Abstract

A longstanding goal of human behavior science is to model and predict how humans interact with each other or with other systems. Such models are beneficial and have many applications, including designing and implementing assistive technologies, improving users' experiences and quality of life and making better decisions to create public policies. Behavior is highly complex due to uncertainties and a lack of scientific tools to measure it. Hence prediction of human behavior cannot be 100% accurate. However, prediction is also not hopeless because the biological needs, as well as cultural conventions (for instance, regarding meal times) set the general patterns of the humans' daily behavior. Furthermore, while individual humans might adjust these patterns according to their own preferences, they also show some degree of consistency in their daily routine. In this dissertation, we focus on interrelated challenges of improving the prediction models for human daily activities and developing techniques through which intelligent applications can benefit from this improved prediction. We describe techniques for creating predictive models that can help humans in their daily life using deep learning-based models. One of the challenges of learning based approaches in this setting is the scarcity of data. If we are collecting information about a given human in a home, our database will increase with exactly one sample a day – this is insufficient for deep learning algorithms that are often trained on datasets with millions of samples. We investigate three directions through which the paucity of samples can be overcome. First, we discuss techniques through which, starting from a small number of representative samples, we can generate much larger synthetic datasets that capture the statistical properties of the real world data, and can be used in training. We consider an application where we apply human behavior prediction to the practical problem of improving the quality of experience. By learning to predict the experience requested by the user, we are able to perform intelligent pre-caching, and achieve higher average quality of experience for a given available network bandwidth. Another direction we investigate is the collection of data from multiple users. This creates multiple challenges. First, users would prefer to minimize the shared personal data. This requires us to investigate techniques that learn predictive models from multiple user experiences without requiring the users to upload their data to a common repository. We adapt the technique of federated learning, which requires the users to only share the training gradients on a model that had been sent by a central server, but not raw data. We investigate procedures that allow the user to obtain the best possible model for her own prediction while minimizing the amount of data disclosed. The second challenge is that not all the users benefit to the same degree from creating a central learning model; by investigating how much the user can benefit, we can stop the learning process and implicit privacy loss earlier. Finally, we developed predictive models for the spread of pandemics and techniques that use these predictions to recommend Non-Pharmaceutical Interventions (NPIs) to local stakeholders. We find that the prediction of pandemics is also conditioned on the behavior of individual humans and the actions taken by the governments and, especially in the early phases of the pandemic, suffers from a lack of data. We used a combination of a deep learning-based predictive model with a compartmental model, which is trained on the months elapsed from the pandemic and predicts infection rates for the next months. We used cultural and geographical attributes as constant features along with the history of cases and deaths as context features and NPIs as action features to train a single predictive model that can predict both the infection rate and the stringency of the NPIs deployed by policymakers for all countries / regions. We found that the stringency is not always aligned with the number of cases but also depends on political, economic and cultural factors.

Notes

If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu.

Graduation Date

2021

Semester

Fall

Advisor

Boloni, Ladislau

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Degree Program

Computer Science

Format

application/pdf

Identifier

CFE0008916

Language

English

Release Date

December 2021

Length of Campus-only Access

None

Access Status

Doctoral Dissertation (Open Access)

Share

COinS