Advanced Data Mining and Applications

Event sequences play as a general fine-grained representation for temporal asynchronous event streams. However, in practice, event sequences are often fragmentary and incomplete with censored intervals or missing data, making it hard for downstream prediction and decision-making tasks. In this work, we propose a fresh extension on the definition of the temporal point process, which conventionally characterizes chronological prediction based on historical events, and introduce inverse point process that characterizes counter-chronological attribution based on future events. These two point process models allow us to impute missing events for one partially observed sequence with conditional intensities in two symmetric directions. We further design a peer imitation learning algorithm that lets two models cooperatively learn from each other, leveraging imputed sequences given by the counterpart as the supervised signal. The training process consists of iterative learning of two models and facilitates them to achieve a consensus. We conduct extensive experiments on both synthetic and real-world datasets, which demonstrate that our model can recover incomplete event sequences very close to the ground-truth, with averagely 49.40% improvement compared with related competitors measured by normalized optimal transport distance.

Multivariate time series classification is crucial for various applications such as activity recognition, disease diagnosis, and brain-computer interfaces. Deep learning methods have recently achieved promising performance thanks to their powerful representation learning capacity. However, existing deep learning-based classifiers rely solely on temporal information while disregarding clues from the frequency perspective. In this regard, we propose a novel method for classifying multivariate time series leveraging both temporal and frequency information. We first apply Short-Time Fourier Transform (STFT) to transform time series into spectrograms, which contain a 2D representation of frequency components and their temporal positions. In particular, for each variable, we generate spectrograms with varying frequencies and temporal resolutions under different window sizes. The transformation essentially adds a new modality to 1D time series and converts the multivariate time series classification into a multi-modality data classification task, making it possible to bring powerful backbones from computer vision fields to solve the time series classification problem. We then construct a dual-stream network based on the ResNet architecture that takes in both 1D and 2D representations for accurate multivariate time series classification. Our extensive experiments on 30 public datasets show our method outperforms multiple competitive state-of-the-art baselines.

Positional embedding is an effective means of injecting position information into sequential data to make the vanilla Transformer position-sensitive. Current Transformer-based models routinely use positional embedding for their position-sensitive modules while no efforts are paid to evaluating its effectiveness in specific problems. In this paper, we explore the impact of positional embedding on the vanilla Transformer and six Transformer-based variants. Since multivariate time series classification requires distinguishing the differences between time series sequences with different labels, it risks causing performance degradation to inject the same content-irrelevant position token into all sequences. Our experiments on 30 public multivariate time series classification datasets show positional embedding positively impacts the vanilla Transformer’s performance yet negatively impacts Transformer-based variants. Our findings reveal the varying effectiveness of positional embedding on different model architectures, highlighting the significance of using positional embedding cautiously in Transformer-based models.

Predicting the time for an event to occur while simultaneously exploring the coexisting effects of various risk factors has captivated considerable research interest. However, the profusion of repeated measurements involving a diverse array of risk factors has outpaced the capabilities of current methods for analyzing time-to-event data. In this paper, we propose a novel approach that entails the conversion of the time-to-event analysis conundrum into a sequence of discrete survival learning and prediction tasks, each approached autonomously. Our innovative strategy for modeling repeated measures facilitates the quantification of measurement impacts on projected outcomes at distinct junctures. When extrapolating the trajectory of health status over time, our method harnesses both censored and uncensored data to refine logistic regression parameters. Through a series of comparative experiments and meticulous ablation studies conducted on two real-life health datasets, we underscore the intrinsic practical promise of our method. Notably, our approach showcases its efficacy in prognosticating the temporal aspects of breast cancer patient mortality and the onset of disabilities among the elderly.

Timeliness is one of the important indicators of data quality. In industrial production processes, a large amount of dependent data is generated, often resulting in unclear timestamps. Therefore, this article combines the conclusion dependency graph into a process dependency graph to determine the identification order of the timeliness of each process data; By constructing a weighted timeliness graph (WTG) and path single flux, a data timeliness identification method that does not completely rely on timestamps is proposed. Finally, a time-effectiveness identification method based on weighted time-effectiveness graph was discussed through an example and 9 dependency rules, and the effectiveness of the method was verified through a set of experiments.

Anomaly detection for multivariate time series is a very complex problem that requires models not only to accurately identify anomalies, but also to provide explanations for the detected anomalies. However, the majority of existing models focus solely on the temporal relationships of multivariate time series, while ignoring the spatial relationships among them, which leads to the decrease of detection accuracy and the defects of anomaly interpretation. To address these limitations, we propose a novel model, named spatio-temporal relationship anomaly detection (STAD). This model employs a novel graph structure learning strategy to discover spatial features among multivariate time series. Specifically, Graph Attention Networks (GAT) and graph structure are used to integrate each time series with its neighboring series. The temporal features of multivariate time series are jointly modeled by using Transformers. Furthermore, we incorporate an anomaly amplification strategy to enhance the detection of anomalies. Experimental results on four public datasets demonstrate the superiority of our proposed model in terms of anomaly detection and interpretation.

Recently, because of the remarkable performance in alleviating the data sparseness problem in recommender systems, Graph Convolutional (Neural) Networks (GCNs) have drawn wide attention as an effective recommendation approach. By modeling the user-item interaction graph, GCN iteratively aggregates neighboring nodes into embeddings of different depths according to the importance of each node. However, the existing GCN-based methods face the common issues that, they do not consider the node information and graph structure during aggregating nodes, such that they cannot assign reasonable weights to the neighboring nodes. Additionally, they ignore the differences in node types in the user-item interaction graph and thus, cannot explore the complex relationship between users and items, resulting in a suboptimal result. To solve these problems, a novel GCN-based framework called RNT-GCN is proposed in this paper. RNT-GCN integrates the structure of the graph and node information to assign reasonable importance to different nodes. In addition, RNT-GCN refines the node types, such that the heterogeneous properties of the user-item interaction graph can be better preserved, and the collaborative information of users and items can be effectively extracted. Extensive experiments prove the RNT-GCN achieved significant performance compared to SOTA methods.

