Recent work has shown that diffusion models can serve as powerful neural rendering engines that can be leveraged for inserting virtual objects into images. However, unlike typical physics-based renderers, these neural rendering engines are limited by the lack of manual control over the lighting, which is often essential for improving or personalizing the desired image outcome. In this paper, we show that precise and controllable lighting can be achieved without any additional training, simply by supplying a coarse shadow hint for the object. Indeed, we show that injecting only the desired shadow of the object into a pre-trained diffusionbased neural renderer enables it to accurately shade the object according to the desired light position, while properly harmonizing the object (and its shadow) within the target background image. Our method, SpotLight, is entirely training-free and leverages existing neural rendering approaches to achieve controllable relighting. We show that SpotLight achieves superior object compositing results, both quantitatively and perceptually, as confirmed by a user study, outperforming existing diffusion-based models specifically designed for relighting. We also demonstrate other applications, such as hand-scribbling shadows and full-image relighting, demonstrating its versatility.

Non-terrestrial networks (NTNs) have been proposed by 3GPP as a next-gen infrastructure paradigm to enhance terrestrial network coverage and capacity. Among various NTNs, ultra-dense LEO satellite networks are holding immense potential to offload terrestrial traffic and support increasing demands of emerging services, such as real-time applications with stringent delay requirements. However, some challenges in ultra-dense LEO satellite networks can severely degrade transmission performance, such as satellite movement, handover, sun outage and electromagnetic interference. Multipath transport protocol Multipath QUIC (MPQUIC) can be a viable solution to provide aggregated bandwidth and robust transmission in such environments. While the Out-of-Order (OFO) problem of MPQUIC can significantly degrade throughput and even block the entire transmission when path characteristics are much diverse. This issue is further amplified in LEO satellite networks due to the inherent orbital dynamics and space environment effects, which makes it exceptionally difficult to guarantee the demands of real-time services. In this paper, we propose a Prediction-Based Multipath QUIC Scheduler (PBMS) to enhance the transmission performance of real-time applications in NTNs. It primarily comprises two integral components:Path ManagerandBlock Scheduler.Path Managerdynamically allocates packets across available paths by considering packet loss rates, round-trip time (RTT), and congestion window states. AndBlock Schedulercalculates the estimated delivery time for each block in sender buffer according to the allocation scheme, and prioritizes block transmission based on the real-time requirements. We implement PBMS in Kuiper K3 shell network simulated by NS-3. The experiments show that PBMS can significantly guarantee and improve the transmission performance of real-time applications in NTNs.

In the field of computational pathology, multiple instance learning (MIL) has become a critical modeling framework for analyzing gigapixel whole-slide images (WSIs). Data scarcity, particularly in survival analyses requiring long-term follow-up, is a prevalent challenge that is often mitigated through data augmentation. Although data augmentation methods for multi-instance classification have progressed, they remain insufficient for multi-instance regression in survival analysis. Compared to classification tasks, data augmentation for survival analysis faces two major challenges: i) higher quality requirements for pseudo-labels on newly generated samples, and ii) imbalanced data distribution, exhibiting the long-tail distribution. To address these challenges, this paper introduces InterMix, a novel plug-and-play multi-instance data augmentation scheme inspired by the concept of intervention. To provide high-quality pseudo-labels, sub-bag assessment rules are proposed to screen out stable and consistent sub-bags before mixing. To alleviate the long-tail problem, this paper proposes the complementary mix strategy to generate minority-augmented instances with high diversity based on the risk distribution of patients. Extensive experiments are conducted on three datasets to demonstrate the superiority of InterMix. In addition, it enables the model to focus more on key areas of concern for pathologists. Our source code has been made available at github.com/lingzhi-T/InterMIx.

