Point-of-interest (POI) recommendation plays a crucial role in location-based services by predicting users’ next visiting locations from historical check-in data. In real-world scenarios, user behaviors exhibit significant temporal heterogeneity, referring to the distribution mismatch of preferences across different time contexts, which is further exacerbated in federated learning due to decentralized and non-IID data. To address this issue, we propose FedTAAR, a federated POI recommendation framework that models temporal heterogeneity at multiple levels. Specifically, to capture dynamic time-aware preferences, we design a temporal decoupled adapter for flexible temporal feature modulation. To mitigate noisy and entangled signals in user sequences, we introduce a semantic anchor routing module that learns stable high-level behavioral patterns. Furthermore, to alleviate parameter conflicts across heterogeneous clients, we develop an anchor-aware aggregation strategy that adaptively adjusts client contributions. Extensive experiments on real-world datasets demonstrate that FedTAAR consistently outperforms state-of-the-art methods and effectively improves recommendation performance under heterogeneous federated settings.

Deep reinforcement learning (DRL) is widely used in home energy management for its ability to handle nonlinearity and uncertainty. However, its reliance on trial-and-error interaction makes early unsafe and inefficient behaviors impractical in real households. To address this, we propose RB-ZeroHEM, a zero-shot knowledge transfer framework based on representation learning and behavioral cloning. RB-ZeroHEM employs contrastive learning to extract stable, physics-aligned representations of household energy dynamics from historical control trajectories, enabling clustering-based similarity measurement without hand-crafted features. For a new household, it identifies the most similar source cases and clones a deployable policy requiring zero target-environment interactions. Experiments on 12 simulated households driven by real-world energy data (San Diego smart meter and weather records spanning four seasons) demonstrate that when transferring to physically similar environments, RB-ZeroHEM reduces discomfort ratio by 33% (from 12.34% to 8.24%) and electricity costs by 15% (from $228 to $193) compared to rule-based control, while achieving 81% of the grid energy savings of the best online DRL method, soft actor-critic (SAC), with zero interactions versus SAC’s 24,192-step exploration budget. The learned representations exhibit strong physics alignment and remain robust across seasonal shifts and deployment noise. With minimal fine-tuning (576 interactions, 2.4% of SAC’s budget), performance matches or surpasses online DRL baselines. When source-target physical dynamics are mismatched, zero-shot performance degrades gracefully, indicating that similarity-based retrieval reliability is critical. These results establish RB-ZeroHEM as a practical, low-risk solution for deploying intelligent energy management in real households, enabling rapid scaling without the safety and economic penalties of online exploration.

Deep reinforcement learning (DRL) has emerged as a promising paradigm for home energy management systems (HEMSs) due to its model-free nature and ability to handle complex dynamics. However, existing DRL-based approaches typically employ a unified reward function that aggregates multiple objectives into a single scalar, failing to account for the heterogeneous roles of controllable energy devices (CEDs) and their distinct causal relationships with control objectives. This leads to reward misattribution, where CEDs with simpler constraints dominate the optimization process while critical components such as battery storage remain underutilized. To address this challenge, we propose a causally aligned multiagent reinforcement learning (MARL) framework that explicitly models CED-objective causal pathways using a structural causal model (SCM). A causal surgery procedure decomposes shared objectives into CED-specific variants, enabling individualized reward signals aligned with each CED's causal responsibility. The proposed heterogeneous reward-aware multiagent soft actor-critic (HR-MASAC) algorithm features a multihead centralized critic for learning vectorized Q-values and agent-specific entropy coefficients for heterogeneous exploration. Experiments across diverse home scenarios demonstrate that our method achieves 40.1% cost reduction and 57.1% comfort improvement over unified-reward baselines, with robust performance under sensor noise and household heterogeneity.

