Caimao Li’s research while affiliated with Hainan University and other places

Graph-level clustering is a fundamental and significant task in data mining. The advancement of graph neural networks has provided substantial impetus to this area of research. However, existing graph-level clustering methods often focus exclusively on either graph structure or node attributes, which limits their ability to comprehensively capture graph-level features. To address this issue, we propose a Multi-view Enhancement Graph-level Clustering Network (ME-GCN), which consists of a Multi-view Feature Extraction Strategy (MFES) and a Graph-level Heterogeneous Enhancement Mechanism (GHEM) to generate high-quality graph-level representations, thus improving clustering performance. Specifically, the network extracts node attributes, subgraph structure, and global structure information through three personalized encoders with different receptive fields, respectively, to enrich feature representations from different views. In addition, it perceives and reasons heterogeneous features through GHEM, including multi-source information and hierarchical enhancement, which improves the compactness of multi-source representations. Extensive experiments on five benchmark datasets have demonstrated the superiority of ME-GCN, highlighting its effectiveness in leveraging multiple types of structure information for deep graph-level clustering.

Click-through-rate (CTR) prediction is a crucial task in recommendation systems. The accuracy of CTR prediction is strongly influenced by the precise extraction of essential data and the modeling strategy chosen. The data of the CTR task are often very sparse, and Factorization Machines (FMs) are a class of general predictors working effectively with it. However, the performance of FMs can be limited by the fixed feature representation and the same weight of different features. In this work, we propose an improved Bitwise Feature Importance Factorization Machine (BFIFM) to improve the accuracy. The necessity of learning the degree of effect of the same feature under various situations is learned through the low-order intersection method, and the deep neural network (DNN) in our model is used in parallel to study high-order intersections. According to the final results obtained, the BFIFM model significantly outperforms other state-of-the-art models. KeywordsFactorization machinesDeep learningRecommendationSparse data

The user’s age and gender play a vital role within the user portrait. In view of the lack of basic attribute information, such as the age and gender of users, this paper constructs an attribute prediction method based on stacking multimodel integration. The user’s browsing and clicking history is analyzed to predict the user’s basic attributes. First, LR, RF, XGBoost, and ExtraTree were selected as the base classifiers for the first layer of the stacking framework, and the training results of the first layer were input as new training data into the second layer LightGBM for training. Experiments show that the proposed model can improve the accuracy of prediction results.KeywordsMachine learningAttribute predictionModel fusionLightGBM

Traditional CTR recommendation models have concentrated on how to learn low-order and high-order characteristics. The majority of them make many efforts at combining low-order and high-order functions. However, they ignore the importance of the attention mechanism for learning input features. The ECABiNet model is proposed in this article to enhance the performance of CTR. On the one hand, the ECABiNet model can learn the importance of features dynamically via the LayerNorm and ECANET layers. On the other hand, through the use of a bi-interaction layer and a DNN layer, it is capable of effectively learning the feature interactions. According to the experimental results on two public datasets, the ECABiNet model is more effective than the previous CTR model.