Considering the mobility of high-speed railway (HSR) trains, we propose an interruption-avoiding transmission paradigm for content delivery in HSR communication systems, enabled by a space-terrestrial integrated network (STIN) that supports both out-carriage and in-carriage communications. Each content delivery task must be fully transmitted from a satellite or terrestrial base station to the corresponding carriage femto-cell access point (FAP) before handoff occurs. To minimize the maximum transmission latency, we formulate the Content delivery task Scheduling and Resource Allocation problem for both Out-carriage and In-carriage communications (CSRA-OI) over a long-term scheduling period. To solve the CSRA-OI problem, we decompose it into two interrelated sub-problems, i.e., a transmission rate control and in-carriage bandwidth allocation (RCIBA) sub-problem, and a content delivery task scheduling and out-carriage resource management (CSORM) sub-problem. We show that the optimal solution to the RCIBA sub-problem can be obtained by solving a series of linear programming problems. To solve the CSORM sub-problem, we propose a multi-agent deep reinforcement learning (DRL)-based framework assisted by a graph attention network (GAT). We present a novel training scheme for the CSORM framework, where the policy networks are updated rapidly in a distributed manner while the critic network and GAT are updated jointly at a slower rate. Extensive experiments validate the advantages of our approach.

The ultra-dense low earth orbit (LEO) satellite networks are becoming indispensable infrastructures for future sixth-generation (6 G) architectures by providing low-latency and high-speed communication. A LEO satellite network can be viewed as an autonomous system (AS), which necessitates seamless integration with terrestrial ASes using Border Gateway Protocol (BGP). However, traditional BGP and existing optimization have not sufficiently considered the significant overhead of BGP and its adaptability to high dynamics. In this paper, we propose a Proactive Hierarchical Border Gateway Protocol (PH-BGP) routing scheme and system tailored for ultra-dense LEO networks. Specifically, we develop an inter-domain routing scheme leveraging hierarchical convergence and proactive updating to balance routing overhead and network stability, which accelerates route convergence. Based on previous work, we develop a modular routing system, which utilizes a prediction module, decision module, hierarchical convergence mechanism module, and proactive routing update mechanism module to collaborate. Our simulation results on real-world typical Walkerdelta type LEO constellations demonstrate that PH-BGP is effective and superior compared to other routing schemes.

Ultra-dense low earth orbit (LEO) satellite networks (UDLSNs) present a viable solution to deliver high-speed, lowlatency Internet services. Data transmission scheduling stands as a key technology for guaranteeing high-speed Internet services. Nonetheless, data transmission scheduling in UDLSNs encounters some critical challenges, predominantly related to the complexity of network scale, resource contention and the stringency of user demands. In this paper, we focus on the typical application scenario of satellite network, namely Airborne Internet, and delve into the data delivery scheduling problem in UDLSN. We aim to maximize the number of successfully scheduled flows by jointly optimizing satellite-to-aircraft downlink subchannel access, intersatellite link path planning, and flow scheduling based on timeexpanded graphs. Given the large-scale nature of the network, we decouple the proposed problem into a downlink subchannel access problem and a flow scheduling problem. We further design the matching-based subchannel allocation (MSA) algorithm and$A^{*}$-based path planning and flow scheduling (APPFS) algorithm to solve the two subproblems, respectively. We use authentic civil aviation flight trajectories and$\mathbf{1 1, 9 2 6}$LEO satellites in the entire Starlink phase as simulation data to evaluate the effectiveness of the proposed algorithm. Simulation results validate that the proposed algorithm can satisfy stringent user latency requirements and improve the number of successfully transmitted data.

While Large Language Models (LLMs) have exhibited remarkable emergent capabilities through extensive pre-training, they still face critical limitations in generalizing to specialized domains and handling diverse linguistic variations, known as distribution shifts. In this paper, we propose a Test-Time Learning (TTL) paradigm for LLMs, namely TLM, which dynamically adapts LLMs to target domains using only unlabeled test data during testing. Specifically, we first provide empirical evidence and theoretical insights to reveal that more accurate predictions from LLMs can be achieved by minimizing the input perplexity of the unlabeled test data. Based on this insight, we formulate the Test-Time Learning process of LLMs as input perplexity minimization, enabling self-supervised enhancement of LLM performance. Furthermore, we observe that high-perplexity samples tend to be more informative for model optimization. Accordingly, we introduce a Sample Efficient Learning Strategy that actively selects and emphasizes these high-perplexity samples for test-time updates. Lastly, to mitigate catastrophic forgetting and ensure adaptation stability, we adopt Low-Rank Adaptation (LoRA) instead of full-parameter optimization, which allows lightweight model updates while preserving more original knowledge from the model. We introduce the AdaptEval benchmark for TTL and demonstrate through experiments that TLM improves performance by at least 20% compared to original LLMs on domain knowledge adaptation.

