Abstract
Neural networks in one form or another are common precision agriculture artificial intelligence techniques for making predictions based on data. However, neural networks are computationally intensive to train and to run, and are typically " black-box " models without explainable output. This paper investigates an alternative artificial intelligence prediction technique, genetic algorithm (GA) quantitative association rule mining (QARM), which creates explainable output with impacts directly quantified in the existing dataset. QARM takes one or more data features of a dataset and restricts the value of the feature between two bounds. These feature(s) are then associated with a particular outcome from the data (such as frost). The resulting rule's correlation can then be quantified in terms of the support (how often it is seen in the dataset), confidence (how often it co-occurs with the outcomes), and lift (how much more or less often we see this than expected). The genetic algorithm component finds the optimal features and value bounds to maximize the significance of the correlation. Generating quantitative association rules with genetic algorithms is not a new method, however, it is not commonly used and likely deserves more attention in the explainable AI realm. Additionally, this paper extends the technique by adding a sequence to each feature to analyze time data. Time steps were added to value bounds to determine what time range in the past was most significant to the correlation. This technique was compared with neural network predictors for multivariate time-sequence weather data in two scenarios: the open Jornada Basin LTR dataset for the purpose of predicting frost one day ahead, and a custom-collected dataset from Laurel Grove Wine Farm in Winchester, Virginia to predict frost in 5 minute intervals. The QARM GA technique had comparable performance to the neural network methods in the Jornada Basin dataset (0.803 F1 statistic score on the dataset compared to 0.847 for the neural network) while generating highly interpretable and computationally cheap-to-implement prediction rules. For the Laurel Grove Wine Farm study, both techniques were limited in overall results, but the genetic algorithm outperformed the neural network method (0.489 F1 score for the QARM GA method compared to 0.217 F1 score for the neural network). The results of these experiments indicate quantitative association rule mining is worth further investigation for artificial intelligence in precision agriculture.