Cache‐Assisted Offloading Optimization for Edge Computing Tasks

Abstract

ABSTRACT

Mobile edge computing (MEC) serves as a feasible architecture that brings computation closer to the edge, enabling rapid response to user demands. However, most research on task offloading (TO) overlooks the scenario of repetitive requests for the same computing tasks during long time slots, and the spatiotemporal disparities in user demands. To address this gap, in this paper, we first introduce edge caching into TO and then divide base stations (BSs) into different communities based on the regional characteristics of user demands and activity areas, enabling collaborative caching among BSs within the same community. Subsequently, we design a dual timescale to update task popularity within both short and long‐term time slots. To maximize cache benefits, we construct a model that transforms the caching issue into a 0–1 knapsack problem, and employ dynamic programming to obtain offloading strategies. Simulation results confirm the efficiency of the proposed task caching policy algorithm, and it effectively reduces the offloading cost and improves cache resource utilization compared to the other three baseline algorithms.In this paper, we first introduce edge caching into TO and then divide BSs into different communities based on the regional characteristics of user demands and activity areas, enabling collaborative caching among BSs within the same community. Subsequently, we design a dual timescale to update task popularity within both short and long‐term time slots. To maximize cache benefits, we construct a model that transforms the caching issue into a 0–1 knapsack problem and employ dynamic programming to obtain offloading strategies.

Introduction

With the fast-growing mobile networks, smart devices such as smartphones, wearables, and vehicular terminals are becoming increasingly ubiquitous. Along with this trend, there is a rapid advancement in new types of applications, including computer vision, voice control, and object recognition [1]. These applications pose significant challenges to mobile networks in terms of real-time and high-performance processing requirements. To ensure the quality of experience for users, the operation of these services requires greater bandwidth, processing power, and storage capacity. However, the computing, storage, and energy resources of mobile terminals are typically limited, making it difficult to meet these demands. Mobile cloud computing (MCC) has been proposed to leverage the abundant resources of the cloud to satisfy the needs of mobile devices. However, the distance between mobile devices and the cloud may result in high latency and jitter [2]. Additionally, uploading massive tasks to the core network can lead to network congestion. Considering the limitations of MCC, mobile edge computing (MEC) brings cloud service capabilities closer to the network edge, such as BSs, routers, and other locations. By offering network-side computing services near the data source, MEC can more securely and promptly respond to various business demands [3], effectively reducing latency and energy costs.

Although MEC fills the gap between the constrained capacities of mobile devices and the growing needs of apps requiring a lot of resources [4], the computing resources of MEC servers and the communication resources of wireless links are still limited compared to cloud computing. The shortfall in single-dimensional resources can be compensated for by integrating multidimensional resources [5, 6], leveraging caching resources on MEC servers and introducing edge caching in TO. Leveraging caching mechanisms to store popular tasks proximate to a user-dense environment can effectively shorten the transmission of redundant data, decrease task upload time and energy costs, and enhance user experience. As users become more dense and diverse in their requests, limiting caching decisions to a single BS area [7, 8] makes it difficult to adapt to user mobility and time-varying demands. Considering the task request situations in different regions and diversifying cache contents accordingly becomes necessary. Additionally, the storage resources of MEC servers are limited, making it impossible to cache all tasks. In such scenarios, rational resource allocation to enhance TO performance has become a critical bottleneck in MEC systems.

Related Works

In edge networks, the enhancement of mobile device performance through MEC servers primarily revolves around two aspects. First, MEC servers facilitate edge computing by bringing computation closer to users' locations, which effectively alleviates the computational resource pressure on terminals, thereby reducing energy consumption and response time [9]. However, it is obviously unreasonable to offload all tasks when there are numerous mobile and edge devices. Intense resource competition among users would be counterproductive. Therefore, the issue of TO has always been a research focus in edge computing and can be summarized as the process of offloading the right tasks to the right computing nodes at the right time, which is evidently a multi-objective optimization problem [10, 11]. Extensive research efforts have been dedicated to seeking rational task-scheduling decisions. In [12], a dynamic programming problem of single-user, multi-fog nodes for delay-sensitive tasks was explored further addressing the “one for all” strategy for multi-user, multi-fog groups to achieve a win-win situation among users. However, the aforementioned research was assumed to offload all tasks, leading to intense resource competition among users. In [13], the authors studied a partial offloading resource allocation algorithm combining tabu search and graph coloring to address the inefficiencies of full offloading in MEC-based V2X networks under long-distance communication and resource constraints, effectively reducing task delay and energy consumption. The authors of [14] developed a joint multi-task offloading and resource allocation framework for satellite IoT-based multi-task MEC systems and designed an attention-enhanced proximal policy optimization algorithm to learn optimal offloading strategies. In [15], a deep reinforcement learning (DRL)-based framework was considered to jointly optimize TO and resource allocation in 6G-enabled IIoT systems with MEC support. By introducing novel techniques such as isotone action generation and adaptive action aggregation update, the original problem was decomposed and partially transformed into a convex form. In [16], a decentralized DRL-based TO framework was proposed for multi-user MEC systems, addressing limitations of centralized methods by enabling independent decision-making without global network information, which considers non-divisible, delay-sensitive tasks and concurrent execution at edge servers, aiming to minimize long-term latency and the task drop ratio. The authors of [17] considered a robust RL-based task scheduling algorithm for MEC-enabled IoT networks to minimize worst-case power consumption under deadline constraints. By formulating the problem as a robust-return constrained MDP, the approach addresses uncertainties in task arrival rates and ensures policy robustness. These studies mentioned above have overlooked the potential benefits of edge caching in reducing redundant transmissions and computation overhead, which failed to exploit the repetitive nature of computational task requests and the spatiotemporal locality of user behavior. Existing approaches typically treat TO and caching as decoupled problems, without considering how cache hit rates can directly influence offloading efficiency and system cost. Furthermore, community-level collaboration and regional request patterns remain underexplored in cache-aware MEC systems. These limitations hinder the full potential of MEC in latency-critical and cost-sensitive multi-user scenarios.

The caching function of MEC servers improves user experience by pre-storing resources at nodes near the users. When cache hits occur, they reduces content delivery costs. Reference [18] analysed the expression of the average successful delivery probability (i.e., the probability of receiving content within a tolerable delay) by considering the distribution of wireless fading channels, and proposed a suboptimal caching placement algorithm to optimize this probability. The authors of [19] proposed a content freshness-aware caching framework for content distribution networks with front-end local caches and back-end databases, aiming to minimize the total cost from content staleness and cache updates. The study distinguishes between push-based and pull-based caching under asymmetric information availability. The authors of [20] formulated a delay-energy tradeoff problem by jointly optimizing user association, power control, and content caching for cache-enabled heterogeneous cellular networks, proposing an iterative algorithm that used benders decomposition to minimize both delay and energy consumption. To minimize data staleness and delivery delay, a two-stage algorithm was developed in [21] for connected vehicular networks based on Markov decision processes and Lyapunov optimization. This framework leveraged RSUs and MBSs to maintain road environment information freshness. However, the caching and computing resources of MEC are limited. Combining caching and computing resources is a feasible solution, addressing the shortcomings and limitations of single-dimensional resources with multi-dimensional resources. This approach effectively enhances system performance but requires flexible configuration of the limited network resources.

