Theses and Dissertations
This collection includes both ASU Theses and Dissertations, submitted by graduate students, and the Barrett, Honors College theses submitted by undergraduate students.
Displaying 1 - 10 of 204
Filtering by
- Creators: Liu, Huan
- Creators: Glenberg, Arthur
Description
The increasing popularity of Twitter renders improved trustworthiness and relevance assessment of tweets much more important for search. However, given the limitations on the size of tweets, it is hard to extract measures for ranking from the tweet's content alone. I propose a method of ranking tweets by generating a reputation score for each tweet that is based not just on content, but also additional information from the Twitter ecosystem that consists of users, tweets, and the web pages that tweets link to. This information is obtained by modeling the Twitter ecosystem as a three-layer graph. The reputation score is used to power two novel methods of ranking tweets by propagating the reputation over an agreement graph based on tweets' content similarity. Additionally, I show how the agreement graph helps counter tweet spam. An evaluation of my method on 16~million tweets from the TREC 2011 Microblog Dataset shows that it doubles the precision over baseline Twitter Search and achieves higher precision than current state of the art method. I present a detailed internal empirical evaluation of RAProp in comparison to several alternative approaches proposed by me, as well as external evaluation in comparison to the current state of the art method.
ContributorsRavikumar, Srijith (Author) / Kambhampati, Subbarao (Thesis advisor) / Davulcu, Hasan (Committee member) / Liu, Huan (Committee member) / Arizona State University (Publisher)
Created2013
Description
Automating aspects of biocuration through biomedical information extraction could significantly impact biomedical research by enabling greater biocuration throughput and improving the feasibility of a wider scope. An important step in biomedical information extraction systems is named entity recognition (NER), where mentions of entities such as proteins and diseases are located within natural-language text and their semantic type is determined. This step is critical for later tasks in an information extraction pipeline, including normalization and relationship extraction. BANNER is a benchmark biomedical NER system using linear-chain conditional random fields and the rich feature set approach. A case study with BANNER locating genes and proteins in biomedical literature is described. The first corpus for disease NER adequate for use as training data is introduced, and employed in a case study of disease NER. The first corpus locating adverse drug reactions (ADRs) in user posts to a health-related social website is also described, and a system to locate and identify ADRs in social media text is created and evaluated. The rich feature set approach to creating NER feature sets is argued to be subject to diminishing returns, implying that additional improvements may require more sophisticated methods for creating the feature set. This motivates the first application of multivariate feature selection with filters and false discovery rate analysis to biomedical NER, resulting in a feature set at least 3 orders of magnitude smaller than the set created by the rich feature set approach. Finally, two novel approaches to NER by modeling the semantics of token sequences are introduced. The first method focuses on the sequence content by using language models to determine whether a sequence resembles entries in a lexicon of entity names or text from an unlabeled corpus more closely. The second method models the distributional semantics of token sequences, determining the similarity between a potential mention and the token sequences from the training data by analyzing the contexts where each sequence appears in a large unlabeled corpus. The second method is shown to improve the performance of BANNER on multiple data sets.
ContributorsLeaman, James Robert (Author) / Gonzalez, Graciela (Thesis advisor) / Baral, Chitta (Thesis advisor) / Cohen, Kevin B (Committee member) / Liu, Huan (Committee member) / Ye, Jieping (Committee member) / Arizona State University (Publisher)
Created2013
Description
Most people are experts in some area of information; however, they may not be knowledgeable about other closely related areas. How knowledge is generalized to hierarchically related categories was explored. Past work has found little to no generalization to categories closely related to learned categories. These results do not fit well with other work focusing on attention during and after category learning. The current work attempted to merge these two areas of by creating a category structure with the best chance to detect generalization. Participants learned order level bird categories and family level wading bird categories. Then participants completed multiple measures to test generalization to old wading bird categories, new wading bird categories, owl and raptor categories, and lizard categories. As expected, the generalization measures converged on a single overall pattern of generalization. No generalization was found, except for already learned categories. This pattern fits well with past work on generalization within a hierarchy, but do not fit well with theories of dimensional attention. Reasons why these findings do not match are discussed, as well as directions for future research.
