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 .

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.

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, aDouble-enhancement framework forlong-tailRecommendation. 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. Our code is available athttps://github.com/leo0481/DeltaRec/E3D7.

Dynamic recommendation systems, where users interact with items continuously over time, have been widely deployed in real-world online streaming applications. The burst of interaction stream causes a rapid evolution of both users and items. To update representations dynamically, existing studies have investigated event-level and history-level dynamics by modeling the newly arrived interactions and aggregating historical interactions, respectively. However, most of them directly learn the representation evolution as new interactions occur, without exploring the collaboration between the newly arrived and historical interactions, thus failing to scrutinize whether those new interactions would benefit the evolution learning process when generating dynamic representations. Moreover, most of them model the two levels of dynamics independently, explicitly ignoring the inherent co-evolving correlation between them. In this work, we propose the Temporal Collaboration-Aware Graph Co-Evolution Learning (TCGC) for the dynamic recommendation scenario. First, we explore the effectiveness of collaborative information and devise the collaboration-aware indicator to guide the evolution learning process. Second, we design a temporal co-evolving graph network, enabling our framework to capture the correlation between event and history dynamics. Third, we leverage the evolution task and recommendation task together for joint training. Extensive experiments on four public datasets demonstrate the superiority and effectiveness of our proposed TCGC.

Federated point-of-interest (POI) recommendation achieves global model training and optimization through a distributed training framework, ensuring that user data is always stored locally on the client side, thereby effectively protecting user data privacy. However, this distributed nature of data also severely limits the performance improvement of POI recommendation models under federated learning. On the one hand, in such a distributed data environment, the semantic fragmentation of feature spaces in local client data hinders the learning of user personalized preference representations, thereby limiting the improvement of local model personalization performance. On the other hand, the heterogeneity in data quantity, quality, and distribution among clients leads to significant differences in local model training.The traditional average aggregation strategy is difficult to effectively alleviate this difference, affecting the global aggregation efficiency and resulting in insufficient performance of the global model. To address these challenges, this paper proposes a federated POI recommendation method based on heterogeneous graph semantics (FedHGS). This method enhances the representation of local user preferences and improves the personalization of local models by learning personalized user preference features through heterogeneous graph semantic mining. Additionally, it introduces local model distillation alignment and performance-aware aggregation to balance training differences across clients and improve the performance of the global POI recommendation model. Finally, extensive experiments are conducted to verify both the global and local performance advantages of FedHGS.

The rise of smart devices has accelerated location-based services, creating vast trajectory data for Point-of-Interest (POI) recommendations. However, user omissions or privacy concerns often result in incomplete trajectory data, compromising sequential patterns and spatiotemporal relationships. Existing solutions using graph structures, self-supervised learning (SSL), or spatiotemporal contexts still face two limitations: (1) graph-based and SSL methods produce suboptimal trajectory representations due to respective inherent constraints; (2) noise interference persists when modeling distorted spatiotemporal signals. To mitigate these issues, we propose aDynamic Structure-aware GraphMasked Autoencoder andSpatiotemporalDiffusion for Next-POIRecommendation (DMSDRec). Specifically, we introduce a dynamic structure-aware improved graph masked autoencoder that adaptively and dynamically distills global transitional information for self-supervised augmentation. It naturally avoids the noise introduced by existing SSL methods' dependency on manual views augmentation. Meanwhile, the masked reconstruction task synergistically enhances trajectory representations by capturing deeper cross-sequence dependencies. Additionally, we propose an effective latent-space spatiotemporal diffusion denoising method. First, we employ graph structures to model spatiotemporal relationships, utilizing higher-order structural information to alleviate the linear spatiotemporal relationship deviations caused by incomplete trajectories. Building on this, we implement diffusion models in the latent space to systematically identify and remove noise from spatiotemporal representations. Through experimental results on two real-world datasets, we demonstrate the superiority of our proposed DMSDRec in terms of recommendation accuracy and robustness.