Multi-behavior recommendation leverages multiple user-item interaction information to alleviate data sparsity. Although different types of user-item interactions are temporally mutually exclusive, the sequence of behavioral interactions consisting of multi-level positive feedback signals contains rich information. However, most existing studies have unilaterally focused on the positive utility of auxiliary behaviors, ignoring multi-level user preference information. Effectively fusing multi-behavioral data and better modeling behavioral dependencies are urgent problems that need to be addressed for multi-behavior recommendation. We propose the parallel learning of positive and negative interests with an auxiliary-view representation enhancement (PPN-ARE) scheme for multi-level user interest learning based on multi-behavioral interaction sequences. Specifically, multi-level positive and negative feedback view chains are constructed from multi-behavioral sequence data to learn multi-level user interests. User preference evolution is simulated during multi-behavior interactions using residual connections, and the shortcomings of the cascading structure used for higher-order graph learning are analytically highlighted. The influence of low-quality embeddings of auxiliary behaviors is filtered, and the learning of target behaviors is optimized by designing a representation enhancement layer. Finally, the model is optimized using a multi-task training framework. The experimental results indicate that PPN-ARE significantly improved over the state-of-the-art (SOTA). The open source code is available at https://github.com/lhybq/PPN-ARE.

Session-based recommendation predicts the next item a user will interact with based on their short-term session behavior, typically without long-term user profiles. Existing approaches often fail to capture the hierarchical nature of user preferences, leading to suboptimal personalization and limited recommendation accuracy. In this work, we argue that user preferences exhibit coarse-grained and fine-grained characteristics, and item features should be modeled accordingly across these two levels to capture users' preference signals more accurately. To this end, we propose a novel method, Multi-Granularity Preference Enhancement with Hierarchical Feature Extraction (MPEHFE), for session-based recommendation. MPEHFE explicitly captures semantic item relationships at each granularity and enhances fine-grained preference modeling through a differentiable architecture search mechanism. It also identifies interactions inconsistent with the user's general intent as noise, leveraging contrastive learning to reinforce the representation of coarse-grained preferences. Moreover, experiments on three real-world benchmark datasets demonstrate that MPEHFE consistently outperforms state-of-the-art baselines, achieving relative improvements of 3%-9% in P@20 and 11%-56% in MRR@20.

Sequential recommendation focuses on modeling and predicting a user's next actions based on their sequential behavior patterns, using the temporal order and dynamics of user actions to provide more personalized and contextual suggestions. Sequential recommendation models rely on limited temporal scales, making it challenging to explicitly capture diverse user behaviors spanning multiple scales. Motivated by this challenge, this paper introduces ScaleRec, an advanced Multi-Scale Transformer architecture augmented with dual attention mechanisms and adaptive masking for sequential recommendation. ScaleRec integrates interaction granularity and context through multi-scale division, segmenting user behavior sequences into patches of varying lengths. Dual attention explicitly models fine-grained interests and coarse-grained preferences, including intra-patch cross-attention and inter-patch self-attention. Specifically, intra-patch cross-attention employs a learnable Gaussian kernel to introduce locality-based inductive biases, capturing fine-grained behavioral dynamics. The inter-patch self-attention is further enhanced by a Context-adaptive Preferences Aggregator, which dynamically selects and integrates relevant long-term user preferences. Additionally, we introduce an adaptive masking fusion strategy to filter redundant information dynamically. Extensive experiments on six benchmark datasets show that ScaleRec achieves state-of-the-art performance, improving the recommendation performance by up to 24.95% in terms of HR@5. The code of the proposed model is available at: https://github.com/gangtann/ScaleRec.