Open-domain dialogue systems aim to generate natural and engaging conversations, providing significant practical value in real applications such as social robotics and personal assistants. The advent of large language models (LLMs) has greatly advanced this field by improving context understanding and conversational fluency. However, existing LLM-based dialogue systems often fall short in proactively understanding the user's chatting preferences and guiding conversations toward user-centered topics. This lack of user-oriented proactivity can lead users to feel unappreciated, reducing their satisfaction and willingness to continue the conversation in human-computer interactions. To address this issue, we propose a User-oriented Proactive Chatbot (UPC) to enhance the user-oriented proactivity. Specifically, we first construct a critic to evaluate this proactivity inspired by the LLM-as-a-judge strategy. Given the scarcity of high-quality training data, we then employ the critic to guide dialogues between the chatbot and user agents, generating a corpus with enhanced user-oriented proactivity. To ensure the diversity of the user backgrounds, we introduce the ISCO-800, a diverse user background dataset for constructing user agents. Moreover, considering the communication difficulty varies among users, we propose an iterative curriculum learning method that trains the chatbot from easy-to-communicate users to more challenging ones, thereby gradually enhancing its performance. Experiments demonstrate that our proposed training method is applicable to different LLMs, improving user-oriented proactivity and attractiveness in open-domain dialogues. Code and appendix are available at github.com/wang678/LLM-UPC.

We present Zerocomp,an effective zero-shot 3D object compositing approach that does not require paired composite-scene images during training. Our method leverages ControlNet to condition from intrinsic images and combines it with a Stable Diffusion model to utilize its scene priors, together operating as an effective rendering engine. During training, Zerocompuses intrinsic images based on geometry, albedo, and masked shading, all without the need for paired images of scenes with and without compos-ite objects. Once trained, it seamlessly integrates virtual 3D objects into scenes, adjusting shading to create realistic composites. We develop a high-quality evaluation dataset and demonstrate that Zerocompoutperforms methods using explicit lighting estimations and generative techniques in quantitative and human perception benchmarks. Additionally, Zerocompextends to real and outdoor image compositing, even when trained solely on synthetic indoor data, showcasing its effectiveness in image compositing.

The ultradense low-Earth orbit (UD-LEO) satellite network has attracted significant attention recently due to its great potential in providing global Internet coverage and services. For the sake of improving performance and reliability, multiple network paths can be utilized for coordinated transmission. However, state-of-the-art multipath routing algorithms face the challenge when dealing with highly dynamic network characteristics (i.e., high-speed node movement, frequent topology changes) in such emerging networks. In this article, we propose a deep-reinforcement-learning-based multipath cooperative routing (DRL-MPCR) scheme for UD-LEO satellite networks, with the aim of enhancing routing discovery capability and improving multipath transmission performance. Two main building blocks of multipath transport protocol are considered: 1) routing discovery and 2) multipath scheduling. On the one hand, in order to cope with the highly dynamic satellite network, a DRL-based multipath routing discovery algorithm is proposed, where satellite agents independently make routing decisions according to the perceived local network state, so that multiple available paths can be obtained. On the other hand, to promptly make traffic scheduling according to the varying path conditions, a water filling algorithm-based multipath scheduling policy is designed, which aims to optimize the maximum path cost when multiple paths are utilized for cooperative transmission. Extensive simulation results demonstrate that the proposed DRL-MPCR scheme achieves more efficient routing discovery and better multipath transmission performance than existing ones.

