In music recommendation systems, multimodal interest learning is pivotal, which allows the model to capture nuanced preferences, including textual elements such as lyrics and various musical attributes such as different instruments and melodies. Recently, methods that incorporate multimodal content features through semantic IDs have achieved promising results. However, existing methods suffer from two critical limitations: 1) intra-modal semantic degradation, where residual-based quantization processes gradually decouple discrete IDs from original content semantics, leading to semantic drift; and 2) inter-modal modeling gaps, where traditional fusion strategies either overlook modal-specific details or fail to capture cross-modal correlations, hindering comprehensive user interest modeling. To address these challenges, we propose a novel multimodal recommendation framework with two stages. In the first stage, our Progressive Semantic Residual Quantization (PSRQ) method generates modal-specific and modal-joint semantic IDs by explicitly preserving the prefix semantic feature. In the second stage, to model multimodal interest of users, a Multi-Codebook Cross-Attention (MCCA) network is designed to enable the model to simultaneously capture modal-specific interests and perceive cross-modal correlations. Extensive experiments on multiple real-world datasets demonstrate that our framework outperforms state-of-the-art baselines. This framework has been deployed on one of China's largest music streaming platforms, and online A/B tests confirm significant improvements in commercial metrics, underscoring its practical value for industrial-scale recommendation systems.

Transformer-based generative models have achieved remarkable success across domains with various scaling law manifestations. However, our extensive experiments reveal persistent challenges when applying Transformer to recommendation systems: (1) Transformer scaling is not ideal with increased computational resources, due to structural incompatibilities with recommendation-specific features such as multi-source data heterogeneity; (2) critical online inference latency constraints (tens of milliseconds) that intensify with longer user behavior sequences and growing computational demands. We propose Climber, an efficient recommendation framework comprising two synergistic components: the model architecture for efficient scaling and the co-designed acceleration techniques. Our proposed model adopts two core innovations: (1) multi-scale sequence extraction that achieves a time complexity reduction by a constant factor, enabling more efficient scaling with sequence length; (2) dynamic temperature modulation adapting attention distributions to the multi-scenario and multi-behavior patterns. Complemented by acceleration techniques, Climber achieves a 5.15× throughput gain without performance degradation by adopting a ''single user, multiple item'' batched processing and memory-efficient Key-Value caching.

With the progressive evolution of large language models (LLMs) and the increasing need of computing for$\mathbf{6 G}$, it becomes crucial for multiple network nodes with limited computing resources to share the need for large model inference. Model partition methods have been proposed to enable computation-intensive artificial intelligence (AI) services by splitting an AI model across multi-edge and cloud nodes. In this paper, a distributed inference optimization framework for transformer decoder-only based LLMs (DIO-LLMs) is proposed in edge-cloud collaborative networks. The partitioning and offloading strategy is determined based on the computing workload and network status. DIO-LLMs specifically accounts for the parallel execution capabilities of the transformer architecture. It employs a two-phase model partitioning strategy, comprising inter-layer and intra-layer partitions, to effectively distribute LLMs across edge and cloud nodes. Additionally, to mitigate inference latency under resource limitations, a Greedy Proximal Policy Optimization (GPPO) based algorithm has been developed to devise optimal strategies. Simulation results indicate that under memory constraints, the proposed algorithm can reduce inference latency more effectively than other baseline algorithms.

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.

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 3.1× improvement in compute speed compared to state-of-the-art (SOTA) models on the widely adopted ETT benchmark dataset.