Knowledge Graph (KG) can provide semantic information about items, which can be used to mitigate the sparsity problem in recommendation systems. In recent years, the trend in knowledge-aware recommendation methods has been to leverage Graph Neural Networks (GNNs) to aggregate node information in KG. However, many of these methods focus on mining the item knowledge association on KG, but ignore the potential item auxiliary information in user’s history interaction outside KG. Furthermore, these methods equally aggregate all neighbor entities of the item on KG, which will inevitably introduce irrelevant entity-interaction behaviors. To address these issues, we propose a novel model, called Multi-level Noise Filtering and Preference Propagation Enhanced Recommendation (MNFP). Technically, we employ self-attention mechanisms to model the user’s interaction sequence to mine the item’s auxiliary information. Then, we design a twin-tower preference propagation mechanism that iteratively expands item auxiliary information on KG. Additionally, we propose a multi-level noise filtering mechanism. By learning the relationship consistency between the item and its neighbor entities, the model can guide the item to selectively link highly related neighbors in preference propagation, thus reducing the introduction of noise. We evaluate MNFP on three real-world datasets: MovieLens-1M, Last.FM and Book-Crossing. Results show that MNFP significantly outperforms state-of-the-art methods on AUC and F1.

Knowledge graph serves as a side information, bringing diversity and interpretability to the recommendation. A well-developed recommender system can efficiently capture user and item characteristics, accurately reflecting user preferences. However, supervised signals with graph structure are extraordinarily sparse, and the collaborative and knowledge graphs contain irrelevant edges, exacerbating noise propagation and reducing the robustness of recommendations. To address the above issues, we propose a model for enhancing Knowledge-aware Recommendation with Contrastive Learning (KRCL), including two contrastive learning tasks and three functional modules. Specifically, we construct two views, using TransR and TATEC to optimize knowledge representations from distance and semantic aspects, respectively. After the item-side knowledge is augmented, we remove unreliable interaction edges from collaborative graph to reduce noise propagation. We then perform contrastive learning on the output node representations of different views through graph propagation. To further tap the latent interest of users, we consider users/items that exhibit similar representations as semantic neighbors, treating them as positive pairs in contrastive learning. The structural and semantic contrastive tasks are eventually integrated in a multi-task learning manner to jointly boost the recommendation performance. To validate the effectiveness of our method, we conduct extensive experiments on three benchmark datasets. Experimental results demonstrate that our KRCL significantly outperforms previous state-of-the-art baselines.

One of the biggest challenges in Recommendation Algorithms (RA) is how to obtain user and item embeddings from sparse interaction history. To take this challenge, most graph neural network based RAs explicitly incorporate high-order collaborative filtering signals on the user-item bipartite graph with either multi-layer semantics on the Knowledge Graph (KG) or multi-level neighbors on the social network. However, none of them fully integrate these three types of graph-structured data, which decreases embeddings’ precision. Based on this consideration, this paper integrates the three types of data by proposing a knowledge-rich influence propagation RA based on the graph attention mechanism. Specifically, in the semantic propagation, we categorize user preferences into deep interest obtained by multiple graph attention message propagations on related KG parts, and shallow interest generated from the interaction history. Moreover, the influence weight between items is determined by the number of co-interactions and the semantic similarity. These two factors as well as social relations together decide the influence weight between users. With these influence weights, final user and item embeddings are calculated through multi-layer message propagation. The experimental results show that the proposed recommendation algorithm outperforms several compelling baselines on six scaled-down real-world datasets. This work has confirmed the effectiveness of combining these three types of data to increase RAs’ coverage and accuracy.

Variational Autoencoder (VAE) has been extended as a representative nonlinear latent method for collaborative filtering recommendation. As a high-dimensional representation of data, latent vectors play a vital role in the transmission of important information in a VAE model. However, VAE-based models suffer from a common limitation that the transmission ability of the latent vectors’ important information is limited, resulting in lower quality of global information representation. To address this, we present a novel VAE model with multi-position latent self-attention and reinforcement learning’ actor-critic algorithm. We first build a multi-position latent self-attention model, which can learn richer and more complex latent vectors and strengthens the transmission of important information at different positions. At the same time, we use reinforcement learning to enhance the interactive learning process of collaborative filtering recommendation training. Specifically, our model is stable and can be easy applied in the recommendation. We observed significant improvements over the previous state-of-the-art baselines on three social media datasets, where the largest improvement can reach $$26.10\%$$ 26.10 % .

The successful application of graph neural networks in recommendation scenarios causes serious exposure of sensitive information of users. Research shows that social bias such as sexism and ageism are prevalent in recommendations, and the use of multi-graph information even makes it worse. Existing fair recommendation algorithms only concentrate on users’ sensitive attributes in user-item graph, failing to fully remove those attributes from multiple graphs. In addition, merely hiding sensitive information is not enough, there is still a gap in recommendation utility for different user groups. In this work, we propose a novel fair re-ranking recommendation model based on debiased multi-graph representations, which contains three functional layers. Multi-graph embedding layer iteratively propagates and aggregates both topological and interactive information on multiple graphs. Attribute hiding layer uses generative adversarial networks to hide user sensitive information and thus debias users’ representations. Fair ranking layer adopts a re-ranking strategy with our proposed unfairness metric to further optimize the final recommendation list. Extensive experiments on real-world datasets demonstrate the performance of our proposed model in both recommendation utility and fairness, outperforming state-of-the-art models.

BERT and its variants are the most performing models for named entity recognition (NER), a fundamental information extraction task. We must apply inference speedup methods for BERT-based NER models to be deployed in the industrial setting. Early exiting allows the model to use only the shallow layers to process easy samples, thus reducing the average latency. In this work, we introduce FastNER, a novel framework for early exiting with a BERT biaffine NER model, which supports both flat NER tasks and nested NER tasks. First, we introduce a convolutional bypass module to provide suitable features for the current layer’s biaffine prediction head. This way, an intermediate layer can focus more on delivering high-quality semantic representations for the next layer. Second, we introduce a series of early exiting mechanisms for BERT biaffine model, which is the first in the literature. We conduct extensive experiments on 6 benchmark NER datasets, 3 of which are nested NER tasks. The experiments show that: (a) Our proposed convolutional bypass method can significantly improve the overall performances of the multi-exit BERT biaffine NER model. (b) our proposed early exiting mechanisms can effectively speed up the inference of BERT biaffine model. Comprehensive ablation studies are conducted and demonstrate the validity of our design for our FastNER framework.

Chinese Nested Named Entity Recognition (CNNER) faces several challenges due to the language diversity phenomena, the complexity of the language, and the imbalanced distribution of entity types in Chinese text. To address these challenges in CNNER, we propose a new method called CPMFA (Character Pair-based method with Multi-feature representation and Attention mechanism). The CPMFA method predicts the predefined relations of character pairs in a sentence, and identifies nested named entities based on these relations. First, our method utilizes the pre-trained language model LERT (Linguistically-motivated Bidirectional Encoder Representation from Transformer), and Bidirectional Long Short-Term Memory (BiLSTM) to generate comprehensive and precise character representations. Second, our method uses multi-feature representation to capture complex semantic information within the text, and employs the Pyramid Squeeze Attention (PSA) module to emphasize key features. Finally, to overcome the challenge of the imbalanced distribution of entity types, PolyLoss function is integrated into our model training process. Results of experiments show that the proposed CPMFA method achieves an F1 score of 83.79%. Compared to other mainstream span-based methods, the proposed CPMFA method has excellent performance in CNNER.