ContributorsLancaster, Matthew E (Author) / Homa, Donald (Thesis advisor) / Glenberg, Arthur (Committee member) / Chi, Michelene (Committee member) / Brewer, Gene (Committee member) / Arizona State University (Publisher)
Created2013
Description
In most social networking websites, users are allowed to perform interactive activities. One of the fundamental features that these sites provide is to connecting with users of their kind. On one hand, this activity makes online connections visible and tangible; on the other hand, it enables the exploration of our connections and the expansion of our social networks easier. The aggregation of people who share common interests forms social groups, which are fundamental parts of our social lives. Social behavioral analysis at a group level is an active research area and attracts many interests from the industry. Challenges of my work mainly arise from the scale and complexity of user generated behavioral data. The multiple types of interactions, highly dynamic nature of social networking and the volatile user behavior suggest that these data are complex and big in general. Effective and efficient approaches are required to analyze and interpret such data. My work provide effective channels to help connect the like-minded and, furthermore, understand user behavior at a group level. The contributions of this dissertation are in threefold: (1) proposing novel representation of collective tagging knowledge via tag networks; (2) proposing the new information spreader identification problem in egocentric soical networks; (3) defining group profiling as a systematic approach to understanding social groups. In sum, the research proposes novel concepts and approaches for connecting the like-minded, enables the understanding of user groups, and exposes interesting research opportunities.
ContributorsWang, Xufei (Author) / Liu, Huan (Thesis advisor) / Kambhampati, Subbarao (Committee member) / Sundaram, Hari (Committee member) / Ye, Jieping (Committee member) / Arizona State University (Publisher)
Created2013
Description
Data mining is increasing in importance in solving a variety of industry problems. Our initiative involves the estimation of resource requirements by skill set for future projects by mining and analyzing actual resource consumption data from past projects in the semiconductor industry. To achieve this goal we face difficulties like data with relevant consumption information but stored in different format and insufficient data about project attributes to interpret consumption data. Our first goal is to clean the historical data and organize it into meaningful structures for analysis. Once the preprocessing on data is completed, different data mining techniques like clustering is applied to find projects which involve resources of similar skillsets and which involve similar complexities and size. This results in "resource utilization templates" for groups of related projects from a resource consumption perspective. Then project characteristics are identified which generate this diversity in headcounts and skillsets. These characteristics are not currently contained in the data base and are elicited from the managers of historical projects. This represents an opportunity to improve the usefulness of the data collection system for the future. The ultimate goal is to match the product technical features with the resource requirement for projects in the past as a model to forecast resource requirements by skill set for future projects. The forecasting model is developed using linear regression with cross validation of the training data as the past project execution are relatively few in number. Acceptable levels of forecast accuracy are achieved relative to human experts' results and the tool is applied to forecast some future projects' resource demand.