The computing power network (CPN) offers exceptional computational capabilities and reliable network services, with significant potential for future applications. To achieve ubiquitous coverage and efficient computational resource allocation, CPN can be seamlessly coordinated with low-cost autonomous aerial vehicle (AAV)-based mobile computing platforms. This article investigates an efficient low-altitude AAV-assisted computing power and resource allocation mechanism tailored for urban environments. The aim is to ensure seamless scheduling and efficient processing of computational tasks across various computing devices at different layers of the CPN, while minimizing AAV energy consumption and ensuring flight safety. First, this article proposes an Urban AAV-assisted CPN task-allocation and AAV trajectory-management decision-making problem. The AAV works until it safely lands, aiming to minimize overall task processing delay and AAV energy consumption while ensuring fairness in task allocation. Then, a novel AAV-protection-based multiagent deep deterministic policy gradient (UP-MADDPG) algorithm is introduced. It offers dynamic management of secure computing and communication flight paths when facing building blockages. Finally, we compared the proposed algorithm with baseline algorithms across various metrics. Experimental results demonstrate that the proposed algorithm achieves lower and more balanced task execution delay and AAV energy consumption while also improving fairness.

This paper proposes a novel meta-learning based hyper-parameter optimization framework for wireless network traffic prediction (NTP) models. The primary objective is to accumulate and leverage the acquired hyper-parameter optimization experience, enabling the rapid determination of optimal hyper-parameters for new tasks. In this paper, an attention-based deep neural network (ADNN) is employed as the base-learner to address specific NTP tasks. The meta-learner is an innovative framework that integrates meta-learning with the k-nearest neighbor algorithm (KNN), genetic algorithm (GA), and gated residual network (GRN). Specifically, KNN is utilized to identify a set of candidate hyper-parameter selection strategies for a new task, which then serves as the initial population for GA, while a GRN-based chromosome screening module accelerates the validation of offspring chromosomes, ultimately determining the optimal hyper-parameters. Experimental results demonstrate that, compared to traditional methods such as Bayesian optimization (BO), GA, and particle swarm optimization (PSO), the proposed framework determines optimal hyper-parameters more rapidly, significantly reduces optimization time, and enhances the performance of the base-learner. It achieves an optimal balance between optimization efficiency and prediction accuracy.

The emerging drones enabled air-to-ground (A2G) communications in millimeter-wave (mm-wave) bands are facing the Doppler effect that arises from the inevitable wobbling of the drones. The fast time-varying channel for drones A2G communications can make the channel state information (CSI) from channel estimation outdated, i.e., imperfect CSI. In this article, we introduce two detectors to demodulate the received signal and get the instantaneous bit error probability (BEP) of a mm-wave drones A2G link under imperfect CSI. Based on the designed detectors, we propose an adaptive modulation scheme to maximize the average transmission rate under imperfect CSI by optimizing the signal transmission time subject to the maximum tolerable BEP. A power control policy is proposed in conjunction with the adaptive modulation to minimize the transmission power while maintaining both the BEP under the threshold and the maximized average transmission rate. Numerical results show that the proposed adaptive modulation scheme in conjunction with the power control policy can maximize the temporally averaged transmission rate while reducing the power consumption by up to 50% for wobbling mm-wave drones A2G links.

Ultradense low-Earth orbit (LEO) satellite-terrestrial network (ULSN) has evolved as a new paradigm to provide ubiquitous and high-capacity communications in next generation wireless networks. However, the direct LEO satellite broadband connectivity faces significant challenges in urban environments due to the masking effect, which limits the reliability and availability of communication links in ULSNs. To address this, reconfigurable intelligent surface (RIS) is emerging as a promising solution in ULSNs. In this article, we investigate RIS-aided downlink data transmission in urban environments of multiusers in ULSNs. We set up a mixed-integer programming (MIP) model for maximizing the sum rate of terrestrial users in ULSNs. To solve the complex MIP problem, we propose a two-phase joint optimization algorithm with a deep learning phase and an alternative optimization (AO) phase. In the deep learning phase, a deep neural network (DNN) algorithm is employed to obtain the optimal user association matrix based on the positions of terrestrial users and LEO satellites. Then in the AO phase, successive convex approximation is utilized to transform the nonconvex subproblems of beamforming and RIS phase design into convex formulations and iteratively solve them. Simulation results demonstrate that the proposed algorithm outperforms other baseline algorithms.

