A proof of concept: a model trained on Wael's and Z's Beli rankings, reading how Yelp reviewers describe each place. It scores places neither of them has been to, explains why, and combines both tastes into one list. It rests on two people's lists, so the results show that the approach works, not that it is finished.
Correlation between how a place is described and where it sits on each person's list. Bars right of the line: the more reviewers say this, the higher the place is ranked. Solid bars are unlikely to be chance (p < 0.05).
Groups found from review language alone, overlapping (a place can be partly in two). The numbers are each person's average percentile for places in that group, so it shows which kinds each person favours.
166 topics found in 37,509 review sentences, split by positive, neutral and negative. The biggest in each aspect:
How the model is built, from the Beli lists and Yelp reviews to the group picks.
Each method ranks places it has not seen (5-fold, repeated 5 times). The number is the Spearman correlation between predicted and actual rank. The dashed line is what random guessing reaches 5% of the time; a method has to clear it to count.