Jiaquan Zhang, Chaoning Zhang, Shuxu Chen, Zhenzhen Huang, Pengcheng Zheng, Zhicheng Wang, Ping Guo, Fan Mo, Sung-Ho Bae, Jie Zou, Jiwei Wei, Yang Yang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Sequential recommendation(SR) aims to provide personalized recommendations by capturing behavioral intents from existing user interaction sequences. Most previous studies are based on attention mechanisms; however, these approaches suffer from inherent over-smoothing issues that limit their ability to capture transient behavioral signals reflecting the user's immediate intents in interaction sequences. Recently, frequency-domain analysis methods based on the Fourier transform have garnered significant attention in the sequential recommendation domain. By applying the Fourier transform, interaction sequences can be mapped to the frequency domain, enabling direct analysis and targeted manipulation of distinct frequency components. In addition to the inherent limitations of self-attention mechanisms, sequential recommendation faces persistent challenges such as data sparsity and noise. To address these issues, we propose Frequency-Domain Disentanglement-Fusion and Dual Contrastive Learning for Sequential Recommendation (FDCLRec). FDCLRec replaces self-attention mechanisms with a frequency-domain adaptive filtering module, which decouples sequence patterns into distinct high-/low-frequency components and synthesizes comprehensive sequence representations through adaptively weighted fusion. In addition, two auxiliary contrastive learning tasks(augmented-view contrasting and same-target sequence contrastive learning) are strategically integrated to alleviate data sparsity and interaction noise. Extensive experiments on four real-world datasets demonstrate that our model outperforms baseline methods.

With the increasing awareness of health, using wearable sensors to monitor individual activities and accurately estimate energy expenditure has become a current research focus. However, existing research encounters challenges including low estimation accuracy, a deficiency of frequency domain features, and difficulty in integrating time domain and frequency domain features. To address these issues, we propose an innovative framework called the Dual-Stream Fusion Network (DSFN). This framework combines the Time Domain Encoding (TDE) module, the Frequency Domain Hierarchical-Split Encoding (FDHSE) module, and a Two-Stage Feature Fusion (TSF) module. Specifically, the temporal stream of the framework employs the TDE module to capture deep temporal features that reflect the complex dynamic variations in time-series data. The frequency domain stream introduces the FDHSE module, which extracts frequency domain features using a multi-level, multi-scale approach, ensuring a comprehensive and diverse representation of frequency information. Through this dual-stream architecture, our model effectively learns both time and frequency domain features, addressing the limitations of frequency domain features observed in prior studies. Additionally, we propose the TSF module to fully integrate time and frequency domain features, effectively overcoming the challenge of fusing these two types of features. We conducted experiments on two public datasets, namely the GOTOV dataset (elderly people) and the JSI dataset (young people). Experimental results demonstrate that our method achieves excellent performance across different age groups. Compared to the baseline models, the proposed DSFN significantly improves the accuracy of human energy expenditure estimation.

The task of steel surface defect recognition is an industrial problem with great industry values. The data insufficiency is the major challenge in training a robust defect recognition network. Existing methods have investigated to enlarge the dataset by generating samples with generative models. However, their generation quality is still limited by the insufficiency of defect image samples. To this end, we propose Stable Surface Defect Generation (StableSDG), which transfers the vast generation distribution embedded in Stable Diffusion model for steel surface defect image generation. To tackle with the distinctive distribution gap between steel surface images and generated images of the diffusion model, we propose two processes. First, we align the distribution by adapting parameters of the diffusion model, adopted both in the token embedding space and network parameter space. Besides, in the generation process, we propose image-oriented generation rather than from pure Gaussian noises. We conduct extensive experiments on steel surface defect dataset, demonstrating state-of-the-art performance on generating high-quality samples and training recognition models, and both designed processes are significant for the performance. Note to Practitioners—This article introduces StableSDG, a method that generates realistic defect images even with limited data. It overcomes the shortcomings of current deep learning approaches that need large datasets to train from scratch. Our solution is to adapt a text-to-image diffusion model for defect generation. The proposed strategy involves two processes: training to adapt token embeddings and model parameters, and generation from partially perturbed defect images. The results show enhanced generation quality and improved accuracy for recognition models trained on the expanded dataset. StableSDG can be practically applied to efficiently enlarge a defect dataset, even when starting with a small amount of data.

