Proactive Edge Caching with Heterogeneous User Demands

28 concepts17 papers1 proposal

In traditional content delivery networks, caching strategies rely heavily on a simple mathematical convenience: the law of large numbers. When a core network router handles millions of requests, individual quirks smooth out into a highly predictable, global popularity distribution (often modeled as a Zipf distribution). Caching the most globally popular items yields a high cache hit rate.

However, as caching moves to the network edge—into small cell base stations, roadside units, and local routers—this statistical aggregation vanishes. An edge node might only serve a handful of users at any given time. If those specific users have niche interests that deviate from global trends, a cache populated with globally popular content becomes highly inefficient.

To solve this, a major shift in research rejects the traditional homogeneous user model (which treats all users as sharing identical demand distributions) in favor of caching frameworks that explicitly model heterogeneous user preferences. By understanding why and how local demands fluctuate, these modern edge caching strategies drastically reduce network latency and backhaul load.


The Failure of Aggregation at the Edge

To understand why a homogeneous user assumption fails, consider how standard cache eviction policies like Least Recently Used (LRU) operate. As Guocong Quan et al. observe in Optimal Edge Caching For Individualized Demand Dynamics, LRU and its variants are optimized for the stable, aggregated request streams of core networks. At the network edge, however, caches serve individual users whose interests shift rapidly and idiosyncratically.

When an edge node relies solely on aggregate popularity metrics, it suffers from a severe content popularity mismatch. As highlighted by Haoye Lu et al. in A GNN-based Proactive Caching Strategy in NDN Networks, ignoring localized community preferences causes edge nodes to waste storage on irrelevant content while repeatedly missing requests for specialized local content. This insight is echoed by Xing Chen et al. in Hit Ratio Driven Mobile Edge Caching Scheme for Video on Demand Services, who demonstrate that incorporating historical, individual user preferences into Video-on-Demand caching significantly outperforms traditional, blind edge storage.


Scaling Personalization via Content-Aware Clustering

While tracking the exact preference profile of every single user maximizes accuracy, it introduces massive computational overhead and privacy concerns. To make personalized edge caching practical, researchers group users into distinct behavioral archetypes using content-aware clustering.

Instead of clustering users by their physical location, systems cluster them by what they consume.

[User Requests] ──> [Spectral Clustering / AIC] ──> [User Clusters (Archetypes)]
                                                            │
                                                            ▼
                                                [Tailored Proactive Cache]

Mohammed S. Elbamby et al. pioneered this approach in Content-Aware User Clustering and Caching in Wireless Small Cell Networks. By pairing spectral clustering with a regret-minimization learning algorithm, small cell base stations can predictively prefetch content tailored to the specific sub-groups active within their footprint, cutting service delays by up to 42%.

A key challenge in clustering is determining the right granularity: too few clusters replicate the flaws of the homogeneous model, while too many clusters overfit and waste resources. To solve this, Salah Eddine Hajri developed an adaptive strategy in Caching Improvement Using Adaptive User Clustering that uses the Akaike Information Criterion (AIC) to dynamically calculate the optimal number of user groups in real time. In a follow-up work, Salah Eddine Hajri demonstrated in Energy Efficiency in Cache-Enabled Small Cell Networks With Adaptive User Clustering that this preference-based grouping drastically improves the network's overall energy efficiency by maximizing the cache hit probability per transmission.


Restoring Mathematical Symmetry to Coded Caching

The transition to heterogeneous demands does not just change what files are stored; it fundamentally alters the mathematics of how data is transmitted. In coded caching, a single multicast transmission can simultaneously satisfy different user requests by mathematically combining (XOR-ing) files that users have partially cached.

The foundational coded caching framework (the Maddah-Ali and Niesen, or MN scheme) relies on perfect symmetry: it assumes all users share identical file preferences and cache sizes. When users have heterogeneous demands, this symmetry breaks, rendering standard multicast coding combinations impossible.

To fix this mathematical bottleneck, Ciyuan Zhang et al. proposed new coding variations in Coded Caching With Heterogeneous User Profiles. Their schemes adapt the MN framework by systematically categorizing files into "common" and "unique" sets across different user classes, restoring the structural alignment needed to execute coded multicast delivery.

Traditional MN Scheme (Symmetric Demands):
[User 1 wants A] ──┐
                   ├──> Multicast [A ⊕ B] (Perfect alignment)
[User 2 wants B] ──┘

Heterogeneous Demands (Asymmetric):
[User 1 wants Niche X] ──┐
                         ├──> Symmetry breaks; cannot easily code together.
[User 2 wants Niche Y] ──┘

Strategy β (Uniform Sub-packetization):
[File X (Niche)] ──> Partitioned into uniform sub-packets ──┐
                                                            ├──> Restores algebraic coding structure
[File Y (Niche)] ──> Partitioned into uniform sub-packets ──┘

Similarly, Pierre Quinton et al. introduced Strategy $\beta$ in A Novel Centralized Strategy for Coded Caching with Non-uniform Demands. Strategy $\beta$ enforces a uniform "sub-packetization" structure across all files regardless of their individual popularity or how much of them is cached. By maintaining this rigid sub-packet architecture, the network can still find algebraic coding opportunities across entirely different popularity groups, bypassing a long-standing technical barrier in information theory.


Multidimensional Heterogeneity: Locality, Activity, and Urgency

True user heterogeneity extends beyond content preferences. Users also differ in their physical movement patterns, network activity levels, and situational urgency.

Dong Liu et al. formalized this multidimensionality in Caching at Base Stations With Heterogeneous User Demands and Spatial Locality. They noted that a few hyper-active users can easily dominate an edge cache, forcing out content needed by less active users. By formulating an optimization problem that balances individual preferences, spatial locality (where users request content), and varying activity levels, the authors proved in When Exploiting Individual User Preference Is Beneficial for Caching at Base Stations that accounting for these differences drastically improves both average download delays and network fairness.

This philosophy is increasingly vital in highly dynamic edge environments like the Internet of Vehicles (IoV):

Ultimately, whether routing video packets or allocating physical charging slots, moving away from the uniform user assumption allows edge networks to transition from reactive, congested pipelines into proactive, context-aware delivery systems.

Key Foundational Work

Content-aware user clustering and caching in wireless small cell networks (2014) — contributed 1 concept elements to this direction.
Energy efficiency of on-demand video caching systems and user behavior (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_homogeneous_user_model_79173af0,
  title = {homogeneous user model},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/homogeneous-user-model/rdr_79173af0fc1f4a2d99a93f14f8f9d42b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_homogeneous_user_model_79173af0_v1,
  title = {homogeneous user model},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/homogeneous-user-model/rdr_79173af0fc1f4a2d99a93f14f8f9d42b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake