Leveraging Vivino Experts Via Machine Learning

Abstract

Rating systems like Vivino offer a democratic way of assessing the quality of wines, wine regions, varietals, and styles. However, not all ratings are created equally; the ratings of experts and knowledgeable enthusiasts can be drowned out in a sea of biased ratings. Current methods of filtering out less trustworthy or downright fraudulent ratings are ad-hoc and, at the same time, it can be hard to justify throwing out any data at all. This motivates us to assess the quality of rater sub-populations in terms of their ability to predict the opinions of others. Using machine learning techniques on a Vivino social network public dataset, we find that the ratings from users who have global and local expertise (i.e., those who have rated more than 200 wines and those who have rated more than 10 wines from a particular wine region and style) have higher out-of-sample precision, compared to ratings sampled from the entire population.

Keywords: wine ratings, online rating systems, wine industry, machine learning