The current sequential recommendation systems mainly focus on mining information related to users to make personalized recommendations. However, there are two subjects in the user historical interaction sequence: users and items. We believe that mining sequence information only from the users' perspective is limited, ignoring effective information from the perspective of items, which is not conducive to alleviating the data sparsity problem. To explore potential links between items and use them for recommendation, we propose Intent-guided Bilateral Long and Short-Term Information Mining with Contrastive Learning for Sequential Recommendation (IBLSRec), which interpretively integrates three kinds of information mined from the sequence: user preferences, user intentions, and potential relationships between items. Specifically, we model the potential relationships between interactive items from a long-term and short-term perspective. The short-term relationship between items is regarded as noise; the long-term relationship between items is regarded as a stable common relationship and integrated with the user's personalized preferences. In addition, user intent is used to guide the modeling of user preferences to refine the representation of user preferences further. A large number of experiments on four real data sets validate the superiority of our model.

In recent years, next Point-of-Interest (POI) recommendation is essential for many location-based services, aiming to predict the most likely POI a user will visit next. Current research employs graph-based and sequential methods, which have significantly improved performance. However, there are still limitations: numerous methods overlook the fact that user intent is constantly changing and complex. Furthermore, prior studies have seldom addressed spatiotemporal correlations while considering differences in user behavior patterns. Additionally, implicit feedback contains noise. To address these issues, we propose a recommender model named HGDRec for the next POI recommendation. Specifically, we introduce an approach for extracting trajectory intent by integrating multi-dimensional trajectory representations to achieve a multi-level understanding of user trajectories. Then, by analyzing users' long trajectories, we construct global hypergraph structures across spatiotemporal regions to comprehensively capture user behavior patterns. Additionally, to further optimize trajectory intent representation, we employ a feature optimization method based on the improved diffusion model. Extensive experiments on three real-world datasets validate the superiority of HGDRec over the state-of-the-art methods.

Users' behaviors on recommendation platforms are typically driven by evolving intentions. Existing sequential recommendation models have two main limitations in capturing these intentions: insufficient modeling of higher-order relationships between prefix sequences and target items, and lack of effective mechanisms for capturing complex temporal dependencies. To address these challenges, we propose a Spatio-Temporal Intent Modeling framework (STIRec) that enhances recommendations through spatial and temporal dimensions. Our key innovations include: (1) a Multi-Hop Intent Aggregation mechanism that constructs a Spatial Intent Graph modeling three types of relationships (prefix-target, prefix-prefix, target-target), capturing common intent patterns through graph neural networks from a global perspective; (2) a Multi-Span Self-Attention module that fuses long and short-term query information to comprehensively model user behaviors and evolving intentions across temporal dimensions. These complementary mechanisms work together to understand user intent better, integrating global contextual patterns and temporal evolution dynamics. Experiments on five public datasets show that STIRec outperforms state-of-the-art methods by an average of 9.78% in recommendation accuracy, with enhanced robustness against noisy data. Source code is available athttps://github.com/theshy877/STIRec.

Addressing the common issues in the cultivation of collaborative innovation abilities for software engineering talents, such as imperfections in the industry-education integration system, inefficiencies in the management mechanism, and poor connections between the education chain, talent chain, and industry chain, this article introduces the implementation measures for cultivating the collaborative innovation abilities of software engineering talents from the perspectives of organizational models, training models, practical systems, and safeguard mechanisms under the context of industry-education integration.

The next Point-of-interest (POI) recommendation task is to predict the next POI that users may be interested in. POI check-in sequence implicitly reflects the user’s location transition patterns, and the sequence modeling relies on the engineering of multiple independent features. Even though traditional POI recommendation models can fulfill the predicting task via entangled features, those black box models fail to mine the intrinsic check-in intention. The disentanglement representation learning method enables models to disentangle targeting intentions and provide interpretability for recommended results. However, existing disentanglement representation studies focus on disentangling user preference but neglect the entangled location characteristic. Besides, unrecorded check-ins result in inconsecutive transition sequences, which may influence their disentanglement qualities. Therefore, we proposed CrossDR, a Cross-sequence Location Disentanglement Representation method for the next POI recommendation, to explore how spatial and temporal factors influence check-in behaviors by a global view of location transitions. We apply disentanglement learning along with the time-masked data augmentation method and frequency-domain learning technique to further alleviate the short trajectory cold start problem caused by consecutive sequence generation. Experiments on two real-world datasets show our model has competitive capability compared to strong baselines.