The long-tail recommendation problem remains a significant challenge in modern recommender systems, primarily due to data sparsity and popularity bias, which hinder the accurate ID representation of users and items. Recent advancements in large language models (LLMs) have enabled the direct modeling of user and item semantic representations, offering potential improvements in representation learning through the alignment of these two types of representations. However, systems relying on LLM representation alignment face two critical challenges: (1) the substantial differences between LLMs and recommendation models in terms of training objectives, phases, and data; (2) the pervasive popularity bias in collaborative data. These challenges create a semantic gap between ID representations and semantic representations. Directly aligning these representations risks introducing recommendation-irrelevant noise, disrupting the collaborative information embedded in ID representations, and ultimately leading to suboptimal recommendation outcomes. To address this gap, we propose DeltaRec, a Double-enhancement framework for long-tail Recommendation. DeltaRec tackles the long-tail recommendation problem through two approaches. First, it incorporates semantic information for all items. Second, it provides additional supervision signals specifically for long-tail items. The framework begins by disentangling ID representations into interest representations and conformity representations. To integrate semantic information from LLMs while preserving popularity information, we design a contrastive learning-based semantic alignment module that aligns interest representations with semantic representations. Furthermore, to enhance the representation learning of unpopular items, we introduce a ranking-based behavior alignment module, which provides additional supervision signals for these items. To avoid introducing recommendation-irrelevant noise and disrupting collaborative semantics due to excessive alignment, we propose a curriculum learning-based training mechanism. Extensive experiments on real-world datasets demonstrate that DeltaRec effectively mitigates popularity bias and significantly improves long-tail recommendation performance without relying on prior knowledge of popularity distributions.

White matter hyperintensities (WMH) are common radiological findings with diverse clinical outcomes, including absence of clinical symptom, cognitive decline and dementia. One potential explanation for these differences is that WMH could differentially disrupt specific white matter tracts and lead to alterations in structural brain networks.We included participants from the UK Biobank who had available information on total volume of WMH and divided them into three groups: healthy controls (HC), individuals at risk for cerebral small vessel disease (RcSVD) and individuals with dementia. We assessed alterations in structural connectivity in the RcSVD and dementia groups relative to the HC group, and examined whether the alteration in structural connectivity is associated with WMH volumes. To further investigate the extent to which between-group differences in structural connectivity can be explained by the location of WMH, we mapped WMH-specific disconnectivity matrices for every individual and compared them with the alterations in structural connectivity identified.We found a widespread disruption of structural brain networks, both in RcSVD and dementia group, compared to healthy individuals. We showed that these disruptions can be partially explained by individual differences in the volume and location of WMH, particularly WMH in periventricular areas. Moreover, we found that WMH are more strongly associated with RcSVD-related structural connectivity changes (average r=-0.32) than with those implicated in dementia (average r=-0.03). This suggest that WMH may impact structural brain network through different pathways in RcSVD and dementia.Our work underscores the important role of WMH lesions in disrupting structural brain networks in RcSVD and dementia, offering insights into the varying clinical impacts of WMH.

Federated Learning (FL) has emerged as a privacy-preserving paradigm for Point-of-Interest (POI) recommendation. However, FL fragments holistic global interactions into isolated client views, causing simple local modeling to fall short in accurately comprehending user personalized preferences. Compounding this, cross-client non-IID data distributions further constrain global recommendation performance. To address this, we propose FedLHSSG: A Federated Framework for POI Recommendation Using Local Hypergraph with Shared Spatial Graph. The framework incorporates a server-side Shared Spatial Graph to explicitly encode public geographical topology, injecting a globally consistent spatial structural prior into local models to complement fragmented contexts. Locally, we devise a Hypergraph-based training paradigm synergized with a Diversity-Guided Diffusion Module for personalized refinement of user representations. Finally, an adaptive edge learning layer is introduced to extract cross-client behavioral patterns and improve the global model. Our main contribution lies in enabling efficient POI recommendation with privacy protection, and improving performance through personalized modeling, spatial utilization, and collaborative learning. Extensive experiments validate the effectiveness of FedLHSSG.

Multimodal Recommender Systems leverage rich side information to capture user preferences more accurately. However, existing methods often fuse different modalities via straightforward operators such as concatenation or attention, overlooking two critical challenges: i)The high-frequency components in features introduce modality-specific noise, which often carries low semantic relevance. ii)The sparse user-item interactions fail to capture vital co-frequency behavioral patterns. To address these issues, we propose MSCF-Net, a Multi-Scale Frequency Denoising and Co-Frequency Enhancement Network. Our key contribution lies in integrating frequency-domain insights throughout the recommendation process. We design an adaptive spectral domain filter with learnable weights to dynamically suppress high-frequency noise while preserving cross-modal semantics. Furthermore, we augment the bipartite interaction graph by constructing User-User and Item-Item Co-frequency Matrices derived from interaction histories, reinforcing collaborative signals through frequency-based affinity measurement. A novel Spectrum Harmonization Loss is also introduced to ensure both cross-modal alignment and modality-specificity preservation. These complementary mechanisms work together to better understand user preferences. On three real-world datasets, MSCF-Net shows encouraging performance compared to state-of-the-art baselines, validating the robustness and effectiveness of our proposed framework.

