In this article we review the existing techniques in group recommender systems and we propose some improvement based on the study of the different individual behaviors when carrying out a decision-making process. Our method includes an analysis of group personality composition and trust between each group member to improve the accuracy of group recommenders. This way we simulate the argumentation process followed by groups of people when agreeing on a common activity in a more realistic way. Moreover, we reflect how they expect the system to behave in a long term recommendation process. This is achieved by including a memory of past recommendations that increases the satisfaction of users whose preferences have not been taken into account in previous recommendations.

Social Media are sensors in the real world that can be used to measure the pulse of societies. However, the massive and unfiltered feed of messages posted in social media is a phenomenon that nowadays raises social alarms, especially when these messages contain hate speech targeted to a specific individual or group. In this context, governments and non-governmental organizations (NGOs) are concerned about the possible negative impact that these messages can have on individuals or on the society. In this paper, we present HaterNet, an intelligent system currently being used by the Spanish National Office Against Hate Crimes of the Spanish State Secretariat for Security that identifies and monitors the evolution of hate speech in Twitter. The contributions of this research are many-fold: (1) It introduces the first intelligent system that monitors and visualizes, using social network analysis techniques, hate speech in Social Media. (2) It introduces a novel public dataset on hate speech in Spanish consisting of 6000 expert-labeled tweets. (3) It compares several classification approaches based on different document representation strategies and text classification models. (4) The best approach consists of a combination of a LTSM+MLP neural network that takes as input the tweet’s word, emoji, and expression tokens’ embeddings enriched by the tf-idf, and obtains an area under the curve (AUC) of 0.828 on our dataset, outperforming previous methods presented in the literature.

No abstract