Tips Sentiment
Opportunity
Currently, there aren't many ways to sort or rank venues against each other. Mostly, it's done by 'checkin's or 'herenow's; occasionally they're delineated by gender or category.
Tips are a sprawling mess of awesomeness. They are micro-reviews of venues that can be parsed in ways that full reviews from services like Yelp, TripAdvisor, and Thrillist can't. If a definitive taxonomy was designed to automatically assign numeric values to tips, the aggregate analysis could rival the depth of existing review sites.
The Hack
In real-time, you can search for venue tip sentiment around you. Venues are color coded to indicate favorable or unfavorable, and weighted by confidence.
How It works
The HTML5 map does an initial venue search from coordinates sent from a web browser or phone. Tips for each venue are looked up, then scored on a favorability scale. Since tips aren't changed very often, venues and tips are cached (indefinitely for now, as a hack).
In this iteration, the algorithm to determine favorability only assumes there's "good" or "bad" associated with a venue tip. To initially train it, we paired thousands of sample tips against a set list of good and bad keyphrases. We did a naive Bayesian analysis on the sample set to guess what additional word patterns are likely to occur in favorable or unfavorable tips. Additional analysis is done as tips are added to the cache.
Caveats and Potential Improvements
- Sentiment isn't a black and white spectrum
- When there is a large variety of venues to choose from, polarizing venues are often more interesting than bland average venues; here they're getting the same score
- Sentiment may not be the best thing to derive from tips, since they're usual calls to action
- Different venue types require different probabilities ("smell" could be a positive at a restaurant, not cool at a gym)
Team
Patrick Butler (@doctorfantastic), Eric Tang (@ericxtang), Heliodor Jalba (@heliodorj)
