Dissertation
USING PROBABILISTIC GRAPHICAL MODELS FOR ACTIVITY PREDICTION
Doctor of Philosophy (PhD), Washington State University
01/2014
Handle:
https://hdl.handle.net/2376/111311
Abstract
Recent advances in the areas of pervasive computing, data mining, and machine learning offer unprecedented opportunities to provide health monitoring and assistance for individuals experiencing difficulties living independently at home. Several components
have to work together to provide health monitoring for smart home residents including activity recognition, activity discovery, activity prediction, and prompting infrastructure. Despite the significant work that has been done to discover and recognize activities in smart
home research, less attention has been paid to predict the future activities that a smart home resident is likely to perform.
We focus on employing directed probabilistic graphical models, often known as Bayesian networks (BNs), to predict the activities that smart home residents are likely to perform. To describe a BN, one would need to specify whether the graph structure is known or not and whether the data is fully observable or contains hidden variables. The mentioned points give rise to four different scenarios.
In this dissertation, we utilize dynamic BNs to address the temporal activity prediction problem for each of the above mentioned scenarios. For the scenario where the BN structure is known and data contains hidden variables, we propose a model based on autoregressive hidden Markov models called POsH. The POsH model is based on a novel inference process that first predicts future observations and then predicts future hidden states. Next, we customize POsH to comply with the case where the BN structure is known and data is fully observable. We refer to the latter model as CRAFFT. We also propose a novel method to predict the time offset between current activity and predicted activity by CRAFFT, based on modeling the time offset as a continuous normal distribution and outlier detection. For the case where the BN structure is unknown and data has full observability, we present a model based on BNs, called SUDO, and compare its performance with some well-known search and score-based and constraint-based BN structure learning algorithms. We evaluate all of the proposed models using real data collected from 20 smart homes set up at two different retirement communities.
Metrics
45 File views/ downloads
17 Record Views
Details
- Title
- USING PROBABILISTIC GRAPHICAL MODELS FOR ACTIVITY PREDICTION
- Creators
- Ehsan Nazerfard
- Contributors
- Diane J. Cook (Advisor)Lawrence B. Holder (Committee Member)Behrooz A. Shirazi (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Electrical Engineering and Computer Science, School of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Number of pages
- 182
- Identifiers
- 99900581445901842
- Language
- English
- Resource Type
- Dissertation