In summary, the application of edge caching in computation offloading is a challenging issue. Motivated by this, this paper investigates the problem of multi-user TO assisted by caching in mobile network edge computing scenarios [22]. Considering the repetitive nature of requested computational tasks and the regional characteristics of user requests, edge caching is introduced in the TO process. When a cache hit occurs, tasks do not need to be uploaded, and the computation results can be directly returned at the BS, effectively improving transmission and latency performance. The main contributions of this paper are summarized as follows:

• We first introduce edge caching in TO by considering the scenario of repetitive computational task requests. Then we model latency and energy consumption, which aim to minimize offloading costs. We use dynamic programming to derive the optimal caching strategy. Based on the strategy, the offloading costs are optimized.

• Considering user mobility and the regional nature of requested tasks, we integrate spatiotemporal characteristics to develop a community collaboration-based cache profit maximization model. We then transform this problem into a 0–1 knapsack problem for solution. Using the derived caching strategy, we redefine the offloading costs by leveraging the complementarity of multi-dimensional resources to optimize TO expenses.

• Extensive experiments are conducted to validate the performance of the proposed strategy. The experimental results consistently demonstrate that the proposed strategy significantly outperforms other existing approaches in multiple key metrics, particularly in terms of offloading costs and cache utilization.

The rest of this paper is as follows. Section 2 gives the system model. Section 3 describes the community collaborative caching model, cache benefit model, and cache-assisted offloading optimization. In Section 4, we provide experimental results and analyses. The conclusions are given in Section 5.

System Model

We consider a multi-user urban scenario, as shown in Figure 1, where BSs deployed along the road are denoted by the set $$\mathcal{M}{\mathrm{ = \{ }}1,\ldots,m,\ldots,M{\mathrm{\} }}$$. Wireless links (user-to-BS) exhibit dynamic channel conditions due to mobility, while fiber links (inter-BS) provide stable high-throughput collaboration. The coverage radius of BSs is $$R$$, and each BS is equipped with a MEC server. The system bandwidth and the computational resources of the MEC are denoted by $$B$$ and $$F$$, respectively. The set of cache resources of the MEC and users is denoted by $$St{\mathrm{ = \{ }}S{{t}_m}|m \in \mathcal{M}{\mathrm{\} }}$$ and $$\mathcal{N} = \{ 1,..n,..N\} $$, respectively.

[IMAGE OMITTED. SEE PDF]

Each user initiates a task $$n$$ at time $$t$$, which is defined as $${{T}_n} = ({{d}_n},{{c}_n},t_n^{\max })$$, where $${{d}_n}$$, $${{c}_n}$$ and $$t_n^{\max }$$ denote the computational resources required by the task $${{T}_n}$$, the task magnitude $${{T}_n}$$, and the longest tolerable delay for executing the task, respectively. Assume that each task has two processing modes, local execution and edge offloading. The TO decision of a user is shown as follows: 1 $$$\begin{equation}{{a}_n} = \left\{ { \def\eqcellsep{&}\begin{array}{@{}*{2}{l}@{}} 1&{\textit{UE n admitted to offload tasks}}\\ 0&{{{T}_n} \textit{is executed at UE n locally}} \end{array} } \right..\end{equation}$$$

The set of users taking offload is denoted by $$S = \{ n|{{a}_n} = 1,n \in \mathcal{N}\} $$. Assuming that at most $$|\mathcal{N}|$$ users perform the task, so the constraints are as follows: 2 $$$\begin{equation}\sum\limits_{n \in N} {{{a}_n}} \le N,\quad \forall n \in \mathcal{N}.\end{equation}$$$

Delay and Energy Modeling

The total time cost of TO in the compute offloading mode is denoted by $$t_n^{off}$$, including the task upload delay $$t_n^{tr}$$, server execution delay $$t_n^{com}$$, and result transmission delay $$t_n^{bac}$$. The delay when the user is stationary denoted by $$t_n^{bac}$$, can be disregarded [23]. However, user mobility may result in the user moving beyond the coverage radius of the BS while downloading results. In such scenarios, considering the user downloading results from the edge server of adjacent BSs, the result transmission delay becomes non-negligible and can be expressed as follows: 3 $$$\begin{equation}t_n^{bac} = \frac{{{{d}_n}}}{{{{B}_{wired}}}},\end{equation}$$$where $${{B}_{wired}}$$ represents the transmission speed of fibre optics between BSs. Thus, the total time cost $$t_n^{off}$$ of offloading $${{T}_n}$$ can be expressed as follows: 4 $$$\begin{equation}t_n^{off} = t_n^{tr} + t_n^{com} + {{\gamma }_n}t_n^{bac},\end{equation}$$$where $${{\gamma }_n}$$ is a binary variable, $${{\gamma }_n} = 1$$ indicates that the user has moved outside the radio radius of the BS before $${{T}_n}$$ obtained the result and needs to receive the result from a neighboring BS. $${{\gamma }_n} = 0$$ indicates that the user has not left the current BS.

According to Shannon's theorem, the uplink transmission rate $${{r}_n}(t)$$ of $${{T}_n}$$ unloaded to the BS at time $$t$$ can be expressed as: 5 $$$\begin{equation}{{r}_n}(t) = \frac{B}{{\sum\limits_{n \in N} {{{a}_n}} }}log(1 + \frac{{{{p}_n}{{H}_n}(t)}}{{{{N}_0}}}),\end{equation}$$$where $${{p}_n}$$ is the transmission power of $${{T}_n}$$ accessing the BS. $${{H}_n}(t)$$ represents the channel gain from user $$n$$ to the BS at time $$t$$, which can be expressed as follows. 6 $$$\begin{equation}{{H}_n}(t) = dis_n^{ - a}(t)*|{{h}_0}{{|}^2},\end{equation}$$$where $$dis_n(t)$$ denotes the distance between user $$n$$ and the BS at time $$t$$. Assuming that the user moves in a constant direction at a speed $$v$$, the coordinates of the BS are $$({{x}_0},{{y}_0})$$, and the two-dimensional coordinates of the user at time $$t$$ are represented by vectors $$L = \{ ({{x}_1}(t),{{y}_1}(t)),..({{x}_n}(t),{{y}_n}(t)),..({{x}_N}(t),{{y}_N}(t))\} $$. Therefore, the distance from user $$n$$ to the BS at time $$t$$ can be expressed as: 7 $$$\begin{equation}dis_n(t) = \sqrt {{{{({{x}_n}(t) - {{x}_0})}}^2} + {{{({{y}_n}(t) - {{y}_0})}}^2}} ,\end{equation}$$$where $$\alpha $$ represents the path loss exponent, $${{h}_0}$$ is the complex Gaussian channel coefficient that follows the complex normal distribution $$\mathcal{C}\mathcal{N}( {0,1} )$$. $${{N}_0}$$ is the additive white Gaussian noise [24].