Aspect-Based Sentiment Triplet Extraction (ASTE) is a rapidly growing field in sentiment analysis. While most research has focused on processing the ASTE task either in a pipeline or end-to-end manner, both methods have their limitations. Pipeline methods may accumulate errors in practical applications, while sequence labeling methods in end-to-end approaches may overlook important feature information of the three elements themselves. Additionally, various features in sentences and emotional word markers have not been effectively explored in these methods. To address these limitations, we propose a novel solution called Span Tagging Multi-Channel Graph Convolutional Network (STMC-GCN) that explicitly combines multiple prominent features to extract span-level sentiment triplets, where each span may consist of multiple words and play different roles. Specifically, we designed a three-channel graph fusion model that converts sentences into multiple channels of graphs. These channels extract node text features, centrality features, and position features, which are then extracted through cross-channel convolution operations to obtain a common graph representation shared by different channels. To optimize downstream classification with better results, we use consistency and difference constraints to enhance common attributes and independence. Finally, we explore span-level information and constraints to generate more accurate aspect-based sentiment triplet extractions. Experimental results illustrate that STMC-GCN performs well on multiple datasets, proving the effectiveness and robustness of the model.

Entity resolution (ER) finds records that refer to the same entities in the real world. Blocking is an important task in ER, filtering out unnecessary comparisons and speeding up ER. Blocking is usually an unsupervised task. In this paper, we develop an unsupervised blocking framework based on pre-trained language models (B-PLM). B-PLM exploits the powerful linguistic expressiveness of the pre-trained language models. A design space for B-PLM contains two steps. (1) The Record Embedding step generates record embeddings with pre-trained language models like BERT and Sentence-BERT. (2) The Block Generation step generates blocks with clustering algorithms and similarity search methods. We explore multiple combinations in above two dimensions of B-PLM. We evaluate B-PLM on six datasets (Structured + dirty, and Textual). The B-PLM is superior to previous deep learning methods in textual and dirty datasets. We perform sufficient experiments to compare and analyze different combinations of record embedding and block generation. Finally, we recommend some good combinations in B-PLM.

Entity disambiguation (ED) is a critical natural language processing (NLP) task that involves identifying and linking entity mentions in the text to their corresponding real-world entities in reference knowledge graphs (KGs). Most existing efforts perform ED by firstly learning the representations of mention and candidate entities using a variety of features and subsequently assessing the compatibility between mention and candidate entities as well as the coherence between entities. Despite advancements in the field, the limited textual descriptions of mentions and entities still lead to semantic ambiguity, resulting in sub-optimal performance for the entity disambiguation task. In this work, we propose a novel framework LogicED, which considers both Local and global semantics for contrastive Entity Disambiguation. Specifically, we design a local contextual module, which utilizes a candidate-aware self-attention (CASA) model and the contrastive learning strategy, to learn robust and discriminative contextual embeddings for both mentions and candidate entities. Furthermore, we propose a global semantic graph module that takes into account both the local mention-entity compatibility and the global entity-entity coherence to optimize the entity disambiguation from a global perspective. Extensive experiments on benchmark datasets demonstrate that our proposed framework surpasses the state-of-the-art baselines.

User reviews contain many key phrases that are crucial for business understanding, but they are often obscured by the sheer volume of reviews. Extracting key phrases from user reviews could help to understand what users are concerned about and provide timely improvement suggestions. Current pattern-based methods for target phrase extraction usually analyze reviews at a coarse-grained level, making the extracted topics unfocused and useless. Hence, in order to address this issue, we proposed a fine-grained analysis approach (KFEA) to extract, cluster, and visualize key phrases from e-commerce reviews. In order to fully utilize the relevant information from comments, KFEA fuses the information like categories and ratings from a large volume of user reviews, and then extracts key phrases with the help of a pre-trained model. A method is also designed to cluster and visualize the extracted key phrases for business understanding. Our evaluation on 6,088 reviews from 6 products shows that KFEA can effectively extract key phrases and perform clustering and visualization. In particular, KFEA achieved an precision of 76.6% and a recall of 81.8% in extracting key phrases from manually annotated data. KFEA’s cross-categories effectiveness is also validated on 16,772 reviews from products like mobile phones, laptops, and furniture.

This paper focuses on the task of product emotion cause analysis which aims to find essential causes of certain emotions from product reviews. Current works only study the explicit causes which are some spans in the given text. However, some crucial and useful causes may be expressed vaguely. They may not mention but can be inferred from the text semantics. They can capture the deeper reasons to explain some unknown phenomena, which can well support the applications like market research and product optimization. To address this problem, we in this paper propose a new task of Emotion Implicit Cause Discovery (EICD). We develop a novel method that can deduce the implicit causes based on the contexts and commonsense knowledge. Our method first retrieves related knowledge from the large language model to construct reasoning graphs for the emotions and potential causes. We then encode the structural knowledge in the graph and infer the implicit cause by deductive reasoning. To evaluate our method, we construct a large dataset called EICDset, based on Amazon product reviews. Experiments on it demonstrate the effectiveness of our model.

This paper proposes a new task of Multi-modal Multi-emotion Emotional Support Conversation (MMESC), which has great value in various applications, such as counseling, daily chatting, and elderly company. This task aims to fully perceive the users’ emotional states from multiple modalities and generate appropriate responses to provide comfort for improving their feelings. Traditional works mainly focus on textual conversation, while a single-modal cannot accurately reflect the users’ emotions, such as saying fine with an inconsistent disgusting feeling. To address this problem, we propose a new task on multi-modalities and exploit a new method called FEAT for this new task. FEAT can integrate fine-grained emotional knowledge from multiple modalities. It first recognizes the users’ mental states based on an emotion-aware transformer. It then generates supportive responses using a hybrid method with multiple comfort strategies. To evaluate our method, we construct a large-scale dataset named MMESConv. It is almost two times larger than existing single-modal datasets. There are three modalities in this dataset (text, audio, and video) with fine-grained emotion annotations and strategy labels. Extensive experiments on this dataset demonstrate the advantages of our proposed framework.

