Prediction of Wine Grade ECE 539: Intro to Artificial Neural Networks Matthew Kean
Wine Consumable product Various physicochemical properties that contribute to the flavor of it (i.e. ABV, pH, sugar content, etc.) Large market with significant competition to produce the best product
Goal of this Work Classification problem: determine the grade of a particular sample based only on the physicochemical properties Use a MLP to get a better classification rate than KNN If successful, the same method could be used for other consumable products
Data Collection Two different types of wine: one red and one white from a Portuguese vineyard Retrieved from UC-Ivine database (http://archive.ics.uci.edu/ml/datasets/Wine+Quality) 6000 samples total 11 features for each sample, with none missing
KNN Used a KNN classifier to determine the benchmark for a multi-layer perceptron classifier It had a peak value of 55% and an average classifying rate of 43%
Results for the MLP So far I have only been able to train MLPs with accuracies averaging 44% No advantage over the KNN yet
Other Work By reducing the number of classes, a classification rate of over 90% was achieved A different group obtained an 85% classification rate using a MLP, but they also got a slightly better classification rate using support vector machines.