# Predicting Temporal Liking of Food Pairings from Temporal Dominance of Sensations Data via Reservoir Computing on Crackers and Spreads

## TL;DR

This study uses a type of neural network to predict how much people like food pairings over time based on sensory data collected during eating.

## Contribution

The novel use of reservoir computing to predict temporal liking from temporal dominance of sensations data in food pairings.

## Key findings

- Reservoir computing models achieved an RMSE of 0.44 on a 9-point liking scale for predicting temporal liking curves.
- Models with minimal auxiliary inputs performed best in predicting temporal liking from TDS data.
- The approach could reduce the need for extensive sensory evaluations in food pairing studies.

## Abstract

The temporal dominance of sensations (TDS) and temporal liking (TL) methods offer complementary insights into the evolution of sensory and hedonic responses during food consumption. This study investigates the feasibility of predicting TL curves for food pairings from their TDS profiles using reservoir computing, a type of recurrent neural network. Participants evaluated eight samples—two crackers (plain, sesame), two spreads (peanut butter, strawberry jam), and their four binary combinations—performing both TDS and TL evaluations. This process yielded paired time-series data of TDS and TL curves. We trained various reservoir models under different conditions, including varying reservoir sizes (64, 128, 192, or 256 neurons) and the inclusion of auxiliary input dimensions, such as flags indicating the types of foods tasted. Our results show that models with minimal auxiliary inputs achieved the lowest root mean squared errors (RMSEs), with the best performance being an RMSE of 0.44 points on a 9-point liking scale between the observed and predicted TL curves. The ability to predict TL curves for food pairings holds some promise for reducing the need for extensive sensory evaluation, especially when a large number of food combinations are targeted.