BigDataGrapes D5.3 - Trust-aware Decision Support System

This document presents an update of deliverable 5.3 where we design and evaluate a trust-aware decision support system that uses visualisation techniques to explain the influence of input (predictor) variables on prediction outcomes. Research has shown that prediction models currently employed in agricultural decision support systems (DSSs) remain opaque to users and hidden behind the software. This black-box nature can often lead to users not trusting the system’s decisions especially when the system fails to provide meaningful explanations. Previous work has expressed that explaining a model’s predictions is an important approach for earning users’ trust. Visualisation is a powerful technique to address this problem and can effectively communicate uncertainty emerging from both data and prediction models.

To demonstrate our first version of  a decision support system, in the previous deliverable, we used an example wine quality dataset which was based on red variants of the Portuguese "Vinho Verde" wine and contains 1599 instances of 11 physicochemical (inputs) variables and a sensory (the output) variable which is wine quality. In this updated version, we have redesigned the system to integrate a dataset that has been collected by one of the pilot partners, AUA. The dataset contains quality assignments for one grape variety, which are based on four features of the grape: 1) total anthocyanin content, 2) berry fresh weight, 3) total soluble solids and 4) titratable acidity. The system aims to answer some of the most important questions in viticulture: 1) which machine learning (ML) model should I use with the data that is specific to this vineyard/grape variety, 2) how do various grape parameters affect the quality predictions of different ML models, and 3) which of the different ML models produces an output that is in-line with my knowledge? In this document, we describe the new version of our decision support system and the results of a qualitative evaluation which was conducted over several semi-structured interview sessions with the partners.

This document is structured as follows. Section 1 lays out an introduction to the deliverable describing our previous version and motivations. In Section 2, our DSS is described in detail together with the development technology we utilised. In Section 3, the results of the evaluation are presented. In Section 4, we provide a usage manual with instructions on how to obtain the source code. This document concludes with Section 5 where a summary of the deliverable is underlined.