With the development of low-Earth orbit (LEO) satellites, ultra large-scale LEO satellite networks (ULSLSNs) hold immense potential to provide high-speed and reliable services in future communication systems. It offers substantial promise to meet emerging Internet traffic demands, such as remote real-time applications with stringent delay constraints (deadline). However, the complexity of ULSLSNs is significantly amplified by the satellite mobility, limited bandwidth, and more packet losses. The complex and dynamic network can greatly increase block transmission latency and may further degrade user’s Quality of Experience for real-time applications. To reduce block transmission latency and deliver more blocks before deadline for real-time applications, we propose a quick user datagram protocol Internet connection-enabled robust streaming transmission (QRST) framework deployed in ULSLSNs. This framework comprises four components and mainly involves an adaptive forward erasure coding (FEC) scheme and a deadline-driven block scheduler. The FEC scheme dynamically allocates redundancy based on current network conditions to reduce extra retransmission delay and balance bandwidth overhead. The block scheduler selects blocks for transmission to deliver more blocks before deadline, especially for high-priority blocks. Finally, we implement QRST over the network of Kuiper K3 Shell simulated by NS-3 and living video streaming applications are transmitted. The experiment results show that QRST can significantly enhance the transmission performance of video streaming applications in ULSLSNs compared with other mechanisms.

Monocular depth prediction has received significant attention in recent years. However, the impact of illumination variations, which can shift scenes to unseen domains, has often been overlooked. To address this, we introduce the first indoor scene dataset featuring RGB-D images captured under multiple illumination conditions, allowing for a comprehensive exploration of indoor depth prediction. Additionally, we propose a novel method, MI-Transformer, which leverages global illumination understanding through large receptive fields to capture depth-attention contexts. This enables our network to overcome local window limitations and effectively mitigate the influence of changing illumination conditions. To evaluate the performance and robustness, we conduct extensive qualitative and quantitative analyses on both the proposed dataset and existing benchmarks, comparing our method with state-of-the-art approaches. The experimental results demonstrate the superiority of our method across various metrics, making it the first solution to achieve robust monocular depth estimation under diverse illumination conditions. We provide the codes, pre-trained models, and dataset openly accessible athttps://github.com/ViktorLiang/midepth.

Accurately predicting the response to neoadjuvant chemotherapy (NCT) is crucial for gastric cancer treatment planning. Multimodal diagnostic methods enhance prediction accuracy by integrating images and clinical features, but they often overlook intra-class differences, causing feature entanglement, and single models struggle to capture dataset diversity. To address these issues, we propose Clustering Multimodal Ensemble Learning (CMEL) framework, which aligns and fuses image features from different domains with clinical features and performs decision fusion using an ensemble model. Specifically, we design the clustering process using Siamese network and memory bank to partition images in feature space and introduce contrastive clustering loss to enhance clustering properties and reduce feature entanglement across domains. We design Hierarchical Feature Alignment Fusion (HFAF) module to generate multi-level image features at different depths and then separately align and fuse them with clinical features, providing diverse features to cover variations in the dataset. We design Dynamic Multi-Classifier Decision Fusion (DMDF) module to predict on fused features using a gating mechanism supervised by pseudo-labels for dynamic weights, enhancing prediction across different domains. On our collected the Gastric Cancer Chemotherapy Response (GCCR) dataset, CMEL achieves an AUC of 76.09%, accuracy of 72.00%, PPV of 86.02%. These results demonstrate the feasibility of joint image and clinical data in predicting NCT response and provide valuable insights for gastric cancer treatment. The code is available at https://github.com/WHX0259/CMEL.

