World Wide Web最新文献

Heterogeneous information networks have been proven to effectively improve recommendations due to their diverse information content. However, there are still two challenges for recommendation methods based on heterogeneous information networks. To begin with, current methods often depend on experts to manually craft meta-paths, and it can be challenging to define an adequate set of meta-paths for complex task scenarios. Second, most models fail to fully explore user preferences for paths or meta-paths whileimultaneously learning path or meta-path explicit representations. To tackle the aforementioned issues, we propose a model for recommendation utilizing meta-path automatically extracted from heterogeneous information network, called MAERec. Specifically, MAERec employs an automatic approach to extract high-quality path instances from heterogeneous information networks and construct meta-paths. These meta-paths are then utilized by a hierarchical attention network to learn an explicit representation of the meta-path-based context. Extensive experiments conducted on various real-world datasets not only showcase the superior performance of MAERec when compared to state-of-the-art methods but also underscore its capability to automatically discover high-quality path instances for meta-path extraction.

Coverage Centrality is an important metric to evaluate vertex importance in road networks. However, current solutions have to compute the coverage centrality of all the vertices together, which is resource-wasting, especially when only some vertices centrality is required. In addition, they have poor adaption to the dynamic scenario because of the computation inefficiency. In this paper, we focus on the coverage centrality query problem and propose a method that efficiently computes the centrality of single vertices without relying on the underlying graph being static by employing the intra-region pruning, inter-region pruning, and top-down search. We further propose the bottom-up search and mixed search to improve efficiency. Experiments validate the efficiency and effectiveness of our algorithms compared with the state-of-the-art method.

Inductive relation prediction is an important learning task for knowledge graph reasoning that aims to infer new facts from existing ones. Previous graph neural networks (GNNs) based methods have demonstrated great success in inductive relation prediction by capturing more subgraph information. However, they aggregate all reasoning paths which might introduces redundant information. Such redundant information changes with the context of entity and easily outside the training distribution making existing GNN-base methods suffer from poor generalization. In this work, we propose a novel causal knowledge graph reasoning (CKGR) framework for inductive relation prediction task with better generalization. We first take a causal view of inductive relation prediction and construct a structural causal model (SCM) that reveals the relationship between variables. With our assumption, CKGR extracts causal and shortcut subgraphs conditioned on query triplet. Then, we parameter the backdoor adjustment of causality theory by making intervention in representation space. In this way, CKGR can learn stable causal feature and alleviates the confounding effect of shortcut features that are spuriously correlated to relation prediction. Extensive experiments on various tasks with real-world and synthetic datasets demonstrate the effectiveness of CKGR.

Answering complex queries with First-order logical operators over knowledge graphs, such as conjunction ((wedge )), disjunction ((vee )), and negation ((lnot )) is immensely useful for identifying missing knowledge. Recently, neural symbolic reasoning methods have been proposed to map entities and relations into a continuous real vector space and model logical operators as differential neural networks. However, traditional methodss employ negative sampling, which corrupts complex queries to train embeddings. Consequently, these embeddings are susceptible to divergence in the open manifold of (mathbb {R}^n). The appropriate regularization is crucial for addressing the divergence of embeddings. In this paper, we introduces a Lie group as a compact embedding space for complex query embedding, enhancing ability to handle the intricacies of knowledge graphs the foundation model. Our method aims to solve the query of disjunctive and conjunctive problems. Entities and queries are represented as a region of a high-dimensional torus, where the projection, intersection, union, and negation of the torus naturally simulate entities and queries. After simulating the operations on the region of the torus we defined, we found that the resulting geometry remains unchanged. Experiments show that our method achieved a significant improvement on FB15K, FB15K-237, and NELL995. Through extensive experiments on datasets FB15K, FB15K-237, and NELL995, our approach demonstrates significant improvements, leveraging the strengths of knowledge graphs foundation model and complex query processing.

In the context of searching a single data graph G, graph pattern matching is to find all the occurrences of a pattern graph Q in G, specified by a matching rule. It is of paramount importance in many real applications such as social network analysis and cyber security, among others. A wide spectrum of studies target general graph pattern matching. However, to analyze time-relevant services such as studying the spread of diseases and detecting attack patterns, it is attractive to study inexact temporal graph pattern matching. Hence, in this paper, we propose a relaxed matching rule called constrained temporal dual simulation, and study simulation-based constrained temporal graph pattern matching which guarantees that the matching result (i) preserves the ancestor and descendant temporal connectivities; and (ii) implements edge-to-temporal path mapping. We devise a decomposition-based matching method, which first decomposes the data graph into Source Temporal Connected Components, and then performs matching on decomposed subgraphs. To speed up the matching, we define child/parent dependency relation tables and propose an efficient double hierarchical traverse strategy. Considering that the temporal graphs are naturally dynamic, we further propose update algorithms. An extensive empirical study over real-world and synthetic temporal graphs has demonstrated the effectiveness and efficiency of our approach.

