Latest ArticlesCross-domain recommendation (CDR) can effectively alleviate the data sparsity problem suffered by the traditional recommendation systems via leveraging additional knowledge from other domains. How to model the interaction information of users and items from the source to target domains is a key issue in CDR. In the current CDR methods, the higher-order information implied by the user-item interaction graph is ignored. To this end, we propose a new framework called graph convolutional broad cross-domain recommender system (GBCD). Specifically, we extend the traditional bipartite graph of user-item interactions to a (D + 1)-partite graph to model the relationship between users and items in each domain, and then use common users as a bridge between the source domain and target domain to transfer information. The higher-order relationships between users and items are learned by graph convolutional network (GCN) to aggregate neighbor information. However, GCN converges very slowly with a large number of nodes and tends to absorb unreliable interaction noise, resulting in poor robustness. Therefore, we feed the domain-aggregated features to broad learning system (BLS), which enhances the robustness of GCN by exploiting the stochastic mapping features of BLS, achieving superior recommendation performance. Experiments conducted on two real datasets show that GBCD outperforms the existing state-of-the-art cross-domain recommendation methods.
Recent advancements in the field of sequential recommender have focused on refining user interests through various methods, such as clustering historical interactions or utilizing graph convolutional neural networks to capture multi-level correlations among interactions. However, while these approaches have significantly advanced the field, they often overlook the interactions between users with similar behavioral patterns and the impact of irregular time intervals within interaction sequences on user interests. Based on the above problems, a multi-interest aware sequential recommender model (MIRec) based on contrastive learning is proposed. This model takes into account both local preference information, including item dependence and location dependence within a sequence, and global preference information obtained through a graph information aggregation mechanism among similar users. The user representations, which incorporate both local and global preferences, are fed into a capsule network to learn multi-interest representations within the user interaction sequence. Subsequently, the user’s historical interaction sequences are brought closer to enhanced interaction sequences through contrastive learning. This process results in the generation of a user’s multi-interest representation that is insensitive to time intervals, ultimately leading to more accurate recommendations for users. The effectiveness of this model is verified on two real datasets, and the experimental results verify the effectiveness of the model.
Combining super-resolution localization microscopy with panoramic digital pathology provides a powerful tool for biomedical researchers to observe the subcellar structure of the entire sample, but also brings the challenge of visualization of panoramic super-resolution image. However, current super-resolution image visualization methods are unable to process large-scale localization data, provide high-resolution panoramic images, and interactively visualize panoramic images. In response to the above problems, an interactive visualization framework for panoramic super-resolution images based on localization data is proposed, called PNanoViewer, which aims to achieve rapid interactive visualization of large-scale localization data on ordinary computers. This framework constructs the multi-resolution hierarchical structure of localization data based on the random sampling strategy, and visualizes multi-scale panoramic super-resolution image in an interactive way. Meanwhile, the block strategy and multi-thread parallel strategy are used to process large-scale localization data in batches to prevent memory overflow and speed up processing. Experimental results on tens to hundreds of millions of localization datasets show that the proposed PNanoViewer framework can visualize localization data with any size. Compared this framework with three currently popular super-resolution image visualization methods: PALMsiever, ThunderSTORM, and QC-STORM, PNanoViewer has obvious advantages in terms of data volume, resolution, and speed. At the same time, PNanoViewer also provides a useful exploration for the visualization of large-scale localization data.