Spaceborne LiDAR provides a promising method for large-scale characterizing LAI. However, the quality of point cloud data from spaceborne LiDAR, especially ICESat-2, is susceptible to atmosphere and background noise, introducing considerable uncertainty in LAI retrieval. Thus, efficiently screening out the high-quality point cloud is a significant guarantee for high-quality LAI retrieval. In this study, we proposed a quality control (QC) method that employed the number of 10 m windows without ground points in the ICESat-2 100 m segment as the QC flag. This method divided segments into 11 QC flags from 0 to 10 and was applied to LAI retrieval across Chinese forests from 2019 to 2020. The field measurements at locations identical to ICESat-2 ground tracks were used to validate the ICESat-2 LAI at different QC flags. The results showed that the proposed method effectively improved point cloud quality recognition and LAI accuracy, with ICESat-2 LAI (QC < 3) reducing RMSE by 26.36% compared to all ICESat-2 LAIs. It also showed good agreement with MODIS and GLASS LAI and mitigated saturation issues in passive optical imagery. The ICESat-2 LAI with QC < 3 performed better in deciduous broadleaved, evergreen needle-leaved, deciduous needle-leaved, and mixed forests, but not in evergreen broadleaved forests. ICESat-2 LAI was particularly adept at capturing high LAI values, which had the highest proportion of LAI values over 6.0 compared to MODIS and GLASS LAI. The proposed method has the potential for large-scale and high-quality LAI retrieval using ICESat-2 data on a global scale.

Representation learning remains one of the most important but challenging tasks within industrial music rec-ommendation systems. In the context of the Matthew effect, item exposure frequency demonstrates substantial inequality, leading to the Harry Potter problem for popular items and the long-tail issue for less interacted items, collectively impairing the adequacy and accuracy of representation learning. In this paper, to alleviate the negative impact of bias on representation learning in music recommendation systems, we propose a unified model based on introducing the unbiased Cascading Multimodal Feature, called CMF4Rec. Specifically, with our cascading feature enhancement module, we implement a dual-stage representation enhancement strategy. In the first stage, the pivotal subsequence is extracted from the coarse-grained similarity sequence derived from cascading multimodal features, which is subsequently ag-gregated to generate the enhanced representation of the candidate item. Moreover, in the feature interaction module, the enhanced representation is crossed with user behaviors to capture the diverse and dynamic interests of users. Furthermore, we employ contrastive learning and design an auxiliary contrastive task to provide high-quality gradients for the main recommendation task. We demonstrate the effectiveness of this model with extensive experiments on public and industrial datasets. Moreover, the deployment of CMF4Rec in a real music recommendation system has also yielded significant improvements.

Leaf area parameters are crucial in ecosystem studies. As ecophysiological models advance toward finer detail, accurately estimating LA at various scales becomes essential, particularly for diverse units like urban individual trees. Several algorithms based on terrestrial laser scanning (TLS) data have been developed to obtain the LA of individual trees. However, their use at the stand level needs further research. In this study, the comparative shortest-path algorithm (CSP) is introduced for the automatic individual tree segmentation, thereby facilitating the application of the path length distribution model (PATH) for leaf area estimation at the stand level. Using high-density TLS data, we presented a bottom-up estimation of stand leaf area index (LAI) from 50 individual tree measurements and validated the results at different scales. At the tree scale, the LA derived from TLS and allometric model were highly correlated, with an R-value of 0.83. At the stand scale, the proposed method provides consistent results with the allometric and TRAC instrument measurements, performing better than vertical upward photography. Generally, 23 shared stations under the forest are enough to accurately obtain the LA of 50 trees and the LAI in an urban forest stand. Sensitivity analysis shows that the method is not sensitive to TLS scan resolution and parameters used in tree crown envelope reconstruction. The proposed bottom-up approach provides a new way of estimating the LAI at stand level using TLS and has the advantage of providing multi-level leaf area information and avoiding the scale effect.