Mobile Edge Computing (MEC) revolutionizes the traditional cloud-based computing paradigm by moving resources in proximity to the network edge, aiming to cater to the rigorous requirements of emerging latency-sensitive applications. However, the escalating resource demands intensify the competition among user devices (UDs). Thus, it is essential to coordinate task offloading and resource scheduling while ensuring fairness among users in MEC. Despite the crucial role of user fairness in motivating task offloading in MEC, it is often overlooked in existing literature. Therefore, we in this paper propose a caching-enhanced MEC framework and formulate a collaborative service caching, task offloading, and multi-resource allocation problem to maximize average user satisfaction. Multiple factors contribute to the difficulty in solving the optimization problem, including constrained resource capabilities, user mobility, service heterogeneity, and spatial demand coupling. Consequently, we transform the origin problem into two distinct subproblems – the service caching and task offloading problem, and the multi-resource allocation problem, respectively. Then, the Advantage Actor-Critic (A2C) based approach is proposed to address the former problem, while a Lagrangian duality-based approach is adopted to tackle the latter problem. The simulation results demonstrate the superior performance of the proposed solution in comparison to several baseline methods.

Physiological signals have been widely used for emotion recognition, but current works seldom apply the feature fusion and attention technologies to ECG emotion recognition. In this paper, we propose a novel ECG emotion recognition method, which adopts a spatial and temporal ECG emotion recognition model based on dynamic feature fusion (DFF-STM) to learn spatial-temporal representations of different ECG areas. Considering the difference in roles played by the different ECG areas in ECG emotion recognition, a dynamic weight distribution layer is introduced into DFF-STM to extract ECG temporal features and learn weights to adjust (e.g., enhance or weaken) the contribution of the ECG areas at the same time. Finally, we conduct experiments using real ECG data on the AMIGOS dataset to evaluate the performance of the DFF-STM on valence and arousal labels. Experiments show that dynamic feature fusion for ECG emotion recognition is much better than those using only handcraft features and deep features.

Abstract Cross-domain recommendation (CDR) has become an important research direction in the field of recommender systems due to the increasing demand for personalized recommendations across different domains. However, CDR faces multiple challenges, including data sparsity, popularity bias, and long-tail problems. To address these challenges, we propose a novel framework that combines graph contrastive embedding and multi-head cross-attention transfer for cross-domain recommendation, called GCE-MCAT. Specifically, in the pre-training process, we generate more uniform user and item embeddings through contrastive learning, effectively solving the problem of inconsistent data embedding space distribution and recommendation popularity bias. Moreover, we propose a multi-head cross-attention transfer mechanism that allows the model to extract user common and specific domain features from multiple perspectives and perform cross-domain bidirectional knowledge transfer. Finally, we propose a cross-domain feature fusion mechanism that dynamically assigns weights to common user features and specific domain features. This enables the model to more effectively learn common user interests. We evaluate the proposed framework on three real-world CDR datasets and show that GCE-MCAT consistently and significantly improves recommendation performance compared to state-of-the-art methods. In particular, the proposed framework has demonstrated remarkable effectiveness in addressing long-tail distribution and enhancing recommendation novelty, providing users with more diversified recommendations and reducing popularity bias.

Using wearable devices to realize the mining and application of human behavior patterns has become a hotspot in the field of intelligent positioning. Wearable devices provide an analyzable data foundation for indoor spatial distribution and human behavior pattern prediction. The development of the intelligent positioning system based on RSSI has encountered a bottleneck that it is difficult to improve the positioning accuracy. Therefore, some research works started emphasizing location technology based on channel state information (CSI). In this paper, the principle used by Wi-Fi channel state information to realize intelligent positioning is described, the characteristics of CSI are analyzed, and an intelligent positioning algorithm based on CSI is proposed. Specifically, the algorithm first estimates the angle of arrival (AoA) based on the MUSIC algorithm, separates the reflected paths in the multipath components, and accurately estimates the AoA of each path. Second, phase estimation with channel state information is achieved by forming different antenna subarray measurements under the consideration of a subset of antennas and subcarriers. Then, the phase response linear fitting of the data packet CSI is eliminated using the ToF purification algorithm to obtain the corrected phase response and realize the elimination of the STO noise of the channel state information. Finally, the target position is calculated by effectively filtering the reflection path through the likelihood value, and the accurate target positioning function is achieved. The experimental results demonstrate that the intelligent positioning algorithm proposed in this paper can achieve decimeter-level positioning accuracy under the condition of a fixed number of APs, and the average error is better than that of deep learning-based and SVM-based positioning algorithms. In other words, the accuracy of intelligent positioning is improved.

