Unsupervised Online Bayesian Autonomic Happy Internet-of-Things Management

Rossi Kamal, Choong Seon HongarXiv 2015 · arXiv

Introduction

The rapid expansion of the Internet-of-Things (IoT) has created a complex ecosystem where service providers and users must find a mutually beneficial equilibrium. This paper introduces Whiz, an unsupervised online Bayesian framework designed for "Happy IoT" management. Whiz addresses the fundamental challenge of synchronizing service provider revenue with the often unobservable and dynamic usage contexts of IoT users—such as emotions or environmental factors. By treating context estimation as an "Online Stochastic Fact-Finding" problem, Whiz enables more responsive and personalized service delivery.

The "Happy IoT" Problem

In a "Happy IoT" ecosystem, two primary actors exist: the service provider (Player A) and the user (Player B). The provider is "happy" when revenue is maximized, while the user is "happy" when the services provided align perfectly with their current needs. The core difficulty lies in the fact that a user's "Choice-of-Interest" (COI)—their immediate context like emotion, location, or intent—is frequently unobservable to the provider.

The interaction model between Service Providers (Player A) and Users (Player B) in the Happy IoT framework.

Providers must infer this hidden context from "Incentive Filters" (IF), which are sensed data points like usage duration, frequency, or location. This inference is complicated by several factors:

  • Noisy Sensor Data: IoT sensors can be unreliable, poorly calibrated, or provide deceptive signals.
  • Dynamic Contexts: User interests and environmental factors shift rapidly over time.
  • Algorithmic Limitations: Traditional fact-finding algorithms, such as PageRank or simple voting schemes, often lack the necessary convergence guarantees or global performance metrics required for real-time autonomic management.

The Whiz Framework

The authors propose Whiz to bridge the gap between noisy sensor data and accurate context estimation. Whiz integrates into the classic MAPE (Monitor, Analyze, Plan, Execute) loop but replaces manual policy-making with an automated Bayesian inference engine. The framework decomposes the management problem into three integrated sub-models:

  1. Measurement Model (Recruitment): Employs K-means clustering to identify and attract initial user groups sharing common interests.
  2. Learning Model (Revenue-renewal): An iterative algorithm based on Expectation-Maximization (EM) that treats usage context as a latent variable. It estimates the likelihood of a "fact" (the true user context) from incomplete and unreliable sensed data.
  3. Filtering Model (Stochastic Monetization): Utilizes Bayesian online filtering to adapt to frequent changes in user interests in real-time.
The Whiz ecosystem architecture, illustrating the integration of various sensors, business goals, and the core inference engine.

A key innovation in Whiz is the application of the Bayesian Cramer-Rao Lower Bound (BCRLB). Unlike standard EM algorithms that may settle for local likelihood, Whiz uses BCRLB to quantify the global confidence of its estimation results. This allows the system to self-regulate and improve its accuracy as prior knowledge evolves.

Evaluation and Results

The Whiz framework was validated using a combination of synthetic benchmarks and real-world datasets, including e-mail surveys of smart-device usage, YouTube group dynamics, Cenceme accelerometer/GPS traces, and Higgs Twitter social trends.

Key findings include:

  • Superior Accuracy: In synthetic testing, Whiz maintained a significantly lower Root Mean Square Error (RMSE) compared to traditional fact-finders like PageRank, TruthFinder, and standard EM models.
  • Faster Convergence: Whiz demonstrated quicker convergence to optimal solutions, particularly as the number of available "Incentive Filters" (sensors) increased.
  • Real-world Utility: A prototype developed for Android and Web platforms showed that the "Online" version of Whiz outperformed "Offline" versions in processing intermittent GPS samples for user activity estimation.

Conclusion

Whiz represents a significant step toward truly autonomic IoT management. By reframing context estimation as a non-linear optimization problem and leveraging Bayesian inference, it provides a robust method for synchronizing provider goals with user needs. While challenges remain—such as the computational overhead on edge devices and the difficulty of inferring complex states like "emotion" solely from network traffic—Whiz offers a scalable foundation for future "Happy IoT" ecosystems.

Summary updated

BibTeX
@misc{kamal2015unsupervised,
  author = {Rossi Kamal and Choong Seon Hong},
  title = {{Unsupervised Online Bayesian Autonomic Happy Internet-of-Things Management}},
  year = {2015},
  eprint = {1509.06856},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1509.06856},
}
Figure