With the extensive accumulation of conversational data on the Internet, emotion recognition in conversations (ERC) has received increasing attention. Previous efforts of this task mainly focus on leveraging contextual and speaker-specific features, or integrating heterogeneous external commonsense knowledge. Among them, some heavily rely on future contexts, which, however, are not always available in real-life scenarios. This fact inspires us to generate pseudo future contexts to improve ERC. Specifically, for an utterance, we generate its future context with pre-trained language models, potentially containing extra beneficial knowledge in a conversational form homogeneous with the historical ones. These characteristics make pseudo future contexts easily fused with historical contexts and historical speaker-specific contexts, yielding a conceptually simple framework systematically integrating multi-contexts. Experimental results on four ERC datasets demonstrate our method’s superiority. Further in-depth analyses reveal that pseudo future contexts can rival real ones to some extent, especially in relatively context-independent conversations.

Emotional support conversation aims to provide comfort and suggestions to users and gradually reduce their negative emotions such as anxiety. It is a valuable topic for many applications, including mental health support and customer service chats. However, due to the lack of enough expert knowledge, existing methods fail to provide enlightened suggestions to reverse users’ worries. Additionally, these methods neglect to grasp the mental state evolution of users. To address these problems, we propose a novel method that considers Mental State Evolution to provide Knowledge-grounded Suggestions (MEKS). In detail, we first create a suggestion corpus called MentalQA to grasp the psychological knowledge by resorting to the mental health forum. The relevant passages are selected based on both the context and the original response. Then we leverage graph structure to enrich the context with the inferred user’s mental state evolution. Furthermore, we introduce a gate to combine textual expert knowledge with the mental state evolution graph, so as to facilitate the generation of supportive responses. Experimental results show that this method can provide reasonable solutions to help the users.

The abundance of user-generated online content has presented significant challenges in handling big data. One challenge involves analyzing short posts on social media, ranging from sentiment identification to abusive content detection. Despite recent advancements in pre-trained language models and transfer learning for textual data analysis, the classification performance is hindered by imbalanced data, where anomalous data represents only a small portion of the dataset. To address this, we propose Deep One-Class Fine-Tuning (DOCFT), a versatile method for fine-tuning transfer learning-based textual classifiers. DOCFT uses a one-class SVM-style hyperplane to encapsulate anomalous data. This approach involves a two-step fine-tuning process and utilizes an alternating optimization method based on a custom OC-SVM loss function and quantile regression. Through evaluations on four different hate-speech datasets, we observe that significant performance improvements can be achieved by our method.

Content-based methods are inadequate for detecting fake news related to COVID-19 due to the complexity of this domain. Some studies integrate the social context information of the news to improve performance. However, such information is not consistently available and sometimes not helpful regarding COVID-19, as most users lack professional knowledge about it and may be unable to respond accurately. Additionally, fake news often employs emotional manipulation to exploit people’s emotions to shape their beliefs and actions. Therefore, we propose EmoKnow, an emotion- and knowledge-oriented model, for detecting fake news about COVID-19. Our proposed method incorporates language modeling, emotion feature extraction, and external knowledge sources to provide an informative representation of news. Experimental results on four COVID-19-related datasets show that EmoKnow significantly outperforms state-of-the-art approaches.

Music plays an important role in our daily life. It can have a powerful effect on our emotions, mental health, and even the community we live in. Although numerous studies have been conducted to prove the great impact music has on humans, few investigations place an emphasis on the exploration of the relationship between music and listener’s sentiment. To this end, we first examined three song demographics: Beats Per Minute, Key, and Length, and six song metrics: Danceability, Energy, Speechiness, Acousticness, Liveness, and Valence of popular songs, and then conducted an empirical study to examine the potential correlation between song demographics/metrics and the sentiment expressed as in written text (such as social media). To accomplish this, we scraped around 20 million tweets referencing the most popular songs from 2018 to 2022 as shown on Spotify’s Top Global chart, as well as the immediate surrounding tweets, and performed a double sentiment analysis on the data. Our study concludes that there exists a significant correlation between all the pairs of song metrics. From the sentiment analysis of tweets, our results indicate that there may not be a significant correlation between the sentiment expressed in tweets of a song’s listeners and the song itself. Our study provides empirical evidence for a deeper understanding of popular songs using data mining techniques.

As more and more netizens participate in financial market transactions, online discussions on asset price movements are becoming more comprehensive and timely. Online text, especially from social media, has the potential to be an important data source for financial opinion mining. Market sentiment analysis mainly includes direct analysis methods in the form of text-based surveys and indirect inference methods based on structured data such as price, trading volume, and volatility. In theory, the former is helpful for us to understand investor sentiment earlier, but due to the difficulty of obtaining a sufficient number of objective survey samples, its obtained research attentions are far less than the latter. To combine the advantages and offset the weakness of these two approaches, this paper uses Valence Aware Dictionary and Sentiment Reasoner (VADER) and Fast Fourier Transform (FFT) to construct social media sentiment indexes based on plenty of daily discussion texts about Bitcoin (BTC) and S &P500 (SPX) from Reddit for analyzing their interaction with prices. We also propose a new time series synchronization verification method called Rolling Time-lagged Cross-correlation (RTLCC) surface, and corresponding feature constructing methods, in which RTLCC helps us observe Time-lagged Cross-correlation from the perspective of Rolling Correlation while determining the hyperparameters (Window Size & Time Offset) for features construction. Finally, based on these features, we use four machine learning classifiers for modeling and verify the effectiveness of the proposed market sentiment analysis pipeline, in which on the prediction of 10-day price movements, the best model achieves 89.9% in accuracy (ACC) and 92.5% in AUC.

A high utility co-location pattern (HUCP) is a set of spatial features, which is supported by groups of neighboring spatial instances, and the pattern utility ratios (PUR) of the spatial feature set are greater than a minimal utility threshold assigned by users, can reveal hidden relationships between spatial features in spatial datasets, is one of the most important branches of spatial data mining. The current algorithm for mining HUCPs adopts a level-wise search style. That is, it first generates candidates, then tests these candidates, and finally determines whether the candidates are HUCPs. It performs mining from the smallest size candidate and gradually expands until no more candidates are generated. However, in mining HUCPs, the UPR measurement scale does not hold the downward-closure property. If the level-wise search style is adopted, unnecessary candidates cannot be effectively pruned in advance, and the mining efficiency is extremely low, especially in large-scale and dense spatial datasets. To overcome this, this paper proposes a mining algorithm based on a query strategy. First, the neighboring spatial instances are obtained by enumerating maximal cliques, and then these maximal cliques are stored in a hash map structure. Neighboring spatial instances that support candidates can be quickly queried from the hash structure. Finally, the UPR of the candidate is calculated and a decision is made on it. A series of experiments are implemented on both synthetic and real datasets. Experimental results show that the proposed algorithm gives better mining performance than existing algorithms.

