Pirc net poster
1 like248 views
This research aims to characterize HIV at-risk populations among men who have sex with men (MSM) in San Diego by analyzing social media data. The researchers collect tweets from San Diego and classify them based on risk categories like drug use, sex venues, etc. They build a social network graph of Twitter users and their connections and compare the structure to the real-world HIV transmission network. Exploratory analysis of the social graph reveals patterns in topics of discussion and network structures that can help predict HIV transmission risk and enable prevention efforts. Future work includes further data collection, interactive visualizations, and computational models to understand how the social network evolves and relates to the sexual network transmitting HIV.
Pirc net poster
- 1. Analyzing social media to characterize HIV at-risk populations among MSM in San Diego Narendran Thangarajan1, Dr. Nella Green3, Dr. Amarnath Gupta2, Dr. Susan Little3, Dr. Nadir Weibel1 Digital Epidemiology This research is funded by Frontier of Innovative Scholars Program, UCSD and Center for AIDS Research, UCSD 1 Department of CSE, UC San Diego, 2 San Diego Supercomputer Center, 3 School of Medicine, UC San Diego naren@ucsd.edu 35 MILLION people with AIDS worldwide. 1.2 MILLION people with AIDS in US. 660,000 total deaths caused by AIDS in US. 78% of the new infections in 2010 were MSM. California (along with Florida) had the highest number of HIV diagnoses in 2013. Interesting recent trend - Proliferation of social networks and real-time communication capabilities. FISP CFAR + = “Just treated a HIV infected person from location X.We should probably conduct a PrEP intervention at X.” “We should deploy peer education in locationY, most of our patients are from there.” Ineffective prevention strategies: 50,000 new HIV infections each year.Problem Characterize and identify HIV at-risk MSM populations by studying user sentiments and behaviors on social networks. 2015 2012 Salathé et. al. published “Digital Epidemiology” in PLoS Computational Biology Journal Solution 2014 Ginsberg et. al. published “Detecting influenza epidemics using search engine query data” in Nature journal. 2008 Methods of using real-time social media technologies for detection and remote monitoring of HIV outcomes - Sean D. Young et. al., Elsevier Preventive Medicine, 2014. Unraveling Abstinence and Relapse: Smoking Cessation Reflected in Social Media - Dr. Elizabeth Murnane, CHI 2014. 1. Data collection, classification and refinementMethod • Tweets are collected in real-time through theTwitter Streaming API. Twitter’s “filter hose” is used to collect tweets from San Diego county. • Each tweet is cleaned by removing stop words, punctuations and converting to lower case. III. Migration from raw twitter data to social network graph II. Improving the accuracy of HIV risk tweets classification using machine learning To improve the accuracy of HIV risk tweets classified, we evaluated two linear classifiers - SupportVector Machines (SVM) and Logistic Regression with different sets of features. Feature Set SVM Logistic Regression Bag of Words 15.73% 15.72% Stop Word Removal 12.9% 12.98% Domain Specific Terms 11.37% 7.42% Tweeter information 17.12% 15.23% Error rates using different linear classifiers • The property graph model was adopted as the data model for HIV at- risk MSM twitter social network. • 7 node types and 9 edge types were identified as shown. • Ontologies (shown in green) are used to infer indirect relationships between entities. For instance, it allows us to query for users who post tweets related to meth and sex venues. • The resulting graph was materialized in a graph database called Neo4J. Results obtained using EDA queriesAnalysis Exploratory Data Analysis queries helped understand the hidden patterns in the HIV at-risk social network. Querying the social graph to identify interesting communication structuresResults Currently, we have a query-able HIV at-risk twitter network graph. Proximity: How close are drug bucket users to other homosexual bucket users in terms of hop count? Topics of interest: What are the main topics in the discussions among people who are at a one-hop following distance from their sub- graph’s hubs? Conversations: How many conversations are happening among drug bucket users alone , sex bucket users alone and across drug bucket users and sex bucket users?” Preferences: Identify two drug bucket users who are most consulted by homosexual people. Current status and future worksFuture (0) Drug (1) Homosexual (2) STI (3) Sex (4) SexVenues The HIV at-risk MSM social network coupled with the real-world HIV transmission network inferred using phylodynamics from SD PIC will help us understand if the actual sexual network can be reconstructed using the social network. Ultimately, this social network could predict an individual’s future HIV transmission risk enabling us to prevent it in real-time. • Each tweet is classified as a HIV risk tweet if it falls in one of the five HIV risk categories - Drug, SexVenues, Sex, Homosexual, SexuallyTransmitted Infections. • Classified tweets are refined further using exclusion and inclusion lists of co-occurring words. e.g.“ice cold” doesn't refer to meth (a drug commonly called “ice") • After getting a refined set of HIV risk tweets, the relevant metadata (like tweeters and the mentioned users) were fetched usingTwitter’s public APIs. • Retweet and reply chains were pulled in recursively to ensure the original tweet and the corresponding tweeter were part of the resulting social network graph. Most active time of the day Most active day of the week Power-law distribution of tweets Length of HIV risk tweets Tweets distribution across risk buckets Most co-occurring risk categories • IRB approval and recruitment - Currently, we are collecting twitter handles of people in the HIV transmission network and those at risk of acquiring HIV. This enables us to compare the structural similarities in the sexual network and the twitter social network. • Interactive data visualizations to enable visualizing the evolving HIV at-risk social network to decipher underlying patterns in network structure evolution and the corresponding changes in SNA metrics. • Computational model that captures the behavior of a HIV at- risk user onTwitter. Social Network Sexual Network • Collaboration with Harvard to identify change-points in the social network structure.