Computer Science Theses and Dissertations
Permanent URI for this collectionhttp://hdl.handle.net/1903/2756
Browse
2 results
Search Results
Item Rich and Scalable Models for Text(2019) nguyen, thang dai; Boyd-Graber, Jordan; Resnik, Philip; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Topic models have become essential tools for uncovering hidden structures in big data. However, the most popular topic model algorithm—Latent Dirichlet Allocation (LDA)— and its extensions suffer from sluggish performance on big datasets. Recently, the machine learning community has attacked this problem using spectral learning approaches such as the moment method with tensor decomposition or matrix factorization. The anchor word algorithm by Arora et al. [2013] has emerged as a more efficient approach to solve a large class of topic modeling problems. The anchor word algorithm is high-speed, and it has a provable theoretical guarantee: it will converge to a global solution given enough number of documents. In this thesis, we present a series of spectral models based on the anchor word algorithm to serve a broader class of datasets and to provide more abundant and more flexible modeling capacity. First, we improve the anchor word algorithm by incorporating various rich priors in the form of appropriate regularization terms. Our new regularized anchor word algorithms produce higher topic quality and provide flexibility to incorporate informed priors, creating the ability to discover topics more suited for external knowledge. Second, we enrich the anchor word algorithm with metadata-based word representation for labeled datasets. Our new supervised anchor word algorithm runs very fast and predicts better than supervised topic models such as Supervised LDA on three sentiment datasets. Also, sentiment anchor words, which play a vital role in generating sentiment topics, provide cues to understand sentiment datasets better than unsupervised topic models. Lastly, we examine ALTO, an active learning framework with a static topic overview, and investigate the usability of supervised topic models for active learning. We develop a new, dynamic, active learning framework that combines the concept of informativeness and representativeness of documents using dynamically updating topics from our fast supervised anchor word algorithm. Experiments using three multi-class datasets show that our new framework consistently improves classification accuracy over ALTO.Item Cost-sensitive Information Acquisition in Structured Domains(2010) Bilgic, Mustafa; Getoor, Lise C; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Many real-world prediction tasks require collecting information about the domain entities to achieve better predictive performance. Collecting the additional information is often a costly process that involves acquiring the features describing the entities and annotating the entities with target labels. For example, document collections need to be manually annotated for classification and lab tests need to be ordered for medical diagnosis. Annotating the whole document collection and ordering all possible lab tests might be infeasible due to limited resources. In this thesis, I explore effective and efficient ways of choosing the right features and labels to acquire under limited resources. For the problem of feature acquisition, we are given entities with missing features and the task is to classify them with minimum cost. The likelihood of misclassification can be reduced by acquiring features but acquiring features incurs costs as well. The objective is to acquire the right set of features that balance acquisition and misclassification cost. I introduce a technique that can reduce the space of possible sets of features to consider for acquisition by exploiting the conditional independence properties in the underlying probability distribution. For the problem of label acquisition, I consider two real-world scenarios. In the first one, we are given a previously trained model and a budget determining how many labels we can acquire, and the objective is to determine the right set of labels to acquire so that the accuracy on the remaining ones is maximized. I describe a system that can automatically learn and predict on which entities the underlying classifier is likely to make mistakes and it suggests acquiring the labels of the entities that lie in a high density potentially-misclassified region. In the second scenario, we are given a network of entities that are unlabeled and our objective is to learn a classification model that will have the least future expected error by acquiring minimum number of labels. I describe an active learning technique that can exploit the relationships in the network both to select informative entities to label and to learn a collective classifier that utilizes the label correlations in the network.