The rapid development of 5G technology has led to the emergence of various Configuration and Optimization Parameters (COPs), which significantly influence network Key Performance Indicators (KPIs). This paper aims to accurately predict long-term KPI changes for parameter adjustments in cellular networks. Existing studies have employed causal inference to estimate the short-term effects of COP adjustments but often overlook the low-frequency nature of COP adjustments in real-world networks. Furthermore, prior research on multivariate time series (MTS) has primarily relied on time-domain and frequency-domain methods, which struggle to effectively capture the dynamic characteristics of non-stationary data. To address these limitations, Wave-CMNet (Wavelet-domain Causal Modeling Network) is proposed, an innovative causal modeling framework for predicting multivariate KPIs. This framework models the effects of COP adjustments on KPIs as Individual Treatment Effects (ITEs) within a point treatment setting in causal inference. Wave-CMNet leverages the time-frequency localization properties and multi-resolution analysis capabilities of the wavelet domain to effectively capture both frequency-domain and time-domain features. Subsequently, the model decomposes the MTS representation into time-dependent and inter-variable features to facilitate targeted learning. Extensive evaluations on real-world cellular network datasets confirm the effectiveness of Wave-CMNet.

Multi-domain multivariate time series (MTS) forecasting is increasingly important for large-scale and transferable time-series modeling. However, heterogeneous datasets usually contain different numbers of variables, mak ing scalable inter-series dependency modeling challenging. Existing scalable forecasting paradigms often rely on channel-independent (CI) strategies to accommodate variable-dimensional datasets. Nevertheless, by sharing global parameters across independently processed channels, CI models may confuse heterogeneous dependency structures under multi-domain joint training, leading to the dependency confusion problem. In this paper, we argue that effective multi-domain MTS forecasting requires variable-number-agnostic channel-dependence mod eling that can explicitly capture fine-grained inter-series dependencies (FID), including both local inter-series dependencies and cross-temporal inter-series dependencies. To this end, we propose the Fine-grained Inter-series Dependency Enhanced (FIDE) framework, a plug-and-play module for patch-based CI forecasters. FIDE introduces dependency prototypes to dynamically perceive input-specific inter-series dependency patterns and employs dependency transmission to propagate dependency information across temporal segments while remaining ag nostic to the number of variables. Extensive experiments on eight real-world benchmarks with six representative CI forecasting models demonstrate that FIDE consistently improves forecasting accuracy, effectively mitigates dependency confusion, and achieves statistically significant gains over baseline methods.

Accurate wireless network traffic prediction is pivotal for Quality of Service and enabling proactive network optimization. However, real-world traffic data exhibits complex dynamics, and communication networks across different regions often vary significantly in node density and traffic distribution. Existing spatiotemporal forecasting methods, particularly graph-based approaches, are typically coupled with specific number of nodes, hindering their transferability across regions with varying node counts and limiting the extraction of generalizable spatial patterns from large-scale heterogeneous data. To address these challenges, this paper proposes the Decoupled Generalizable SpatioTemporal (DeG-ST) framework. First, we introduce a representation framework independent of spatiotemporal node counts, utilizing channel-independent strategy for temporal modeling and a three-stage attention mechanism for spatial encoding. Second, to alleviate the low-rank bottleneck in spatial generalization, we design a generalized spatial learning scheme based on a Sparse Mixture-of-Experts strategy with multi-token routing to capture richer details. Experiments on real-world telecommunication traffic data demonstrate that DeG-ST benefits significantly from large-scale joint learning and exhibits superior performance in Zero-Shot transfer scenarios.

The explosive proliferation of mobile devices and the advent of mobile networks have triggered an unprecedented surge in data traffic, rendering accurate traffic prediction indispensable for network resource management. Despite the success of deep learning in this domain, existing spatiotemporal forecasting methods face critical limitations. They often neglect unobserved confounders—such as implicit user activity patterns and collective mobility dynamics—that introduce systematic prediction bias. Moreover, while timestamp data is ubiquitous, prior works predominantly utilize it merely as positional encodings or auxiliary features, failing to exploit its causal capacity to uncover these hidden latent factors. To address the aforementioned challenges, the Exogenous-aware Causal Invariant Spatio-Temporal model (ECIST) is proposed. This framework explicitly models latent confounders by reconstructing their distributions from hierarchical timestamps (e.g., hour-of-day) as exogenous variables via a Variational Autoencoder (VAE). Coupled with mutual information minimization, this approach eliminates spurious correlations from unobserved confounders to ensure unbiased predictions. Additionally, a patch-based Spatiotemporal Transformer is introduced to capture dynamics through decoupled temporal and spatial attention modules. Incorporating Dynamic Routing Sparse Attention (DRSA) overcomes quadratic complexity by adaptively selecting relevant base stations, suppressing noise, and achieving linear $O(N M K)$ complexity. Extensive experiments on real-world cellular network datasets demonstrate that ECIST significantly outperforms SOTA baselines, reducing MAE by 11.11%.