Home energy management systems (HEMS) must balance competing objectives-electricity cost, thermal com fort, and carbon emissions-according to user preferences that are personalized, evolving, and often expressed in natural language. Conventional deep reinforcement learning (DRL) methods cannot directly interpret such preferences, while existing reinforcement learning (RL)+LLM integrations either invoke large language models (LLMs) at every control timestep or rely on one-shot preference parsing without feedback. This paper proposes LA-UPAHEM, an iterative LLM-augmented framework in which three specialized agents collaborate in a closed loop: a Code Generation Agent translates preferences into reward and state modification functions, a Result Analysis Agent diagnoses policy-preference misalignment from evaluation metrics, and an Optimal Performance Search Agent identifies the best-performing policy as the warm-start for the next iteration. LLMs are invoked only during this offline refinement phase; the deployed policy operates independently without LLM inference. Experiments on real residential energy datasets show that LA-UPAHEM outperforms classical DRL and existing RL+LLM base lines in both macro-level key performance indicator (KPI) alignment (cost, comfort, emissions) and micro-level rule compliance, achieving a Weighted Improvement Ratio (WIR) of 0.13 and a Rule Compliance Rate (RCR) of 0.84, while reducing the failure rate from 34.3% (static parsing) to 5.4%. The framework is robust to environ mental noise, evaluation weight perturbation, and preference paraphrasing, and generalizes across multiple DRL algorithms, LLM backbones, and preference languages.

In Point-of-Interest (POI) recommendation tasks, users’ trajectory data is scattered across different geographical regions, leading to spatial heterogeneity in Federated Learning (FL) data distribution, which significantly hampers model performance. While most personalized FL algorithms primarily address non-Independent and Identically Distributed (non-IID) data, they frequently neglect the fact that isolated clients can disrupt spatial patterns within local models, resulting in incomplete or biased spatial correlations. To tackle this, we explore the impact of spatial heterogeneity in federated POI recommendation and propose FedKG, an optimization framework that alleviates performance degradation through a combination of generative modeling and knowledge distillation. Specifically, FedKG introduces ST-GAN (Spatiotemporal Generative Adversarial Network), a generative adversarial model that reconstructs trajectory sequences to restore disrupted spatial patterns. Furthermore, PKD (Personalized Knowledge Distillation) and GKD (Global Knowledge Distillation) modules leverage these generated samples to enhance local model training and facilitate global knowledge distillation. Extensive comparisons with eight personalized federated learning algorithms on four real-world datasets demonstrate that FedKG significantly enhances model performance in both non-IID and spatially heterogeneous settings. Notably, FedKG achieves an 18.21% improvement over the strongest baseline under spatial heterogeneity, and a 7.59% gain in non-IID settings. The code for FedKG has been open-sourced at: https://github.com/Chips98/FedKG.

As one of the core functions of Location-Based Social Networks, next Point-of-Interest (POI) recommendation plays a crucial role in improving user experience in applications such as urban navigation, tourism, and personalized location services. However, existing models mainly rely on users’ fine-grained POI-level historical check-in behaviors, focusing on precise spatio-temporal transitions between individual locations, while often neglecting users’ coarse-grained semantic preferences, such as category-level activity tendencies and region-level mobility patterns, which capture higher-level abstractions of user intent. This limitation restricts their predictive capability. To address this issue, we propose MCMDRec, a Multi-Channel next POI recommendation framework that comprehensively models users’ multi-level preferences across heterogeneous granularities. Specifically, MCMDRec jointly captures fine-grained long- and short-term POI-level temporal preferences and coarse-grained category- and multi-level region preferences, and integrates them within a multi-channel joint prediction framework to enhance preference complementarity. Furthermore, a spatially aware Transformer is introduced to encode geographic dependencies via a distance-based spatial bias, improving spatial reasoning and contextual adaptability. Extensive experiments on three real-world datasets demonstrate that MCMDRec achieves performance improvements of up to 13.66% in HR@10 and 22.11% in NDCG@10 compared to state-of-the-art baselines. The proposed framework provides insights for urban mobility prediction, intelligent navigation, and personalized tourism recommendation, offering a generalizable solution for multi-granularity spatio-temporal user modeling in LBSN applications.