The transmission delay $${{T}_n}$$ can be defined as follows: 8 $$$\begin{equation}t_n^{tr} = \frac{{{{d}_n}}}{{{{r}_n}(t)}},\end{equation}$$$where $${{f}_n}$$ denotes the computation resources allocated to $${{T}_n}$$ by the MEC. The computing delay can be given by 9 $$$\begin{equation}t_n^{com} = \frac{{{{c}_n}}}{{{{f}_n}}},\end{equation}$$$where $${{p}_n}$$ represents the user $$n$$’s transmission power, and the offloading energy consumption $${{T}_n}$$ can be defined as follows: 10 $$$\begin{equation}{{e}_n} = t_n^{tr}{{p}_n}.\end{equation}$$$

Therefore, the offloading cost of offloading $${{T}_n}$$ can be given by 11 $$$\begin{equation}{{\mathcal{C}}_n} = {{\theta }_t}*t_n^{off} + {{\theta }_e}{{e}_n} \forall n \in \mathcal{S},\end{equation}$$$where $${{\theta }_e}$$ and $${{\theta }_t}$$ are the weight factors of delay and energy consumption, respectively.

Cache-Assisted TO

Introducing edge caching in the system model, we study the optimization problem of cache-enhanced edge computing task offloading. When the cache hits, the transmission delay and energy consumption of the task will be greatly improved by deploying popular request tasks at the edge, which can achieve the purpose of improving the offloading cost. However, the huge data scale and limited cache capacity make it impractical for edge servers to store all content. Therefore, this section proposes an efficient collaborative caching strategy to further optimize the offloading cost.

Community Collaborative Caching Model

Given the regionality of user preferences, BSs are divided into different communities according to user request preferences. The user preferences of BS services in the same community are similar [25]. This paper assumes that BSs in each community can perform collaborative caching. The community collaborative caching model is shown in Figure 2. The BSs are divided into $$K$$ communities $${{C}_{BS}} = \{ {{C}_1},..,{{C}_k},..,{{C}_K}\} $$, where $${{C}_k}$$ represents the set of BSs in the community $$k$$. Each BS can only belong to one community and should satisfy the constraints as follows, 12 $$$\begin{equation}\sum\limits_{k \in K} {|{{C}_k}|} \le |\mathcal{M}|.\end{equation}$$$

[IMAGE OMITTED. SEE PDF]

The BSs are connected by optical fibers. The tasks initiated by users in each BS are recorded in a file library with dynamically updated content, which is denoted as $$\mathcal{J}_m^t$$. The file library collection that records all requested tasks in the BS is represented by $$J$$. When a new task type arrives, it will be added in $$\mathcal{J}_m^t$$ and $$J$$. The size of each cached task $${{d}_n}$$ is the data size of the task itself. A portion of the computing tasks of popular requests are cached in the MEC server. $${{\mathcal{D}}^t} = \{ {{D}^t}_1,..,D_m^t,..,D_M^t\} $$ is a set of cache files collected for each BS, where $$D_m^t = \{ l_{jm}^t = 1|j \in {{\mathcal{J}}_m},m \in M\} $$ represents the task collection cached by BS $$m$$. The cache variable $$l_{nm}^t \in \{ 0,1\} $$ is a binary variable, where $$l_{nm}^t = 1$$ means that BS $$m$$ has a cache task $$n$$ at time $$t$$, and $$l_{nm}^t{\mathrm{ = 0}}$$ vice versa. The capacity constraint of the cache resource is expressed as follows [26]: 13 $$$\begin{equation}\sum\limits_{n \in \mathcal{J}_m^t} {l_{nm}^t{{d}_n}} \le S{{t}_m}{\mathrm{, }}\forall m \in \mathcal{M}, t \in T.\end{equation}$$$

Cache Benefit Model

This subsection defines the benefit value $$Pr{{o}_{n,m}}$$ of cache task $$n$$ for BS $$m$$, which is mainly related to three factors, i.e., task popularity, transmission delay, and transmission energy consumption. Even the same task may be in different locations when requested at different times. Therefore, it can be seen that even for the same tasks, their data sizes will still be slightly different. Due to the differences in location, the transmission delay and energy consumption of the same task may also be different, resulting in different cache benefits. In order to make the cache benefits of all the same computing tasks the same, the unit size benefit of the task is used to represent the entire task. The transmission rate $$r_{n,m}^{tr}$$ and transmission power $$p$$ are set to the mean $${{r}^0},p$$, respectively. They are updated in each iteration $${{r}^0}$$. 14 $$$\begin{equation}Pr{{o}_{n,m}} = p_{n,m}^t\frac{{1 + p}}{{{{r}^0}}},\end{equation}$$$where $$p_{n,m}^t$$ is the popularity of task $$n$$ at BS $$m$$. Due to the limitation of cache resources, the storage capacity of a single BS cache is relatively small. In addition, $$p_{n,m}^t$$ is the popularity of tasks in different areas often varies. Relying only on the content of the local cache will affect the cache hit performance. Through collaborative caching, content sharing between different entities can improve cache performance.

Dual-Time Scale Popularity Model

This subsection introduces a dual-time scale popularity update mechanism to update the task popularity in the BS, the community, and the community model in different time periods. Therefore, $${{\tau }_{long}}$$ is used as the long-term update time interval in the community, while $${{\tau }_{short}}$$ is used as the short-term update time interval in the BS. The popularity of tasks $$n$$ at different spatial scales is initialized according to the Zipf distribution as follows: 15 $$$\begin{equation}p(r\_f) = \frac{{r\_{{f}^{ - a}}}}{{\sum\limits_{r\_f = 1}^\mathcal{F} {r\_{{f}^{ - a}}} }}.\end{equation}$$$

The Zipf distribution reasonably follows the rule that a very small number of contents are provided to most requests in the network. Where $$\alpha $$ is the Zipf factor, it controls the degree of popularity difference between different tasks. When $$\alpha $$ increases, the popularity difference between tasks increases. $$\mathcal{F}$$ and $$r\_f$$ represents the popularity ranking of tasks and the ranking of the task in historical data, respectively [27]. In addition, the coverage of the community is much larger than that of a single BS, and the changes in the popularity of tasks within a single BS are usually much faster than the changes in the popularity of content within the community [28]. Updating the global content popularity within the community $${{C}_k}$$ at each interval $${{\tau }_{long}}$$ can be given by: 16 $$$\begin{equation}p_{k,n}^t = {{\beta }_1}p_{k,n}^{t - 1} + (1 - {{\beta }_1})\frac{{D_n^{t - 1}}}{{\sum\limits_{n \in J} {D_n^{t - 1}} }}, t = x{{\tau }_{long}},x = 0,1,2\ldots,\end{equation}$$$where $$p_{k,n}^t$$ represents the global popularity of task $$n$$ in the community $${{C}_k}$$ at time $$t$$, $$D_n^{t - 1}$$ is the number of times of the task $$n$$ has been requested up to moment $$t - 1$$. $$p$$ is the weight coefficient, which reflects the influence of the overall number of requests and the popularity of historical tasks on the global popularity to a certain extent.

