Predicting response waiting time in a chat room

In a chat room, the Response Waiting Time (RWT) provides a useful insight on user's behaviour. The RWT is the time difference between successive messages between two people. This responding time could either encourage or discourage the online conversation. In this paper, we investigate which properties of a chat can be used to predict the response waiting time. The properties considered are the number of messages exchanged between individuals, the statistical model that best describes the RWT of a conversation, the sentiment of the utterances and the topics discussed. These properties were used as inputs in a Neural Network and a Support Vector Machine to predict the duration of a chat. To classify which of these properties are more relevant for the RWT prediction we used a decision tree algorithm. We observed that the best predictors are the total number of messages exchanged by pairs of individuals and the statistical model that describes the RWT of a conversation. The topics are the next relevant and the sentiment of the utterances are the least relevant.