Boston University
Methods for learning with limited data in media and neuroscience research
Abstract
dc:description.abstractMachine learning (ML) models are critically sensitive to limitations in both data quality and data quantity. The quality issues include but are not limited to noise, bias, missing values, and inaccurate labels. The quantity issue is simply the lack of samples, which can prevent the models from learning complex patterns and generalizing to new situations. This thesis confronts these challenges by exploring how such data deficiencies can compromise ML outcomes and proposes strategies addressing these challenges. The first half of this thesis examines data quality challenges inherent in labeled datasets, considering both the labels and the human annotation process. Here, "quality" is evaluated through a data science lens, emphasizing suitability for ML applications. We first address the problem where pre-trained ML models are unavailable for an unconventional task, while acquiring sufficient labeled data for a new model is impractical. Our proposed solution leverages an existing ML model to perform simulations, thereby overcoming data limitations. In this thesis, we demonstrate this approach with an example from crowdsourcing, estimating the difficulty of correctly labeling data samples. Secondly, we tackle the issue of manual labeling being not only time-consuming but also susceptible to noise introduced by human subjectivity. To mitigate this, we introduce an automated labeling methodology employing a rich knowledge base to produce objective labels. This thesis presents an application of this automated approach in the field of communication research, where Wikipedia serves as the knowledge base for optimizing news labeling. The second half of this thesis examines the issue of data scarcity often encountered in medical research, where datasets typically contain numerous features but few samples. This imbalance poses a significant challenge such as model overfitting. We first scrutinize feature dimension through the investigation of feature selection techniques aimed at identifying the most salient predictors. Second, we tackle the sample size constraint by employing transfer learning, thereby harnessing the capabilities of pre-existing ML models trained on larger, often public, datasets to enhance performance despite limited specific data. In the thesis, the practical application of both strategies is demonstrated using functional neuroimaging data of post-stroke aphasia patients.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lai, Sha
- Advisors dc:contributor.advisor
-
- Betke, Margrit
- Ishwar, Prakash
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53298