Weighting the past: an extended relational event model for negative and positive events

In relational event networks, the sentiment of each social interaction describes a qualitative characteristic of the relational event. The additional information about the sentiment of an event allows the researcher to better understand social interaction in temporal social networks. To achieve this, this paper introduces a modelling framework called SentiREM, which extends the standard relational event model by (i) including a logistic regression model for the type (or sentiment) of the next event given the observed dyad, (ii) including typed endogenous statistics which summarize the past event history including their type, and (iii) including memory parameters, which capture the decay of the weight of past events as a function of their transpired time and their type/sentiment, which are estimated from the data. We discuss how to estimate the model parameters, test hypotheses on the memory parameters and model coefficients of different event types, and learn how long past events are 'remembered' depending on their type/sentiment and transpired time. The proposed SentiREM is applied to an empirical case study to analyse social interactions between players in an online strategy game where positive and negative relational events (i.e. trades and attacks, respectively) were observed among players.