Cross-project defect prediction (CPDP) uses the historical defect dataset collected from source projects to train a model and then applies it to the target project. However, almost all existing CPDP methods are developed for offline scenarios where the trained models are fixed and cannot be updated along with the incoming labeled target instances after training. Actually, the label of target instances usually arrives online in a streaming manner which can be used to update CPDP models for better defect prediction performance on the next unlabeled target instance. To bridge these gaps, we propose a novel effective online cross-project defect prediction framework named ALOGO. ALOGO includes two essential phases: offline cross-project defect prediction phase and online within-project defect prediction (WPDP) phase which are combined by an adaptive weighted adjustment mechanism. In the offline CPDP phase, the global offline defect knowledge is learned by minimizing the difference between the source and target datasets based on an offline CPDP model. In the online WPDP phase, the local online defect knowledge is learned based on an online WPDP model. These two kinds of defect knowledge are then combined to obtain the latest and most valuable defect knowledge. Experimental results on 27 defect datasets show that ALOGO improves the performance over the existing state-of-the-art online CPDP model by 31.2% in terms of the Matthews correlation coefficient (MCC) and also outperforms the baseline in terms of other four well-known measures. It can be concluded that 1) it is necessary to consider building online CPDP models; 2) ALOGO is a more promising alternative for online CPDP.

Design automation and optimization for analog integrated circuits (ICs) are challenging, especially for transistor sizing. Given certain design specifications and circuit topology, circuit designers need to size various components to achieve the desired performance, possibly involving many optimization iterations. Recently, reinforcement learning (RL) has been applied to optimize analog circuits. The trained RL agents can achieve very high sample efficiency over evolutionary-based algorithms. By using the ability of transfer learning, the trained agent can be applied to optimize the same circuit across different technology nodes and even the circuits with different topologies. However, a significant bottleneck in applying machine learning (ML) techniques to analog IC design is the non-disclosure agreement (NDA) of the process development kit (PDK), which makes reproducibility of the prior art a big challenge. This work presents an RL framework that leverages the open-source SKY130 PDK to address the limitation above. We apply a novel heterogeneous graph neural network (GNN) called relational graph convolutional network (RGCN) as the function approximator of RL to capture more topological information about a circuit. As a proof-of-concept, low-dropout voltage regulators (LDO) are optimized by our proposed RL circuit optimizer framework to show its feasibility, achieving promising results.

Multi-source cross-project defect prediction (MSCPDP) refers to transferring defect knowledge from multiple source projects to the target project. MSCPDP has drawn increasing attention of academic and industry communities owing to its advantages compared with single-source cross-project defect prediction (SSCPDP) and some MSCPDP models have been proposed. However, to the best of our knowledge, there are no empirical studies to investigate the effect of different MSCPCP models on the performance of MSCPDP. To comprehensively investigate the performance of different MSCPDP models, we first conduct the literature research about MSCPDP studies, and then identify and compare 7 state-of-the-art MSCPDP models in terms of multiple performance measures including PD, PF, area under ROC curve (AUC), F1, precision, Matthews correlation coefficient (MCC), and Popt20% on 20 publicly available defect datasets. Furthermore, a robust multiple comparison method, i.e., the Scott-Knott effect-size difference (ESD) test, is used for statistical test. The experiment results show that 1) Burak’s Filter always performs best in terms of precision, AUC, MCC, Popt20% except for F1;2) MSCPDP models outperform the mean performance of SSCPDP models on most datasets; 3) the performance of MSCPDP models still needs to be further improved. We suggest software engineers use MSCPDP models but not SSCPDP models for CPDP and pay more attention to both the distribution difference of different datasets and the problems of sample similarity and weight when building MSCPDP models.

This paper proposes a NPEL (Nine-Pipeline & Ensemble Learning) strategy based on machine Learning algorithm to predict protein-protein interaction hotspots by training amino acid composition, surface area, amino acid chains and other complex/interface-related structural information. We applied Random Forest, Linear Svm, KNN, Gaussian Naive Bayes, Multi-layer Perceptron Neural Network, Adaboost, XGBoost etc. nine machine learning algorithms combination into an independent pipeline to predict protein hot spots, and the final results are optimized through voting and stacking scheme. In the stacking result of XGBoost and Logistic Regression, the highest accuracy is 0.8462 and improve the indicators of the pipeline results greatly.

Large-scale point cloud semantic segmentation has wide applications. Current popular researches mainly focus on fully supervised learning which demands expensive and tedious manual point-wise annotation. Weakly supervised learning is an alternative way to avoid this exhausting an-notation. However, for large-scale point clouds with few labeled points, the network is difficult to extract discriminative features for unlabeled points, as well as the regularization of topology between labeled and unlabeled points is usually ignored, resulting in incorrect segmentation results.To address this problem, we propose a perturbed self-distillation (PSD) framework. Specifically, inspired by self-supervised learning, we construct the perturbed branch and enforce the predictive consistency among the perturbed branch and original branch. In this way, the graph topology of the whole point cloud can be effectively established by the introduced auxiliary supervision, such that the in-formation propagation between the labeled and unlabeled points will be realized. Besides point-level supervision, we present a well-integrated context-aware module to explicitly regularize the affinity correlation of labeled points. Therefore, the graph topology of the point cloud can be further refined. The experimental results evaluated on three large-scale datasets show the large gain (3.0% on average) against recent weakly supervised methods and comparable results to some fully supervised methods.