Rong Chen

Despite significant progress in point cloud upsampling, most existing methods rely heavily on supervised training with paired data, which are often difficult to acquire. Moreover, the inherent lack of explicit connectivity in point clouds makes it difficult to achieve both continuous and uniform densification while accurately recovering fine geometric structures. To address these problems, we propose an upsampling model that treats this task as saliency-aware implicit surface sampling, enabling self-supervised and fine-grained point densification. Central to our idea is correlating implicit surface reconstruction with salient point identification, and carrying out sampling on the saliency-aware surface representation. Motivated by this, we introduce a salient point detector along with a corresponding implicit surface-based interpolator, and a geometry filter, upon which we develop a three-stage architecture consisting of a pre-trained saliency guidance block, a saliency-aware enhancer, and an upsampler. The guidance block captures meaningful shape patterns to prevent detail loss and incomplete recovery, while the enhancer facilitates detail enhancement in complex salient regions. These components are integrated with the upsampler to generate dense results that accurately retain both global shape and meticulous structures. In comparison with state-of-the-art methods, our model significantly improves the upsampling quality. Extensive experiments conducted on various datasets, comprising both synthetic and real-world captured shapes, demonstrate the flexibility and availability of our method in processing the point clouds across different scales and distributions.

While federated learning (FL) is promising for collaborative training that preserves privacy on the mobile Internet, it remains vulnerable to data heterogeneity, model personalization, and user mobility. A viable paradigm for tackling these challenges is knowledge distillation (KD) based FL, which enables knowledge sharing between user-customized models. However, misleading and ambiguous knowledge sharing arises due to discrepancies in personal data distribution, the absence of a public dataset or a well-trained teacher model, which results in significant degradation in convergence and model performance. To address this issue, we propose a novel scheme named “coordinated learning through peer teaching,” which integrates conceptual alignments (CAs) from cognitive science into mainstream FL systems. The proposed method involves a novel rotational CA coordinator mechanism for model aggregation, for which the server takes turns choosing a CA coordinator from participating clients and arranges them to align concepts via asymmetric KD on the party side. By incorporating this mechanism into iterative training, we introduce FedRC, a new federated learning framework that combines recent advances in channel distillation and decoupled knowledge distillation for FL-assisted mobile computing. We evaluated our approach through extensive experiments on three real-world datasets that feature degrees of data and model heterogeneity. Empirical results indicate the efficacy of FedRC in improving transfer efficiency while notably improving the accuracy of the client population without using public data. Compared with state-of-the-art FL methods, FedRC outperforms baselines in standard and realistic mobile FL settings.

Software defect prediction aims to identify potentially defective software modules using various techniques, while fine-grained line-level defect prediction can pinpoint defective lines of code. This helps developers promptly discover and fix errors, thereby enhancing the efficiency of testing and code review. However, previous studies often overlook the impact of characterizing noise and the skewed distribution of defect knowledge in software projects, making it difficult for current methods to achieve satisfactory accuracy and cost-effectiveness in software defect prediction. To address these challenges, we propose a model named EU-LLDP, which effectively resolves the issue of low cost-effectiveness in line-level defect prediction models. Specifically, the EU-LLDP model consists of two main components: the defect mining component mines the most valuable defect knowledge from numerous software defects using prediction probability matrices, noise labels, and the borderline information of code vectorizations. The adaptive resampling component samples valuable defect knowledge through the density distribution of defect knowledge, thereby making full use of existing defect knowledge and improving the cost-effectiveness of line-level software defect prediction models. Seven comprehensive experiments were conducted on 32 defect datasets from 9 Java open source systems using file-level prediction models and line-level defect prediction models to evaluate the effectiveness of the EU-LLDP model. The EU-LLDP model improves the state-of-the-art file-level defect prediction model in terms of Balanced Accuracy by 9.87%, the MCC by 38.09%, and enhances the state-of-the-art line-level defect prediction method in terms of Recall@Top20%LOC by 44.16%, and Effort@Top20%Recall by 17.62%. These results fully demonstrate the effectiveness of EU-LLDP in improving the accuracy and cost-effectiveness of Software defect prediction.

Crowdsourcing campaigns in the real world often face challenges such as low participation, trust deficits, insufficient interaction, and a lack of incentive mechanisms, especially for collaborative tasks that require skills and commitment. Researchers have explored ways of deep reinforcement learning that incentivize individuals in accordance with their different social influences, predominantly characterized by node embedding of social networks. However, they overlook the role of group influence and the dynamics of social networks. This paper proposes an incentive mechanism GraphTD3 that, under limited budget constraints, encourages effective cooperation among self-interested workers by using deep reinforcement learning with group-aware embeddings. Notably, the graph learning module can adapt to changes in the network over time, and accordingly the policy module adjusts individual incentives to maximize the long-term pay-offs. Experimental results on four real-world network datasets demonstrate that GraphTD3 consistently outperforms existing methods. It achieves an average performance improvement of over 10%, with gains of up to 0.25 on the Twitter dataset, while also proving highly effective and stable in allocating incentives within dynamically changing social networks.

The Just-In-Time defect prediction model helps development teams improve software quality and efficiency by assessing whether code changes submitted by developers are likely to introduce defects in real-time, allowing timely identification of potential issues during the commit stage. In the meanwhile, identifying and labeling whether code change commits introduce defects are crucial for training the JIT model. However, two main challenges exist in current work due to the reality that all deleted and added lines in bug-fixing commits may be related to the root cause of the introduced bug: 1) lack of effective integration of heterogeneous graph information, and 2) lack of semantic relationships between changed code lines. To address these challenges, we propose a method called JIT-Finder, which utilizes relational graph convolutional network to capture the semantic relationships between changed code lines. JIT-Finder is used to detect root-cause deletion lines in changed code lines, thereby identifying the root cause of introduced bugs in bug-fixing commits. Specifically, the JIT-Finder consists of three components: the graph construction component, the graph type conversion component, and the root cause detection component. The graph construction component analyzes the source code of bugfixing commits to construct a heterogeneous graph representation by extracting added/deleted nodes based on added/deleted lines and extracting edges according to the relationships between the nodes. Next, to address the challenge of varying feature dimensions and the difficulty of integrating information in the heterogeneous graph of changed code lines, the graph type conversion component merges different types of nodes/edges into a unified set of nodes/edges and the type information of each node/edge is encoded as an additional vector. This process unifies the heterogeneous graph data into homogeneous graph data while preserving the type characteristics of different nodes and edges, thereby facilitating the integration of information from various nodes and edges. Finally, the root cause detection module uses a node embedding layer to obtain embedding vectors for the corresponding code statements, followed by a relational graph convolutional layer to capture the semantic relationships between the changed code lines and generate prediction labels. Ultimately, the root cause deletion lines in the bug-fixing commit are identified through a ranking layer applied to the deleted nodes. To evaluate the effectiveness of JIT-Finder, we used three datasets that contain high-quality bug-fixing and bug-introducing commits. Extensive experiments were conducted to evaluate the performance of our model by collecting data from 87 open-source projects, including 675 bug-fix commits. The experimental results show that, compared to the most advanced root cause detection methods, JIT-Finder improved Recall@1, Recall@2, Recall@3, and MFR by at 4.11%, 5.11%, 4.29%, and 14.41%, respectively.