BS $$m$$ will update the task popularity in the area according to the change of global popularity. It is determined by the regional popularity of the adjacent BSs at the previous time and the global popularity at the previous time. Therefore, $${{\tau }_{short}}$$ updates the local popularity in its area at each short time slot, and the update method is shown as follows: 17 $$$\begin{equation} \def\eqcellsep{&}\begin{array}{l} p_{m,n}^t = {{\beta }_2}{{\mathbb{E}}_k}\left[ {\sum\nolimits_{i \in {{\mathcal{C}}_k}} {\zeta _i^tp_{i,n}^{t - 1}} } \right] + (1 - {{\beta }_2})p_{k,n}^{t - 1}{\mathrm{, }}\\ t = x{{\tau }_{short}}, x = 0,1,2\ldots. \end{array} \end{equation}$$$where $$\zeta _i^t \in [0,1]$$ is the weight factor, which represents the influence of neighboring BSs on task $$n$$ popularity. $${{\mathbb{E}}_k}[ \cdot ]$$ represents the statistical average of content popularity between neighboring BSs of the BS $$m$$. $${{\beta }_2}$$ is the weight factor, which describes the influence of global popularity and neighboring BS task popularity on local popularity.

In addition, the task popularity update formula in the normal time slot $$t$$ is as follows, 18 $$$\begin{equation}\hat{p}_{m,n}^t = {{\beta }_3}\frac{{N_{n,m}^{t - 1}}}{{N_n^{t - 1}}} + (1 - {{\beta }_3})p_{m,n}^{{{\tau }_{short}} - 1},\end{equation}$$$where $$N_{n,m}^{t - 1}$$ and $$N_n^{t - 1}$$ represent the number of times of the task $$n$$ requested in the current BS $$m$$ and the total number of tasks requested in BS $$m$$, respectively.

Optimization Formulation

For the determination of cache placement strategy, we set the optimization goal as maximizing cache benefits. The specific optimization problem is formulated as follows: 19 $$$\begin{equation} \def\eqcellsep{&}\begin{array}{l} {\mathrm{P0 : }}\mathop {\max }\limits_{{{\mathcal{D}}^t}} \sum\limits_{n \in {{\mathcal{J}}_m}} {\sum\limits_{m \in {{C}_k}} {l_{mn}^tPr{{o}_{n,m}}} } \\ {\mathrm{s}}{\mathrm{.t}}{\mathrm{. }} {\mathrm{C1:}}\sum\limits_{k \in K} {|{{C}_k}|} \le |\mathcal{M}| \\ {\mathrm{ C2: }}l_{mn}^t \in \left\{ {0,1} \right\}, \forall m \in \mathcal{M},{\mathrm{ n}} \in {{\mathcal{J}}_m}, t \in T,\\ {\mathrm{ C3:}}\sum\limits_{n \in \mathcal{D}_m^t} {l_{mn}^t{{d}_n}} \le S{{t}_m}{\mathrm{, }}\forall t \in T, \forall m \in \mathcal{M}, \end{array} \end{equation}$$$where the optimization variable is the cache decision matrix $$\mathcal{D}$$. The number of users in the community is restricted to not exceed the total number of users in $${\mathrm{C1}}$$; $${\mathrm{C2}}$$ is a binary variable constraint; $${\mathrm{C3}}$$ is the BS cache capacity constraint.

We provide a logical block diagram of collaborative caching under dual time scales as shown in Figure 3.

[IMAGE OMITTED. SEE PDF]

Cache-Assisted Offloading Optimization

In cache-assisted offloading, caching can reduce the transmission of repetitive data and thus shorten part of the task processing time. However, even with a cache hit, tasks still require transmission, coordination among BSs, and computation on MEC servers, all of which consume communication and computing resources. If these processes are not carefully optimized, the system may experience increased delays or energy consumption, especially when fetching data from collaborative BSs. Therefore, offloading cost—defined as the weighted sum of delay and energy consumption—remains a key performance metric in cache-assisted scenarios. Optimizing the offloading cost ensures that caching strategies are aligned with TO decisions, so that the benefit of cache hits is not offset by high collaborative retrieval delays or excessive resource usage.

Offloading Cost Optimization

As shown in Figure 4, each user can choose to execute locally or take offloading in the cache-assisted MEC system architecture. When a user initiates a request, the BS will check whether it has cached the task. If not, it will go to its community to find out whether it has cached the task. If so, only data of negligible size needs to be uploaded to be calculated by the BS. At this point, the transmission delay and transmission energy consumption of the task will be greatly improved. These two performance indicators, $$t_{n,m}^{opt}$$ and $$e_{n,m}^{opt}$$ are redefined in this section. Therefore, the offloading cost $${{\mathcal{C}}_n}$$ can be further rewritten as follows: 20 $$$\begin{equation}{{\mathcal{C}}_n} = {{\theta }_t}*(t_{n,m}^{opt} + t_{n,m}^{com} + \gamma t_{n,m}^{bac}) + {{\theta }_e}e_{n,m}^{opt}\; \forall n \in \mathcal{N},\; m \in M.\end{equation}$$$

[IMAGE OMITTED. SEE PDF]

When a user $$n$$ initiates a request to BS $$m$$, the offloaded task may have the following three cache situations, as shown in Figure 5.

• The task $$n$$ is cached at BS $$m$$: The task $$n$$ does not need to be uploaded, and the task-related data can be directly obtained from the local BS without uploading. It can be expressed as $$t_{n,m}^{opt} = 0, e_{n,m}^{opt} = 0$$.

• There is no cached task $$n$$ at BS $$m$$: The community $${{C}_k}$$ to which the BS $$m$$ belongs has cached the task $$n$$, and the task $$n$$ can be obtained from a cooperative BS $$z$$. In this case, we have21 $$$\begin{equation}t_{n,m}^{opt} = \min ({{t}_{m,z}}), e_{n,m}^{opt} = \min ({{t}_{m,z}})*{{p}_n}, \forall z \ne m, m{\mathrm{,}}z \in {{C}_k},\end{equation}$$$

Where the shortest distance $$\min ({{t}_{m,z}})$$ between BS $$m$$ and cooperative BS $$z$$ can be calculated by the Dijkstra algorithm. Since some BSs are not connected and need to communicate through other BSs, it is necessary to calculate the minimum path between BSs to obtain the smallest possible transmission delay.