Abstract Nowadays, wireless communication system is facing the problems of spectrum resource shortage. Cognitive radio technology allows cognitive users to use the spectrums authorized to primary users to improve the spectrum utilization. In this paper, a cognitive network model based on hybrid overlay–underlay spectrum access mode is established. To solve the resource allocation problem, a multi‐agent resource allocation algorithm based on graph convolution reinforcement learning which combines deep Q network (DQN) and graph attention network is proposed. DQN is used for action selection and graph attention network is used to obtain the information about neighbours, so as to achieve local cooperation. The proposed algorithm can adaptively optimize cognitive network throughput, spectrum efficiency, or power efficiency by controlling the transmission power and channel selection of cognitive users. To improve the information interaction efficiency, the agent's states are divided into two categories, whether it needs to interact with neighbours or not, which shortens training time and improves convergence speed. Simulation results show that the proposed algorithm can effectively improve the power efficiency of cognitive networks. Compared with Q‐learning, DQN and exiting graph convolutional reinforcement learning algorithm, the proposed algorithm has faster convergence speed and higher stability, and obtains higher network power efficiency.

The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) employs a unique multibeam photon counting approach to acquire a near-continuously sampled profile and provides more precise technology for mapping the leaf area index (LAI) at the global scale. The inversion accuracy of LAI is affected by the clumping effect, which has been an open question for spaceborne laser scanning (SLS). Here, we present a segmented method based on the path length distribution model to calculate the clumping-corrected LAI independently using ICESat-2 data. The results showed that the LAI derived by the proposed method with a 200 m segment was consistent with the airborne laser scanning (ALS)-derived LAI, with a root mean squared error (RMSE) of 0.37. A satisfactory agreement (RMSE$=1.03$) was also shown between moderate resolution imaging spectroradiometer (MODIS) LAI and ICESat-2 LAI. Moreover, the LAI derived by the proposed method was on average 31.72% higher than the LAIe derived by Beer’s law, which indicated that the proposed method achieved the purpose of correcting the clumping effect. The gap probability was calculated by the 200 m moving window and the path length distribution was obtained by the 1 m moving window as the model input had the highest accuracy. In addition, the limitation of the point cloud data and the time lag of ICESat-2 acquisitions and ALS observations may affect the inversion accuracy of LAI. This study proposed a feasible way to correct the clumping effect and invert LAI independently using ICESat-2 data, which has the potential to characterize vegetation structure precisely at regional and global scales.

The temperature and relative humidity (hereafter called humidity) in the indoor environment is closely related to the operation of its control system. The centralized control system, which has identical air inlets not only leads to uneven indoor temperatures and humidity but also highly controlled latency when interference occurs. To address this challenge and improve the precision and uniformity of temperatures and humidity in the indoor environment, we propose a distributed temperature and humidity control (DTHC) framework based on deep reinforcement learning (DRL). In this work, we use a constant temperature and humidity air-conditioning (CTHA) system for a museum as a case study to validate the optimization performance of the proposed controller. The state–action space, reward function, and DRL network structure are proposed. The air flow rate of multiple air inlets of CTHA is adjusted according to the feedback from the distributed temperature and humidity sensors. We develop a DRL-based computational fluid dynamics (CFD) experiment platform to evaluate the proposed mechanism. The experiment results show that our approach can improve the precision and uniformity of the temperature and humidity while enhancing the anti-interference capability of the control system. The adjustment time and energy consumed to reach the desired indoor air temperatures and humidity are reduced compared with rule-based methods.

