Abstract Multimodal sentiment analysis utilizes different types of information to improve sentiment recognition, driving research at the intersection of natural language processing and computer vision. However, practical applications often encounter missing image modal data due to collection restrictions, privacy protection, and other issues, which reduces the accuracy of sentiment analysis. Recently, diffusion models, as an emerging generative model, have demonstrated remarkable capabilities in handling modality completion and generation tasks. To this end, this paper presents an expert feature-aligned diffusion model (EFADM) integrated with curriculum learning, aiming to efficiently solve multimodal sentiment analysis tasks by dynamically generating missing modality data and optimizing feature alignment between modalities. EFADM consists of three primary modules: the Adaptive Noise Suppression Module (ANSM), the Expert Alignment Module (EAM), and a progressive generation framework based on curriculum learning. First, ANSM dynamically adjusts noise suppression during the diffusion process to enhance the quality of generated features. EAM strengthens semantic consistency between generated and existing modalities by aligning expert features. Additionally, the progressive generation strategy utilizing curriculum learning gradually increases task difficulty, thereby optimizing the model’s performance across different noise levels. Compared with existing methods, the proposed approach achieves leading performance on the publicly available MVSA series datasets.

Abstract Recently, click-through rate (CTR) prediction research based on empirical risk minimization has achieved remarkable results, which assumes that training and test data follow the same distribution and optimizes CTR prediction models by minimizing prediction errors in the training data. However, this assumption does not hold in the real world. Specifically, user preferences change over time, leading to ‘temporal drift,’ where the distribution of test data differs from that of the training data. In this paper, we propose a ‘temporal drift adaptation framework’ (TDAF) for CTR prediction to cope with temporal drift. In TDAF, we devise a feature embedding predictor to learn the evolution of user preferences from historical feature embeddings and simulate feature embeddings after temporal drift. The model’s performance on simulated and real feature embeddings is improved through a novel bi-level optimization based on meta-learning, enhancing its ability to cope with temporal drift. TDAF is model-agnostic and can be widely applied to common CTR prediction models. We conduct extensive experiments, and the results show that TDAF improves the performance of CTR prediction models by an average of 0.55% across multiple datasets. Theoretical analysis and ablation studies further validate the effectiveness of TDAF. The code is available at https://github.com/lgxccc/TDAF-for-CTR-prediction/tree/main.

Image segmentation remains a pivotal challenge in computer vision, particularly in complex scenarios requiring fine-grained feature discrimination. Current approaches often suffer from inefficient feature utilization and local detail loss during semantic segmentation. To address these limitations, we propose a novel deep neural network with multi-scale attention fusion for accurate fine-grained image segmentation and the lightweight architecture ensures computational efficiency without sacrificing accuracy. Our approach integrates three key components: the Dynamic Spatial-Atrous Spatial Pyramid Pooling (DSA-ASPP) module, which combines depthwise separable convolution with adaptive dilation rates to reduce parameters; a multi-scale attention fusion mechanism which hierarchically integrates features to enhance local texture discriminability and minimizing computational overhead. and the PreactResNet-ECA, a pre-activated residual network with channel-wise attention optimized for fine-grained feature interaction. Experimental results on CamVid and Cityscapes datasets demonstrate the superior performance of our proposed model, achieving mean intersection-over-union (mIoU) scores of 69.6% and 73.6%, respectively, with inference speeds reaching 255.8 FPS. Furthermore, evaluations on fine-grained datasets (CUB-200-2011 and Stanford Dogs) reveal that our PreactResNet-based model outperforms state-of-the-art approaches, attaining accuracies of 93.0% and 97.0%. The framework effectively preserves local texture details, reduces pixel-level misclassification, and offers a balanced trade-off between accuracy and computational efficiency.

With the continued growth of the online video platform market, finding an effective business model has become one of the main issues for video providers. This paper proposes an advertising incentive model to maximize video providers' revenue and investigate users' motivation to obtain premium services based on value differentiation and snob effects. Considering the impact of ad loss on advertisers, we establish the existence of an equilibrium in our proposed hybrid revenue model using a two-stage Stackelberg model. We find that more aggressive advertising incentives lead to an increase in the proportion of free users, and the threshold for free users to become paid users is closely related to the disutility of ads and the strength of advertising incentives. Experimental results show that video providers can use personalized recommendations to enhance user utility. When the advertising incentives are large enough, users should choose the free mode. This provides theoretical support for optimizing the revenue model of online video platforms.