[IMAGE OMITTED. SEE PDF]

The calculation of the shortest path to find the cache is shown in Algorithm 1. First, an undirected weighted graph $$G2 = \{ V2,E2\} $$ is constructed as shown in Figure 6. Each BS is a vertex, and the distance between BSs is the edge weight between vertices. When a task request is initiated at BS $$m$$, it first checks whether the task is cached locally. If not, BS $$m$$ searches for the BS it can reach in the community to see if the task is cached. If such a BS is found, it directly returns the result closest to BS $$m$$. If not, select a BS $${{j}^*}$$ closest to the BS $$m$$ and update it as the source node, and then repeat the previous steps until the cached task is found. Finally, the minimum collaborative delay is obtained $$\min ({{t}_{m,z}})$$. When the collaborative caching delay of the task is greater than the delay of uploading the task to the BS, it will be meaningless to use collaborative caching at this time, so the constraints must be met as follows: 22 $$$\begin{equation}\min ({{t}_{m,z}}) < t_{n,m}^{tr}.\end{equation}$$$

[IMAGE OMITTED. SEE PDF]

1

Algorithm

Shortest collaborative path algorithm.

Input: The community where the target BS is located, the shortest distance $$dist[|{{C}_k}|][|{{C}_k}|] = \infty $$ between BSs is initialized.
Output: Minimum collaboration delay $$\min ({{t}_{m,z}})$$
1:	Construct an undirected weighted graph $$G2$$
2:	Start from the current source node and traverse other reachable base stations to find out whether there is a cache in the reachable base stations. If the task exists, return $$\mathop {\arg \min }\limits_j dist[i][j]$$ directly; if not, proceed to the next step.
3:	Update $$dist$$ and source nodes $${{j}^*} = \mathop {\arg \min }\limits_j dist[i][j]$$
4:	Repeat step 2
5:	$$\min ({{t}_{m,z}}) = \frac{{\min (dis[m][z])}}{{{{B}_{wire}}}}$$

Neither the target BS nor other BSs in the community have cached tasks $$n$$: At this time, the task $$n$$ can only be uploaded to the BS for processing via the wireless link $$t_{n,m}^{opt} = t_{n,m}^{tr}, e_{n,m}^{opt} = e_{n,m}^{tr}$$.

Problem Modeling

In this subsection, we optimize TO and introduce edge caching based on the previous subsection. Then we study the problem of cache-enhanced offloading optimization. Based on the proposed cache strategy, the final optimization goal is still to minimize the offloading cost, i.e., 23 $$$\begin{equation} \def\eqcellsep{&}\begin{array}{l} \mathop {\min }\limits_{{{D}^t}} \sum\limits_{n \in \mathcal{S}} {{{\mathcal{C}}_\mathcal{N}}} \\ {\mathrm{s}}{\mathrm{.t}}{\mathrm{.}}\;\;\;\sum\limits_{n \in \mathcal{J}_m^t} {l_{nm}^t{{d}_n}} \le S{{t}_m}{\mathrm{, }}\forall m \in \mathcal{M}, t \in T. \end{array} \end{equation}$$$

Cache Strategy to Maximize Benefits

Louvain-based community clustering algorithm

Although there exist differences in terms of the locations of BSs, the number of users, and the types of requested tasks in their service locations. There is a certain degree of overlap in the types of requests of users in adjacent local areas. For example, users in a business district prefer entertainment and fashion information. To address this, we divide BSs into different communities and group areas where users with similar interests and preferences belong to the same community. Within each community, BSs cooperate with each other to respond to needs more quickly. Although user mobility and user preferences change rapidly, most users are active in a limited area within the BS range every day, and the user preferences within similar BSs will not change much overall. Therefore, it is still used $${{\tau }_{long}}$$ as the interval for similarity update. Define the similarity $$\varphi _{m,z}^t$$ between BS $$m$$ and BS $$z$$ at time instant $${{\tau }_{long}}$$. The similarity between BSs is measured by the Jaccard coefficient [29], which can be used to describe the similarity between two sets. $$\mathcal{J}_m^t$$ and $$\mathcal{J}_z^t$$ represent the task request sets of BS $$m$$ and BS $$z$$ at time instants $$t$$, respectively. $${{\mathbb{E}}_k}[ \cdot ]$$ represents the statistical average of the similarity of BSs from the $${{\tau }_{long}}$$ of x−1^th to the $${{\tau }_{long}}$$ of x^th. 24 $$$\begin{equation}\varphi _{m,z}^{{{\tau }_{long}}} = {{E}_k}[\sum\limits_{(n - 1){{\tau }_{long}}}^{n{{\tau }_{long}}} {\frac{{|\mathcal{J}_m^t \cap \mathcal{J}_z^t|}}{{\mathcal{J}_m^t \cup \mathcal{J}_z^t}}} ], \forall t \in T,\end{equation}$$$where $$\mathcal{J}_m^t \cap \mathcal{J}_z^t$$ represents the same request between BS $$m$$ and $$z$$. The larger the value $${{\varphi }_{m,z}}$$, the higher the similarity between the two BSs. Thus, a similarity matrix $$\psi $$ of $$M*M$$ dimension between BSs is constructed, which is denoted as $$\psi = \{ {{\varphi }_{m,z}}|m,z \in \mathcal{M}\} $$. An undirected weighted graph $$G = \{ V,E\} $$ is constructed based on the similarity matrix $$\psi $$. Each vertex of the graph represents a BS, and the edge weight between BSs is represented by the similarity between BSs. The BSs are divided into different communities using the Louvain-based clustering algorithm [30] to obtain a set $$Social$$. Louvain is a community discovery algorithm based on modularity. Its basic idea is to traverse all neighbor communities that can be reached by the node and select the community that maximizes the modular gain. At the beginning, each vertex is a community. Each time a vertex is merged, only the adjacent communities that have not been merged and are reachable by distance are considered. The modular gain brought by merging the vertex $$i$$ into the community is calculated $$A$$ and expressed as follows: 25 $$$\begin{equation}\Delta Q = \frac{1}{{2sum}}({{k}_{i,in}} - \frac{{\sum {_{tot}{{k}_i}} }}{{sum}}),\end{equation}$$$where the brackets $${{k}_{i,in}}$$ represent the sum of the edge weights between the vertex $$i$$ and the community $$A$$, the numerator of the second term $$\sum {_{tot}{{k}_i}} $$ denotes the product of the sum of all edge weights connected to the node $$i$$ and the sum of the edge weights of all external edges connected to the community $$A$$, the denominator $$sum$$ is the sum of all edge weights in the graph. Finally, $$i$$ choose to move to the community with the largest value $$\Delta Q$$. When the largest value $$\Delta Q$$ is less than 0, the node $$i$$ remains stationary. Combine every community into a new super node once stage 1 is finished. The super node's edge weight is the total of all the nodes' edge weights from the original community, forming a new network. Repeat the above steps until the new move does not bring any module gain. Algorithm 2 stops.

2

Algorithm

Louvain-based community clustering algorithm.