The sequential recommendation task based on the multi-interest framework aims to model multiple interests of users from different aspects to predict their future interactions. However, researchers rarely consider the differences in features between the interests generated by the model. In extreme cases, all interest capsules have the same meaning, leading to the failure of modeling users with multiple interests. To address this issue, we propose the High-level Preferences as positive examples in Contrastive Learning for multi-interest Sequence Recommendation framework (HPCL4SR), which uses contrastive learning to distinguish differences in interests based on user item interaction information. In order to find high-quality comparative examples, this paper introduces the category information to construct a global graph, learning the association between categories for high-level preference interest of users. Then, a multi-layer perceptron is used to adaptively fuse the low-level preference interest features of the user’s items and the high-level preference interest features of the categories. Finally, user multi-interest contrastive samples are obtained through item sequence information and corresponding categories, which are fed into contrastive learning to optimize model parameters and generate multi-interest representations that are more in line with the user sequence. In addition, when modeling the user’s item sequence information, in order to increase the differentiation between item representations, the category of the item is used to supervise the learning process. Extensive experiments on three real datasets demonstrate that our method outperforms existing multi-interest recommendation models.

Abstract

It is challenging to deploy convolutional neural networks (CNNs) for image denoising on low-power devices which can suffer from computational and memory constraints. To address this limitation, a simple yet effective and efficient feature redundancy reduction-based network (FRRN) is proposed in this paper, which integrates a feature refinement block (FRB), an attention fusion block (AFB), and an enhancement block (EB). Specifically, the FRB distills structural information via two parallel sub-networks, selecting representative feature representations while suppressing spatial-channel redundancy. The AFB absorbs an attentive fusion mechanism to facilitate diverse features extracted from two sub-networks, emphasizing texture and structure details but alleviating harmful features from problematic regions. The subsequent EB further boosts the feature representation abilities. Aiming to enhance denoising performance at both pixel level and semantic level, a multi-loss scheme comprising three popular loss functions is leveraged to improve the robustness of the denoiser. Comprehensive quantitative and qualitative analyses demonstrate the superiority of the proposed FRRN.

Abstract

Nowadays, with the development of recommender systems, an emerging recommendation scenario called group-to-group recommendation has played a vital role in information acquisition for users. The new recommendation scenario seeks to recommend a group of related items to users with similar interests. To some extent, it alleviates the problem of point-to-point recommendations getting trapped in an information cocoon due to an over-reliance on user behaviors. For the new recommendation scenario, the existing recommendation methods cannot model the complex interactions between user groups and item groups, thus affecting the accuracy of the group-to-group recommendation. In this paper, we propose a group-to-group recommendation method, which abstracts user groups and item groups into graphs and calculates the similarity between two graphs based on graph matching, dubbed as GMRec. Specifically, we construct the graph of user groups and item groups and then calculate the graph similarity scores between user groups and item groups from two perspectives of feature matching and structure matching. Experimental results show that our model achieves higher accuracy than state-of-the-art models on three industrial datasets with different group sizes, with a maximum improvement of 8.2%.

The huge amount of graph data necessitates sampling methods to support graph-based analysis applications. Node influence is to count the influential nodes with a given node in large graphs that has wide applications including product promotion and information diffusion in social networks. However, existing sampling methods mainly consider node degree to compute the node influence while ignoring the important connections in terms of groups in which nodes participate, resulting in inaccuracy of influence estimations. To this end, this paper proposes group sampling, called GVRW, to count the groups along with node degrees to evaluate node influence in large graphs. Specifically, GVRW changes the way of random walker traversing a large graph from one node to a random neighbor node of the groups to enlarge the sampling space for the sake of characterizing the nodes and groups simultaneously. Furthermore, we carefully design the corresponding estimated method to employ the samples to estimate the specific distributions of groups and node degrees to compute the node influence. Experimental results on real-world graph datasets show that our proposed sampling and estimating methods can accurately obtain the properties and approximate the node influences closer to the real values than existing methods.

Online question matching is the process of comparing user queries with system questions to find appropriate answers. This task has become increasingly important with the popularity of knowledge sharing social networks in product search and intelligent Q &A in customer service. Many previous studies have focused on designing complex semantic structures through the questions themselves. In fact, the online user’s queries accumulate a large number of similar sentences, which have been grouped by semantics in the retrieval system. However, how to use these sentences to enhance the understanding of system questions is rarely studied. In this paper, we propose a novel Relation-aware Semantic Enhancement Network (RSEN) model. Specifically, we leverage the labels of the history records to identify different semantically related sentences. Then, we construct an expanded relation network to integrate the representation of different semantic relations. Furthermore, we interact we integrate the features of the system question with the semantically related sentences to augment the semantic information. Finally, we evaluate our proposed RSEN on two publicly available datasets. The results demonstrate the effectiveness of our proposed RSEN method compared to the advanced baselines.