Recently, dense retrieval method has significantly outperformed sparse retrieval technology. It becomes a main-stream approach of relevant passage retrieving task. Dense retrieval tasks encode query and passage into dense representation space and apply contrastive learning to obtain more representative vectors, then retrieve the most similar passage by the inner product of the dense vectors. However, previous achievements rely on extremely large batch size, epoch and large-scaled pre-trained models, which results in tremendous computational resource consumption. Also, the expensive equipment prerequisite and low training efficiency constrain the development of dense retrieval community. Therefore, we present an alternative solution to improve training efficiency and quality by syntactical contrastive learning methods with specially designed masking strategy. To alleviate the computational consumption problem, this paper proposes query-based and passage-based masking strategies to obtain syntactical-isolated representations. Besides, instead of only considering query-to-passage similarity while conducting contrastive learning, we additionally consider query-to-query and passage-to-passage similarity when training the dual-encoder retriever. The experiments show that the proposed approach achieved competitive results in small batch size and epoch comparing to previous state-of-the-art dense retrieval methods, and also to strong baseline.

Extractive text summarization is a well-studied downstream task of natural language processing that aims to select sentences as a summary of the document's critical information. For the systems built based on pre-trained language models, the large volume of parameters can cause the model performance gets significantly degradation when the training data is insufficient. However, acquiring high-quality labeled data is a time-consuming and laborious task. Previous works mainly focus on accuracy, and few of them pay attention to the robustness of the model. Hence training a robust and high-quality model is the concern of this work. In this work, we propose RBPSum, a robust extractive summarization model based on the pre-trained language model. Through the experiments, we find that under the situation of restricted data size, sentence position information plays a critical role in extractive summarization. In terms of ROUGE metrics, our model outperforms the previous state-of-the-art approaches when using the entire training set, and around a third of the training set produces competitive results.

Abstract The text matching is a basic task of NLP and is important for tasks such as text retrieval, question answering, and so forth. The development of pre‐trained language models has promoted the progress of text matching tasks. However, due to the natural particularity of Chinese characters and expressions, the Chinese text matching tasks still have problems such as word segmentation difficulty, serious semantic loss, and model instability. In this paper, we propose the DAINet model, which includes DMM, AA and IO modules. We use the Dynamic Multi‐Mask module (DMM) to enhance the completeness of word segmentation. Then we use the Augmented Adversarial module (AA) to further extraction of semantic information. Finally, we use the Integrated Output module (IO) for a more stable output. We conducted experiments on LCQMC, BQ and Xiaobu datasets and compared the results with seven strong baseline models. The results showed that DAINet model made great improvement, including improving ACC value of BQ dataset to , AUC value to , ACC value of LCQMC dataset to and AUC value to . The ACC value of Xiaobu dataset was improved to and the AUC value was improved to . Further ablation experiment results show that the proposed DMM, AA and IO modules have good adaptability and improvement over existing models.

Dense text retrieval has become a widely used paradigm for recalling existing language models , and efficient dense text retrieval is essential for obtaining sufficiently accurate candidate samples. However, the existing methods for dense text retrieval, which typically use dual-encoder architectures to match similar queries and documents, suffer from a lack of information interaction at low data volumes, resulting in suboptimal performance. Additionally, existing research relies on negative sampling techniques that do not take into account the negative effects of single negative sampling bias on the robustness of the model. These limitations hinder the development of more robust dense text retrieval models . In this paper, we propose a multi-view contrast learning architecture, named MvCR, to address these issues. MvCR improves the performance of dense text retrieval by performing contrast learning with multiple views while significantly increasing the model’s ability to discriminate between positive and negative samples. Additionally, we propose a data augmentation method that focuses on increasing the number of hard negative samples with accurate and semantic matching features. The experimental results have shown that MvCR can perform as well as strong baseline models even when the data volume is small. Furthermore, MvCR achieved better results on two popular retrieval benchmarks with comparable amounts of data. Specifically, MRR@10 was 39 . 1 ( + 0 . 9 % ) and Recall@50 was 87 . 8 ( + 1 . 4 % ) on the MS-MARCO dataset. And Recall@5 increased to 77 . 2 ( + 1 . 8 % ) and Recall@50 increased to 85 . 3 ( + 1 . 0 % ) on the Natural Questions dataset.