Input: Similarity weighted undirected graph $$G1$$
Output: Community division $${{C}_{BS}}$$
1:	First stage node merging
2:	Initialize each node as a community
3:	for $$i = 1:V1$$
4:	for $$j = 1:V1\& \& j \ne i$$
5:	if $$visit[i][j]\& \& visit[j][i] = = 0\& \& dis[i][j] \ne \infty $$
6:	$$visit[i][j]\& \& visit[j][i] = 1$$
7:	0-1
8:	end
9:	end
10:	$${{j}^*} = \mathop {\arg \max }\limits_j (\Delta Q)$$
11:	if $$\max (\Delta Q) > 0$$
12:	Nodes $$i$$ join new communities $${{j}^*}$$
13:	end
14:	end
15:	Second stage community aggregation
16:	Merge multiple nodes in the same community into a new node and update the edge weight
17:	Repeat steps 3–8 until $$Q < 0$$

Task Caching Strategy

Determining the cache benefit after obtaining a clear community division $${{C}_{BS}}$$. To achieve efficient use of cache resources, the cache scheme has a crucial impact on the cache effect. This section transforms the cache placement problem into a knapsack problem, which can describe how to pack a group of items with the highest benefit in a limited knapsack capacity [31]. Each item has two attributes: price and weight. There are generally four methods to solve the knapsack problem: brute force traversal of all feasible solutions. Although it can find the optimal solution, its time complexity is often exponential; recursion and divide and conquer decompose large problems into multiple identical small problems. When the recursive level is too deep, it has high time complexity; greedy algorithms can often only obtain local optimal solutions. Finally, we consider using low-complexity dynamic programming to solve. We describe the cache placement problem as a 01 knapsack problem, which puts the items with the best cache benefit in a limited cache capacity. The limited capacity of the knapsack is analogous to the cache capacity $$S{{t}_m}$$ of BS $$m$$. The benefit of the cache task can be expressed as the price of the item, and the weight of each item can be represented by the size of the task. Each item has two states, namely, being placed and not placed. When the benefit of adding the task minus the cost of caching the task makes the overall benefit better than not placing it ($$Value[i] < Value[i - d[n]] + Pr{{o}_n}$$), the task will be cached, i.e., set $$dp[i] = 1$$, otherwise set $$dp[i] = 0$$. Dynamic programming can effectively and with low complexity backtrack the cache placement strategy from the knapsack problem $$D_m^t$$.

Simulation Experiment and Result Analysis

Simulation Parameter Settings

All experiments were run on Matlab R2019a. Table 1 lists the main simulation parameters [24, 32]. We assume that there are 4 BSs in the scenario, and each BS is a circle with coverage radius $$R = 150{\mathrm{m}}$$. The linear overlap spacing between BSs is 70 m. BS and user locations are randomly scattered in the network using Poisson distribution, and users move in a straight line at a uniform speed. The user's local computing power $$f_n^{loc}$$, task size $${{d}_n}$$ and computing resources required by the task $${{c}_n}$$ are all random variables that obey uniform distribution. Parameters related to cache are added. The cache size of the BS is set to 50 MB, and the Zipf distribution parameter is set to 0.5 [33]. This parameter reflects the difference in popularity changes. A content library $$\Psi $$ of 500 tasks is randomly generated, and 100 tasks are randomly initiated in each iteration. At each time, each user initiates a task $${{T}_n}, \forall {{T}_n} \in \Psi $$. The number of iterations is used instead of the time slot $${{\tau }_{short}} = 5, {{\tau }_{long}} = 15$$. Since the focus of this paper is not on the formulation of offloading decisions, but on the optimization brought by cache to TO. Therefore, the offloading decision in this paper is consistent with the cache strategy $$S = \{ l_{nm}^t = 1|\forall n \in N,m \in M\} $$. The computing resource allocation is obtained by the Lagrange multiplier method [34]. In addition, this paper refers to the proposed scheme as TO-backpack. We use the following three algorithms to verify the effectiveness of the algorithm:

• TO: We denote the task offload solution without caching mechanism as TO.

• TO-greedy: When determining the cache strategy, put the content with the greatest benefit into the cache in sequence until it is full.

• TO no collaborative cache (TO-NCC): We only consider the cache in each BS. The caches between BSs do not affect each other. The cache decision scheme is the same as the algorithm in this paper, and the local popularity and global popularity are no longer divided.

TABLE 1 Simulation parameters.

Parameter	Parameter value
$$R$$	150 m
$${{B}_m}$$	10 MHz
α	4
$${{N}_0}$$	−174 dBm/Hz
$$F$$	20 GHZ
$$f_n^{loc}$$	[0.5, 1.5]GHz
$${{p}_n}$$	$$23{\mathrm{dBm}}$$
$$p$$	$$23{\mathrm{dBm}}$$
$$\theta $$	0.5
$${{c}_n}$$	[1, 3]$${\mathrm{GCycles}}$$
$${{d}_n}$$	(2, 3] MB
$${{\tau }_{short}}$$	5
$${{\tau }_{long}}$$	15
$${{B}_{weird}}$$	$$0.1{\mathrm{Gbps}}$$
$$S{{t}_m}$$	50 MB
$${{r}^0}$$	5.2 × 10^7

Analysis of Simulation Results

Figure 7 shows the average offloading cost when the number of users in the BS increases from 10 to 50. The bar chart clearly shows that the offloading cost is significantly reduced after the edge cache is introduced. The offloading cost of the TO-backpack algorithm is reduced by about 53% compared to the TO algorithm without cache, because it uses a cache placement strategy based on the backpack idea. In this scheme, while considering the popularity of the cache task and the benefits of caching the task, it also considers the cost of caching the task and obtains a trade-off between cost and benefit. The cache benefit is maximized under limited cache capacity. When the offloaded task is cached in the community, it does not need to be uploaded, and the calculated result can be directly returned to the user, which greatly improves the offloading delay and energy consumption and achieves the purpose of optimizing the offloading cost. The TO-greedy algorithm determines the cache strategy only based on the cache benefit, ignores the limited cache capacity, and is difficult to accurately track content changes in limited cache resources. Therefore, the overall offloading cost will be higher than the TO-backpack algorithm, which brings less than 40% optimization. Due to the limitations of the scenario, TO-NCC only updates the popularity locally, and often focuses on the characteristics of the requested content in the region. However, the user's location and preference are time-varying, so the cache strategy within a single BS lacks diversity and does not consider the cooperative relationship between neighboring BSs. The TO algorithm does not consider caching, so its offloading cost is the largest. The overall offloading cost reduction is decreasing, which is due to the fact that computing resources and communication resources limit the reduction of offloading costs.

[IMAGE OMITTED. SEE PDF]

In order to use the control variable method from the perspective of multi-dimensional resources, the following will analyse the performance impact of different algorithms on offloading from the three perspectives of storage space, computing resources, and communication resources. As can be seen from Figure 8, cache resources have a significant impact on offloading costs, and the offloading costs of all algorithms decrease rapidly as cache resources increase. In the TO-knapsack algorithm, the collaborative cache solution is the best, because the backpack strategy can make good use of cache resources to make the cache hit rate higher, thereby reducing transmission delay and energy consumption and optimizing offloading costs. The second is TO-greedy. As cache resources become more abundant, this solution makes up for the shortcomings of only considering cache benefits and can cache more tasks, thereby reducing offloading costs. TO-NCC makes up for the locality of the non-collaborative cache strategy as cache resources increase. The TO algorithm does not consider cache, so changes in cache resources have no effect on its performance.