The derivation of microwave land surface emissivity (MLSE) under various weather conditions from the microwave radiometer plays a crucial role in acquiring land surface and atmospheric parameters. Nevertheless, currently, most existing studies mainly focus on the clear-sky scenarios owing to a lack of cloudy-sky land surface temperature (LST) and uncertainties in simulating the scattering and emission properties of atmospheric hydrometeors. Under this background, with satellite observations and the random forest (RF) model, this study proposes a method to estimate the MLSE under cloudy skies. First, clear-sky MLSEs with satisfactory accuracy are retrieved by using the brightness temperatures (BTs) from the Advanced Microwave Scanning Radiometer-Earth sensor, LSTs from the Moderate Resolution Imaging Spectroradiometer, and atmospheric profiles from the ERA5 reanalysis. Then, the relation among the clear-sky MLSE and related impact factors is built with the RF and extended to the cloudy-sky environment for generating all-weather MLSEs with a 0.25°. The results show that the input datasets present a considerable impact on the calculation of instantaneous MLSE, and a 5.73 K bias of ERA5 LST may generate a 0.014-0.021 error in the MLSE from 6.9 to 89 GHz horizontal polarization, while the impacts of BT and profile uncertainties on the MLSE are smaller. The retrieved clear-sky MLSE is coincident with the existing MLSE for the spatiotemporal variations, and there is an average difference range from -0.035 to 0.035 in January 2008. Meanwhile, the constructed RF model can successfully apply to cloudy-sky status and recover the MLSE image gaps affected by cloud contamination.

Investigating the relations between land surface temperature (LST) and biophysical compositions can help the understanding of the surface biophysical process. However, there are still uncertainties in determining the impacts of biophysical compositions on LST due to the atmospheric effects. In this article, four atmospheric correction algorithms were used to correct 12 Landsat 8 images in Xi'an, Beijing, Wuhan, and Guangzhou, China, including the Atmospheric Correction for Flat Terrain (ATCOR2), Quick Atmospheric Correction (QUAC), Fast Line-of-sight Atmospheric Analysis of Spectral Hypercube (FLAASH), and Second Simulation of Satellite Signal in the Solar Spectrum (6S). Then, geodetector was used to investigate the atmospheric correction differences in the spatial heterogeneity relationships between LST and normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI), and bare soil index (BSI). Results indicate that the selected composition factors were greatly improved after atmospheric correction, and the relations between LST and three factors were characterized by obvious atmospheric correction differences in four study areas. On the whole, the 6S algorithm performed the best in improving the factor values and impacting the spatial heterogeneity relations between LST and biophysical compositions, followed by FLAASH, QUAC, and ATCOR2 algorithms. Except for Wuhan, 6S, FLAASH, and QUAC algorithms significantly enhanced the correlation between LST and NDVI. However, all algorithms weakened the correlations between LST, NDVI, and BSI, except Guangzhou. These findings have been verified using the regression analysis. In addition, with geodetector, combinations of any two composition factors all had strongly enhanced impacts on LST, and a combination between NDVI and NDBI performed the strongest in most cases.

This paper investigates the network slicing in the virtualized wireless network. We consider a downlink orthogonal frequency division multiple access system in which physical resources of base stations are virtualized and divided into enhanced mobile broadband (eMBB) and ultrareliable low latency communication (URLLC) slices. We take the network slicing technology to solve the problems of network spectral efficiency and URLLC reliability. A mixed-integer programming problem is formulated by maximizing the spectral efficiency of the system in the constraint of users’ requirements for two slices, i.e., the requirement of the eMBB slice and the requirement of the URLLC slice with a high probability for each user. By transforming and relaxing integer variables, the original problem is approximated to a convex optimization problem. Then, we combine the objective function and the constraint conditions through dual variables to form an augmented Lagrangian function, and the optimal solution of this function is the upper bound of the original problem. In addition, we propose a resource allocation algorithm that allocates the network slicing by applying the Powell–Hestenes–Rockafellar method and the branch and bound method, obtaining the optimal solution. The simulation results show that the proposed resource allocation algorithm can significantly improve the spectral efficiency of the system and URLLC reliability, compared with the adaptive particle swarm optimization (APSO), the equal power allocation (EPA), and the equal subcarrier allocation (ESA) algorithm. Furthermore, we analyze the spectral efficiency of the proposed algorithm with the users’ requirements change of two slices and get better spectral efficiency performance.

