.png){kind=icon}
.png){kind=icon}
Return
Behavior analysis of elderly using topic models
.png){kind=icon}
PRE
AI
DOI：10.1016/j.pmcj.2014.07.001.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper describes two new topic models for the analysis of human behavior in homes that are equipped with sensor networks. The models are based on Latent Dirichlet Allocation (LDA) topic models and can detect patterns in sensor data in an unsupervised manner. LDA-Gaussian, the first variation of the model, is a combination of a Gaussian Mixture Model and the LDA model. Here the multinomial distribution that is normally used in the LDA model is replaced by a set of Gaussian distributions. LDA-Poisson, the second variation of the model, uses a set of Poisson distribution to model the observations. The Poisson distribution is better suited to handle counts of stochastic events but less well-suited to model time. For this we use the von Mises distribution, resulting in 'LDA-Poisson-von-Mises'. The parameters of the models are determined with an EM-algorithm. The models are evaluated on more than 450 days of real-world sensor data, gathered in the homes of five elderly people, and are compared with a baseline approach where standard k-means clustering is used to quantize the data. We show that the new models find more meaningful topics than the baseline and that a semantic description of these topics can be given. We also evaluated the models quantitatively, using perplexity as measure for the model fit. Both LDA-Gaussian and LDA-Poisson result in much better models than the baseline, and our experiments show that, of the proposed models, the LDA-Poisson-von-Mises model performs best. (C) 2014 Elsevier B.V. All rights reserved.
Keywords:
Activity discovery
Sensor homes
Pervasive computing
Sequential patterns
LDA
Topic models
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
3.5
Papers:
1.5K
Citations:
2.2K