An Autonomous Reputation Framework for Physical Locations based on WiFi Signals

Afra J. Mashhadi, Geert Vanderhulst, Utku Günay Acer, Fahim KawsarWpa@mobisys 2015 · DOI

TL;DR

This paper proposes a novel framework to automatically assess the reputation of physical locations (like restaurants or shops) by analyzing WiFi probe requests from smartphones. Instead of relying on user reviews, it leverages the Preferred Network Lists (PNLs) of devices to infer a "spatial score" that reflects how similar visitors perceive a place, offering a more passive and potentially broader measure of reputation.

What Problem Does This Paper Address?

Online review platforms like Yelp and TripAdvisor are popular for gauging the quality of urban venues. However, they suffer from several limitations:

  • Limited Participation: Only a small fraction of venue visitors contribute reviews, leading to potentially biased or incomplete information.
  • Lack of Social Context: Reviews don't reveal whether the opinion comes from a trusted social circle or a stranger.
  • Pre-planning Required: Users must actively seek information before visiting, diminishing the spontaneity of discovery.

This paper aims to overcome these issues by creating an "offline spatial scoring mechanism" that passively collects data from all visitors with WiFi-enabled devices, not just those who choose to review.

Methods

The core of the framework relies on analyzing WiFi probe requests, which are unencrypted management frames sent by devices to discover available WiFi networks.

  • Preferred Network List (PNL) Extraction: When a device sends a probe request, it often includes a list of SSIDs (network names) it has previously connected to. This list, the PNL, is captured.
  • Homophily Assumption: The paper leverages the theory of homophily, which suggests that similar people tend to engage in similar activities. Therefore, users with similar PNLs are assumed to have similar preferences and experiences.
  • Spatial Score Calculation:
    • A "virtual PNL" is constructed for each user, combining their device's PNL with inferred visit duration and frequency.
  • A pairwise similarity score is computed between a user's PNL and all previously stored PNLs for a given venue.
  • This score is weighted using TF-IDF (Term Frequency-Inverse Document Frequency) to account for the uniqueness of APs (private APs are more unique than public ones) and a "recurrence weight" to factor in how often a user has visited.
  • The final spatial score for a user at a venue is an average of these weighted similarity scores.
  • Beacon Stuffing: To communicate the spatial score back to the user without requiring an active internet connection, the paper proposes using "beacon stuffing." This involves embedding the score in WiFi probe response frames, which can then be interpreted by the device's OS or applications.

Datasets

Not discussed in the paper. The paper describes an early prototype and an evaluation plan involving 50 users, but does not detail specific datasets used for initial development or testing.

Key Insights and Ideas

  • Reputation from Passive Sensing: The most significant insight is the potential to derive meaningful reputation scores from ubiquitous, passively collected WiFi signals, bypassing the need for active user input.
  • PNLs as Social Proxies: The PNL, often seen as a technical detail for network management, is repurposed as a rich source of information about a user's past experiences and potential social connections.
  • Homophily in WiFi Usage: The paper posits that WiFi network preferences can serve as a proxy for social homophily, allowing for the inference of shared experiences among users.
  • Personalized, Contextual Information Delivery: The use of "beacon stuffing" offers a novel way to deliver personalized location-based information directly to users' devices without requiring an internet connection, enhancing the spontaneity of discovery.
  • Spatial Score Distribution: The authors hypothesize that the distribution of spatial scores for a venue can reveal its nature – a normal distribution for private, socially connected spaces versus a long-tail (Pareto-like) distribution for public spaces visited by a wider, less connected audience.

Key Results

Not discussed in the paper. This is a position paper outlining a framework and an early prototype, with an evaluation plan for the future. No quantitative results are presented.

Limitations and Future Work

  • MAC Address Randomization: The paper acknowledges that MAC address randomization by devices poses a challenge for long-term device identification. They propose using the PNL itself as a more stable identifier.
  • Inferring Visit Duration: While they infer visit duration from probe request timing, short visits (seconds) might be filtered out, potentially missing brief but significant interactions.
  • Privacy Concerns: Although the PNL is anonymized, the potential for inferring user behavior and social connections raises privacy considerations that need careful handling.
  • Future Work:
    • Implementing and evaluating the framework in a real-world setting with 50 users.
  • Developing an Android application to visualize the spatial score using icons and semantic labels.
  • Assessing usability through surveys and interviews.
  • Further exploring the spatial score distribution for different venue types.

Summary updated

BibTeX
@inproceedings{mashhadi2015autonomous,
  author = {Afra J. Mashhadi and Geert Vanderhulst and Utku Günay Acer and Fahim Kawsar},
  title = {{An Autonomous Reputation Framework for Physical Locations based on WiFi Signals}},
  booktitle = {Wpa@mobisys},
  year = {2015},
  doi = {10.1145/2753497.2753498},
}
Figure