3D point cloud data semantic mining plays a key role in 3D scene understanding. Although recent point cloud semantic mining methods have achieved great success, they require large amounts of expensive manual annotated data. More importantly, the lack of large-scale annotated datasets limits those approaches in many real-world applications, especially for point-level semantic mining tasks such as point cloud semantic segmentation. In this work, we propose a novel point cloud segmentation framework, called Point-level Label-free Segmentation framework (PLS), that does not require point-level annotations. In this framework, the point cloud semantic mining task is formulated as a clustering problem based on mutual information. Meanwhile, our method can directly predict clusters that correspond to the given semantic classes in a single feed-forward pass of a neural network. We apply the proposed PLS to the shape part segmentation task. Experiments on the benchmark ShapeNetPart dataset demonstrate that our method has the ability to discover clusters that match semantic classes, and it can produce comparable results with methods using incomplete labels on several categories.

Causal discovery involves learning Directed Acyclic Graphs (DAGs) from observational data and has widespread applications in various fields. Recent advancements in the structural equation model (SEM) have successfully applied continuous optimization techniques to causal discovery. These methods introduce acyclicity constraints to tackle the challenge of exploring the exponentially large search space that arises as the number of graph nodes increases. However, these methods often rely on point estimates that fail to fully account for the inherent uncertainty present in the data. This limitation can lead to inaccurate causal graph inference. In this paper, we propose a novel method for causal discovery with Bayesian Neural Networks (CD-BNN). CD-BNN incorporates a Bayesian Neural Network to explicitly model and quantify uncertainty in the data while reducing the influence of noise through model averaging. Moreover, we explore the extraction of the final DAG from the BNN using partial derivatives. We conduct a comprehensive set of experiments on both real-world and synthetic data to evaluate the performance of our approach. The results demonstrate that our proposed method surpasses related baselines in accurately identifying causal graphs, particularly when faced with data uncertainty.

Heuristics have been effective in solving computationally difficult optimization issues, but because they are often created for certain problem domains, they perform poorly when the challenges are significantly altered. The currently available techniques are either designed to address single- or multi-objective optimization issues solely, or they perform poorly with the same parameters. The multi-domain approach known as hyper-heuristics (HHs) can be used to solve optimization issues with minor variations. Motivated by the notion of utilizing the benefits of low-level heuristics (LLHs) in order to obtain well-distributed and convergent optimum solutions along with taking into account the shortcomings of the work completed in many-objective HHs. For many-objective optimization problems, this paper develops a high-level selection approach that employs indicators by preference and offers a unique selection hyper-heuristic called Preference-based Indicator Selection Hyper-heuristic (PBI-HH). In order to establish fairness between exploration and exploitation, the method makes use of a randomization mechanism and a greedy strategy to address a significant problem faced by HHs. Three well-known many-objective evolutionary algorithms are combined in the unique technique that is being proposed. The efficacy of the proposed strategy is assessed by contrasting it with cutting-edge HHs. PBI-HH performs better or equal to the state-of-the-art HHs on 155 out of 160 cases employing the HV indicator and has the optimal $$\mu $$ μ norm mean values across all datasets.

In distributed stream computing systems, dynamic data skew and cluster heterogeneity can lead to major load imbalance among multiple instances of stateful operators. Existing stream grouping schemes mainly focus on data load balancing for stateful operators, but they are not considered to be sufficiently elastic scalable, which directly affects the latency and throughput. We propose an elastic scalable grouping (called Es-Stream) for stateful operators. This paper discusses the following aspects: (1) Investigating the dynamic grouping of real-time data stream, proposing a general data stream graph model and a data stream grouping model, as well as formalizing the problem of load balancing optimization and data stream grouping. (2) Utilizing key splitting to solve the bottleneck problem caused by high-frequency keys in the data streams, and lightweight weight adjustment strategy to dynamically change the data tuple allocation probability of the instance according to the network cost, data stream rate and processing rate. (3) Implementing Es-Stream in Apache Storm platform and evaluating the system using metrics such as latency, throughput and load imbalance. Experimental results showed that Es-Stream reduces latency by up to 72%, increases throughput by up to 44% and reduces load imbalance by up to 75%, compared with existing state-of-the-art grouping schemes.

The natural gradient boosting method for probabilistic regression $$(\mathrm {{\textbf {NGBoost}}})$$ ( NGBoost ) is capable of predicting not only point estimates but also target distributions under sample conditions, thereby quantifying prediction uncertainty. However, NGBoost is designed only for batch settings, which are not well-suited for data stream learning. In this paper, we present an incremental natural gradient boosting method for probabilistic regression $$(\mathrm {{\textbf {INGBoost}}})$$ ( INGBoost ) . The proposed method employs scoring rule reduction as a metric and applies the Hoeffding inequality incrementally to construct decision trees that fit the natural gradient, thus achieving incremental natural gradient boosting. Experimental results demonstrate that INGBoost performs well in both point regression and probabilistic regression tasks while maintaining the interpretability of the tree model. Furthermore, the model size of INGBoost is significantly smaller than that of NGBoost.

As an extended version of frequent itemset patterns, periodic itemset patterns concern both the frequency and periodicity of itemsets at the same time, so they contain more information than frequent itemset patterns, which only concern the frequency. With further research, we found that, in some cases, the periodic itemset patterns with higher frequency, or with optimal periodicity, or with both higher frequency and optimal periodicity have higher application value. However, there is currently no work focusing on such a kind of periodic itemset patterns. In view of this, this paper first proposes a new concept of skyline periodic itemset patterns, and states the problem of skyline periodic itemset pattern mining, then presents an algorithm called SLPIM (SkyLine Periodic Itemset pattern Miner) for skyline periodic itemset pattern mining. SLPIM first adopts the well-known FP-Growth algorithm to mine all frequent itemset patterns, and then uses an effective judgment strategy to determine which frequent itemset patterns are skyline periodic itemset patterns. Finally, experiments are conducted on two real-world and two simulated datasets. The results show that SLPIM is competent for mining skyline periodic itemset patterns.

Automation control, which is one functional core of industrial control system, has become the prime attack target due to its vulnerabilities. Furthermore, many industrial cyber threats can disturb or destroy the correctness of control operation data to cause industrial accidents, when one normal production process is running smoothly and orderly. In order to effectively identify abnormal activities in various control operation data, this paper proposes one BP (Back Propagation) neural network anomaly prediction model based on the double-optimized CS (Cuckoo Search) algorithm. By using the exponential decline strategy and Gaussian perturbation to improve the traditional CS algorithm, this model can obtain one effective anomaly prediction engine based on the optimized BP neural network: for one thing, it can quickly enter the local search through the exponential decline strategy; for another, the information exchange between all local positions and the global optimal positions is realized by Gaussian perturbation. Moreover, the double-optimized CS algorithm not only solves the problem that the traditional BP neural network is prone to fall into local optimal solution, but also eliminates the defect of low vitality in the traditional CS algorithm. Consequently, this model can realize the high-precision prediction of abnormal control operation data. The experimental results show that, compared with other approaches, this model has better prediction performance under both normal and attack states, and can ensure the security of automation control in industrial production.

