Chances are you’ve watched a short video online. That dunk in the basketball game, a tour POV video from your favorite singer, an island filled with cats in Japan, or instructions on changing the string cartridge on your lawn trimmer – are all examples of a unique form of communication. When you watch one, you find yourself directed to more.
Siebel School of Computing and Data Science in The Grainger College of Engineering at the University of Illinois Urbana-Champaign computer science professor Indranil Gupta and assistant professor Deepak Vasisht say that “this delivery of a sequence of short videos, driven by a highly-personalized recommendation algorithm, is pretty unique.”
Their team has received an NSF grant to develop new video delivery techniques. A $1.1M NSF Medium project was awarded for collaborative research: CSR: Medium: LANDS – Learning-Based Adaptive Networked Systems for Delivery of Short Videos, aimed at transforming video delivery techniques, optimizing compute and network resource use, and reducing resource consumption on user devices, while offering targeted educational course modules focusing on short-video streaming.
Gupta is the lead principal investigator, joined by co-principal investigators Vasisht and Deepti Raghavan, an assistant professor at Brown University. As the Illinois duo describes it, “Content Distribution Networks (CDNs) have been around for decades, and they serve different classes of traffic, from web to multimedia to Internet of Things traffic classes. Short videos are a new, quickly growing class of traffic that CDNs have to increasingly contend with.”
They add, “Traditional long form video includes server-based streaming (think Netflix), peer-to-peer traffic (think multi-player games), and recent AR/VR/XR workloads. Short form videos include TikTok, YouTube Shorts, and Instagram Reels. Their popularity has even forced social media platforms like X and Facebook to include short video offerings.”
“The last few years have seen an explosion of short video delivery systems—TikTok, YouTube Shorts, Instagram Reels, Triller, and regional variants like Moj (India), Douyin (China), etc. Already useful well beyond entertainment media, these short video systems are increasingly beneficial to society, for microlearning (e.g., quizzes, school tutorials, how-to videos), citizen reporting, advertising, user-generated content, testimonials, sports, etc. In 2021, TikTok overtook Google as the world’s most popular website, and has remained popular ever since, with 1 billion monthly active users in 2023, of whom 170M are in the US. In 2024, TikTok was already responsible for 12% (and growing) of internet traffic, with the average TikTok user consuming 173MB TikTok data per day over cellular networks and 1.1 GB TikTok data per day overall.
The key difference is that the algorithm is deciding what you watch, unlike, say, YouTube, where the algorithm merely recommends a set of videos to choose from. In other words, short video systems “push” a (personalized) sequence of videos to each user, rather than merely recommending choices for the user to choose from.”
Photo Credit: LANDS project team
LANDS project proposal showing a short video description.
Gupta and Vasisht note that “we researchers always want to be working at the cutting edge, looking at new systems that are quickly gaining in popularity. Short video systems are one of these. It is an area that is relatively unexplored, and in which papers have just started to be published by other researchers. We at Illinois want to be part of that vanguard of seminal research in that area. It is heartening that federal funding agencies recognize the impact that these short video systems are having on Americans, and the yearning for the community to build better networked systems that cater to this new paradigm.”
Gupta and Vasisht have been working with PhD student Maleeha Masood, who is the lead student on the project. Masood is a big TikTok fan and a big Taylor Swift fan, who is popular on the platform. They say that “Students often strongly influence faculty decisions to move into new research areas!”
In conversation with Deepti Raghavan at Brown University, Gupta and Vasisht discovered that her skills and expertise were complementary matches for some aspects of the proposal for which they were looking for help.
The project will have multifaceted contributions including designing new CDN caching techniques that do not affect user’s perceived experience, using Large Language Models (LLMs) to predict user behavior, and building ontologies for short video systems. Rather uniquely, the project team will combine systems research with human use research, which Gupta and Vasisht call “the most exciting part of our work. Today, a major issue that networked systems designers face is that they are unable to measure impacts of systems/networking design decisions on real human users, without resorting to user studies which can be expensive and time-consuming. These user studies can slow down the design cycle. Papers in systems/networking conferences typically do not include any user studies. Nor do papers in HCI (Human Computer Interaction) conferences study the effects of new systems/network design ideas. Our work falls exactly in that overlap area—such inter-disciplinary work is much-needed in today’s era of computing, where increasingly integrated systems are used by humans. So the way we build these systems has to be holistic rather than piecemeal.”
Grainger Engineering Affiliations
Indranil Gupta is an Illinois Grainger Engineering professor of computer science.
Deepak Vasisht is an Illinois Grainger Engineering assistant professor of computer science.