Information systems usually use data management systems to manage data, among which SQL has been the mainstream query language for data management because of its ease of use and flexibility, and users can write SQL statements and submit them to the data management system to get query results. The efficiency of the execution model determines whether the system can quickly respond to user queries. The existing execution models mainly adopt interpreted execution and compiled execution. Interpreted execution is used by most systems due to its scalability and maintainability. Unlike interpreted execution, compiled execution generates efficient custom code to speed up queries that should have been processed by interpreted execution, and the significant performance gains have attracted a number of database systems to implement the technology. However, generating the corresponding custom code for a query is a complex process that requires a number of considerations, even in some cases, the performance of using compiled execution may not be as good as the traditional volcano model. We provide a systematic review of the progress of compiled execution techniques from conceptual and technical perspectives. Firstly, we outline the basic concepts of query compilation and introduce the relevant terminology and background knowledge. Secondly, we introduce the relevant techniques from three perspectives: intermediate code generation, intermediate representation, machine code generation and running. Finally, we look at the future development direction of compiled execution technology in the context of current research trends in data management systems and recent research work.
Due to the rapid development of information technology, society is in a critical period of digitalization and information transformation, and the demand for information systems based on database technology in various industries is becoming increasingly prominent. Location-based services rely on massive real-time generated trajectory data. In the processing of hundreds of millions of continuously changing trajectory data, dimensionality reduction algorithm and query technology have been the key to research. By reducing the scale of trajectory data and reducing the time of data processing during query operations, the performance of query can be effectively improved, and whether high-quality and efficient query can be achieved is very important for the database. In this paper, a UGC(uniform grid code) and a NGDR(non-uniform grid dimensionality reduction algorithm) for trajectory data are proposed, which convert the coordinates of trajectory data into one-dimensional string storage, merge the grids that do not meet the requirements, fully retain the complex interrelationship between trajectory data through spatial position mapping, and use range query and nearest neighbor query to test the performance of the reduced data. The real trajectory and virtual generated trajectory data in different cities are used as datasets, and the uniform grid code algorithm, non-uniform grid algorithm proposed in this paper are compared with three benchmark methods. Experiments show that the spatial position relationship similarity of the data after NGDR can be up to 82.5%. The range query time of NGDR is improved at least by 73.86% compared with the other queries, and the nearest neighbour query time is improved at least by 52.26%, which achieves better results than other benchmark methods.
In recent years, many research efforts have focused on combining data-driven machine learning with knowledge-driven logical reasoning, aiming to improve the performance of machine learning systems. Many works have attempted to use abductive reasoning to integrate machine learning and logical reasoning into a unified framework. These methods typically first generate pseudo-labels through machine learning models, and then use abductive reasoning to revise the inconsistent pseudo-labels, which serves to update the machine learning model and the above routine repeats. However, there may be incorrect labels in the abduced labels, which could have a negative impact on the training of the machine learning model and are challenging to detect. We propose abductive learning with rejection reasoning, a method that takes into account both model uncertainty and reasoning uncertainty of the abduced labels. This strategy comprehensively evaluates the reliability of abductive results from the perspective of both the data and knowledge levels, and avoids the negative impact of unreliable abduced labels on model training by rejecting some of the abductive reasoning results. Empirical studies show that the proposed method can reduce the proportion of incorrect abduced labels. In turn, this facilitates a faster training process for abductive learning and improves its overall performance.
In the era of big data, there is a large amount of streaming data emerging. Concept drift, as the most typical and difficult problem in streaming data mining, has received increasing attention. Ensemble learning is a common method for handling concept drift in streaming data. However, after drift occurs, learning models often cannot timely respond to the distribution changes of streaming data and cannot effectively handle different types of concept drift, leading to the decrease in model generalization performance. Aiming at this problem, we propose a two-stage adaptive ensemble learning method for different types of concept drift (TAEL). Firstly, the concept drift type is determined by detecting the drift span. Then, based on different drift types, a “filtering-expansion” two-stage sample processing mechanism is proposed to dynamically select appropriate sample processing strategy. Specifically, during the filtering stage, different non-critical sample filters are created for different drift types to extract key samples from historical sample blocks, making the historical data distribution closer to the latest data distribution and improving the effectiveness of the base learners. During the expansion stage, a block-priority sampling method is proposed, which sets an appropriate sampling scale for the drift type and sets the sampling priority according to the size proportion of the class in the current sample block to which the historical key sample belongs. Then, the sampling probability is determined based on the sampling priority, and a subset of key samples is extracted from the historical key sample blocks according to the sampling probability to expand the current sample block. This alleviates the class imbalance phenomenon after sample expansion, solves the underfitting problem of the current base learner and enhances its stability. Experimental results show that the proposed method can timely respond to different concept drift types, accelerate the convergence speed of online ensemble models after drift occurs, and improve the overall generalization performance of the model.
