You can find my most up-to-date publication list at my Google Scholar profile.
preprint, 2023
In our study on Twitter, we analyze how users change their behavior after adding social identity signals to their profiles, revealing they tend to tweet more in line with their identity and connect with users sharing similar identities. Additionally, we found that openly disclosing identity through tweets and profiles correlates with receiving fewer offensive replies.
Download here
The International AAAI Conference on Web and Social Media (ICWSM), 2023
We analyze Twitter data to study how responses to online shocks vary based on user relationships, using a dataset of over 13K self-reported shock events. Our findings highlight distinct patterns in response levels and topic shifts across 110K replies, showing that relationship dynamics significantly influence shock responsiveness online.
Download here
Empirical Methods in Natural Language Processing (EMNLP), 2023
We curate a benchmark for NLP tasks on their ability to understand tasks requiring social knowledge and contextual appropriateness, which we name SocKET. We perform analyses on popular open-source LLMs and show that these tasks remain a challenge at its current state.
Download here
The International AAAI Conference on Web and Social Media (ICWSM), 2021
We analyzed over 9.6M dyads of Twitter users to explore how relationship types affect language use, topic diversity, communication frequencies, and diurnal conversation patterns. Our findings demonstrate that incorporating relationship types enhances predictive models for user relationships and future retweet behavior, indicating their potential to offer fresh insights into communication dynamics and information diffusion in social networks.
Download here
Proceedings of The Web Conference (WWW), 2020
In this paper, based on 10 different social dimensions built on several literature, we collected crowdsourced data and trained classifiers for extracting social dimensions from textual conversation data. We used our models to examine the existence of different dimensions in various datasets of text conversations, and how they relate to actual community-wide or societal outcomes.
Download here
IEEE Transactions on Visualization and Computer Graphics (TVCG), 2019
Electronic medical records (EMRs) are increasingly being coupled with deep learning models for predicting the future condition of a patient. In this study, we introduce RetainVis, a visual analytics tool that can be used for interpreting how each individual risk code in EMR data can contribute to the predicted outcomes such as heart failure or cataract symptoms. We provide both quantitative results of the model performance as well as qualitative use cases.
Download here
IEEE/ CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2018
Existing approaches on image-to-image translation, while capable of producing high-quality translation images, were limited to one-on-one translations and also less robust on multiple domains. To address this limitation, we introduced StarGAN, a model that uses a simple yet effective approach, one-hot masking, and show through facial image translation tasks that our model is capable of performing multiple forms of image translations.
Download here
Journal of Biomedical Informatics (JBI), 2016
Online health communities (OHCs) are utilized by researchers and practitioners to influence health behavior and offer social support, yet they encounter attrition issues akin to other social media platforms. To address this, personas were developed based on qualitative interviews and survey responses from 184 OHC users, resulting in four personas—Caretakers, Opportunists, Scientists, and Adventurers—informing tailored support strategies to enhance user engagement and satisfaction within OHCs.
Download here