World Wide Web最新文献

Supervised cross-modal hashing has attracted many researchers. In these studies, they seek a common semantic space or directly regress the zero-one label information into the Hamming space. Although they achieve many achievements, they neglect some issues: 1) some methods of the classification task are not suitable for retrieval tasks, since they are lack of learning personalized features of sample; 2) the outcomes of hash retrieval are related to both the length and encoding method of hash codes. Because a sample possess more personalized features than label semantics, in this paper, we propose a novel supervised cross-modal hashing collaboration learning method called discrete Cross-modal Hashing with Relaxation and Label Semantic Guidance (CHRLSG). First, we introduce two relaxation variables as latent spaces. One is used to extract text features and label semantic information collaboratively, and the other is used to extract image features and label semantics collaboratively. Second, the more accurate hash codes are generated from latent spaces, since CHRLSG learns collaboratively feature semantics and label semantics by using labels as the domination and features as the auxiliary. Third, we utilize labels to strengthen the similar relationship of inter-modal samples via keeping the pairwise closeness. Label semantics are made full use of to avoid classification error. Fourth, we introduce class weight to further increase the discrimination of samples that belong to different classes in intra-modal and keep the similarity of samples unchanged. Therefore, CHRLSG model preserves not only the relationship between samples, but also maintains the consistency of label semantic during collaboration optimization. Experimental results of three common benchmark datasets demonstrate that the proposed model is superior to the existing advanced methods.

Online social networks facilitate the spread of information, while rumors can also propagate widely and fast, which may mislead some users. Therefore, suppressing the spread of rumors has become a daunting task. One of the widely used approaches is to select users in the social network to spread the truth and compete against the rumor, so that users who receive the truth before receiving rumors will not trust or propagate the rumor. However, the existing works only aim to speed up blocking rumors without considering the pertinency of users. For example, consider a social media platform operator aiming to enhance user online safety. Based on the user’s online behavior, the users who are at high risk should be alerted first. Motivated by this, we formally define the rumor blocking with pertinence set (RBP) problem, which aims to find a truth seed set that maximizes the number of nodes affected by truth and ensures that the number of influenced nodes within the pertinence set reaches at least a given threshold. To solve this problem, we design a hybrid greedy framework (HGF) algorithm with local and global phases. We prove that HGF can provide a ((1-1/e-epsilon ))-approximate solution with high probability while reducing the cost of the sampling process. Extensive experiments on 8 real social networks demonstrate the efficiency and effectiveness of our proposed algorithms.

Abstract

Because of the processing of continuous unstructured large streams of data, mining real-time streaming data is a more challenging research issue than mining static data. The privacy issue persists when sensitive data is included in streaming data. In recent years, there has been significant progress in research on the anonymization of static data. For the anonymization of quasi-identifiers, two typical strategies are generalization and suppression. However, the high dynamicity and potential infinite properties of the streaming data make it a challenging task. To end this, we propose a novel Efficient Approximation and Privacy Preservation Algorithms (EAPPA) framework in this paper to achieve efficient data pre-processing from the live streaming and its privacy preservation with minimum Information Loss (IL) and computational requirements. As the existing privacy preservation solutions for streaming data suffer from the challenges of redundant data, we first propose the efficient technique of data approximation with data pre-processing. We design the Flajolet Martin (FM) algorithm for robust and efficient approximation of unique elements in the data stream with a data cleaning mechanism. We fed the periodically approximated and pre-processed streaming data to the anonymization algorithm. Using adaptive clustering, we propose innovative k-anonymization and l-diversity privacy principles for data streams. The proposed approach scans a stream to detect and reuse clusters that fulfill the k-anonymity and l-diversity criteria for reducing anonymization time and IL. The experimental results reveal the efficiency of the EAPPA framework compared to state-of-art methods.

Sequential recommendation aims to predict the next items that users will interact with according to the sequential dependencies within historical user interactions. Recently, self-attention based sequence modeling methods have become the mainstream method due to their competitive accuracy. Despite their effectiveness, these methods still have non-trivial limitations: (1) they mainly take the transition patterns between items into consideration but ignore the semantic associations between items, and (2) they mostly focus on dynamic short-term user preferences and fail to consider user static long-term preferences explicitly. To address these limitations, we propose a Knowledge Enhanced Personalized Hierarchical Attention Network (KPHAN), which can incorporate the semantic associations among items by learning from knowledge graphs and capture the fine-grained long- and short-term interests of users through a novel personalized hierarchical attention network. Specifically, we employ the entities and relationships in the knowledge graph to enrich semantic information for items while preserving the structural information of the knowledge graph. The self-attention mechanism then captures semantic associations among items to obtain short-term user preferences more accurately. Finally, a personalized hierarchical attention network is developed to generate the final user preference representations, which can fully capture user static long-term preferences while fusing dynamic short-term preferences. Experimental results on three real-world datasets demonstrate that our method can outperform prior works by 2.7% - 35.5% on HR metrics and 6.7% - 27.9% on NDCG metrics.