Neural ordinary differential equations (NODEs) have achieved remarkable performance in many data mining applications that involve multivariate time series data. Its adoption in the data-driven discovery of dynamic systems, however, was hindered by the lack of interpretability due to the black-box nature of neural networks. In this study, we propose a simple yet effective NODE architecture inspired by the highly successful generalised additive models. Our proposed model combines linear and nonlinear components to capture interpretable evolution rules with only a marginal loss of model expressiveness. Experiments show that our model can effectively recover interactions among variables in a complex dynamic system from observation data.

This paper presents an approach to detect outliers present in a data set, also called aberration. These outliers often cause problems to the learning algorithms by deviating their behavior, which makes them less efficient. It is therefore necessary to identify and remove them during the cleaning data step before the learning process. For this purpose, a method that detects if data is an outlier from its k nearest neighbors is proposed for multidimensional data sets. In order to make the method more accurate, the number of k nearest neighbors chosen is adaptive for each class present in the data set, and each neighbor has a different weight in the decision, depending on their respective proximity. The proposed method is called Makoto for Multidimensional Adaptative kNN Over Tracking Outliers. The effectiveness of this method is compared with four other known methods based on different principles: LOF (Local Outlier Factor), Isolation forest, One Class SVM and Inter Quartile Range (IQR). Thus, on the basis of 406 synthetic data sets and 17 real data sets with distinct characteristics, the Makoto method appears to be more efficient.

Complete and accurate road network information is an important basis in the detection of EHV transmission lines, and regular updates of road distribution near transmission lines are necessary and meaningful. However, no relevant research has been found for this application area, and coupled with the fact that roads themselves are significantly challenging, extracting roads with good connectivity and integrity in remote sensing images remains a problem to be solved. Therefore, in this paper, we develop a new end-to-end road extraction network, Multiple Attention Networks (MANet). Specifically, by fusing convolutional and self-attentive approaches, we focus on global contextual features to obtain an effective feature map. In addition, the Strip Multi-scale Channel Attention (SMCA) module is specifically designed for the line features of roads, focusing on extracting row and column features, while the Edge-aware Module (EAM) is used to extract connected and complete roads, aided by edge information. Meanwhile, in order to enhance the practicality of the study, a Mengxi Transmission Line Road Dataset was constructed independently following the processing process of remote sensing images in industrial production. By conducting relevant quantitative and qualitative experiments on this dataset and the publicly available CHN6-CUG dataset, it is fully verified that the method in this paper is superior to other advanced methods and can still extract roads with strong connectivity in complex backgrounds, which has good potential and outstanding advantages in practical applications.

The Trip Planning Query (TPQ), which returns the optimal path from the starting point to the destination that satisfies multiple types of points of interest (POIs) specified by the user, has attracted more and more attention. The most straightforward approach is to enumerate all the POI combinations that meet the user’s needs, and then select the path with the shortest distance. So this problem can be regarded as a combinatorial optimization problem and solved with deep reinforcement learning. Hence, in this paper, we explore the application of deep reinforcement learning in solving TPQ problem. Since the selection of POI can be considered as a sequence decision problem, we model it as a seq2seq problem. Firstly, to help the model reduce the difficulty of selection, we remove POIs that can not be the result, and propose a candidate set generation method. Its nodes are enough to meet the query POIs for the model to select different node sequences. Secondly, we use the encoder-decoder model base on attention mechanism. We concatenate the embedding of the start point, the end point and the selected nodes as the query part of the attention mechanism. We mask the same poi after select so that the model can solve the TPQ problem. Finally, we employ the REINFORCE method for training with a greedy baseline. Our model has a good performance on different maps, different POI densities, and different numbers of required POIs.

Detecting small, multi-scale, and easily obscured traffic signs in real-world scenarios presents a persistent challenge. This paper proposes an approach that utilizes a multi-scale feature pyramid module to capture hierarchical features, facilitating robust detection of traffic signs across varying viewing angles and scales. To aggregate features at different scales and eliminate background interference, we employ a superposition of null convolution kernels with varying dilation rates, expanding the perceptual field from small to large. This effectively covers the object distribution across multiple scales while enhancing the resolution of the final output feature map for improved small target localization. Our method has demonstrated its effectiveness and superiority over several state-of-the-art approaches through extensive experiments conducted on two public traffic sign detection datasets.

Traffic congestion occurs frequently and concurrently on urban road networks, and may cause widespread traffic paralysis if not controlled promptly. To relieve traffic congestion and avoid traffic paralysis, it is significant to identify critical congested roads with great propagation influence on others. Existing studies mainly focus on topological measures and statistical approaches to evaluate the criticality of road segments. However, critical congested roads are generated by dynamic changes in traffic flow, so that identifying them involves both the network structure and dynamic traffic flows is required. In this paper, we propose a novel road network embedding model, called Seg2Vec, to learn comprehensive features of road segments considering both the road structural information and traffic flow distribution. The Seg2Vec model combines a Markov Chain-based random walk with the Skip-gram model. The random walk is conducted on the road network based on the transition probabilities computed from historical trajectory data. Moreover, we define the propagation influence of a congested road by a score function based on the learned road representation. The goal is to find the critical congested roads with top-K propagation influences. Evaluation experiments are conducted to verify the effectiveness and efficiency of the proposed method. A case study of identifying critical congested roads from a congestion cluster is also demonstrated. The identified critical congested roads can facilitate decision-making for traffic management.

With the rapid development of mobile Internet and sharing economy, carsharing has attracted a lot of attention around the globe. Many popular taxi-calling service platforms, such as DiDi and Uber, have provided carsharing service to the passengers. Such carpooling service reduces the energy consumption while meeting passengers’ convenience and economic benefits. Although numbers of algorithms have been proposed to support carsharing, the computing efficiency and matching quality of these existing algorithms are all sensitive to the distribution of passengers. In many cases, they cannot effectively and efficiently support carsharing in an on-line way. Motivated from the aforementioned issues and challenges, in this paper, we propose a novel framework, namely, Cross-Region-based Task Matching (CRTM) for supporting carsharing for smart city. Compared with existing algorithms, CRTM analyzes and monitors regions having multitudes of tasks for car sharing among users. In order to achieve this goal, we first propose a new machine learning-based algorithm to find a group of regions which contain many tasks. Then, we propose a novel index, namely, Included Angle Partition-based B-tree (IAPB), for maintaining tasks such as (i)whose pick-up points are contained in these regions, (ii) that may pass this kind of regions. Thirdly, we propose three buffer-based matching algorithms for cross-region-based task matching. Experiment results demonstrate the significant superior performance of the proposed algorithms in terms of energy saving and overall cost minimization.

