Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Junfeng Zhou has a total of 54 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Bitruss Maintenance on Dynamic Bipartite Graphs

    by Jing Bai, Junfeng Zhou, Ming Du, Ziyang Chen

    Bitruss maintenance updates the hierarchy of cohesive subgraphs in a bipartite graph as edges are inserted or deleted. Existing support-driven methods repeatedly search the graph for butterflies during candidate expansion and support recomputation. We propose an index-based approach that maintains butterfly associations in a compressed bloom-edge index. By fixing the vertex priority order, we preserve the assignment of surviving butterflies and update only the affected blooms. On this index, we develop four successive algorithms. BM performs level-by-level maintenance; SA shares qualified-path counts and aggregates support losses; SAP reduces candidate expansion through structural filters and early eviction; and SAP+ maintains exact own-level support across updates. We prove correctness for both insertion and deletion using candidate coverage, threshold peeling, and support invariants. Experiments on nine real-world bipartite graphs evaluate independent single-edge updates that create or destroy butterflies. Relative to the single-threaded PI*/PR* baselines on matching workloads, SAP+ achieves geometric mean speedups of 50.44× for insertion and 114.56× for deletion. Further experiments evaluate each refinement and verify exact maintenance with an index size close to fresh rebuilding over the evaluated accumulating-update sequences.

  • Open Access

    Accelerating (k,l,η)-Core Query Processing in Directed Uncertain Graphs

    by Xian Tang, Guo Chen, Junfeng Zhou

    Uncertain graphs are commonly used to model the uncertain relationships between entities that arise from experimental or measurement errors. In recent years, the analysis of uncertain graphs has attracted significant research attention, with the computation of (k,η)-cores emerging as a fundamental problem. However, existing studies on (k,η)-cores often neglect edge directions, resulting in weak correlations among vertices in the resulting subgraph. To address this limitation, we propose a direction-aware (k,l,η)-core model. Specifically, a (k,l,η)-core is defined as a maximal connected subgraph in which every vertex has a probability of at least η of having in-degree ≥k and out-degree ≥l. We first present an online algorithm based on a peeling strategy to compute (k,l,η)-cores. To improve query performance, we develop two indexing mechanisms, DUCS-E and DUCS, that accelerate query processing. DUCS-E stores probability information for all possible (k,l,η)-cores, enabling it to completely avoid redundant computations during query processing, but at the cost of large storage space. To mitigate this issue, we propose the lightweight DUCS index, which stores directional probability information separately, reducing storage overhead while still pruning many irrelevant vertices; however, it requires additional verification. To balance efficiency and storage, we further design a hybrid index that combines the strengths of both approaches. Finally, experimental evaluations on real-world datasets demonstrate the effectiveness of the proposed (k,l,η)-core model as well as the efficiency and scalability of our methods.

  • Efficient k-step Weighted Reachability Queries Processing Algorithms

    by Congquan Mei, Junfeng Zhou, Ming Du, Lian Chen, Sheng Yu, Xian Tang, Ziyang Chen

  • Open Access

    Hybrid Intersection Over Union Loss for a Robust Small Object Detection in Low-Light Conditions

    by Twahir Kiobya, Junfeng Zhou, Baraka Maiseli, Maqbool Khan

    In computer vision, most existing works about object detection focus on detecting objects in the good lighting conditions instead of low-light conditions. Even the few existing works that are centered on object detection in the low-light conditions, predominantly focus on the general object detection rather than the detection of small objects. The main challenges affecting small object detection accuracy in low-light conditions are occlusion caused by the low light, shadows, and darkness that adversely affect the surrounding context leading to poor object classification and the insufficient spatial information that negatively affect object localization resulting in poor small object detection. To address the challenge of poor small object detection in low-light conditions we propose the Hybrid Intersection over Union (HIoU) localization loss to enhance the detection accuracy of small objects in these conditions. This loss utilizes the top-bottom distances of the targeted and predicted bounding boxes and the manhattan distance of the boxes’ centres to deal with the issue of misalignment that negatively affect the small object detection accuracy. Also, it jointly works with the classification loss to offer a joint optimization that facilitates a network to learn features that are important for both localization and classification. Experimental results show that the proposed loss enhances the detection accuracy of small objects in low-light conditions.


  • $(\alpha ,\beta )$-Butterfly Computation on Bipartite Graphs

    by Jing Bai, Junfeng Zhou, Ming Du, Ziyang Chen

  • Open Access

    Efficient Visual-Aware Fashion Recommendation Using Compressed Node Features and Graph-Based Learning

