Sadari Jayawardena(The University of New South Wales, Sydney), Julien Epps(School of Electrical Engineering and Telecommunications, UNSW Australia) and Zhaocheng Huang(School of Electrical Engineering and Telecommunications, UNSW Australia)
Many affective computing datasets are annotated using ordinal scales, as are many other forms of ground truth involving subjectivity, e.g. depression severity. When investigating these datasets, the speech processing community has chosen classification problems in some cases, and regression in others, while ordinal regression may also arguably be the correct approach for some. However, there is currently essentially no guidance on selecting a suitable machine learning and evaluation method. To investigate this problem, this paper proposes a neural network-based framework which can transition between different modelling methods with the help of a novel multi-term loss function. Experiments on synthetic datasets show that the proposed framework is empirically well-behaved and able to correctly identify classification-like, ordinal regression-like and regression-like properties within multidimensional datasets. Application of the proposed framework to six real datasets widely used in affective computing and related fields suggests that more focus should be placed on ordinal regression instead of classifying or predicting, which are the common practices to date.