This thesis addresses several problems that can be treated through models of opinion dynamics, according to which individuals, connected according to complex networks, interact through rules that shape their preferences and opinions in relation to a particular issue. The methodology used to investigate the patterns that emerge from those interactions relies on the use of various techniques of statistical physics. The thesis is organized around four distinct problems, with a particular question to be answered in each case, always looking for empirical validation of the theoretical and computational results. In the first work, it is answered the following basic question about network properties that can have impact on the spreading processes: what are the typical values of the distances, clustering coefficient and other structural quantities, when considering the ensemble of random networks with fix assortativity? In the second study, we investigated the patterns that emerge in the ratings of films, considering as source IMDb (Internet Movie Database). We found that the distribution of votes has a scale-free behavior with a exponent close to 3/2. Interestingly, this pattern is robust, independently of movie attributes such as average note, age or gender. The empirical analysis points to a simple mechanism of adoption propagation, that generates mean-field avalanches. In the third study, we discuss the problem of multiple choices by means of a model which includes the possibility of indecision and where the choices of individuals evolve according to a plurality rule. We show that this dynamics on top of networks with the small-world property produces different stationary states that depend on the number of alternatives and on the degree distribution: consensus, wide adoption distributions similar to actual ones and situations where indecision prevails when the number of alternatives is large enough. Finally, we investigate the appearance of extreme positions in society, through the polls on a wide variety of questions. The increase of extreme opinions has as precursor a non-linear relationship between the fraction of extremists and that of moderates. We propose a model with activation rules, based on the stubbornness of the individuals, which enables interpreting the beginning of the non-linearity in terms of an abrupt transition of the class of bootstrap percolation, where activation cascades occur. As a general conclusion, we emphasize that this thesis illustrates how opinion models, combined with huge databases, provide results with power of interpretation and prediction of empirical patterns.