Network Representation Learning (NRL), which aims to embed nodes into a latent, low-dimensional vector space while preserving some network properties, facilitates the network analysis tasks. The goal of most NRL methods is to make similar nodes represented similarly in the embedding space. Many methods adopt the skip-gram model to achieve such goal by maximizing the predictive probability among the context nodes for each center node. The context nodes are usually determined based on the concept of \textit{proximity} which is defined based on some explicit network features. However, these proximities may result in a loss of training samples and have limited discriminative power. We propose a general and unified generative adversarial learning framework to address the problems. The proposed framework can handle almost all kinds of networks in a unified way, including homogeneous plain networks, attribute augmented networks and heterogeneous networks. It can improve the performances of the most of the state-of-the-art skip-gram based NRL methods. Moreover, another unified and general NRL method is extended from the framework. It can learn the network representation independently. Extensive experiments on proximity preserving evaluation and two network analysis tasks, i.e., link prediction and node classifications, demonstrate the superiority and versatility of our framework.

Penalty prediction is one of the main tracks of legal judgement prediction (LJP) which is to apply artificial intelligence methods to predicting the court’s judgement based on case descriptions. It is still far from effective, as the elements affecting the penalty are numerous but sparsely distributed in case descriptions, which may make them covered by noise in contexts. Allocating reasonable attention to these elements related to the penalty is the key to improving the effect of the penalty prediction. To this end, we propose a novel model called MHA-AR which learns to focus on the key elements by a modulated hierarchical attention mechanism and a legal attribute recognition subtask. Besides benefiting the prediction accuracy, it also improves the reliability and credibility of predictions for users by providing comprehensible legal attributes related to the penalty. A series of experiments conducted on the real-world datasets demonstrate the correctness of our assumptions and the superiority of MHA-AR. The implementation of our proposed model will be available at https://github.com/realcatking/penaltyprediction.

Modern learnable collaborative filtering recommendation models generate user and item representations by deep learning methods (e.g. graph neural networks) for modeling user-item interactions. However, most of them may still have unsatisfied performances due to two issues. Firstly, some models assume that the representations of users or items are fixed when modeling interactions with different objects. However, a user may have different interests in different items, and an item may also have different attractions to different users. Thus the representations of users and items should depend on their contexts to some extent. Secondly, existing models learn representations for user and item by symmetrical dual methods which have identical or similar operations. Symmetrical methods may fail to sufficiently and reasonably extract the features of user and item as their interaction data have diverse semantic properties. To address the above issues, a novel model called Asymmetrical context-awaRe modulation for collaBorative filtering REcommendation (ARBRE) is proposed. It adopts simplified GNNs on collaborative graphs to capture homogeneous user preferences and item attributes, then designs two asymmetrical context-aware modulation models to learn dynamic user interests and item attractions, respectively. The learned representations from user domain and item domain are input pair-wisely into 4 Multi-Layer Perceptrons in different combinations to model user-item interactions. Experimental results on three real-world datasets demonstrate the superiority of ARBRE over various state-of-the-arts.

Knowledge graph (KG) representation learning aims at embedding triples in the form of vectors. Their semantic similarity can be expressed through the distance of those vectors, and thus easily be computed for further application, such as knowledge completion. Early KG embedding methods, such as the TransE model, were mainly training with relationship of each individual triple, and ignore the relationships between multiple triples. Thus their representation result is subject to the integrity of the triples. Recently, some path-enhanced methods, such as PTransE and RPJE, adopt the path information composed of multiple triples as supplement relation information for training, which achieves better effect than those triples based methods. On the other hand, the performance of path-enhanced methods is still affected by the fact that they are hard to learn the symmetric relation pattern in both triples and paths. Thus we propose Path Rotation based KG Representation Learning method (PRRL), which maps entities and relations into complex vector space and defines both relations and paths (composed by sequence of relations) as a rotation from source entity to target entity. PRRL can model and infer a variety of relation patterns, including symmetry/antisymmetry, inversion and composition relations by representing the path with Hadamard product of relations. The results of experiment on multiple datasets show that PRRL is better than the baseline in the completion of the task of KG.

Clustering is the basis of many complex network analysis and application tasks. Preserving the clustering properties in network representation space contributes to a better clustering performance. In this paper, an Attribute Network Embedding method based on Joint Clustering of representation and network (ANEJC) is proposed. Based on variational graph auto-encoder, ANEJC jointly reconstructs adjacency matrix and attribute matrix. In order to preserve the clustering property of the network, ANEJC clusters network structure and hidden layer representations of variational graph auto-encoder, simultaneously. Extensive experiments carried out on four synthetic datasets and three real-world datasets demonstrate a superior clustering performance of ANEJC over the state-of-art methods.

Information diffusion model aims to understand the process of information diffusion in the network. Currently, state-of-the-art methods utilize vector representation of users to encode these factors. Apart from personal factors, decisions of others of the local community can also affect a user’s decision on propagation. Recently, a multiscale information diffusion model called HID applies hierarchical clustering to improve the performance of many existing diffusion models. Though extensive experiments have proven the effectiveness of the model, it fails to encode diffusion time into the representation space, and the adopted clustering algorithms are independent of the diffusion model. Thus, we propose a multiscale clustering based diffusion representation method that incorporates diffusion time into diffusion proximity matrix and adopts a hierarchical clustering method suitable for multiscale diffusion learning. Experiments show the effectiveness of the proposed method.

Most of overlapping community detection algorithms cannot be applied to networks with highly overlapping community such as online social networks where individuals belong to many communities. One important reason is that many algorithms detect communities based on the explicit borders where nodes have more connections inside the communities, however, when the vertices' membership number gets large, the explicit borders between communities will fade away. To overcome this disadvantage, a new algorithm named MCNLPA is proposed by expanding the traditional Label Propagation Algorithm (LPA) based on the Maximal Clique Network for highly overlapping community detection. By finding all maximal cliques in networks and defining reasonable edges between them, the maximal clique network is established. Then the updated rule of classic LPA is modified to apply to the maximal network. Experiments show that MCNLPA has a relatively good performance in highly overlapping community detection and overlapping nodes identification.