Shlomit Gur

PUBLICATIONS

PATENet: Pairwise Alignment of Time Evolving Networks

Gur S., and Honava V.J.
in Machine Learning and Data Mining in Pattern Recognition [MLDM 2018]
Key words: Network science, Multilayer networks, Temporal alignment

Networks that change over time, e.g. functional brain networks that change their structure due to processes such as development or aging, are naturally modeled by time-evolving networks. In this paper we present PATENet, a novel method for aligning time-evolving networks. PATENet offers a mathematically-sound approach to aligning time evolving networks. PATENet leverages existing similarity measures for networks with fixed topologies to define well-behaved similarity measures for time evolving networks. We empirically explore the behavior of PATENet through synthetic time evolving networks under a variety of conditions.
Age-Related Differences in Task-Specific Functional Connectivity in the Context of Phonological and Semantic Cognitive Tasks with Distracting Words

Gur S., El-Manzalawy Y., Diaz M.T., and Honava V.J.
[In Preparation]
Key words: functional connectivity, aging, semantics, phonology, machine learning, attention

Language-related cognitive processes are affected differently by aging: phonological processes display deficits, whereas semantic processes do not. The inhibition deficit hypothesis of aging suggests that older adults find it more difficult to ignore irrelevant information compared to younger adults. We investigated the interactions between these processes with distracting words in younger and older adults. Specifically, we used task-specific functional connectivity from phonological and semantic conditions with distracting words to train machine learning classifiers and identify graph topological features that reliably discriminate between the conditions in younger and older adults, separately and combined.
 In our analysis, measures of strong links in the functional connectivity were especially effective (as a group) in condition discrimination. Our results revealed greater similarity in functional connectivity between the two conditions in older adults as compared to younger adults, a result that is consistent with the dedifferentiation hypothesis. Analysis of the identified features showed that while some features’ brain regions were age-independent (left thalamus and caudate nucleus), others were largely age-dependent. Our results suggest that condition-related difference in older adults may be explained in part by differences in connectivity of brain regions that have been implicated in intrinsic alertness, a prerequisite for selective attention to aspects of the tasks.
Our results also demonstrate the potential of data-driven machine learning algorithms in identifying topological features of functional connectivity between conditions, and in finding evidence to support, refute, or generate novel hypotheses regarding differences in functional connectivity or activation of different brain regions in relation to task and age.
Improving the Reliability of Sentiment Prediction in Online Health Communities

El-Manzalawy Y., Lee S., Gur S., Le T., Bui N., Yen J., and Honava V.J.
[In Preparation]
Key words: cancer survivors, sentiment classification, machine learning, mining online health communities, online discussion boards

Background: Online health communities (OHCs) offer a powerful internet-based platform for providing social support to participants through the sharing of knowledge, experience, and feelings with other participants. Understanding how online social interactions impact individuals’ health require effective methods for analyzing usergenerated content.
Objective: The major goals of this research are to: i) Develop a methodology for effective application of machine learning algorithms to OHC data: ii) Demonstrate the viability of the methodology via a case study for developing improved models for classification of sentiment expressed in user-generated content in Cancer Survivors Network (CSN), an online community of cancer survivors that is operated by the American Cancer Society.
Methods: We introduce an approach for extracting representative subset of data based on randomly sampled threads (as opposed to randomly sampled posts used in previous related studies). Four annotators assign labels to each post, considering the context of the post in the thread in which it appears. We partition the labeled data into crossvalidation and independent test sets, and run a competition between four data scientists to obtain the best sentiment classifier using cross-validation data. Classifiers submitted by participants are evaluated using the independent test data.
Results: Comparison of our competition winning classifier with the original study by Qui et al., using their dataset of 298 posts shows that our best classifier achieves an AUC of 0.92 and an accuracy of 84.6% as compared to an AUC of 0.83 and an accuracy of 79.2% reported in the original study.