This paper presents our approach for the group cohesion prediction sub-challenge in the EmotiW 2019. The task is to predict group cohesiveness in images. We mainly explore several regularizations with three types of visual cues, namely face, body ,and global image. Our main contribution is two-fold. First, we jointly train the group cohesion prediction task and group emotion recognition task using multi-task learning strategy with all visual cues. Second, we elaborately design two regularizations, namely a rank loss and a hourglass loss, where the former aims to give a margin between the distance of distant categories and near categories and the later to avoid centralization predictions with only MSE loss. With careful evaluations, we finally achieve the second place in this sub-challenge with MSE of 0.43821 on the testing set. https://github.com/DaleAG/Group_Cohesion_Prediction

This paper presents our approach for the engagement intensity regression task of EmotiW 2019. The task is to predict the engagement intensity value of a student when he or she is watching an online MOOCs video in various conditions. Based on our winner solution last year, we mainly explore head features and body features with a bootstrap strategy and two novel loss functions in this paper. We maintain the framework of multi-instance learning with long short-term memory (LSTM) network, and make three contributions. First, besides of the gaze and head pose features, we explore facial landmark features in our framework. Second, inspired by the fact that engagement intensity can be ranked in values, we design a rank loss as a regularization which enforces a distance margin between the features of distant category pairs and adjacent category pairs. Third, we use the classical bootstrap aggregation method to perform model ensemble which randomly samples a certain training data by several times and then averages the model predictions. We evaluate the performance of our method and discuss the influence of each part on the validation dataset. Our methods finally win 3rd place with MSE of 0.0626 on the testing set. https://github.com/kaiwang960112/EmotiW_2019_ engagement_regression

In this paper, multi-cell power allocation approach is researched. Different from the traditional optimization decomposition method, Deep Reinforcement Learning (DRL) method is employed to solve the power allocation issue which is an NP-hard problem. The objective of our work is to maximize the overall capacity of the entire network in the scenario where the base stations are randomly and densely distributed. We propose a wireless resource mapping method and a deep neural network for multi-cell power allocation named as Deep-Q-Full-Connected-Network (DQFCNet). Compared with the water-filling power allocation and Q-learning method, DQFCNet can achieve a higher overall capacity. Furthermore, the simulation results show that DQFCNet has significant improvement in convergence speed and stability.

Nowadays, with the increasing awareness of environmental and economic issues, energy efficiency has received an enormous amount of attention. In this paper, we investigate an energy-saving algorithm called Greedy Base Station Sleeping (G- BSS), based on clustering for Device-to-Device (D2D) communication in cellular network and derive the energy utility function to evaluate energy consumption of the implementation scenario. In addition, a novel user association scheme is developed for the first time to solve the communication problem of users in sleep cells, in which the D2D clusters in sleep cells associate to the selected inter-CH (Cluster Head) in neighbor cells according to location information and residual energy, namely forming a merger cluster. Extensive simulation results confirm that the proposed algorithm reduces efficiently energy consumption and achieves low delay.

The purpose of this study was to determine whether acupuncture is effective in reducing pain and swelling around the knee and improving range of motion (ROM) during the early rehabilitation after anterior cruciate ligament(ACL) reconstruction. Following ACL reconstruction, 80 patients were randomly assigned to either an acupuncture treatment group (Group A) or a control group (Group C). In Group A, the complementary treatment of acupuncture was performed three times week from postoperative day 7 until postoperative day 21. Outcome measures were:1) pain as assessed by a visual analog scale; 2) reduction of swelling around the knee as indicated by its circumference at the center of the patella; and 3) ROM of the affected knee.

