TL;DR
This paper introduces VisitSense, a smartphone framework that accurately detects and predicts users' movements within shopping malls using ambient radio signals and accelerometer data. This enables highly personalized mobile advertising by understanding user visit patterns, offering a more effective and less intrusive experience for both users and advertisers.
What Problem Does This Paper Address?
The paper tackles the challenge of providing effective and relevant mobile advertising within large, complex environments like shopping malls. Traditional location-based advertising (LBA) is often too simplistic, relying only on current proximity, which can lead to irrelevant ads. Behavioral targeting, successful in online environments, is difficult to implement offline due to the complexity of accurately tracking user movements and understanding their intent within a physical space. Existing methods struggle with the noisy and dynamic radio environments of shopping malls, leading to inaccurate visit detection and prediction.
Methods
VisitSense employs a two-pronged approach:
- Change-Based Visit Detection: Instead of relying on stable radio signals (which are rare in malls), VisitSense detects transitions. It monitors changes in ambient radio stability (using a "Stable AP List" or SAL) and user mobility (via accelerometer data). Entrance is detected when radio becomes stable and mobility decreases, while departure is signaled by radio instability and increased mobility. This method is designed to be robust to the noisy radio environments of shopping malls.
- Causality-Based Visit Prediction: To predict where a user might go next, VisitSense uses Bayesian Networks. This model learns sequential visit patterns and the inherent causality between them (e.g., after dinner, a user is more likely to visit a cafe than another restaurant). It incorporates features like place, time, duration, age, and gender to make predictions.
- Noise-Filtered Wi-Fi Fingerprinting: For place recognition, VisitSense filters out unreliable Wi-Fi access points (APs) that are intermittent or have weak signals. This creates more robust fingerprints that are better at distinguishing between adjacent places.
Datasets
- COEX Mall Dataset: A Wi-Fi fingerprint database was generated for the COEX Mall, covering approximately 120 places.
- Real-world Deployment Data: Sensing logs (Wi-Fi scans and accelerometer data) were collected from tours within the COEX Mall to evaluate visit detection and place recognition.
- User Survey Data: Data from approximately 130 participants who listed their visited places, visit times, and durations in the COEX Mall was used to evaluate visit prediction accuracy.
- User Study Participants: 15 participants were recruited for a preliminary user study to compare VisitSense-based advertising (VAA) with traditional LBA.
Key Insights and Ideas
- Noise is the Enemy: The paper strongly emphasizes that the noisy and dynamic nature of Wi-Fi signals in shopping malls is a major hurdle for existing location-sensing techniques.
- Transitions, Not Stability: Focusing on changes in radio signals and mobility, rather than assuming stability, is a more effective strategy for detecting place visits in challenging environments.
- Causality Matters: Understanding the sequential nature of human behavior (e.g., what people do after visiting a particular place) is crucial for accurate prediction and, consequently, for effective targeted advertising.
- Privacy-Centric Design: By processing data locally on the smartphone and only pulling targeted ads, VisitSense offers a more privacy-conscious approach compared to systems that send raw sensor data to a server.
Key Results
- Visit Detection: VisitSense achieved significantly better visit detection accuracy than SensLoc, a state-of-the-art system, in noisy radio environments. Entrance detection precision was 0.83 (recall 1), and visit detection precision was 0.75 (recall 0.9). Visit sequence detection precision was 0.70 (recall 0.86).
- Place Recognition: Noise-filtered Wi-Fi fingerprinting improved place recognition accuracy across various similarity algorithms, with Pearson correlation increasing from 68% to 85%.
- Visit Prediction: The Bayesian Network model achieved the highest prediction accuracy (59.45% for top-1 prediction, 70.27% for top-2) when all features (place, time, duration, gender, age) were included. This outperformed decision trees and CRFs.
- User Study: Participants rated VAA ads as more useful (2.50) than LBA ads (2.13), despite receiving fewer ads from VAA (26.6 vs. 33.66). This suggests VAA is more efficient for advertisers and less intrusive for users.
Limitations and Future Work
- Power Consumption: The paper acknowledges that power consumption for accelerometer and Wi-Fi sensing was not addressed and is an important practical issue. Future work will focus on optimizing this through adaptive duty cycling and hierarchical sensing.
- User Acceptance: While promising, more in-depth study is needed to fully understand user acceptance of visit-pattern-aware mobile ads and the specific benefits users expect.
- Group Visit Patterns: The current study did not explicitly classify visitors into groups. Future work could analyze group visit patterns (e.g., couples, families) and tailor ads accordingly.
- Comprehensive User Study: The current user study was preliminary and limited in participant demographics, sample size, and study duration. A more comprehensive study is planned.
- Smartphone Model Diversity: The evaluation was conducted on two Android smartphone models, and further investigation into performance variations across different devices is needed.
Summary updated