Elara
  • Home
  • Research
  • Projects
  • CV
  • Home
  • Research
  • Projects
  • CV
GitHubLinkedInInstagramORCIDBlueskyRSS Feed

© 2025 Elara Liu | All rights reserved.

Themed by EnjuFolio · Crafted by Elara Liu

Not All Places Are The Same: Computationally Surfacing Regional Experience Differences from Online Reviews

Zhuoran Liu, Medini Chopra, Haoqi Zhang

Advised by:Dr. Haoqi Zhang (Northwestern University)

ICWSM 2026Apr, 2025

Link
Keywords:
Experiential computingPlace-aware recommender systemsOnline reviewsGeographic variationLocation-based servicesNatural language processingHuman–computer interaction

Abstract

I led an HCI/NLP project that uses Yelp reviews and guidebook priors to model how the same place categories afford different experiences across U.S. regions. The work surfaces "regional experiential signatures"—like king cake, picon punch, or tax-free outlet shopping—that current category-based recommenders flatten.

1 Overview

Working with my advisor Haoqi, I led the development of a project that treats location-based recommendation as a question of how place imprints experience, not just where venues are located. Drawing on travel guidebooks and HCI theories of "space versus place," I operationalized narrative claims—such as King Cake being routine in Louisiana but "imported" in Pennsylvania—into a concrete hypothesis set. We compared how often and how concretely specific experience phrases appear in reviews for the same Yelp category across different states. This framing allowed us to hold categories constant to reveal geographic meaning shifts, using a framework of environmental, cultural, social, and institutional drivers to interpret the results.

To test this, I engineered an end-to-end text pipeline using the Yelp Open Dataset across thirteen U.S. states. I cleaned millions of reviews, extracted n-gram activity phrases, and computed TF–IDF salience before applying a stabilized log-ratio to quantify regional distinctiveness. To ensure data quality, I iteratively refined thresholds and blacklist rules to filter out brand chatter and implausible phrases. The system pairs every numerical contrast with ranked review snippets, allowing us to audit whether terms like "no sales tax" truly reflected local behavior.

The resulting case portfolio demonstrates that review language reliably encodes regional experiential signatures. However, it also highlighted limitations where purely affective claims (e.g., a beach feeling "personal") were difficult to capture. These findings informed our design implications for future systems: interfaces that expose experiential facets alongside ratings, context-aware defaults, and transparent textual evidence.

2 Selected visuals

CategoryRegionSurfaced phrase (experience)Dominant driver
BakeriesLouisianaking cake around Mardi Gras seasonCultural
Outlet mallsDelawareno sales tax / cross-border shoppingInstitutional
Basque restaurantsNevadapicon punch as a pre-dinner ritualSocial

Each row shows how, by holding the Yelp category fixed and contrasting review phrases across states, we recover distinct regional experiential signatures interpreted through our four-driver vocabulary.