Multimodal time series representation learning has been extensively applied in healthcare, financial forecasting, and autonomous driving. However, heterogeneity among modalities has challenges for robust representation learning and multimodal task adaptation. To address the limitations of existing methods that lack sufficient consideration of modality differences and modality-task dependent interactions, we propose a multimodal learning framework named the Modality-Task Dependent Heterogeneous Mixture-of-Experts (MTD-HMoE) for multimodal time series representation learning. The proposed framework designs an Adaptive Heterogeneous Mixture-of-Experts model (AHMoE). AHMoE employs modality-specific expert sub-networks for time-series, image, and text data to enable semantic feature extraction and learns independent representations within each modality. A pretraining and joint optimization collaborative strategy explicitly models modality heterogeneity by coordinating the optimization of modality-specific output tokens and multimodal self-supervised objectives. In the fine-tuning stage, modality-task dependencies are leveraged to adaptively regulate inter-modal interaction weights, enabling efficient adaptation to diverse downstream tasks under limited labeled data. Extensive experiments on multiple real-world tasks demonstrate that MTD-HMoE consistently outperforms existing approaches, validating its effectiveness in capturing modality heterogeneity and improving task adaptability and scalability.

Future 6G networks will enable pervasive Generative Artificial Intelligence (GenAI) services. Migrating Large Language Model (LLM) inference to the network edge is essential to meet the low-latency and privacy demands of these services. However, LLM inference offloading faces a mismatch between resource-limited edge infrastructure and volatile computational demands. Conventional offloading schemes for static bit-level tasks cannot sufficiently handle the prefill/decoding phase difference and the uncertain output length of LLM requests. To bridge this gap, we derive a stochastic workload model based on decoder-only Transformer mechanics and profiled input/output token statistics, characterizing expected workload and task-dependent fluctuations. We embed these statistics into edge-cloud queue dynamics and formulate the massive-scale collaborative inference problem as a Mean Field Game (MFG). We approximate the high-concurrency queue dynamics using Stochastic Differential Equations (SDEs) and develop a Mean Field State Deep Q-Network (MFSDQN) algorithm. MFSDQN enables edge agents to learn decentralized offloading-ratio decisions under Quality of Service (QoS) constraints using local observations and neigh-borassisted mean-field states, avoiding global edge-state exchange. Extensive simulations with up to 500 edge agents demonstrate stable convergence and scalability. MFSDQN reduces the average system cost and unfinished workload by 8.0% and 6.3%, respectively, relative to the best-performing baseline.

Time series forecasting plays a pivotal role in edge intelligence. Current research predominantly focuses on exploring complex model architectures, such as Transformer and Graph Neural Network (GNN), which demonstrate remarkable advantages in capturing high-dimensional complex features. However, these models suffer from inherent limitations in computational efficiency and deployment resource, resulting in significant constraints in temporal efficiency and edge computing compatibility. To address these challenges, this paper proposes a lightweight linear model architecture termed ELinear. By introducing the Linear Channel Fusion module (LCF), the Frequency-domain Multi-Period Awareness mechanism (FMPA) and the Residual Period Fusion module (RPF), we enhance the prediction accuracy of linear models. Experimental results demonstrate that ELinear achieves a 4.7% reduction in Mean Absolute Error (MAE) and a 3.1x improvement in compute speed compared to state-of-the-art (SOTA) models on the widely adopted ETT benchmark dataset.