Abstract Through knowledge distillation method, a student model can imitate the output of a teacher model to improve its generalization ability without changing the computational complexity. However, in existing knowledge distillation research, the efficiency of knowledge transfer is still not satisfactory, especially from pre-trained language models (PTMs) like Robustly optimized BERT approach (RoBERTa) to another structure student model. To address this issue, this paper proposes a prediction framework (RTLSTM) for Chinese emotion classification based on knowledge distillation. In RTLSTM, a new triple loss strategy is proposed for training a student ‘BiLSTM’, which combines supervised learning, distillation and word vector losses. This strategy enables the student to learn more fully from a teacher model RoBERTa and retains 99% of the teacher models’ language understanding capability. We carried out emotion classification experiments on five Chinese datasets to compare RTLSTM with baseline models. The experiment results show that RTLSTM outperforms the baseline models belonging to the RNN group in terms of prediction performance under similar numbers of parameters. Moreover, RTLSTM is superior to the PTMs group baseline models through 92% fewer parameters and 83% less prediction time under comparable prediction performance.

Data-driven crowd evacuation learning methods are often used to enhance the realism of crowd simulation. However, the learning results of traditional methods cannot adapt to the dynamic changes of the simple scene, and thus have the disadvantage of poor generalization. To solve this problem, we propose a data-driven crowd evacuation framework based on hierarchical deep reinforcement learning. The framework consists of: a macro-control layer with path programming function and a micro control layer with collision avoidance function. In this paper, a path programming method combining data-driven and deep reinforcement learning is proposed in the macro-control layer. The method combines the pedestrian motion attributes in the video with the DDPG algorithm to learn the pedestrian track in the video from a macro perspective. In the micro-control layer, the track sequence learned in the macro-control layer is used as the motion target to learn the collision-free motion velocity of individuals using the multiple agent deep reinforcement learning method. When the scene changes, the micro-control layer adaptively adjusts the motion speed without the need for the macro-control layer to repeat the path programming learning. The experimental results demonstrate that the proposed hierarchical crowd evacuation framework can not only simulate the real crowd movement behavior and improve the simulation fidelity, but also flexibly adapt to the dynamic changes of the simple scene and enhance the generalization.

In a peer-to-peer (P2P) video-on-demand (VoD) system, each peer contributes a limited disc storage and stores some watched movies to offload the servers when these movies are requested. When the contributed disc storage of a peer is full, to minimize the server load, which movie should be replaced is a key design problem for P2P VoD systems. This problem is a P2P replication problem. Previous studies on this problem mainly consider content availability, but fail to consider bandwidth availability of peers on the system level. In this paper, assuming that movie popularity is known, we first analyze bandwidth availability of peers on the system level, and then formulate the replication problem as a minimization problem. Based on the minimization problem, we derive two design guidelines. According to the two design guidelines, we propose a bandwidth-availability-based replication algorithm aiming at minimizing the server load, called BAB algorithm. BAB algorithm can make the replicas’ distribution towards the optimal distribution in a distributed way. Furthermore, we consider some practical implementation issues of BAB algorithm. Through extensive simulations, we demonstrate that BAB algorithm outperforms the previously proposed algorithms in terms of reducing the server load and improving the streaming quality, in stable environment and dynamic environment, respectively.

The peer churn rate in the peer-to-peer (P2P) video-on-demand streaming service is much higher than in the P2P live streaming service, which makes the data scheduling problem more challenging. First, the available upload bandwidth information used in the data scheduling scheme is often inaccurate due to the peer churn, which lets a peer make bad scheduling decisions and leads to load imbalance. The higher peer churn makes this problem worse. Secondly, the higher peer churn exacerbates the bandwidth contention problem which occurs between a newly joined peer and some already-existing peers. To tackle the above two challenges, we propose an event-driven high-priority first data scheduling scheme, called EHPF scheme. To tackle the first challenge, we design a piggyback mechanism based on the event-driven mechanism. To tackle the second challenge, we design a priority calculation strategy to differentiate the requests from the newly joined peers and those from the already-existing peers, and use the high-priority first policy to allocate the upload bandwidths of peers. Through simulations and a real environment experiment, we demonstrate that the EHPF scheme outperforms the periodical data scheduling scheme in terms of startup delay, streaming quality and load balancing.