|
Title:
|
FACTORING THE HABITS: COMPARING METHODS FOR DISCOVERING BEHAVIOR PATTERNS FROM LARGE SCALE ACTIVITY DATASETS |
|
Author(s):
|
Onur Yürüten, Pearl Pu |
|
ISBN:
|
978-989-8533-54-8 |
|
Editors:
|
Piet Kommers and Guo Chao Peng |
|
Year:
|
2016 |
|
Edition:
|
Single |
|
Keywords:
|
Time series analysis; activities of daily living; behavior profiling |
|
Type:
|
Full Paper |
|
First Page:
|
179 |
|
Last Page:
|
186 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
click to dowload 
|
|
Paper Abstract:
|
The abundance of wearable sensors leads to massive growth of datasets of Activities of Daily Living (ADLs). Many ADL-based applications will thus need to incorporate scalable and efficient methods to organize this data. This requires an in-depth understanding of the empirical properties of alternative clustering methods. In this spirit, we present a comparative analysis of two powerful candidates based on matrix factorization. While the first approach stems from computer vision applications (Low Rank and Sparse Decomposition - LRSD), the second approach is a collaborative filtering method used in recommender systems to model temporal trends (Time-SVD++). We describe the necessary modifications to adapt these approaches to ADL datasets, and then compare and contrast them in two major aspects: scalability and clustering quality. We quantify our comparisons with run-time complexity analysis and clustering quality measurements on two different datasets (one ADL dataset and one ratings dataset) with statistical significance. Our results not only confirm these methods superiority over basic clustering approaches, but also demonstrate notable differences between ADL datasets and customer ratings datasets. We conclude that in contrast to traditional recommender systems approaches, ADL clustering methods should specifically handle the density and noise in ADL datasets. |
|
|
|
|
|
|