Sequential recommendation leverages interaction sequences to model users’ interests and provide personalized recommendations. However, most existing models focus on modeling users’ long-term preferences (low-frequency information) and struggle to perceive sudden changes in users’ interests, limiting the model’s ability to capture users’ short-term preferences (high-frequency information). Furthermore, they rely on the sequential relationship of users’ interaction sequences for modeling, failing to fully capture effective users’ preference information when interaction sequences are sparse, consequently affecting the accurate modeling of users’ latent intents. In this article, we propose an intent-aware spectrally enhanced contrastive learning sequential recommendation model (ISECLRec). Specifically, we design a frequency adaptive reconstruction module to adjust the frequency-domain representation of the sequence to enhance the model’s ability to capture users’ short-term preferences. Meanwhile, we introduce a calibration factor to linearly combine the reconstructed frequency-domain representation with the original sequence representation, comprehensively modeling users’ long-term and short-term preferences. In addition, we construct shared latent intent prototypes and align users’ interaction sequences with their target intents, enabling ISECLRec to accurately capture users’ latent intents. Extensive experimental results on six real-world datasets demonstrate that ISECLRec outperforms advanced baseline models by an average of 4.34% and 3.71% in HR and NDCG, respectively.

Next point-of-interest (POI) recommendation has emerged as a key research topic within location-based social networks. Prior studies suggest that combining users' enduring preferences with their short-term behaviors can substantially improve recommendation quality. However, since manually annotated labels distinguishing these two forms of interest are usually unavailable, many existing models tend to merge them during training, resulting in less accurate representations. In addition, the interaction between stable and dynamic interests is complex, and current approaches often struggle to model the evolving dependency. To overcome these challenges, we introduce a novel recommendation framework. Specifically, we adopt two bidirectional Transformer encoders to separately model general tastes and dynamic preferences. Preference labels are derived from users' historical trajectories, and contrastive learning is applied to realize self-supervised separation of long- and short-term interests. Moreover, an attention mechanism with adaptive weighting is designed to integrate these two types of preferences to capture their dynamically changing dependencies. By leveraging spatio-temporal correlations across short-term and long-term sequences, our method generates more robust and distinctive preference representations. According to our review of prior studies, this work is the first POI recommendation study to employ a self-supervised strategy to explicitly distinguish between long- and short-term user preferences. Extensive experiments on three real-world datasets demonstrate that our method consistently outperforms both classical recommendation models and recent deep learning-based state-of-the-art baselines, achieving significant gains in HR@K and NDCG@K with average improvements of 21.73% (HR@5), 16.69% (HR@10), 12.97% (NDCG@5), and 5.90% (NDCG@10).

With the rapid development of location-based applications and mobile internet technologies, improving the performance of next Point-of-Interest (PoI) recommendation has become crucial for enhancing user experience and advancing the intelligence of applications. Addressing the limitations of existing POI recommendation methods in modeling spatio-temporal coupling features, multi-view feature fusion, and high-order relational capture, this paper proposes a novel Channel Attention-enhanced Multi-Scale Hypergraph Learning for PoI recommendation CAMSHG. CAMSHG first constructs a multi-semantic hypergraph and introduces a channel attention mechanism to dynamically model the importance of different semantic channels, thereby enhancing high-order feature representations. Second, a view-consistent alignment mechanism based on contrastive learning is employed to improve the consistency of multi-view representations. Third, a multi-scale convolutional feature fusion module is designed to fully integrate multi-view information at local and global granularities, further boosting recommendation performance. Extensive experiments and evaluations on three publicly available datasets demonstrate that CAMSHG consistently outperforms various advanced baseline methods across multiple metrics. Ablation studies and hyperparameter sensitivity analyses further validate the effectiveness of each designed module and the robustness of the overall framework. The results confirm that CAMSHG offers a unified and efficient solution for next POI recommendation in complex application scenarios, showing strong potential for practical deployment and further development. To facilitate future research, we release the code at https://github.com/Z70rain/CAMSHG.