The ultra-dense low earth orbit (UD-LEO) satellite network is being vigorously developed due to its great potential in providing global coverage and services. For the sake of improved network performance in resource-constrained satellite networks, multipath schemes are being explored. However, state-of-the-art multipath routing algorithms face the challenge when dealing with highly dynamic satellite network features (i.e., frequent traffic variation, link failures) and fail to exploit the simple grid topology to design fast yet efficient traffic engineering (TE) approaches. In this paper, we propose a novel distributed TE scheme called Flexible Satellite Traffic Engineering (FlexSATE), which leverages global path computation coupled with distributed local routing decisions to improve the overall load balancing performance for ultra-dense LEO satellite networks. By constructing a minimum-hop binary tree (MHBT), we propose an MHBT-based k-segment Routing algorithm, which is capable of promptly discovering routing paths with low latency, high diversity, and good load balancing. To further enhance network transmission performance, we employ supervised learning into dynamic rate adaption, where FlexSATE employs centralized offline learning to derive insights from the globally optimal routing strategy and utilizes distributed deployment to predict the optimal distribution of traffic in real time. Our simulation results on a real-world typical Walker-delta type LEO constellation with 720 satellites show that FlexSATE outperforms some existing approaches with superior robustness and flexibility.

Ultra-dense low earth orbit (LEO) satellite terrestrial networks (ULSNs) are considered as a crucial component of future six generation (6G) networks, offering ubiquitous and massive services for various applications. However, for the development of advanced physical layer technologies for ULSNs, a comprehensive link-level simulation tool that integrates up-to-date satellite communication protocols becomes paramount and is urgently needed. In this paper, we develop a versatile simulator for the link-level performance analysis of ULSNs under the prevalent digital video broadcasting (DVB) standards. We first establish a complete satellite-terrestrial microwave channel model, taking practical factors such as rain attenuation, cloud attenuation, and Doppler frequency shift into consideration. Subsequently, the whole physical layer modules tailored for satellite-terrestrial microwave communication are implemented, including diverse physical layer modulation and coding schemes (MCSs). Furthermore, we realize adaptive coding and modulation (ACM) for adaptive channel performance simulation. Finally, comparative performance analysis using the established channel model is conducted to demonstrate the effectiveness of different MCSs of DVB standards. The complete link-level performance analysis based on our self-developed simulator can advance the field of satellite-terrestrial microwave communication and provide valuable insights for further exploration of ULSNs.

Colorectal cancer is a prevalent and life-threatening disease, where colorectal cancer liver metastasis (CRLM) exhibits the highest mortality rate. Currently, surgery stands as the most effective curative option for eligible patients. However, due to the insufficient performance of traditional methods and the lack of multi-modality MRI feature complementarity in existing deep learning methods, the prognosis of CRLM surgical resection has not been fully explored. This paper proposes a new method, multi-modal guided complementary network (MGCNet), which employs multi-sequence MRI to predict 1-year recurrence and recurrence-free survival in patients after CRLM resection. In light of the complexity and redundancy of features in the liver region, we designed the multi-modal guided local feature fusion module to utilize the tumor features to guide the dynamic fusion of prognostically relevant local features within the liver. On the other hand, to solve the loss of spatial information during multi-sequence MRI fusion, the cross-modal complementary external attention module designed an external mask branch to establish inter-layer correlation. The results show that the model has accuracy (ACC) of 0.79, the area under the curve (AUC) of 0.84, C-Index of 0.73, and hazard ratio (HR) of 4.0, which is a significant improvement over state-of-the-art methods. Additionally, MGCNet exhibits good interpretability.

The Ultra-Dense LEO Satellite Networks (UDLSN) has immense potential to provide low-latency and high-reliability services in future communication networks, owing to its global coverage, high capacity and reliable connectivity. However, the LEO networks usually suffer relatively high and variable transmission errors due to multipath, shadowing and handover. For delay-constrained video transmission, existing packet protection mechanisms frequently violate the constraint and degrade quality in such environments. In this paper, we propose a QUIC-Enabled Reliable Video Transmission Scheme (QRVTS) with adaptive Forward Error Correction (FEC) to reduce loss recovery time and enhance transmission performance especially for delay-constrained videos. Specifically, QRVTS incorporates an adaptive mechanism to dynamically adjust FEC redundancy based on the prevailing channel loss conditions and frame types. We evaluate our mechanism under multiple satellite scenarios with different network characteristics. The simulation shows significant gains in overview completion time and frame-level delivery delay for delay-constrained video transmission in LEO satellite scenarios.

