Utilising Target Adjacency Information for Multi-target Prediction

Abstract

In this paper, we explored how information on the cost of misprediction can be used to train supervised learners for multi-target prediction (MTP). In particular, our work uses depression, anxiety and stress severity level prediction as the case study. MTP describes proposals which results require the concurrent prediction of multiple targets. There is an increasing number of practical applications that involve MTP. They include global weather forecasting, social network users’ interaction and the thriving of different species in a single habitat. Recent work in MTP suggests the utilization of “side information” to improve prediction performance. Side information has been used in other areas, such as recommender systems, information retrieval and computer vision. Existing side information includes matrices, rules, feature representations, etc. In this work, we review very recent work on MTP with side information and propose the use of knowledge on the cost of incorrect prediction as side information. We apply this notion in predicting depression, anxiety and stress of 270,322 anonymous respondents to the DASS-21 psychometric scale in Malaysia. Predicting depression, anxiety and stress based on the DASS-21 fit an MTP problem. Often, a patient experiences anxiety as well as depression at the same time. This is not unusual since it has been discovered that both tend to co-exist at different degrees depending on a patient’s experience. By using existing machine learning algorithms to predict the severity levels of each category (i.e., depression, anxiety and stress), the result shows improved precision with the use of cost matrix as side information in MTP.