Read on Twitter
I'm excited to share a project long in the making. We designed algorithms to find influential nodes in social networks, applied to HIV prevention for homeless youth. A trial with 713 youth over 2 years showed significant benefits. Paper just posted, https://bit.ly/3iWbMbR (1/9)
Homeless youth have up to 10x HIV prevalence vs general population. One intervention is to recruit peer leaders from the youth to promote protective behaviors. But how to choose the most influential peer leaders? (2/9)
There's tons of computer science work on finding influential nodes in a social network ("influence maximization"). But, mostly targeted at advertising/online social networks...not easily applicable to community health. (3/9)
What are the new challenges? In a word, data. Who's connected to who? How will information diffuse? None of this is known. Gathering network structure = time consuming, face to face interviews with youth. (4/9)
We developed algorithms to efficiently subsample the network, only requiring about 20% of the effort in data collection. Then, we designed a robust optimization algorithm to identify influential nodes even under uncertainty. (5/9)
It worked in simulation but what about reality? We ran a clinical trial at centers for homeless youth in LA. Trial compared three arms: interventions with our algorithm, selecting highest-degree youth (standard baseline), and no intervention. 713 youth total over 2 years. (6/9)
The results just out: in the algorithm arm, statistically significant reduction in key outcome, condomless anal sex (OR = 0.69). No significant change for the other arms. AI helped! (7/9)
Key takeaways in the paper ( https://arxiv.org/pdf/2009.09559.pdf): simple, robust, data-efficient algorithms are critical for public health domains. Beyond the algorithm though, always requires community trust. (8/9)
It was truly amazing to work with this close-knit team of social work/AI researchers: @MilindTambe_AI, @EricRicePhD, @onasch_vera, Graham Diguiseppe, @AmulyaYadav19 and many more at @CAIS_USC and @HCRCS. (9/9)