The global situation regarding mental health problems is a matter of serious concern. Unfortunately, in many countries, the shortage of mental health practitioners often leaves individuals with mental health problems without access to professional services. The retrieval-based community question answering for mental health can provide self-service information for those who are suffering from mental health problems. There are two different matching criteria to consider when matching question post and support response in this task, namely semantic matching and mental health matching. To address the issue of excessive non-ideal support responses in a large pool of candidate support responses, we propose a two-stage text matching model that integrates characters’ mental portraits using a multi-layer attention mechanism. This method divides our task into two sequential subtasks: a retrieval task and a matching task, respectively. In the first stage, a retrieval sub-method is used to retrieve a set of candidate support responses with a certain semantic similarity for each question post. In the second stage, the matching sub-method performs matching judgments based on the results generated in the previous stage. This sub-method applies a multi-layer attention mechanism to effectively integrate characters’ mental portraits and semantics. This process can modify the results of simple semantic matching. Experimental results show that T2CMP proposed in this paper is superior to state-of-the-art methods.
Graph convolutional network (GCN) has been widely used in recommendation systems due to its unique advantages in processing graph data. GCN propagates node attribute information by exploiting the dependencies between nodes in the graph, which greatly improves the accuracy of node representation and thus improves recommendation performance. However, existing GCN-based recommendation methods still have difficulty in modeling deeper layers due to the over-smoothing problem, which limits the representation of higher-order relationships between users and items. To this end, we propose a novel link prediction method driven by high-order relations among items in complete graph (short for LinkCG). The LinkCG method utilizes a heterogeneous graph consisting of a user-item interaction graph and a global graph representing implicit item associations to predict user-item links. By directly modeling item associations based on user interactions, LinkCG captures higher-order relationships and mitigates the issue of data sparsity. In addition, unlike node embedding-based methods, LinkCG improves the training process and enhances recommendation accuracy by assigning link weights that represent the degree of item association. Experimental results on three publicly available datasets show that LinkCG method, as a non-deep learning model containing only two hyperparameters, provides better performance compared with the deep learning-based baseline methods. The application on social relationship data further demonstrates LinkCG can capture rich enough information about user preferences from their historical interaction items.
Many existing pruning methods for deep neural network models require modifying the loss function or embedding additional variables in the network, thus they can’t benefit from the pre-trained network directly, and complicate the forward inference and training process. So far, most of the feature-oriented pruning work only use the intra-channel information to analyze the importance of filters, which makes it impossible to use the potential connections among channels during the pruning process. To address these issues, we consider the feature-oriented filter pruning task from an inter-channel perspective. The proposed method uses geometric distance to measure the potential correlation among channels, defines filter pruning as an optimization problem, and applies a greedy strategy to find an approximate solution to the optimal solution. The method achieves the decoupling of pruning from network and pruning from training, thus simplifying the pruning task. Extensive experiments demonstrate that the proposed pruning method achieves high performance for various network structures, for example, on CIFAR-10 dataset, the number of parameters and floating point operations of VGG-16 are reduced by 87.1% and 63.7%, respectively, while still has an accuracy of 93.81%. We also evaluate the proposed method using MobileFaceNets, a lightweight network, on CASIA-WebFace large dataset, and the evaluation results show that, when the number of parameters and floating-point operations are reduced by 58.0% and 63.6%, respectively, MobileFaceNets achieves an accuracy of 99.02% on LFW dataset without loss of inference accuracy (The code is available at: https://github.com/SSriven/FOAD).