ContributorsBhattacharya, Indrani (Author) / Sen, Arunabha (Thesis advisor) / Kempf, Karl G. (Thesis advisor) / Liu, Huan (Committee member) / Arizona State University (Publisher)
Created2013
Description
Contemporary online social platforms present individuals with social signals in the form of news feed on their peers' activities. On networks such as Facebook, Quora, network operator decides how that information is shown to an individual. Then the user, with her own interests and resource constraints selectively acts on a subset of items presented to her. The network operator again, shows that activity to a selection of peers, and thus creating a behavioral loop. That mechanism of interaction and information flow raises some very interesting questions such as: can network operator design social signals to promote a particular activity like sustainability, public health care awareness, or to promote a specific product? The focus of my thesis is to answer that question. In this thesis, I develop a framework to personalize social signals for users to guide their activities on an online platform. As the result, we gradually nudge the activity distribution on the platform from the initial distribution p to the target distribution q. My work is particularly applicable to guiding collaborations, guiding collective actions, and online advertising. In particular, I first propose a probabilistic model on how users behave and how information flows on the platform. The main part of this thesis after that discusses the Influence Individuals through Social Signals (IISS) framework. IISS consists of four main components: (1) Learner: it learns users' interests and characteristics from their historical activities using Bayesian model, (2) Calculator: it uses gradient descent method to compute the intermediate activity distributions, (3) Selector: it selects users who can be influenced to adopt or drop specific activities, (4) Designer: it personalizes social signals for each user. I evaluate the performance of IISS framework by simulation on several network topologies such as preferential attachment, small world, and random. I show that the framework gradually nudges users' activities to approach the target distribution. I use both simulation and mathematical method to analyse convergence properties such as how fast and how close we can approach the target distribution. When the number of activities is 3, I show that for about 45% of target distributions, we can achieve KL-divergence as low as 0.05. But for some other distributions KL-divergence can be as large as 0.5.
ContributorsLe, Tien D (Author) / Sundaram, Hari (Thesis advisor) / Davulcu, Hasan (Thesis advisor) / Liu, Huan (Committee member) / Arizona State University (Publisher)
Created2014
Description
Most data cleaning systems aim to go from a given deterministic dirty database to another deterministic but clean database. Such an enterprise pre–supposes that it is in fact possible for the cleaning process to uniquely recover the clean versions of each dirty data tuple. This is not possible in many cases, where the most a cleaning system can do is to generate a (hopefully small) set of clean candidates for each dirty tuple. When the cleaning system is required to output a deterministic database, it is forced to pick one clean candidate (say the "most likely" candidate) per tuple. Such an approach can lead to loss of information. For example, consider a situation where there are three equally likely clean candidates of a dirty tuple. An appealing alternative that avoids such an information loss is to abandon the requirement that the output database be deterministic. In other words, even though the input (dirty) database is deterministic, I allow the reconstructed database to be probabilistic. Although such an approach does avoid the information loss, it also brings forth several challenges. For example, how many alternatives should be kept per tuple in the reconstructed database? Maintaining too many alternatives increases the size of the reconstructed database, and hence the query processing time. Second, while processing queries on the probabilistic database may well increase recall, how would they affect the precision of the query processing? In this thesis, I investigate these questions. My investigation is done in the context of a data cleaning system called BayesWipe that has the capability of producing multiple clean candidates per each dirty tuple, along with the probability that they are the correct cleaned version. I represent these alternatives as tuples in a tuple disjoint probabilistic database, and use the Mystiq system to process queries on it. This probabilistic reconstruction (called BayesWipe–PDB) is compared to a deterministic reconstruction (called BayesWipe–DET)—where the most likely clean candidate for each tuple is chosen, and the rest of the alternatives discarded.