As Internet of Things (IoT)-based artificial intelligence (AI) applications grow, the surge in computational and communication demands has raised concerns about energy consumption, making it critical for 6G networks to address this challenge. This article examines the joint optimization of AI service placement, task scheduling, and computing resource allocation in an edge-network-cloud system to minimize long-term energy consumption. These problems are interdependent: AI service placement determines service locations, influencing task scheduling, which in turn dictates computing resource allocation. The key challenge lies in the coupling of these variables and the two time-scale nature of the problem, involving long-term (AI service placement) and short-term (task scheduling and computing resource allocation) strategies. To address this, a hierarchical Markov decision process (HMDP) framework is proposed for efficient and coordinated optimization across time scales. A hierarchical mean-field dueling double deep Q-network (HMFD3QN) algorithm is developed within this framework, where the upper layer optimizes AI service placement, and the lower layer manages task scheduling and computing resource allocation. By integrating mean-field theory, the algorithm reduces the complexity of multiagent interactions. The computing resource allocation problem is shown to be convex when other variables are fixed, and an optimal strategy is derived using Karush-Kuhn-Tucker (KKT) conditions to simplify the action space for reinforcement learning. Experimental results demonstrate that the proposed method can reduce energy consumption by up to 34% compared to baseline methods, significantly improve queue stability, and increase the proportion of tasks meeting QoS requirements.

Long-term time series forecasting (LTSF) poses significant challenges due to the need to capture complex long-term dependencies over extended horizons. We revisit the temporal structure of time series data and categorize it into regular periodic pattern and irregular periodic pattern. We argue that a comprehensive modeling of long-term temporal dependencies must simultaneously account for both regular periodic pattern and the latent dependencies present in irregular periodic pattern. However, prior works either focus on regular periodic pattern from a single scale or neglect the rich dependencies within irregular periodic pattern. To address these limitations, we propose PerTime, a multi-scale periodic mixing model for long-term time series forecasting, which consists of two modules: the Periodic Broaden Module (PBM) and the Hierarchical Fusion Module (HFM). PBM enables the extraction of periodic features across multiple time scales, while HFM captures latent long-term dependencies in irregular periodic pattern and integrates information across scales. Extensive experiments on six long-term forecasting benchmark datasets demonstrate that PerTime outperforms existing methods.

Diffusion-based denoising models have demonstrated impressive performance in probabilistic forecasting for multivariate time series (MTS). Nonetheless, existing approaches often model the entire data distribution, neglecting the variability in uncertainty across different components of the time series. This paper introduces a Diffusion-based Decoupled Deterministic and Uncertain ($\mathrm{D^3U}$) framework for probabilistic MTS forecasting. The framework integrates non-probabilistic forecasting with conditional diffusion generation, enabling both accurate point predictions and probabilistic forecasting. $\mathrm{D^3U}$ utilizes a point forecasting model to non-probabilistically model high-certainty components in the time series, generating embedded representations that are conditionally injected into a diffusion model. To better model high-uncertainty components, a patch-based denoising network (PatchDN) is designed in the conditional diffusion model. Designed as a plug-and-play framework, $\mathrm{D^3U}$ can be seamlessly integrated into existing point forecasting models to provide probabilistic forecasting capabilities. It can also be applied to other conditional diffusion methods that incorporate point forecasting models. Experiments on six real-world datasets demonstrate that our method achieves over a 20\% improvement in both point and probabilistic forecasting performance in MTS long-term forecasting compared to state-of-the-art (SOTA) probabilistic forecasting methods. Additionally, extensive ablation studies further validate the effectiveness of the $\mathrm{D^3U}$ framework.

Current time series forecasting models face significant challenges in addressing two key issues inherent in cloud cluster workload forecasting: high missing rates and high dimensionality. To tackle the missing data challenge, we propose a Channel Dependency Pre-learning Module (CD-Block) that combines wavelet decomposition with contrastive learning. This module extracts inter-channel dependencies through self-supervised learning, enhancing the model’s ability to interpret and process incomplete data. For the high dimensionality problem, we introduce a Fourier Graph Network (FGN) that reformulates convolutions in the frequency domain, significantly reducing computational complexity. FGN incorporates a time-frequency alignment loss to align pre-learned channel dependencies with the spectral representations of time series. Building on these innovations, we propose the Pre-learned Dependency Fourier Network (PefNet). Experimental results demonstrate that PefNet achieves superior performance on four high-dimensional benchmark datasets for forecasting tasks and achieves state-of-the-art (SOTA) performance on three real-world cloud cluster workload datasets with different missing data scenarios.

