People-Aware Computing (PAC) Group

Most machine learning techniques today are applied to constrained datasets. These constraints may be explicit, as in the case of controlled user experiments, or implicit, where the structure of user interaction with an online application, e.g., Amazon or Netflix, controls the kind of raw data available for mining. As sensor-equipped mobile devices become ubiquitous, however, the assumptions of constrained datasets no longer apply. To apply learning techniques to large bodies of free-form sensor data and labels, we must relax our constraints and recognize that not only might some users be dishonest or malicious, but all users will have slightly different notions and words for labels. The goal of the community-guided learning project is to investigate techniques that make learning possible in such free-form environments.

Advances in ubiquitous computing have made it possible to gather data about the real-world behavior of entire groups of people. The aim of this work is to build activity-aware systems that can infer a wide range of human behavior and can adapt and evolve over time. We are developing an active learning algorithm for conditional random fields (CRFs) that at opportune moments actively queries the users for additional activity labels. This extends our earlier work on virtual evidence boosting (VEB), which is an efficient techniques for performing parameter estimation and feature selection in CRFs.

The pilot study is aimed at comparing the effectiveness of paper-based survey tools with that of a passive electronic mobile sensing device (MSP) for the measurement of overall physical and social activity of eight senior citizens at Kendal Continuing Care Retirement Community. Before the start of the study, the Yale Physical Activity Survey, SF-36, Friendship Scale, and the CES-D were used to obtain a baseline understanding of subjects’ physical, and social wellbeing. The study length was 15 consecutive days. Each subject was provided a charged, and activated device to wear at the waist during his/her waking hours and was instructed to continue performing their daily activities. Devices were collected at the end of the day for data extraction and charging. We are currently analyzing the results of the device and comparing them with the scores generated by the current paper-based surveys.

It is becoming increasingly easy to collect data that captures the simultaneous, real-world behavior of entire groups of people. Such data sets often capture, either directly or indirectly, interactions between people. Despite that, much of the research on such data considers behavior only at the level of a single person. Models that do consider social behavior typically rise only to the level of the dyad or small interacting group. Conversely, an arsenal of techniques has been developed for social network analysis but most of those methods consider only static, binary networks. Social networks derived from behavioral data will almost always be temporal and will often have finer grained observations about interactions than simple binary indicators. Work on multi-valued or temporal network models has been scarce since such data was previously hard to obtain. 

We are working on a new modeling framework that simultaneously models the dynamics and structural properties (e.g., transitivity, network density) of automatically collected behavioral data. Our model extends curved exponential random graph models to learn the strengths of pairwise interactions and how those interactions evolve over time.

SoundSense is a scalable framework for modeling sound events on mobile phones. The architecture and algorithms are designed for scalability and SoundSense uses a combination of supervised and unsupervised learning techniques to classify both general sound types (e.g., music, voice) and discover novel sound events speci?c to individual users. The system runs solely on the mobile phone with no back-end interactions.

Mobile medical sensors promise to provide an efficient, accurate, and economic way to monitor patients' health outside the hospital. Patient authentication is a necessary security requirement in remote health monitoring scenarios. The monitoring system needs to make sure that the data is coming from the right person before any medical or financial decisions are made based on the data. Credential-based authentication methods (e.g., passwords, certificates) are not well-suited for remote healthcare as patients could hand over credentials to someone else. Furthermore, one-time authentication using credentials or trait-based biometrics (e.g., face, fingerprints, iris) do not cover the entire monitoring period and may lead to unauthorized post-authentication use. Recent studies have shown that the human electrocardiogram (ECG) exhibits unique patterns that can be used to discriminate individuals. However, perturbation of the ECG signal due to physical activity is a major obstacle in applying the technology in real-world situations. As part of this project, we are developing a novel ECG and accelerometer-based system that can authenticate individuals in an ongoing manner under various activity conditions. We have collected data from 17 subjects and our results show that activity awareness leads to significant improvement in ECG-based recognition.