Objective: To analyze the distinguishing feature in use of Chinese drugs on cervical spondylosis patients in perimenopausal period through data mining methods. Methods: Chinese medicine recipes for cervical spondylosis patients in perimenopausal period were collected and recorded in the database, and then the correlation coefficient between herbs, core combinations of herbs were analyzed by using modified mutual information, complex system entropy cluster, respectively. Results: Based on the analysis of Chinese medicine recipes for cervical spondylosis patients in perimenopausal period, drugs with high-frequency occurrence in these recipes, frequently-used herb pairs and core combinations, et al. will been founded. Discussion: From the result, we may find the main Chinese medicines for cervical spondylosis patients in perimenopausal period, and the therapeutic principle for cervical spondylosis patients in perimenopausal period, and the method of treatment for cervical spondylosis patients in perimenopausal period. These data may help to reveals the relationship between cervical spondylosis and the physiological and pathology of the menopausal transition.

To analyze the component law of Chinese medicines for ankylosing spondylitis, and develop new prescriptions for ankylosing spondylitis through mutual information mining methods. Chinese medicine recipes for ankylosing spondylitis were collected and recorded in the database, and then the correlation coefficient between herbs, core combinations of herbs and new prescriptions were analyzed by using modified mutual Information, complex system entropy cluster and unsupervised hierarchical clustering, respectively. Based on the analysis of 61 Chinese medicine recipes for ankylosing spondylitis, 26 drugs with high-frequency and the distribution of Four Qi and Five Flavors of Chinese medicines occurrence in these recipes, 23 frequently-used herb pairs and 12 core combinations were founded, and 6 new recipes for ankylosing spondylitis were developed. Chinese medicines for ankylosing spondylitis were mainly consist of kidney yang-tonifying Medicinal, blood-activating and stasis-dispelling edicinal, wind-dampness dispelling and cold dispersing medicinal. The therapeutic principle for ankylosing spondylitis is to reinforce the healthy qi and eliminate the pathogenic factors. Radix Rehmanniae Praeparata, Radix Glycyrrhizae, Radix Paeoniae Rubra, Rhizoma Chuanxiong, Radix Achyranthis Bidentatae, Radix Angelicae Sinensis, Rhizoma Cibotii, Radix Dipsaci, Herba Taxilli, Corthex Eucommiae, Radix Angelicae Pubescentis and Radix Astragali are the main herbal drugs used in the treatment of ankylosing spondylitis.

To analyze the component law of Chinese medicines for rheumatoid arthritis, and develop new prescriptions for rheumatoid arthritis through mutual Information mining methods. Chinese medicine recipes for rheumatoid arthritis were collected and recorded in the database, and then the correlation coefficient between herbs, core combinations of herbs and new prescriptions were analyzed by using modified mutual information, complex system entropy cluster and unsupervised hierarchical clustering, respectively. Based on the analysis of 63 Chinese medicine recipes for rheumatoid arthritis, 30 drugs with high-frequency and the distribution of Four Qi and Five Flavors of Chinese medicines occurrence in these recipes,34 frequently-used herb pairs and 9 core combinations were founded, and 3 new recipes for rheumatoid arthritis were developed. Chinese medicines for rheumatoid arthritis were mainly consist of blood-activating and stasis-dispelling edicinal, wind-dampnessdispelling and heatclearing medicinal, heat-clearing and detoxicating medicinal, tonifying and replenishing medicinal. So, the therapeutic principle for rheumatoid arthritisis to reinforce the healthy qi and eliminate the pathogenic factors. Radix Gentianae Macrophyllae, Rhizoma Chuanxiong, Radix Glycyrrhizae,Radix Angelicae Pubescentis,Radix Paeoniae Alba,Radix Angelicae Sinensis,Radix Achyranthis Bidentatae,Ramulus Cinnamomi,Rhizoma Et Radix Notopterygii and Radix Astragali are the main herbal drugs used in the treatment of rheumatoid arthritis.

This work presents a novel method to detect an important flow feature, vortices, in the ocean. Our method can detect closed streamlines around vortex cores. Coupled with existing vortex core detection, the entire vortex area, which is the combination of the vortex core and surrounding streamlines, can be detected. A variety of feature extraction methods are presented, and those more pertinent to this study are implemented. Detection results are evaluated in terms of accuracy, clarity and usability.