Toward parsimonious modeling of affect dynamics in daily life: An analysis of continuous time, non-linearities and sudden jumps

To study affect dynamics, people's emotions are typically tracked during a certain period of time - subjects have to carry a smartphone on them which regularly prompts them to fill out a questionnaire regarding their momentary emotional experiences. The result is a time series of emotion ratings per participant. Mathematical modeling is one of the principal methods to extract information from these time series. Models enable us to study affect dynamical properties in a systematic manner. Furthermore, they allow us to make predictions about reality, which make it possible to falsify and improve them. In order for a model to be useful, it has to provide an accurate description of affect dynamics and its parameters have to be inferable from actual affect data. In general, the discovery of affect dynamical features in time series is hampered by the limited number of observations; sample sizes are generally constrained because of the burden the act of measuring imposes on participants. The limited sample size puts a cap on the number of model parameters that can reliably be estimated. As such, the difficulty is constructing models for affect dynamics with the right amount of complexity, the difficulty is the construction of a parsimonious model. This PhD project deals with the parsimonious modeling of affective time series. We start from the vector autoregressive (VAR) model, the current standard for affect dynamical modeling. We discuss a classification of the VAR model in terms of the different dynamical regimes that it encompasses. Then, we show that many of its degrees of freedom can actually be neglected for the description of affective time series. In a second step, we note that although the VAR model is the current standard for affect dynamical modeling, it has a major shortcoming: it disregards time and only considers the measurements' sequence. We introduce the Ornstein-Uhlenbeck (OU) model, a continuous-time model that is closely related to the VAR model. Then, we compare the OU model to the VAR model and show that the OU model does not trivially lead to a better description of affective time series; there are drastic sudden changes in affective time series which obscure the continuous-time dynamics. Both the VAR and the OU model are inherently Gaussian - affect states are always expected to be sampled from some Gaussian distribution. Yet, empirical evidence suggest that this is not the case. In a next step, we develop the Affective Ising Model (AIM), a nonlinear model for affect. It is shown that, unlike the OU model, the AIM is able to capture non-trivial features of the data, like skewness and multi-modality. Because of this, the AIM typically provides a better description of the data and is less influenced by sudden changes. Throughout this project, we frequently rely on cross-validation to assess the predictive performance of models. A bottleneck of cross-validation is that a similar procedure has to be done over and over again. This can result in infeasibly long computation times. To counter this, the synergized bootstrap method is developed, an optimization scheme to speed up the optimization of multiple related objective functions in the context of resampling methods.

Jaar van publicatie:2021

Toegankelijkheid:Open