Graph neural networks (GNNs) have shown considerable promise in multi-behavior recommendation tasks, particularly for target behavior prediction (e.g., purchase conversion), by effectively integrating auxiliary behavioral signals (e.g., item browsing, cart addition). Recent advances in the field have substantially improved the modeling of hierarchical interactions and multi-behavior dependencies, effectively mitigating foundational challenges such as data sparsity. However, these state-of-the-art methods frequently overlook the semantic heterogeneity and inherent noise within auxiliary interactions, often resorting to uniform or simplistic denoising strategies that risk discarding valuable signals. To overcome this persistent limitation, the Denoising Cascade-Enhanced Multi-Behavior Recommendation (DCE-MBR) framework is introduced. DCE-MBR is designed to simultaneously suppress noise and preserve semantically informative interactions through a dual-stage architecture. First, a hierarchical graph denoising module dynamically removes noisy edges by applying behavior-specific thresholds across multiple levels of granularity, thereby preserving essential neighbor relations. Next, a cascade-enhanced module incrementally refines user preferences by propagating target behavior signals through auxiliary behavior paths, leading to improved feature representations. Comprehensive evaluations based on the Taobao as well as Tmall data collections show that DCE-MBR outperforms state-of-the-art baselines, achieving relative gains of 18.66% in Hit@10 and 15.42% in NDCG@10. These results confirm the model's robustness against noisy interactions and its effectiveness in capturing intricate multi-behavior dependencies. The source code is publicly available at DCEMBR.1

Serverless computing is gaining popularity, and serverless workflows stand out as its key workload form. Since resource configurations in serverless workflows are typically manually specified, considerable research has focused on automating resource allocation optimization. However, conventional studies generally follow a coupled CPU–memory allocation, thereby limiting cost-effectiveness, or assume that accurate function performance profiles are known in advance. Moreover, they do not explicitly address the impact of function cold starts on workflow performance. To address these limitations, this paper introduces a serverless workflow provisioning model that jointly supports decoupled resource allocation and function container reuse among workflow tasks implementing identical functionality. Based on this model, we propose a structure-aware trust-region Bayesian optimization (STBO) method to optimize resource configurations for minimizing cost under deadline constraints. The proposed approach is implemented in OpenWhisk, and evaluation experiments demonstrate that, compared with traditional methods, STBO increases the constraint satisfaction rate and reduces execution cost by over 17% on average.

Next point-of-interest (POI) recommendation aims to predict a user’s next destination from historical check-in behavior. Many existing methods handle sequential, spatial, and semantic information in separate modules and combine them at later stages, making it challenging to capture how sequential patterns influence spatial choices, how geographic context shapes semantic preferences, and how category transitions guide temporal decisions. We propose MvHCA-POI, a multi-view framework with hierarchical cross-attention that enables cross-view feature integration during encoding. MvHCA-POI builds three graph-based views: a category sequence graph, a POI sequence graph, and a geographic graph that captures both local proximity and global structural relationships. Unlike traditional methods that integrate view embeddings at a late fusion stage, our hierarchical cross-attention operates in two stages. The category sequence first guides token-level enhancements within each view, emphasizing relevant recent check-ins and nearby POIs, and then serves as a semantic bridge to align view-level representations across different views. Experiments on four real-world LBSN datasets demonstrate that MvHCA-POI achieves 6.5% to 11.2% improvements over state-of-the-art baselines in standard ranking metrics. Ablation studies further confirm the contribution of each component in the hierarchical cross-attention framework.