[IMAGE OMITTED. SEE PDF]

In Figure 9, the relationship between cache resources and cache utilization is shown. Cache utilization is defined as the ratio of the total amount of data that is locally and collaboratively used by cache content to the total cache resources. The cache resources of the BS are set from 30 to 70 MB. The figure illustrates that the TO-backpack algorithm suggested in this paper performs optimally because it uses the community collaborative cache based on the backpack to narrow the scope of collaboration to BSs with similar preferences. The popularity change in dual time scales makes the cache more diverse and can make full use of limited cache resources. In addition, the cache utilization of the remaining algorithms is relatively close when the cache resources are 30 MB, because most of the limited cache resources are used to cache popular content. However, it is not the case that the larger the cache resources, the better. Because as the cache resources increase, the efficiency of collaborative caching is also reduced, and most of the space is still wasted. It can be observed from the TO-backpack and TO-greedy cache algorithms that as cache space increases, the increase in cache resource utilization gradually decreases. The cache utilization increases the most when BS storage space is 50 MB. Weighing the cache resource cost and cache resource utilization, it is obvious that the cache resource is 50 MB when the benefit is the best. Therefore, operators need to find a balance between cache resources and cache benefits. In addition, without considering the collaborative cache, TO-NCC decreases cache utilization faster as cache resources increase. However, the cache utilization of the TO algorithm is always zero because it does not consider the cache.

[IMAGE OMITTED. SEE PDF]

Figure 10 depicts the effect of communication resources on the offloading cost. There is a declining tendency in the offloading costs of different schemes when the communication resources continuously rise. When the communication resources are 20 MHz, the offloading cost changes the most, and the proposed TO-backpack scheme changes the most significantly. Task transmission latency and transmission energy usage in the offloading cost drop off quickly as communication resources rise. When the communication resources are greater than or equal to 30 MHz, the decline in the offloading cost gradually stabilizes. The reason is that when the communication resources reach a certain level, the decline in transmission delay and energy consumption is weak, and it can no longer bring more offloading cost optimization. From the comparison between TO-greedy and TO-NCC, it can be seen that the use of data sharing between BSs can bring a significant reduction in the offloading cost. When the communication resources are greater than or equal to 40 MHz, the offloading costs between TO-NCC and TO are similar. This is because when the communication resources are large enough, the cost optimization brought by the cache and the optimization brought by the reduction of transmission delay and energy consumption will become more similar.

[IMAGE OMITTED. SEE PDF]

Figure 11 shows the impact of computing resources on the offloading cost. It is obvious that as the computing resources increase, the offloading cost of each scheme decreases. All algorithms show a sharp decline when the computing resources are 20 GHz, because the loose computing resources alleviate the fierce resource competition between BSs and effectively reduce the computing delay. TO-backpack has always been the best, and dynamic programming finds the optimal solution under low complexity. TO-greedy and TO-NCC are still ranked second and third, and TOMM without cache still has the worst performance. In the end, the constraints of other resources limit the decline in offloading costs, and all algorithms tend to be stable. Compared with the decline in offloading costs brought about by the looseness of cache resources and communication resources, the increase in computing resources makes the overall decline in offloading costs more obvious. This is because the cache cost is influenced by the limitation on the total number of users that can be offloaded due to restricted computing resources. As computing resources increase, the number of users that can be offloaded also rises. Further, more cached tasks can be offloaded, thereby greatly reducing the offloading cost.

[IMAGE OMITTED. SEE PDF]

Conclusion

This paper applies the concept of edge caching to TO, caches the popular tasks requested by users in the MEC server in advance, and studies the optimization problem of cache-assisted TO. This paper decouples caching from task offloading and related resource optimization problems. For the formulation of caching strategies, the community division of collaborative caching is first carried out according to the diversity of user needs and the regional characteristics of preferences. According to the spatiotemporal characteristics, a long-term and short-term popularity update model is constructed, which can capture the overall content popularity trend and the local rapidly changing content popularity from the spatial level. Then the cache benefit is defined, and the cache placement problem is described as a capacity-constrained 0–1 cache knapsack problem to obtain a cache strategy that maximizes the cache benefit. Finally, the offloading cost is rewritten according to three caching situations. Compared with TO without caching, the cache-assisted TO proposed in this paper reduces the overall offloading cost by about 53%.

Our approach assumes stable community partitions. Dynamic user mobility may require real-time community reconfiguration, which will be studied in future work.

Author Contributions

Hao Liu: conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, software, supervision, validation, visualization, writing – original draft, writing – review and editing. Yan Zhen: conceptualization, data curation, formal analysis, investigation, methodology, resources, software, supervision, validation, visualization, writing – original draft. Libin Zheng: conceptualization, data curation, formal analysis, investigation, methodology, resources, software, supervision, validation, visualization, writing – original draft. Chao Huo: Conceptualization, data curation, formal analysis, investigation, methodology, resources, software, supervision, validation, visualization, writing – original draft. Yu Zhang: conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, software, supervision, validation, visualization, writing – original draft, writing – review and editing.

Acknowledgements

This work was supported by the National Key Research and Development Program of China (2021YFB2401300).

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data used to support the findings of this study are available from the corresponding author upon request.

References

J. Shi, Y. Zhou, Z. Li, Z. Zhao, Z. Chu, and P. Xiao, “Delay Minimization for NOMA‐mmW Scheme‐Based MEC Offloading,” IEEE Internet of Things Journal 10, no. 3 (2023): 2285–2296.

G. S. Paschos, G. Iosifidis, M. Tao, D. Towsley, and G. Caire, “The Role of Caching in Future Communication Systems and Networks,” IEEE Journal on Selected Areas in Communications 36, no. 6 (2018): 1111–1125.

Q. Alghazali, H. Al‐Amaireh, and T. Cinkler, “Energy‐Efficient Resource Allocation in Mobile Edge Computing Using NOMA and Massive MIMO,” IEEE Access 13 (2025): 21456–21470.

B. Premalatha and P. Prakasam, “Optimal Energy‐Efficient Resource Allocation and Fault Tolerance Scheme for Task Offloading in IoT‐FoG Computing Networks,” Computer Networks 238 (2024): 110080.

W. Fan, F. Xiao, Y. Pan, X. Chen, L. Han, and S. Yu, “Latency‐Aware Joint Task Offloading and Energy Control for Cooperative Mobile Edge Computing,” IEEE Transactions on Services Computing 18 (2025): 1515–1528.

Y. Fang, M. Li, F. R. Yu, et al., “Parallel Offloading and Resource Optimization for Multi‐Hop Ad Hoc Network‐Enabled CBTC With Mobile Edge Computing,” IEEE Transactions on Vehicular Technology 73, no. 2 (2023): 2684–2698.

S. Yuan, Y. Sun, and M. Peng, “Cache‐Aware Cooperative Multicast Beamforming in Dynamic Satellite‐Terrestrial Networks,” IEEE Transactions on Vehicular Technology 74, no. 1 (2024): 1433–1445.

