Rituals of Leaving

Rituals of Leaving: Predictive Modelling of Leaving Behaviour in Conversation

van Doorn, Felix (TU Delft Electrical Engineering, Mathematics and Computer Science)

Hung, Hayley (mentor)
Reinders, Marcel (graduation committee)
Liem, Cynthia (graduation committee)

Delft University of Technology

2018-08-28

In this thesis, leaving behaviour in a small group setting is studied. In the past, group conversations have mostly been studied in a static setting. Inclusion of a temporal dimension would be of use in numerous applications. In order to account for temporal elements, we must understand how group conversations evolve. Leaving and joining a conversation are a vital part of social interaction and the driving factors behind the evolution of groups. Understanding group evolution has numerous applications, such as (social) surveillance, socially intelligent robotics and multimedia content analysis. In this thesis, I will focus on researching the nature of leaving behaviour in a mingling setting. A better understanding of leaving behaviour is required to incorporate this behaviour in models of group evolution.

By performing an ethnographic study on video material from the MatchNMingle dataset, 90 minutes of annotations of conversational leaving behaviour are collected. These annotations are subsequently combined with action annotations (either ground truth and automatically extracted) and accelerometer readings from the same dataset to describe the behaviour in each interaction. Using a sliding window approach, subsequences of these interactions are extracted using window sizes between 10 and 30 seconds. From these subsequences, pairwise coordination features are extracted between each pair present in the group.
These coordination behaviours are known to be indicative of conversational involvement, which is hypothesized to be linked to leaving.
These features are used in two different experimental settings.
In the first of these settings, a classifier was trained on subsequences of these interactions. This classifier attempts to predict whether a sequence of interaction with result in a leave within the next n seconds, where n varies between 0 and 30 seconds.
In the second experiments, the classifier is replaced with a regression model. For each of the determined subsequences, the time until the interaction ends is assigned and fit the regression model on these inputs.
This is the first study that has attempted to predict leaving using these data sources.
Using this approach, above random performance was achieved. In the classification setting an AUC of 0.63 +/- 0.12 is achieved, which is significantly better than random performance.
In the regression experiments no satisfactory results were obtained.

From these results we can conclude that it is possible to predict conversational leaving behaviour from coordination features. Although there are many possible improvements, the proposed methodology can serve as solid foundation for future research. This does not only include the proposed methodology for creation of predictive models, but also the new feature sets on which these experiments were conducted.

http://resolver.tudelft.nl/uuid:b13e7c6e-03ee-43b2-8b0a-6b79c5e31e5b

Student theses

master thesis

© 2018 Felix van Doorn