux so that even more complex models than the ones present in this work can be constructed using the proposed features. We describe models and algorithms that mix supervised and unsupervised tasks. Specifically, we construct models that connect Topic Models, unsupervised learning models that aim to summarize big corpora of texts with regression models on time series. We also discuss how summaries of past episodes may be helpfull in predicting future series of observations of same category.