With the development of virtualization technology, serverless computing has been gaining significant attention in recent years, primarily due to its advantages in scalability and a pay-as-you-go pricing model. In edge networks, the deployment of fine-grained function instances to handle massive request data makes it more difficult to optimize the quality of user experience. This paper addresses the scheduling of heterogeneous function processing requests generated by users in serverless edge computing scenarios, considering the constraints of resources on edge nodes, and making decisions regarding the warm and cold start during the scheduling process. The problem is modeled as a constrained multi-objective optimization issue aimed at minimizing latency and energy consumption. A deep reinforcement learning strategy, grounded in Multi-Agent Proximal Policy Optimization (MAPPO), is introduced to address this challenge, with each user being represented as an autonomous agent. Simulations were conducted to assess the impacts of the learning rate, the request volume, and the size of the input data of the function. The experimental results indicate that, compared with Particle Swarm Optimization (PSO) and Genetic Algorithm (GA), under different scenarios of request scales, the average system delay is reduced by at most 31%, and the average energy consumption is reduced by at most 24%.

Local 5G networks represent an emerging architecture in the 5G architecture, where local micro operators (MOs) reuse public mobile networks to support differentiated service transmission capacity and coverage requirements. 5G Radio Access Network (RAN) slicing technology offers a solution that allows for the flexible deployment of heterogeneous services as slices sharing the same infrastructure. In this paper, a multi-micro operators scenario with deployed RAN slicing is considered, where interference price is used to incentivize local micro operators to optimize their transmission power and resource block allocation strategies, reducing interference to mobile network users and enhancing transmission efficiency. We formulate the competitive interaction between mobile network operator (MNO) and local MOs as a two-stage Stackelberg game, with the MNO as the leader and MOs as followers. The MNO is responsible for establishing the interference price that MOs need to pay for their communication. MOs decide on their transmission power strategies to satisfy user-customized slice requirements by solving the game. A resource management scheme based on multi-agent reinforcement learning is proposed, introducing a game-theoretic equilibrium solution to determine resource block allocation strategies, ensuring slice isolation while increasing operator revenue. Experimental results demonstrate that our approach outperforms standalone reinforcement learning strategies in terms of transmission rates and interference price payments.

Multivariate time series (MTS) modeling has become omnipresent in extensive areas. A primary challenge in MTS data modeling is capturing the intricate series dependencies. Mainstream modeling strategies include channel-independent (CI), channel-dependent (CD), and their joint versions. Recently, supervised frameworks based on joint strategies have achieved remarkable success in MTS modeling, but they are typically designed for specific tasks. In contrast, self-supervised pre-training frameworks have shown promise in Masked Time-series Modeling (MTM) for benefiting various tasks. However, existing frameworks often overlook the inter-series dependencies across time, referred to as cross-series dependencies, in MTS. This paper thus presents JCCMTM, a Joint CI and CD (JCC) strategy-based pre-training framework for MTM. JCCMTM leverages both intra-series and cross-series dependencies in MTS data to reconstruct masked time-series segments, encouraging the model to focus on relationships between channels. To effectively model cross-series dependencies, we propose the Time-Series-as-Sentence (TSaS), which incorporates cross-series contextual information of MTS segments. Furthermore, JCCMTM introduces a novel embedding transformation paradigm, the Uni-Mul Transformation, to address the embedding alignment issues that arise when applying JCC to MTM. Additionally, two optimization schemes, based on sparse attention and global tokens, respectively, are proposed to reduce JCCMTM's computational complexity. Experimentally, JCCMTM demonstrates outstanding fine-tuning performance compared to the most advanced time series supervised and pre-training methods in two canonical time series analysis tasks: long-term forecasting and anomaly detection. The code for JCCMTM is available at https://github.com/Torea-L/JCCMTM.

With the vigorous development of the Internet of Things (IoT), the demand for user equipment (UE) computing capacity is increasing. Multiaccess edge computing (MEC) provides users with high-performance and low-latency services by offloading computational tasks to the nearest MEC server-configured 5G radio access network (RAN). However, these computationally intensive tasks may lead to a sharp increase in the energy consumption of UE and cause downtime. In this paper, to address this challenge, we design an intelligent scheduling and management system (ISMS) to jointly optimize the allocation of MEC resources and wireless communication resources. The resource allocation problem is a mixed-integer nonlinear programming problem (MINLP), an NP-hard problem. The ISMS models this problem as an MDP with a state, action, reward, and policy and adopts a modified deep deterministic policy gradient (mDDPG) algorithm to ensure the weighted minimization of the energy consumption, latency, and cost of users. The simulation results show that the ISMS can effectively reduce the system’s energy consumption, latency, and cost. The proposed algorithm can provide more stable and efficient performance than other algorithms.