Traffic forecasting is one of the most fundamental problems in transportation science and artificial intelligence. The key challenge is to effectively model complex spatial-temporal dependencies and correlations in modern traffic data. Existing methods, however, cannot accurately model both long-term and short-term temporal correlations simultaneously, limiting their expressive power on complex spatial-temporal patterns. In this paper, we propose a novel spatial-temporal neural network framework: Attention-based Spatial-Temporal Graph Convolutional Recurrent Network (ASTGCRN), which consists of a graph convolutional recurrent module (GCRN) and a global attention module. In particular, GCRN integrates gated recurrent units and adaptive graph convolutional networks for dynamically learning graph structures and capturing spatial dependencies and local temporal relationships. To effectively extract global temporal dependencies, we design a temporal attention layer and implement it as three independent modules based on multi-head self-attention, transformer, and informer respectively. Extensive experiments on five real traffic datasets have demonstrated the excellent predictive performance of all our three models with all their average MAE, RMSE and MAPE across the test datasets lower than the baseline methods.

The classification of driving behavior, with a particular emphasis on discerning safe from unsafe practices, is a task of paramount importance in the appraisal of drivers, and its significance is escalating in the epoch of autonomous driving. Driving behavior classification typically employs an assortment of features, such as velocity, acceleration, pedal pressure, turn signal utilization, and Global Positioning System (GPS) signals, amongst others. Nonetheless, these features exhibit considerable heterogeneity and do not offer comprehensive coverage. The extant literature pertaining to time series classification grapples with efficaciously addressing the high-dimensional nature, voluminous data, and the complexity of scenarios within the safety classification of driving behavior, especially for new energy vehicles. In this study, we have amassed an extensive corpus of sensor data, generated during the operation of new energy vehicles. Our research focused on the classification of driving behaviors concerning safety within the context of new energy vehicles and was predicated upon self-supervised learning. We proffered a time series model that leverages the Transformer architecture, tailored specifically for the aforementioned scenario, and employed a pre-training framework. To ascertain the efficacy of the proposed model, it was subjected to rigorous validation against a dataset comprising driving data from new energy vehicles. The model exhibited commendable performance and was further assessed through a series of downstream tasks.

The probabilistic estimation for multivariate time series forecasting has recently become a trend in various research fields, such as traffic, climate, and finance. The multivariate time series can be treated as an interrelated system, and it is significant to assume each variable to be independent. However, most existing methods fail to simultaneously consider spatial dependencies and probabilistic temporal dynamics. To address this gap, we introduce the Graph Convolution Recurrent Denoising Diffusion model (GCRDD), a recurrent framework for spatial-temporal forecasting that captures both spatial dependencies and temporal dynamics. Specifically, GCRDD incorporates the structural dependency into a hidden state using the graph-modified gated recurrent unit and samples from the estimated data distribution at each time step by a graph conditional diffusion model. We reveal the comparative experiment performance of state-of-the-art models in two real-world road network traffic datasets to demonstrate it as the competitive probabilistic multivariate temporal forecasting framework.

Working with observational data in the context of geophysics can be challenging, since we often have to deal with missing data. This requires imputation techniques in pre-processing to obtain data-mining-ready samples. Here, we present a convolutional neural network (CNN) approach from the domain of deep learning to reconstruct complete data from sparse inputs. CNN architectures are state-of-the-art for image processing. As data, we use two-dimensional fields of sea level pressure (SLP) and sea surface temperature (SST) anomalies. To have consistent data over a sufficiently long time span, we favor to work with output from control simulations of two Earth System Models (ESMs), namely the Flexible Ocean and Climate Infrastructure and the Community Earth System Model. Our networks can restore complete information from incomplete input samples with varying rates of missing data. Moreover, we present a technique to identify the most relevant grid points of our input samples. Choosing the optimal subset of grid points allows us to successfully reconstruct SLP and SST anomaly fields from ultra sparse inputs. As a proof of concept, the insights obtained from ESMs can be transferred to real world observations to improve reconstruction quality. As uncertainty measure, we compare several climate indices derived from reconstructed versus complete fields.

Click-Through-Rate (CTR) is a fundamental metric used to assess the efficacy of recommendation systems. In the past, most CTR prediction approaches focused on modeling the cross feature of various feature fields to improve the accuracy of CTR prediction. But they only learned the fixed representation of feature and neglected the varying significance of different feature fields in distinct contexts - what we refer to as context sensitive information - leading to suboptimal performance. While recent approaches have attempted to leverage linear transformations and feature interactions to capture context sensitive information, they remain inadequate as they overlook the varying importance of original features or different order cross features. In this paper, we propose a new module called Enhancing Feature Network (EFNet). EFNet has two key components: 1) Information Capture Layer (ICL), that dynamically captures explicit and implicit context sensitive information from original embedding features and digs out their corresponding bit-level weights; 2) Enhancing Feature Layer (EFL) that adaptively combine the context sensitive information with original embedding features according to the weights obtained in ICL. It is worth noting that EFNet can be integrated into existing CTR prediction models as a module to boost their overall performance. We conduct comprehensive experiments on four public datasets and the results demonstrate that models incorporating the EFNet module outperform other state-of-the-art models.

Click-Through Rate (CTR) prediction is a critical problem in recommendation systems since it involves enormous business interest. Most deep CTR model follows an Embedding & Feature Interaction paradigm. However, the feature interaction module cannot work well without a good embedding representation of features. Due to the long-tail phenomenon in real scenes, few samples are provided in the dataset for a large proportion of features. In this paper, we present ProtoMix, a model-agnostic framework for learnable data augmentation on few-shot features in CTR prediction. ProtoMix automatically extracts information from co-occurred features within the same instance to assign prototype embedding with vector quantization for few-shot features and further synthesize the embedding representation of the augmented virtual instance for training. Original embedding, feature interaction module, and the embedding generator are jointly trained on a well-designed objective in an end-to-end manner in ProtoMix. We experimentally validate the effectiveness and compatibility of ProtoMix by comparing it with baseline and other data augmentation methods on different deep CTR models and multiple real-world CTR benchmark datasets.

