Alternative Title
Bayesian Variational Inference in Keyword Identification and Multiple Instance Classification
Subject Area
Statistics
Abstract
This dissertation investigates (1) Variational Bayesian Semi-supervised Keyword Extraction and (2) Variational Bayesian Multimodal Multiple Instance Classification.
The expansion of textual data, stemming from various sources such as online product reviews and scholarly publications on scientific discoveries, has created a demand for the extraction of succinct yet comprehensive information. As a result, in recent years, efforts have been spent in developing novel methodologies for keyword extraction. Although many methods have been proposed to automatically extract keywords in the contexts of both unsupervised and fully supervised learning, how to effectively use partially observed keywords, such as author-specified keywords, remains an under-explored area. In Chapter 1, we propose a novel variational Bayesian semi-supervised (VBSS) keyword extraction approach, built on a recent Bayesian semi-supervised (BSS) technique that uses the information from a small set of known keywords to identify previously undetected ones. Our proposed VBSS method greatly enhances the computational efficiency of BSS via mean-field variational inference, coupled with data augmentation, which brings closed-form solutions at each step of the optimization process. Further, our numerical results show that VBSS offers enhanced accuracy for long texts and improved control over false discovery rates when compared with a list of state-of-the-art keyword extraction methods.
In Chapter 2, we apply mean-field variational inference on multiple instance learning (MIL). In MIL, objects are represented by bags of instances. Each instance shares the same feature set but has unique feature values. MIL aims to train models that predict bag-level outcomes based on these instances, making it a weakly supervised approach due to the lack of instance-level labels. While MIL methods focusing on binary classification are abundant, they often cannot identify which specific instances drive bag labels and have limited or little interpretability. Xiong et al. (2024) introduced MICProB, a Bayesian multiple instance classification (MIC) algorithm that addresses these issues. However, MICProB is computationally intensive and best suited for unimodal instances. To overcome these limitations, we propose a novel variational Bayesian multimodal MIC (vMMIC) algorithm. vMMIC handles diverse instance types and significantly improves computational efficiency through Bayesian variational inference, combined with data augmentation. We benchmark vMMIC against MICProB and many other MIC approaches on both simulated and real-world data. Results demonstrate vMMIC's superior performance, computational efficiency, and interpretability.
Degree Date
Summer 8-6-2024
Document Type
Dissertation
Degree Name
Ph.D.
Department
Statistics and Data Science
Advisor
Xinlei Wang
Number of Pages
80
Format
".pdf"
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Hu, Yaofang, "Bayesian Variational Inference in Keyword Identification and Multiple Instance Classification" (2024). Statistical Science Theses and Dissertations. 46.
https://scholar.smu.edu/hum_sci_statisticalscience_etds/46