ContributorsRihan, Preet Inder Singh (Author) / Kambhampati, Subbarao (Thesis advisor) / Liu, Huan (Committee member) / Davulcu, Hasan (Committee member) / Arizona State University (Publisher)
Created2013
Description
As the size and scope of valuable datasets has exploded across many industries and fields of research in recent years, an increasingly diverse audience has sought out effective tools for their large-scale data analytics needs. Over this period, machine learning researchers have also been very prolific in designing improved algorithms which are capable of finding the hidden structure within these datasets. As consumers of popular Big Data frameworks have sought to apply and benefit from these improved learning algorithms, the problems encountered with the frameworks have motivated a new generation of Big Data tools to address the shortcomings of the previous generation. One important example of this is the improved performance in the newer tools with the large class of machine learning algorithms which are highly iterative in nature. In this thesis project, I set about to implement a low-rank matrix completion algorithm (as an example of a highly iterative algorithm) within a popular Big Data framework, and to evaluate its performance processing the Netflix Prize dataset. I begin by describing several approaches which I attempted, but which did not perform adequately. These include an implementation of the Singular Value Thresholding (SVT) algorithm within the Apache Mahout framework, which runs on top of the Apache Hadoop MapReduce engine. I then describe an approach which uses the Divide-Factor-Combine (DFC) algorithmic framework to parallelize the state-of-the-art low-rank completion algorithm Orthogoal Rank-One Matrix Pursuit (OR1MP) within the Apache Spark engine. I describe the results of a series of tests running this implementation with the Netflix dataset on clusters of various sizes, with various degrees of parallelism. For these experiments, I utilized the Amazon Elastic Compute Cloud (EC2) web service. In the final analysis, I conclude that the Spark DFC + OR1MP implementation does indeed produce competitive results, in both accuracy and performance. In particular, the Spark implementation performs nearly as well as the MATLAB implementation of OR1MP without any parallelism, and improves performance to a significant degree as the parallelism increases. In addition, the experience demonstrates how Spark's flexible programming model makes it straightforward to implement this parallel and iterative machine learning algorithm.
ContributorsKrouse, Brian (Author) / Ye, Jieping (Thesis advisor) / Liu, Huan (Committee member) / Davulcu, Hasan (Committee member) / Arizona State University (Publisher)
Created2014
Description
Recognition memory was investigated for naturalistic dynamic scenes. Although visual recognition for static objects and scenes has been investigated previously and found to be extremely robust in terms of fidelity and retention, visual recognition for dynamic scenes has received much less attention. In four experiments, participants view a number of clips from novel films and are then tasked to complete a recognition test containing frames from the previously viewed films and difficult foil frames. Recognition performance is good when foils are taken from other parts of the same film (Experiment 1), but degrades greatly when foils are taken from unseen gaps from within the viewed footage (Experiments 3 and 4). Removing all non-target frames had a serious effect on recognition performance (Experiment 2). Across all experiments, presenting the films as a random series of clips seemed to have no effect on recognition performance. Patterns of accuracy and response latency in Experiments 3 and 4 appear to be a result of a serial-search process. It is concluded that visual representations of dynamic scenes may be stored as units of events, and participant's old
ew judgments of individual frames were better characterized by a cued-recall paradigm than traditional recognition judgments.
ew judgments of individual frames were better characterized by a cued-recall paradigm than traditional recognition judgments.
ContributorsFerguson, Ryan (Author) / Homa, Donald (Thesis advisor) / Goldinger, Stephen (Committee member) / Glenberg, Arthur (Committee member) / Brewer, Gene (Committee member) / Arizona State University (Publisher)
Created2014
Description
Categories are often defined by rules regarding their features. These rules may be intensely complex yet, despite the complexity of these rules, we are often able to learn them with sufficient practice. A possible explanation for how we arrive at consistent category judgments despite these difficulties would be that we may define these complex categories such as chairs, tables, or stairs by understanding the simpler rules defined by potential interactions with these objects. This concept, called grounding, allows for the learning and transfer of complex categorization rules if said rules are capable of being expressed in a more simple fashion by virtue of meaningful physical interactions. The present experiment tested this hypothesis by having participants engage in either a Rule Based (RB) or Information Integration (II) categorization task with instructions to engage with the stimuli in either a non-interactive or interactive fashion. If participants were capable of grounding the categories, which were defined in the II task with a complex visual rule, to a simpler interactive rule, then participants with interactive instructions should outperform participants with non-interactive instructions. Results indicated that physical interaction with stimuli had a marginally beneficial effect on category learning, but this effect seemed most prevalent in participants were engaged in an II task.
ContributorsCrawford, Thomas (Author) / Homa, Donald (Thesis advisor) / Glenberg, Arthur (Committee member) / McBeath, Michael (Committee member) / Brewer, Gene (Committee member) / Arizona State University (Publisher)
Created2014