Networks in the real world are dynamic and evolving. The most critical process in networks is to determine the structure of the community, based on which we can detect hidden communities in a complex network. The design of strong network structures is of great importance, meaning that a system must maintain its function in the face of attacks and failures and have a strong community structure. In this paper, we proposed the robust memetic algorithm and used the idea to optimize the detection of dynamic communities in complex networks called RDMA_NET (Robust Dynamic Memetic Algorithm). In this method, we work on dynamic data that affects the two main parts of the initial population value and the calculation of the evaluation function of each population, and there is no need to determine the number of communities in advance. We used two sets of real-world networks and the LFR dataset. The results show that our proposed method, RDMA_Net, can find a better solution than modern approaches and provide near-optimal performance in search of network topologies with a strong community structure.

The privacy-preserving data publishing (PPDP) problem has gained substantial attention from research communities, industries, and governments due to the increasing requirements for data publishing and concerns about data privacy. However, achieving a balance between preserving privacy and maintaining data quality remains a challenging task in PPDP. This paper presents an information-driven distributed genetic algorithm (ID-DGA) that aims to achieve optimal anonymization through attribute generalization and record suppression. The proposed algorithm incorporates various components, including an information-driven crossover operator, an information-driven mutation operator, an information-driven improvement operator, and a two-dimensional selection operator. Furthermore, a distributed population model is utilized to improve population diversity while reducing the running time. Experimental results confirm the superiority of ID-DGA in terms of solution accuracy, convergence speed, and the effectiveness of all the proposed components.

Abstract

With Bitcoin being universally recognized as the most popular cryptocurrency, more Bitcoin transactions are expected to be populated to the Bitcoin blockchain system. As a result, many transactions can encounter different confirmation delays. Concerned about this, it becomes vital to help a user understand (if possible) how long it may take for a transaction to be confirmed in the Bitcoin blockchain. In this work, we address the issue of predicting confirmation time within a block interval rather than pinpointing a specific timestamp. After dividing the future into a set of block intervals (i.e., classes), the prediction of a transaction’s confirmation is treated as a classification problem. To solve it, we propose a framework, Hybrid Confirmation Time Estimation Network (Hybrid-CTEN), based on neural networks and XGBoost to predict transaction confirmation time in the Bitcoin blockchain system using three different sources of information: historical transactions in the blockchain, unconfirmed transactions in the mempool, as well as the estimated transaction itself. Finally, experiments on real-world blockchain data demonstrate that, other than XGBoost excelling in the binary classification case (to predict whether a transaction will be confirmed in the next generated block), our proposed framework Hybrid-CTEN outperforms state-of-the-art methods on precision, recall and f1-score on all the multiclass classification cases (4-class, 6-class and 8-class) to predict in which future block interval a transaction will be confirmed.

Knowledge representation models have been extensively studied and they provide an important foundation for artificial intelligence. However, the existing knowledge representation models or related knowledge embedding methods mostly aim at static or temporal knowledge, which are not suitable for highly spatio-temporal relevant knowledge, such as the cyber security knowledge. In this paper, we propose a knowledge embedding framework called TSEE to handle this problem, which builds on the MDATA model to represent and utilize dynamic knowledge for cyber security. TSEE is composed of knowledge extraction module, knowledge representation module, knowledge embedding module, and situational awareness module. There modules can obtain, transform, and embed cyber security knowledge from different sources, improving the detection capabilities of various complicated attacks. We conduct experiments on the cyber range for evaluation, and the experimental results validate the higher prediction accuracy and stronger extendability than existing embedding methods. The framework can effectively improve the cyber security defense capabilities in the future.

Reinforcement learning (RL) has achieved ideal performance in recommendation systems (RSs) by taking care of both immediate and future rewards from users. However, the existing RL-based recommendation methods assume that only a single type of interaction behavior (e.g., clicking) exists between user and item, whereas practical recommendation scenarios involve multiple types of user interaction behaviors (e.g., adding to cart, purchasing). In this paper, we propose a Multi-Task Reinforcement Learning model for multi-behavior Recommendation (MTRL4Rec), which gives different actions for users’ different behaviors with a single agent. Specifically, we first introduce a modular network in which modules can be shared or isolated to capture the commonalities and differences across users’ behaviors. Then a task routing network is used to generate routes in the modular network for each behavior task. We adopt a hierarchical reinforcement learning architecture to improve the efficiency of MTRL4Rec. Finally, a training algorithm and a further improved training algorithm are proposed for our model training. Experiments on two public datasets validated the effectiveness of MTRL4Rec.

Conversation Recommender System (CRS) engage in multi-turn conversations with users and provide recommendations through responses. As user preferences evolve dynamically during the course of the conversation, it is crucial to understand natural interaction utterances to capture the user’s dynamic preference accurately. Existing research has focused on obtaining user preference at the entity level and natural language level, and bridging the semantic gap through techniques such as knowledge augmentation, semantic fusion, and prompt learning. However, the representation of each level remains under-explored. At the entity level, user preference is typically extracted from Knowledge Graphs, while other modal data is often overlooked. At the natural language level, user representation is obtained from a fixed language model, disregarding the relationships between different contexts. In this paper, we propose User-augmented Conversation Recommendation via Multi-modal graph learning and Context Mining (UaMC) to address above limitations. At the entity level, we enrich user preference by leveraging multi-modal knowledge. At the natural language level, we employ contrast learning to extract user preference from similar contexts. By incorporating the enhanced representation of user preference, we utilize prompt learning techniques to generate responses related to recommended items. We conduct experiments on two public CRS benchmarks, demonstrating the effectiveness of our approach in both the recommendation and conversation subtasks.