In this article, a time-efficient mechanism is proposed to solve the joint unmanned aerial vehicle (UAV) base station deployment and user/sensor association (UDUA) problem aiming at maximizing the downlink sum transmission throughput and reducing the time of online computations. In this work, two relevant subproblems are decoupled from the joint UDUA problem: the first one is denoted as the user association subproblem, for certain UAV base station (UAV-BS) positions, this subproblem is settled for finding the strategy which matches aerial and ground nodes optimally. The second subproblem is the positioning of UAV-BSs in order to obtain the best possible solution to the user association subproblem from all possible positioning combinations for the UAV-BSs. In the proposed mechanism, the Kuhn–Munkres algorithm is used to solve the first subproblem as an equivalent bipartite matching problem. For the UAV-BS deployment subproblem, when the two user distributions own a high similarity, we theoretically prove that little performance decline will be introduced when the new user distribution’s optimal strategy is compared with choosing the optimal UAV-BS deployment strategy of stored user distributions. Based on our mathematical analyses, the similarity level between user distributions is well defined and becomes the key to solve the second subproblem. According to experimental findings, the proposed UDUA mechanism, when compared to benchmark approaches, can provide near-optimum system performance with regard to average downlink total transmission throughput and failure rate with significantly decreased computing time.

In this paper, we propose a meta-learning based cell-level network traffic prediction framework (ML-TP), which can provide the proper initial weight vector for the learning model of a new prediction task based on the prediction task’s meta-features. In the ML-TP, each prediction task forms a base-task, and a multi-layer long short-term memory (LSTM) network is constructed as its base-learner. Through fast Fourier transform (FFT) analyses of real-world network traffic data, we find that the five most dominating frequency components can capture the cell-level traffic variations, and hence can be used as a base-task’s meta-features. We prove that the well-trained weight vector of a previous base-task’s base-learner is likely to be a proper initial weight vector of a new base-task’s base-learner if the meta-features of the two base-tasks are close to each other in the Euclidean space. Accordingly, we propose a K-nearest neighbours (KNN) algorithm based meta-learner to deal with the meta-task in the ML-TP. Numerical tests show that the ML-TP can significantly increase the base-learners’ after-training prediction accuracy and learning efficiency in terms of the number of base-samples and the number of epochs needed in each base-learner’s fine-tuning progress.

We consider a heterogeneous network (HetNet), where multiple access points (APs) of potentially different transmission capacities serve users simultaneously via beamforming in the same spectrum band. We propose a beamforming framework that exploits multi-agent deep reinforcement learning (DRL) for the HetNet to maximize the system downlink sum-rate. In our framework, each AP acts as an agent, which is equipped with an online policy deep neural network (DNN) and an online Q-function DNN. The former generates an AP’s beamforming vector based only on local observations in a time slot, while the latter evaluates the appropriateness of this beamforming vector. We present a distributed-updating-centralized-rewarding scheme to train the policy DNNs and Q-function DNNs of all the APs in an online trial-and-error way. Under this scheme, all the APs take the system downlink sum-rate in a recent time slot (informed by a central controller) as their identical one-step reward. Trained by the experience items with centralized rewards in every time slot, the weight vectors of each AP’s local DNNs will be updated in the direction to the global optimum. Simulation results demonstrate that the proposed framework converges fast and outperforms the benchmark beamforming methods in terms of the system downlink sum-rate performance.

Modelling cellular traffic pattern plays a critical role to efficiently satisfy consumers’ demand, which is skewed distributed and fast-varying. Although current methods can indicate the traffic fluctuation of each cell, it is still not enough as optimisation techniques require to know the traffic distribution inside cells. In this paper, Neural Network is used to improve the resolution to intra-cell level by modelling the hotspots of geo-tagged Twitter. This paper is based on our previous work, which has already quantified the linear relationship between Tweets and mobile traffic. Here, a similarity measurement is designed to quantify how two patterns are similar to each other. We applied this measurement on the geo-tagged Tweets and found that in one of three periods (day, evening, and night), the hotspots distributions are more similar than the other periods. Then for each period, Generative Adversarial Networks (GAN) is used to train a generator for modelling the intra-cell hotspots distribution. Such a trained generator can also continuously generate convincing artificial-data to expand the data set. The similarity measurement gives high similarity (above 0.8) between generated artificial-data and the real test data.