skip to main content

Search for: All records

Creators/Authors contains: "Zhou, L."

Note: When clicking on a Digital Object Identifier (DOI) number, you will be taken to an external site maintained by the publisher. Some full text articles may not yet be available without a charge during the embargo (administrative interval).
What is a DOI Number?

Some links on this page may take you to non-federal websites. Their policies may differ from this site.

  1. This paper studies the fundamental problem of learning deep generative models that consist of multiple layers of latent variables organized in top-down architectures. Such models have high expressivity and allow for learning hierarchical representations. Learning such a generative model requires inferring the latent variables for each training example based on the posterior distribution of these latent variables. The inference typically requires Markov chain Monte Caro (MCMC) that can be time consuming. In this paper, we propose to use noise initialized non-persistent short run MCMC, such as nite step Langevin dynamics initialized from the prior distribution of the latent variables, asmore »an approximate inference engine, where the step size of the Langevin dynamics is variationally optimized by minimizing the Kullback-Leibler divergence between the distribution produced by the short run MCMC and the posterior distribution. Our experiments show that the proposed method outperforms variational auto-encoder (VAE) in terms of reconstruction error and synthesis quality. The advantage of the proposed method is that it is simple and automatic without the need to design an inference model.« less
  2. Despite that tremendous progress has been made in mobile user authentication (MUA) in recent years, continuous mobile user authentication (CMUA), in which authentication is performed continuously after initial login, remains under studied. In addition, although one-handed interaction with a mobile device becomes increasingly common, one-handed CMUA has never been investigated in the literature. There is a lack of investigation of the CMUA performance between one-handed and two-handed interactions. To fill the literature gap, we developed a new CMUA method based on touch dynamics of thumb scrolling on the touchscreen of a mobile device. We developed a mobile app of themore »proposed CMUA method and evaluated its effectiveness with data collected from a user study. The findings have implications for the design of effective CMUA using touch dynamics and for improvement of accessibility and usability of MUA mechanisms.« less
  3. Social media has become a powerful and efficient platform for information diffusion. The increasing pervasiveness of social media use, however, has brought about the problems of fraudulent accounts that are intended to diffuse misinformation or malicious contents. Twitter recently released comprehensive archives of fraudulent tweets that are possibly connected to a propaganda effort of Internet Research Agency (IRA) on the 2016 U.S. presidential election. To understand information diffusion in fraudulent networks, we analyze structural properties of the IRA retweet network, and develop deep neural network models to detect fraudulent tweets. The structure analysis reveals key characteristics of the fraudulent network.more »The experiment results demonstrate the superior performance of the deep learning technique to a traditional classification method in detecting fraudulent tweets. The findings have potential implications for curbing online misinformation.« less
  4. Online consumer reviews contain rich yet implicit information regarding consumers’ preferences for specific aspects of products/services. Extracting aspects from online consumer reviews has been recognized as a valuable step in performing fine-grained analytical tasks (e.g. aspect-based sentiment analysis). Extant approaches to aspect extraction are dominated by discrete models. Despite explosive research interests in continuous-space language models in recent years, these models have yet to be explored for the task of extracting product/service aspects from online consumer reviews. In addition, previous continuous-space models for information extraction have largely overlooked the role of semantic information embedded in texts. In this study, wemore »propose an approach of aspect extraction that leverages semantic information from WordNet in conjunction of building continuous-space language models from review texts. The experiment results with online restaurant reviews demonstrate that the WordNet-guided continuous-space language models outperform both discrete models and continuous-space language models without incorporating the semantic information. The research findings have important implications for understanding consumer preferences and improving business performances.« less
  5. Craigslist is a popular online customer-to-customer marketplace, which has attracted millions of consumers for trading and purchasing secondhand items. Because of the high financial return that sellers could gain from using this site and the anonymity option that the website provides to its users, Craigslist is highly subject to fraudulent activities. The primary objective of this study is to detect scam ads on Craigslist. Based on the related literature and our observations of ads collected from the platform, we develop a heuristic method for identifying scam ads. We evaluate the proposed heuristics by conducting an experiment and performing additional datamore »analyses using real data. The results provide preliminary evidence for efficacy of the heuristics developed in this study.« less
  6. Despite the tremendous role of online consumer reviews (OCRs) in facilitating consumer purchase decision making, the potential inconsistency between product ratings and review content could cause the uncertainty and confusions of prospect consumers toward a product. This research is aimed to investigate such inconsistency so as to better assist potential consumers with making purchase decisions. First, this study extracted a reviewer’s sentiments from review text via sentiment analysis. Then, it examined the correlation and inconsistency between product ratings and review sentiments via Pearson correlation coefficients (PCC) and box plots. Next, we compared such inconsistency patterns between fake and authentic reviews.more »Based on an analysis of 24,539 Yelp reviews, we find that although the ratings and sentiments are highly correlated, the inconsistency between the two is more salient in fake reviews than in authentic reviews. The comparison also reveals different inconsistency patterns between the two types of reviews.« less