Predición de preferencia de usuario mediante técnicas de "Soft Computing"

Abstract

The content of this Thesis by Compendium is the grouping of three research articles published in prestigious journals, which shows the need and how to improve the methods of predicting user aesthetic preference using soft computing techniques. An extensive state of the art of the use of artificial neural networks and deep learning is performed. This study shows that there are systems based on neural networks capable of performing artistic tasks with varying degrees of objectivity. From the detection of an object in a pictorial work to the creation of images, the classification according to the artistic style or author, or the estimation of quality and aesthetic value. It also shows that in recent years more work is being done on more complex tasks such as image creation, largely thanks to the use of deep learning techniques such as Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN). From this basis, the use of neural network-based systems is formulated for two relevant tasks in the field of aesthetic preference prediction. On the one hand, a system of artificial neural networks is used for aesthetic prediction, using a dataset explored by the state of the art. Not only a low error in the prediction is sought but also a network whose topology is minimal. The results are analyzed by drawing conclusions about the minimum information relevant to perform this highly subjective and complex task. On the other hand, different alternatives for another task highly related to aesthetic perception are analyzed: the perception of visual complexity. There are numerous psychological studies that propose a direct relationship between complexity and aesthetic value. It is proposed to look for a machine learning method, which obtains better prediction of this value. An analysis of the outlayers is also performed, in order to better understand the processes performed by the prediction mechanism.