    by Umar Subhan Malhi, Junfeng Zhou, Abdur Rasool, Shahbaz Siddeeq

    In fashion e-commerce, predicting item compatibility using visual features remains a significant challenge. Current recommendation systems often struggle to incorporate high-dimensional visual data into graph-based learning models effectively. This limitation presents a substantial opportunity to enhance the precision and effectiveness of fashion recommendations. In this paper, we present the Visual-aware Graph Convolutional Network (VAGCN). This novel framework helps improve how visual features can be incorporated into graph-based learning systems for fashion item compatibility predictions. The VAGCN framework employs a deep-stacked autoencoder to convert the input image’s high-dimensional raw CNN visual features into more manageable low-dimensional representations. In addition to improving feature representation, the GCN can also reason more intelligently about predictions, which would not be possible without this compression. The GCN encoder processes nodes in the graph to capture structural and feature correlation. Following the GCN encoder, the refined embeddings are input to a multi-layer perceptron (MLP) to calculate compatibility scores. The approach extends to using neighborhood information only during the testing phase to help with training efficiency and generalizability in practical scenarios, a key characteristic of our model. By leveraging its ability to capture latent visual features and neighborhood-based learning, VAGCN thoroughly investigates item compatibility across various categories. This method significantly improves predictive accuracy, consistently outperforming existing benchmarks. These contributions tackle significant scalability and computational efficiency challenges, showcasing the potential transformation of recommendation systems through enhanced feature representation, paving the way for further innovations in the fashion domain.


  • Truss community search in uncertain graphs

    by Bo Xing, Yuting Tan, Junfeng Zhou, Ming Du


  • Open Access

    Attentive context and semantic enhancement mechanism for printed circuit board defect detection with two-stage and multi-stage object detectors

    by Twahir Kiobya, Junfeng Zhou, Baraka Maiseli, Maqbool Khan

    Printed Circuit Boards (PCBs) are key devices for the modern-day electronic technologies. During the production of these boards, defects may occur. Several methods have been proposed to detect PCB defects. However, detecting significantly smaller and visually unrecognizable defects has been a long-standing challenge. The existing two-stage and multi-stage object detectors that use only one layer of the backbone, such as Resnet’s third layer (\(C_4\)) or fourth layer (\(C_5\)), suffer from low accuracy, and those that use multi-layer feature maps extractors, such as Feature Pyramid Network (FPN), incur higher computational cost. Founded by these challenges, we propose a robust, less computationally intensive, and plug-and-play Attentive Context and Semantic Enhancement Module (ACASEM) for two-stage and multi-stage detectors to enhance PCB defects detection. This module consists of two main parts, namely adaptable feature fusion and attention sub-modules. The proposed model, ACASEM, takes in feature maps from different layers of the backbone and fuses them in a way that enriches the resulting feature maps with more context and semantic information. We test our module with state-of-the-art two-stage object detectors, Faster R-CNN and Double-Head R-CNN, and with multi-stage Cascade R-CNN detector on DeepPCB and Augmented PCB Defect datasets. Empirical results demonstrate improvement in the accuracy of defect detection.


  • Open Access

    HitSim: An Efficient Algorithm for Single-Source and Top-k SimRank Computation

    by Jing Bai, Junfeng Zhou, Shuotong Chen, Ming Du, Ziyang Chen, Mengtao Min

    SimRank is a widely used metric for evaluating vertex similarity based on graph topology, with diverse applications such as large-scale graph mining and natural language processing. The objective of the single-source and top-k SimRank query problem is to retrieve the kvertices with the largest SimRank to the source vertex. However, existing algorithms suffer from inefficiency as they require computing SimRank for all vertices to retrieve the top-k results. To address this issue, we propose an algorithm named HitSimthat utilizes a branch and bound strategy for the single-source and top-k query. HitSim initially partitions vertices into distinct sets based on their shortest-meeting lengths to the source vertex. Subsequently, it computes an upper bound of SimRank for each set. If the upper bound of a set is no larger than the minimum value of the current top-k results, HitSim efficiently batch-prunes the unpromising vertices within the set. However, in scenarios where the graph becomes dense, certain sets with large upper bounds may contain numerous vertices with small SimRank, leading to redundant overhead when processing these vertices. To address this issue, we propose an optimized algorithm named HitSim-OPT that computes the upper bound of SimRank for each vertex instead of each set, resulting in a fine-grained and efficient pruning process. The experimental results conducted on six real-world datasets demonstrate the performance of our algorithms in efficiently addressing the single-source and top-k query problem.

  • Open Access

    Top-k Graph Similarity Search Algorithm Based on Chi-Square Statistics in Probabilistic Graphs

    by Ziyang Chen, Junhao Zhuang, Xuan Wang, Xian Tang, Kun Yang, Ming Du, Junfeng Zhou

    Top-k graph similarity search on probabilistic graphs is widely used in various scenarios, such as symptom–disease diagnostics, community discovery, visual pattern recognition, and communication networks. The state-of-the-art method uses the chi-square statistics to speed up the process. The effectiveness of the chi-square statistics solution depends on the effectiveness of the sample observation and expectation. The existing method assumes that the labels in the data graphs are subject to uniform distribution and calculate the chi-square value based on this. In fact, however, the actual distribution of the labels does not meet the requirement of uniform distribution, resulting in a low quality of the returned results. To solve this problem, we propose a top-k similar subgraph search algorithm ChiSSA based on chi-square statistics. We propose two ways to calculate the expectation vector according to the actual distribution of labels in the graph, including the local expectation calculation method based on the vertex neighbors and the global expectation calculation method based on the label distribution of the whole graph. Furthermore, we propose two optimization strategies to improve the accuracy of query results and the efficiency of our algorithm. We conduct rich experiments on real datasets. The experimental results on real datasets show that our algorithm improves the quality and accuracy by an average of 1.66× and 1.68× in terms of time overhead, it improves by an average of 3.41×.