Click-Through Rate (CTR) prediction is a core task in recommendation systems. Despite VAE-based models have shown promising accuracy performance, they are still weak in supporting cold-start CTR prediction due to limited personal interactions. To this end, this paper proposes a content-based variational CTR model, which jointly models content information and interactions behaviors in a shared probability space via variational inference. Specifically, a three-step scheme is designed to fully utilize content information for the improved ability of preference modeling toward cold-start users. First, our method adopts VAE to model user preferences from personal interactions by probabilistic distributions, instead of a fixed embedding vector for representing the user’s interest. Then, we transform content information into variational probabilistic distribution to model the implicit preferences of cold-start users. Finally, a variational alignment strategy is applied to maximize the similarity between variational preference distributions obtained from interactions behaviors and content information respectively, so that the interest of the cold user can be recovered. Besides, we adopt a self-attention mechanism to reasonably balance the importance of latent features for CTR prediction. Experiments on two public real datasets show the effectiveness of the proposed approach.

The data encountered by Session-based Recommendation System(SBRS) is typically highly sparse, which also serves as one of the bottlenecks limiting the accuracy of recommendations. So Contrastive Learning(CL) is applied in SBRS owing to its capability of improving embedding learning under the condition of sparse data. However, existing CL strategies are limited in their ability to enforce finer-grained (e.g., factor-level) comparisons and, as a result, are unable to capture subtle differences between instances. More than that, these strategies usually use item or segment dropout as a means of data augmentation which may result in sparser data and thus ineffective self-supervised signals. By addressing the two aforementioned limitations, we introduce a novel dual-granularity CL framework. Specifically, two extra augmentation views with different granularities are constructed and the embeddings learned by them are compared with those learned from original view to complete the CL tasks. At factor-level, we employ Disentangled Representation Learning to obtain finer-grained data, with which we can explore connections of items on latent factor independently and generate factor-level embeddings. At item-level, the star graph is deployed as the augmentation method. By setting an additional satellite node, non-adjacent nodes can establish additional connections through satellite nodes instead of reducing the connections of the original graph, so data sparsity can be avoided. Compare the learned embeddings of these two views with the learned embeddings of the original view to achieve CL at two granularities. Finally, the item-level and factor-level embeddings obtained are referenced to generate personalized recommendations for the user. The proposed model is validated through extensive experiments on two benchmark datasets, showcasing superior performance compared to existing methods.

Recently, Contrastive Learning (CL) is becoming a mainstream approach to reduce the influence of data sparsity in recommendation system. However, existing methods do not fully explore the relationship between the outputs of different Graph Neural Network (GNN) layers and fail to fully utilize the capacity of combining GNN and CL for better recommendation. Within this paper, we introduce a novel approach based on CL, called efficient Graph collaborative filtering with multi-layer output-enhanced Contrastive Learning (GmoCL). It maximizes the benefits derived from the information propagation property of GNN with multi-layer aggregation to obtain better node representations. Specifically, the construction of CL tasks involves considerations from both intra-layer and inter-layer perspectives. The goal of intra-layer CL task is to exploit the semantic similarities of different users (or items) on a certain GNN layer. The inter-layer CL task aims to make the outputs of different GNN layers of the same user (or item) more similar. Additionally, we propose the strategy of negative sampling in the inter-layer CL task to learn the better node representations. The efficacy of the suggested approach is validated through comprehensive experiments conducted on five publicly available datasets.

Given a Social Network how to select a small number of influential users to maximize the influence in the network has been studied extensively in the past two decades and formally referred to as the Influence Maximization Problem. Among most of the existing studies, it has been implicitly assumed that there exists a single probability value that represents the influence probability between the users. However, in reality, the influence probability between any two users is dependent on the context (formally referred to as tag e.g.; a sportsman can influence his friends related to any news related to sports with high probability). In this paper, we bridge the gap by studying the Tag-Based Influence Maximization Problem. In this problem, we are given with a social network where each edge is marked with one probability value for every tag and the goal here is to select k influential users and r influential tags to maximize the influence in the network. First, we define a tag-based influence function and show that this function is bi-submodular. We use the orthent-wise maximization procedure of bi-submodular function which gives a constant factor approximation guarantee. Subsequently, we propose a number of efficient pruning techniques that reduces the computational time significantly. We perform an extensive number of experiments with real-world datasets to show the effectiveness and efficiency of the proposed solution approaches.

Social recommendation endeavors to harness users’ social connections to enhance recommender systems. Graph Neural Network (GNN) has gained traction for its robust capacity in managing graph data. However, previous social recommendation works have failed to fully consider the crucial role of user distinct interests, which hinders their ability to accurately model complex user preferences and negatively affected the modeling of social influence. To tackle this challenge, we introduce a multi-interest approach to GNN-based social recommendation. Specifically, we firstly utilize a dynamic routing algorithm to cluster user interests from the items they have interacted with. Subsequently, we use a similarity-weighted GCN operation to capture user relationships within the social network. Finally, we use the aggregate the multiple interest representations for prediction. Our comprehensive experiments underscore the consistent superiority of our model compared with state-of-the-art competitors on real-world datasets.

Multimedia recommendation aims to predict whether users will interact with multimodal items. A few recent works that explicitly learn the semantic structure between items using multimodal features manifest impressive performance gains. This is mainly attributed to the capability of graph convolutional networks (GCNs) to learn superior item representations by propagating and aggregating information from high-order neighbors on the semantic structure. However, they still suffer from two major challenges: a) the noisy relations (edges) in the item-item semantic structure disrupt information propagation and generate low-quality item representations, which impairs the effectiveness and robustness of existing methods; b) the lack of an optimization objective that exploits informative samples and global preference information leads to suboptimal training of the model, which makes users and items indistinguishable in the embedding space. To overcome these challenges, we propose Enhancing Multi media Recommendation through Item-Item Semantic Denoising and Global Preference Awareness (MMGPA). Specifically, the model contains the following two components: (1) a modal semantic representation network is carefully designed to learn the high-quality multimodal representation of items by modeling the denoised item-item semantic structure, and (2) a global preference-aware optimization objective prioritizes the most informative hard sample pairs while constraining the multiple preference distances to better separate the embedding space. Extensive experimental results demonstrate that the proposed method outperforms various state-of-the-art competitors on three public benchmark datasets.

The objective of community commercial planning is to identify appropriate stores to operate in a community shopping center, catering to the daily needs of residents and enhancing the appeal of the shopping center. However, obtaining data on the characteristics of all residents in the community is a major challenge, and practical methods for selecting suitable stores based on resident characteristics are unavailable. To address these issues, we propose a model that leverages mutual information maximization to learn representations of valuable residents in the shopping area and assess their value. Our key innovation is a value-ranking encoder-decoder that learns the characteristics of all residents in the community and recommends the most suitable store for each storefront. To balance the diversity and competition of businesses within the shopping center, we introduce a diversity loss function. Extensive experimental results show the effectiveness of our model.