Supporting artificial intelligence (AI) models training is one of the visions for future 6th generation (6G) networks. An extensive quantum of data and computational capabilities are necessitated for the training of AI models. However, with the development of AI models, it is evident that the existing edge computing network architectures are inadequate to meet the massive computing power and communication demands of distributed training for models with a growing number of parameters. In this paper, we propose a distributed training framework based on the edge-network-cloud architecture. Considering the architecture of the network and the computing capabilities of network nodes, the framework actively adapts the functional partitioning and allocation of data of the network nodes during the process of distributed training. Specifically, aggregation nodes are responsible for parameter aggregation and updating, while training nodes execute training tasks and transmit model gradients to the aggregation nodes asynchronously. To improve training efficiency and reduce communication time, we introduce a solution based on Deep Reinforcement Learning (DRL). The algorithm intelligently allocates suitable data to nodes and selects node types by task-related information, thus accelerating distributed training across network nodes. Experimental results demonstrate that the proposed algorithm effectively accelerates large-scale model training tasks.

Load balancing technology, involving the adjustment of Configuration and Optimization Parameters (COPs), is crucial for enhancing network performance. Improper parameter adjustment directly impacts network Key Performance Indicators (KPIs). Therefore, it is essential to estimate the effect of parameter adjustment. Recent methods indicate that incorporating causal inference into parameter adjusting effect estimation contributes to building more robust models. However, the causation studies remain difficult. Due to the incomplete observation of the data, unobserved covariates lead to spurious correlations between COPs and KPIs, making the results of the model bias in practical scenarios. In this paper, we define the task of COP-KPI modeling under causal inference and design the Causal Dynamic Variational Deconfounder (CDVD) for estimating the effect of parameter adjustment based on a Variational Autoencoder-Bidirectional Long Short-Term Memory (VAE-BiLSTM) hybrid model. Specifically, leveraging a variational autoencoder, we learn the representation of hidden confounders using observed covariates. Subsequently, we apply this representation to predict KPIs values. Domain adversarial approach is used to mitigate confounding bias, achieving unbiased modeling of COP-KPI. We conduct experiments on synthetic datasets and real-world datasets. Experiments show that the proposed CDVD can improve predictive performance by 65 % on real-world wireless network datasets. We also demonstrate the robustness of our CDVD against paramater adjustment.

In the rapidly evolving landscape of telecommunications, Radio Access Network (RAN) optimization is critical for maintaining high network performance and adapting to diverse service requirements. Traditionally, RAN optimization has relied heavily on manual adjustments by human experts, lacking in intelligent decision-making model. However, applying decision-making model to RAN optimization is challenged by high interaction costs and feedback delays. To address these challenges, we introduce a World Model aided Parameter Adjustment Decision and Evaluation System (WMDE), utilizing a World Model framework with offline reinforcement learning to adjust RAN parameters. WMDE, integrating the Transformer-Informed Adjustment Decision Net (TADNet) and the Causal Adjustment Effect Evaluation Net (CAENet). The WMDE system sidesteps real-time network interaction in model training with CAENet's causal estimation, cutting interaction costs. Meanwhile, TADNet employs its Transformer structure and data processing to provide a long-term, global perspective on adjustment effects, reducing feedback delay issues. Utilizing real-world operational RAN parameter adjustment data, our experiments validate the effectiveness of WMDE in decision-making for RAN parameter adjustment.

Network slicing can provide logically isolated networks on the shared network infrastructure by invoking multiple technologies and administrative domains to fulfill end-to-end (E2E) service level agreements (SLAs). To guarantee the E2E service communication quality in the sliced network, an SLA-based cross-domain orchestration framework is proposed in this paper. The framework includes an E2E cross-domain coordination orchestrator at the upper layer and multiple subordinate domain controllers. Furthermore, we design two latency equalization policies applied to the upper layer orchestrator to divide the latency budget for each lower layer domain. Based on the reinforcement learning approach, Double Deep Q-Network with Prioritized Experience Replay (DDQN-PER) and Pointer Network SFC Mapping (PN-SFC), intra-domain resource allocation/mapping algorithms are designed independently for the lower radio access network (RAN) and core network (CN) domain controllers, respectively. The above algorithms are used to jointly optimize the enhanced mobile broadband (eMBB) users service satisfaction level and maximize the number of E2E accessed users. Simulation results show that our proposed algorithm can effectively guarantee the eMBB users QoS and improve the network capacity.