AI predicts what users want to watch, stores it on nearby devices for speed

What is it about?

Imagine you're watching a popular video on your phone while commuting. That video likely traveled all the way from a distant central server through the mobile network, consuming bandwidth and causing delays for everyone. But what if your phone could grab that video from a nearby friend's phone instead, instantly and without burdening the network? That's the idea behind our work. We developed an intelligent system that predicts what videos, movies, or content users will want to watch next—using machine learning—and then stores those popular files on the devices of willing users who are physically close to each other. When someone requests a file, their device first checks nearby "helper" devices before reaching out to the central network. The challenge is that users have very different tastes, storage space is limited, and people are constantly moving around. Our system solves this by grouping nearby users into clusters, learning each group's unique preferences, and strategically placing content where it's most likely to be needed—while respecting each user's willingness to participate and share their storage. In our simulations, this approach achieved a nearly 97% cache utilization rate, meaning almost all available storage space was used effectively, significantly reducing network congestion and delivering content faster. The result is smoother streaming, lower data costs, and more efficient use of mobile networks—making our digital world faster and greener for everyone.

Featured Image

Photo by Vizag Explore on Unsplash

Photo by Vizag Explore on Unsplash

Why is it important?

What makes this work unique and timely? Today's mobile networks are drowning under an avalanche of video content—from TikTok clips to 4K movies—and the problem is only getting worse. By 2025, video will account for over 80% of all mobile traffic, putting enormous pressure on network backhauls and causing frustrating buffering delays. While caching content on user devices and sharing it via Device-to-Device (D2D) communication is a promising solution, existing approaches suffer from a critical flaw: they assume all users have identical tastes, which is simply not true. In reality, less than 20% of users generate 80% of multimedia traffic, and preferences vary dramatically across users and locations. Our work introduces ML-CPCO, the first framework that simultaneously tackles three previously overlooked challenges: User Heterogeneity: Instead of assuming uniform preferences, our system uses XGBoost machine learning to predict individual user tastes based on their age, occupation, content preferences, and viewing history. User Willingness: Most caching schemes force content onto users' devices. Our framework explicitly incorporates whether users are willing to participate in caching—a critical real-world consideration that existing research has largely ignored. Cluster-Level Intelligence: By grouping nearby users into clusters and predicting content popularity at the cluster level, we ensure that what's stored in each cluster actually reflects the unique preferences of its members. This approach is exceptionally timely for three reasons: The 5G/6G rollout enables unprecedented D2D communication speeds and low latency, making device-to-device content sharing finally practical at scale. Mobile video consumption has exploded, and network operators desperately need offloading solutions that don't require massive infrastructure investments. Privacy-conscious caching is becoming essential, as storing content locally on user devices (rather than tracking individual viewing habits centrally) offers a more privacy-preserving alternative. The difference it makes: Our framework achieves a remarkable 97% cache utilization rate—the highest among all compared methods—significantly outperforming traditional strategies like LFU, FIFO, and popularity-based caching. This translates directly to: Less buffering and faster content delivery for end users Reduced backhaul traffic for network operators Lower energy consumption as content is delivered locally rather than from distant servers Smarter use of device storage—only content that's actually wanted is stored, and only on devices whose owners choose to participate Perhaps most importantly, our work demonstrates that high prediction accuracy alone doesn't guarantee good caching performance. What matters is integrating accurate predictions with smart, network-aware placement strategies—a key insight that challenges conventional thinking in the field. This makes our framework not just theoretically sound, but practically deployable in real-world, heterogeneous, and dynamic network environments.

Perspectives