Temporal Network Maps: a novel approach to overcome the complexity of wine tasting data

Sensory temporal data is one of the biggest challenges within Sensory Science due to its high analytical and interpretability complexity. This work proposes the use of temporal network maps as a novel visualization strategy to address these challenges through a case study using a published dataset by Visalli et al., (2023). The study included the assessment of four wines evaluated by consumers and international wine students using “Free-Comment” at three time points, mimicking the three phases of wine tasting: attack (Time 1), evolution (Time 2), and finish (Time 3).

Text data were tokenised and converted into numerical vectors following a Bag-of-Words approach. The individual time points were considered as a subset of the data and the temporal network dataframe was built from their overlap. Each node represents unique sensory attributes, with the node position being fixed across all three time points; however, the node size is dependent on the frequency of co-occurrence of the individual sensory attribute at each time point. By convention, time points are displayed in different colours so that temporal changes can be identified from a qualitative (colour) and a quantitative (node size) perspective. From a qualitative perspective, temporal network maps revealed a clear qualitative shift in sensory vocabulary across tasting stages. Early in the tasting experience (Time 1), descriptions were concentrated around aroma-related terms whereas by Time 3, the vocabulary broadened considerably to include mouth-feel attributes alongside additional aroma descriptors. From a quantitative perspective, these helped to identify through the node-size (frequency) dimension the evolution of terms during the full tasting experience. For example, specific sensory attributes such as “red fruit” were used across all three-time point became less frequent as tasting progressed. The use of temporal network maps followed an unstructured approach which can help to identify relevant patterns.