In recent years, the study of social-based recommender systems has become an active research topic. We incorporate a combination of static and dynamic interest characteristics to predict users’ real-time dynamic interests, which has rarely been considered in previous studies. In this paper, we propose a graph neural network social recommendation model that integrates static and dynamic feature relationships (FSDFR-GNNSR). The model uses a graph embedding algorithm to extract static features of users and movies, and takes the static features as input to gated recurrent unit (GRU), so that the model can take static features into consideration while modeling user dynamic behavior. Finally, we use graph attention networks to represent the dynamic influence of friends, simplify the update strategy of second-order neighbor nodes. We apply graph pooling operations to improve the generalization ability of the algorithm. Empirical analyses on real datasets show that the proposed approach achieves superior performance to existing approaches.

With the explosively growing amount of online information, recommender system becomes an important tool to help users efficiently find their desired information. In this paper, we propose a Graph Neural Network Social Recommendation Based on Coupled Influence by analyzing the social influence of 2-level friends (CI-GNNSR). First, we mine the user’s historical rating information and second-degree social information. Then, to learn the feature representation of users and movies, multiple Graph Attention Networks (GAT) are used to model the user-movie Graph and social network Graph. Our algorithm uses an attention-based memory network to learn the interest influence representation between users and their collaborative friends, which can distinguish the related factors among different users’ friends. The experiment results show that CI-GNNSR enhances the accuracy of recommendation by considering users’ social influence factors from multiple perspectives.

Wearable devices provide an extremely convenient way to collect a large amount of human motion data. In this paper, the human motion recognition method based on wearable devices is studied. We smooth the data to remove the noise caused by additional motion first. After that, the characteristic values that can distinguish the types of activities can be extracted. Then, we propose a human motion recognition method based on the improved one-dimensional convolutional neural networks(1D-CNNs). Compared with other traditional classification and recognition methods, the recognition rates of 11 human motions have been greatly improved. The average accuracy of each activity identification can reach 92.8%, while the average precision and recall can reach 98.7% and 92.8%.

The development of sensor technologies and smart devices has made it possible to realize real-time data acquisition of human beings. Human behavior monitoring is the process of obtaining activity information with wearables and computer technology. In this paper, we design a data preprocessing method based on the data collected by a single three-axis accelerometer. We first use Butterworth filter as low-pass filtering to remove the noise. Then, we propose a KGA algorithm to remove abnormal data and smooth them at the same time. This method uses genetic algorithm to optimize the parameters of Kalman filter. After that, we use a threshold-based method to identify falls that are harmful to the elderly. The key point of this method is to distinguish falls from people’s daily activities. According to the characteristics of human falls, we extract eigenvalues that can effectively distinguish daily activities from falls. In addition, we use cross-validation to determine the threshold of the method. The results show that in the analysis of 11 kinds of human daily activities and 15 types of falls, our method can distinguish 15 types of falls. The recognition recall rate in our method reaches 99.1%.

Summary To solve the problem that the general power consumption monitoring system only realizes the total amount statistics of energy consumption information, but there is no statistical analysis of energy consumption in items and classification as well as energy saving control, a solution scheme to the power consumption monitoring and energy saving management system based on the Internet of things technology was proposed. Based on the research on the Internet of things architecture and combined with the actual functional requirements of the project, the four‐level architecture of the Internet of things was designed. The database technology was used to achieve an expert system for enterprise energy saving control based on relational model. Finally, according to the actual project, the engineering application of the system was introduced. The research showed that the system proposed could effectively monitor the power consumption of enterprises and analyze and manage energy consumption. To sum up, the power consumption can be well monitored by applying Internet of things.