During the process of cloud manufacturing, service composition is an essential technology to ensure the smooth execution of the task. An important challenge for cloud manufacturing is how to implement effective and accurate service composition strategy, which are to execute the task effectively while also satisfy the requirements of user by maximizing the overall quality of service (QoS). However, the current cloud manufacturing service composition methods generally have the problem of poor solution quality in the large scale environment. To solve this problem, we propose a hybrid optimization algorithm, named hybrid-TC, which is a hybrid of the teaching-learning-based optimization algorithm (TLBO) and the crisscross optimization algorithm (CSO). First, we added Skyline query in the initialization phase to improve the convergence speed and the quality of the solution. Second, the horizontal crossover of CSO is used in the teaching phase of original TLBO, thereby some dimensions in the population that trapped in the local optimum have the chance to jump out of the iteration. Finally, the offspring individuals learned in the teaching-phase will continue to learn to improve the quality of the solution. Experiments show that our proposed method is effective for improving the solution quality of large-scale environmental service composition.

Service discovery aims to address the problem of service information explosion and find and locate services that meet the needs of service requesters. Because service description information is mostly composed of short text with noise and has the characteristics of semantic sparseness, it is difficult to extract the implied context information of service description. This paper proposes a service discovery framework based on Knowledge graphs and neural Similarity Network (KSN). Which uses knowledge graphs to connect entities to obtain rich external information to enhance the semantic information of service descriptions. convolutional neural network and similarity network is utilized to extract context information. Through extensive experiments on a real service data set show that KSN is superior to existing web service discovery methods in terms of multiple evaluation metrics.

Point-of-interest (POI) recommendation that consists of location-based social networks (LBSNs) and provides personal services for users has become an important part in the field of recommendation system.Due to the sparseness of user check-in matrix, POI recommendation faces great challenges.However, most researches just consider of spatial and temporal impact on recommendation and do not solve the problem of sparsity.This paper proposes a POI recommendation model called RBMNMF which is based on sparse matrix of user check-ins.Firstly, by stacking restricted Boltzmann machines (RBM), the potential relationship between users and POIs is learned and multiple user-POI matrices are extracted.Second, fill the original sparse matrix by using non-negative matrix factorization (NMF).Finally, fuse those prediction matrices to generate final POI recommendation for users, which is benefit for solving the problem of sparsity effectively.Experiments on real-world data set prove that the model we propose has a better accuracy than traditional algorithms.

Point of Interest (POI) recommendation aims to recommend places which users have not visited before. In this paper, we proposed a POI recommendation method using user similarity, which assumes that people may be interested in the places that others have been to but they have not visited before. In this paper, one day can be divided into 24 time-slots, thus each hour can be defined as a time slot. The novelty of the method we proposed lies in user features which adopted by the summation of user’s check-in times in each time slot. The check-in times for each user can be collected and then form a vector, and we can take advantage of the summation of these check-in times in each time slot to find out user characteristics. The similarity between any two users can be calculated by cosine similarity method. Then a sorted list of scores which includes all unvisited locations of each user can be obtained according to user similarity. Through these steps, a POI recommendation list can be produced according to the score from high to low. The experimental result indicates that the method we proposed in this paper is effective.

Quality of service (QoS) has been playing an increasingly important role in today's Web service environment. Many techniques have been proposed to recommend personalized Web services to customers. However, existing methods only utilize the QoS information at the client-side and neglect the contextual characteristics of the service. Based on the fact that the quality of Web service is affected by its context feature, this paper proposes a new QoS-aware Web service recommendation system, which considers the contextual feature similarities of different services. The proposed system first extracts the contextual properties from WSDL files to cluster Web services based on their feature similarities, and then utilizes an improved matrix factorization method to recommend services to users. The proposed framework is validated on a real-world dataset consisting of over 1.5 million Web service invocation records from 5825 Web services and 339 users. The experimental results prove the efficiency and accuracy of the proposed method.

When extracting information from a web page, IE systems usually need to perform pattern recognition to identify the elements that have similar patterns. However, most of them are mainly based on analyzing HMTL source code, DOM tree, tag tree or Xpath of web pages. These methods are language-dependent, or more precisely, HTML-dependent. They have some insuperable limitations. In order to overcome these limitations, we propose a notion of layout-tree and a pattern recognition method to identify visual blocks with similar visual pattern using layout tree. In this paper, we call a visible rectangular region in a web page a visual block or block for short. We consider if the elements of two blocks are displayed in a similar layout, we define that the two blocks are visually similar. We first transform the layout into a layout tree. By calculating the similarity of the layout trees of two blocks, we can determine whether the two blocks are visually similar or not. The result of experiment shows that the layout tree is an effective method to identify visually similar blocks.