Media sharing based on colocation prediction in urban transport

Liam McNamara, Cecilia Mascolo, Licia CapraMobiCom 2008 · DOI

TL;DR

This paper proposes a novel media sharing system for urban transport that leverages historical colocation data to predict when users will be near each other. By intelligently selecting content sources based on these predictions, the system significantly improves the success rate and efficiency of peer-to-peer file transfers, even in dynamic, mobile environments.

What Problem Does This Paper Address?

Commuting in urban areas consumes a significant amount of time, and with the prevalence of Bluetooth-enabled devices, there's a growing opportunity for peer-to-peer media sharing on public transport. However, the transient nature of these encounters poses a challenge: how do you ensure a file transfer completes when your potential source might disembark at any moment? Traditional approaches might rely on random selection or simply hope for the best, leading to incomplete transfers, wasted energy, and network congestion. This paper tackles the problem of efficiently and reliably sharing media content in such dynamic, mobile environments by predicting user colocation.

Methods

The core of the proposed solution is a colocation prediction scheme. The system learns from historical data of user movements to predict how long two users are likely to remain in proximity.

Here's a breakdown of the approach:

  • Colocation Pattern Logging: Each device maintains a profile of its encounters with other devices. Initially, this is an anonymous, aggregate record of colocation durations within specific time slices (e.g., hourly). As a device encounters another more frequently, it develops a personalized profile, recording mean colocation lengths for specific time slices. This allows for more accurate predictions based on the time of day and day of the week.
  • Content Source Selection: When a user (Alice) wants to download content, her device first filters potential sources based on her interests. Then, it ranks the remaining peers based on their predicted remaining colocation time. The system connects to the top-ranked peer, but only initiates a transfer if the predicted colocation duration is sufficiently longer than the estimated download time (controlled by an ATTEMPT_THRESH parameter). This prevents initiating transfers that are likely to fail.
  • Bluetooth Integration: The system leverages Bluetooth for device discovery and service discovery (SDP). Genres of interest are encoded as MD5 digests within service records, allowing for efficient filtering of peers based on content categories without direct interaction with their music applications.

Datasets

The paper uses a real-world dataset of anonymized movement traces collected from two lines of a large metropolitan mass transit system over one month. This dataset, derived from RFID card usage, provides user IDs, entry/exit times, and station information. The authors used this data to validate their assumption about the regularity of commuter movements and to inform their simulation parameters. They also generated synthetic content libraries and user interests based on data from Last.fm, which provides information on music popularity and genre tags.

Key Insights and Ideas

  • Human Movement is Predictable: The fundamental assumption that people exhibit regular movement patterns, especially during daily commutes, is strongly supported by the analysis of the real-world transit data. This regularity is the bedrock of the prediction scheme.
  • Time-of-Day Matters: Colocation durations are not uniform throughout the day. A morning commute might last significantly longer than an evening trip to a supermarket. The system's ability to learn and use time-specific colocation patterns is crucial for accurate predictions.
  • Familiar Strangers are Key: While the majority of encounters are with unique individuals, a non-negligible subset of passengers are "familiar strangers" who are encountered regularly. Exploiting the learned patterns of these familiar strangers leads to much better predictions than relying solely on aggregate data.
  • Proactive Failure Avoidance: The ATTEMPT_THRESH parameter is a smart way to avoid wasting resources on doomed transfers. By requiring a significant buffer between predicted colocation time and estimated download time, the system prioritizes successful downloads.
  • Decentralized and Infrastructure-Free: The system operates entirely on peer-to-peer Bluetooth communication, requiring no central infrastructure, making it highly deployable in real-world scenarios.

Key Results

  • Improved Transfer Success Rate: The "Prediction" scheme consistently achieved a significantly higher transfer success rate (at least 50% better) than a "Random" selection method, approaching 80% success at faster transfer speeds.
  • Enhanced Network Efficiency: The proposed method demonstrated at least 18% better network efficiency than "Random," meaning less data was wasted on incomplete transfers.
  • Increased Library Size: Users using the "Prediction" scheme gained 5 to 10 more files per host than "Random" users, highlighting the practical benefit of more successful downloads.
  • Balanced Genre Availability: The system promotes the availability of less popular music genres, helping to create a more diverse media landscape within the network.

Limitations and Future Work

  • No Peer Behavior Consideration: The current system does not account for malicious or misbehaving peers (e.g., sharing corrupted files). Future work will involve integrating a distributed trust management scheme to address this.
  • Focus on Transport: The current validation is limited to public transport. The authors suggest that similar principles could apply to other human contact networks with churn, like coffee shops or pubs.
  • Platform and Feature Expansion: The authors aim to expand the prototype to more platforms and explore additional features based on user feedback for future large-scale deployments.
  • Dynamic Parameter Learning: While the paper uses fixed time slices, it acknowledges that these could be dynamically learned from colocation data for even greater adaptability.

Summary updated

BibTeX
@inproceedings{mcnamara2008media,
  author = {Liam McNamara and Cecilia Mascolo and Licia Capra},
  title = {{Media sharing based on colocation prediction in urban transport}},
  booktitle = {MobiCom},
  year = {2008},
  doi = {10.1145/1409944.1409953},
}

Loading related content...

Figure