F. Jiang, Z. Yuan, C. Sun, and J. Wang, “Deep Q‐Learning‐Based Content Caching With Update Strategy for Fog Radio Access Networks,” IEEE Access 7 (2019): 97505–97514.

P. K. Bishoyi and S. Misra, “Towards Energy‐and Cost‐Efficient Sustainable MEC‐Assisted Healthcare Systems,” IEEE Transactions on Sustainable Computing 7, no. 4 (2022): 958–969.

T. Lyu, H. Xu, F. Liu, M. Li, L. Li, and Z. Han, “Computing Offloading and Resource Allocation of NOMA Based UAV Emergency Communication in the Marine Internet of Thing,” IEEE Internet of Things Journal 11, no. 9 (2024): 15571–15586.

Y. Pochet and L. A. Wolsey, Production Planning by Mixed Integer Programming (New York: Springer, 2006).

T.‐C. Chiu, A.‐C. Pang, W.‐H. Chung, and J. Zhang, “Latency‐Driven Fog Cooperation Approach in Fog Radio Access Networks,” IEEE Transactions on Services Computing 12, no. 5 (2019): 698–711.

H. Zhang, X. Liu, Y. Xu, D. Li, C. Yuen, and Q. Xue, “Partial Offloading and Resource Allocation for MEC‐assisted Vehicular Networks,” IEEE Transactions on Vehicular Technology 73, no. 1 (2023): 1276–1288.

F. Chai, Q. Zhang, H. Yao, X. Xin, R. Gao, and M. Guizani, “Joint Multi‐Task Offloading and Resource Allocation for Mobile Edge Computing Systems in Satellite IoT,” IEEE Transactions on Vehicular Technology 72, no. 6 (2023): 7783–7795.

Y. Gong, H. Yao, J. Wang, M. Li, and S. Guo, “Edge Intelligence‐Driven Joint Offloading and Resource Allocation for Future 6G Industrial Internet of Things,” IEEE Transactions on Network Science and Engineering 11, no. 6 (2022): 5644–5655.

Y. Fan, J. Ge, S. Zhang, J. Wu, and B. Luo, “Decentralized Scheduling for Concurrent Tasks in Mobile Edge Computing via Deep Reinforcement Learning,” IEEE Transactions on Mobile Computing 23, no. 4 (2023): 2765–2779.

V. Masih and A. Roy, “Robust Learning Algorithms for Task Scheduling in MEC‐Enabled IoT Networks,” IEEE Transactions on Green Communications and Networking (2025): 1–1.

Q. T. Ngo, K. T. Phan, W. Xiang, A. Mahmood, and J. Slay, “Two‐Tier Cache‐Aided Full‐Duplex Hybrid Satellite–Terrestrial Communication Networks,” IEEE Transactions on Aerospace and Electronic Systems 58, no. 3 (2021): 1753–1765.

B. Abolhassani, J. Tadrous, A. Eryilmaz, and S. Yüksel, “Optimal Push and Pull‐Based Edge Caching for Dynamic Content,” IEEE/ACM Transactions on Networking 32, no. 4 (2024): 2765–2777.

W. Ni, H. Xiao, X. Liu, A. T. Chronopoulos, and P. A. Ioannou, “Power‐Aware User Association and Content Caching Algorithm for Delay‐Energy Tradeoff in Cache‐Enabled Heterogeneous Cellular Networks,” IEEE Systems Journal 19, no. 1 (2025): 246–257.

S. Park, C. Park, S. Jung, M. Choi, and J. Kim, “Age‐of‐Information Aware Caching and Delivery for Infrastructure‐Assisted Connected Vehicles,” IEEE Transactions on Vehicular Technology 73, no. 7 (2024): 10681–10696.

T. X. Tran and D. Pompili, “Adaptive Bitrate Video Caching and Processing in Mobile‐Edge Computing Network,” IEEE Transactions on Mobile Computing 18, no. 9 (2019): 1965–1978.

F. Fang, Y. Xu, Z. Ding, C. Shen, M. Peng, and G. K. Karagiannidis, “Optimal Resource Allocation for Delay Minimization in NOMA MEC Networks,” IEEE Transactions on Communications 68, no. 12 (2020): 7867–7881.

X. Hou, Z. Ren, J. Wang, et al., “Reliable Computation Offloading for Edge Computing Enabled Software‐defined IOV,” IEEE Internet of Things Journal 7, no. 8 (2020): 7097–7111.

N. Zhong, H. Zhang, and X. Zhang, “Collaborative Partition Caching With Local Popularity,” in IEEE 20th International Conference on Communication Technology (ICCT) (2020): 514–518.

Q. Li, Y. Zhang, Y. Li, Y. Xiao, and X. Ge, “Capacity‐Aware Edge Caching in Fog Computing Networks,” IEEE Transactions on Vehicular Technology 69, no. 8 (2020): 9244–9248.

R.‐J. Reifert, A. A. Ahmad, H. Dahrouj, et al., “Distributed Resource Management in Downlink Cache‐Enabled Multi‐Cloud Radio Access Networks,” IEEE Transactions on Vehicular Technology 71, no. 12 (2022): 13120–13136.

T. X. Tran, D. V. Le, G. Yue, and D. Pompili, “Cooperative Hierarchical Caching and Request Scheduling in a Cloud Radio Access Network,” IEEE Transactions on Mobile Computing 17, no. 12 (2018): 2729–2743.

W. Zhang, G. Zhang, and S. Mao, “Deep Reinforcement Learning Based Joint Caching and Resources Allocation for Cooperative MEC,” IEEE Internet of Things Journal 11, no. 7 (2024): 12203–12215.

The Louvain method for community detection in large networks[EB/OL], (2011‐03‐14) [2023‐ 02‐11]. http://perso.uclouvain.be/vincent.blondel/research/louvain.html.

Y. Fu, J. Liu, J. Ke, J. K. T. Chui, and K. K. F. Hung, “Optimal and Suboptimal Dynamic Cache Update Algorithms for Wireless Cellular Networks,” IEEE Wireless Communications Letters 11, no. 12 (2022): 2610–2614.

W. Zhan, C. Luo, G. Min, C. Wang, Q. Zhu, and H. Duan, “Mobility‐Aware Multiuser Offloading Optimization for Mobile Edge Computing,” IEEE Transactions on Vehicular Technology 69, no. 3 (2020): 3341–3356.

A. Erfanian, H. Amirpour, F. Tashtarian, C. Timmerer, and H. Hellwagner, “Cd‐lwte: Cost‐and Delay‐Aware Light‐Weight Transcoding at the Edge,” IEEE Transactions on Network and Service Management 20, no. 3 (2022): 3104–3118.

X. Lyu, H. Tian, C. Sengul, and P. Zhang, “Multiuser Joint Task Offloading and Resource Optimization in Proximate Clouds,” IEEE Transactions on Vehicular Technology 66, no. 4 (2017): 3435–3447.