A Generative Model for Correlated Graph Signals

A graph signal is a random vector with a partially known statistical description. The observations are usually sufficient to determine marginal distributions of graph node variables and their pairwise correlations representing the graph edges. However, the curse of dimensionality often prevents estimating a full joint distribution of all variables from the available observations. This paper introduces a computationally effective generative model to sample from arbitrary but known marginal distributions with defined pairwise correlations. Numerical experiments show that the proposed generative model is generally accurate for correlation coefficients with magnitudes up to about 0.3, whilst larger correlations can be obtained at the cost of distribution approximation accuracy. The generative models of graph signals can also be used to sample multivariate distributions for which closed-form mathematical expressions are not known or are too complex.

DEMLP: DeepWalk Embedding in MLP for miRNA-Disease Association Prediction

miRNAs significantly affect multifarious biological processes involving human disease. Biological experiments always need enormous financial support and time cost. Taking expense and difficulty into consideration, to predict the potential miRNA-disease associations, a lot of high-efficiency computational methods by computer have been developed, based on a network generated by miRNA-disease association dataset. However, there exist many challenges. Firstly, the association between miRNAs and diseases is intricate. These methods should consider the influence of the neighborhoods of each node from the network. Secondly, how to measure whether there is an association between two nodes of the network is also an important problem. In our study, we innovatively integrate graph node embedding with a multilayer perceptron and propose a method DEMLP. To begin with, we construct a miRNA-disease network by miRNA-disease adjacency matrix (MDA). Then, low-dimensional embedding representation vectors of nodes are learned from the miRNA-disease network by DeepWalk. Finally, we use these low-dimensional embedding representation vectors as input to train the multilayer perceptron. Experiments show that our proposed method that only utilized the miRNA–disease association information can effectively predict miRNA-disease associations. To evaluate the effectiveness of DEMLP in a miRNA-disease network from HMDD v3.2, we apply fivefold crossvalidation in our study. The ROC-AUC computed result value of DEMLP is 0.943, and the PR-AUC value of DEMLP is 0.937. Compared with other state-of-the-art methods, our method shows good performance using only the miRNA-disease interaction network.

Social behavior prediction with graph U-Net+

We focus on the problem of predicting social media user’s future behavior and consider it as a graph node binary classification task. Existing works use graph representation learning methods to give each node an embedding vector, then update the node representations by designing different information passing and aggregation mechanisms, like GCN or GAT methods. In this paper, we follow the fact that social media users have influence on their neighbor area, and extract subgraph structures from real-world social networks. We propose an encoder–decoder architecture based on graph U-Net, known as the graph U-Net+. In order to improve the feature extraction capability in convolutional process and eliminate the effect of over-smoothing problem, we introduce the bilinear information aggregator and NodeNorm normalization approaches into both encoding and decoding blocks. We reuse four datasets from DeepInf and extensive experimental results demonstrate that our methods achieve better performance than previous models.

Structured Self-Supervised Pretraining for Commonsense Knowledge Graph Completion

To develop commonsense-grounded NLP applications, a comprehensive and accurate commonsense knowledge graph (CKG) is needed. It is time-consuming to manually construct CKGs and many research efforts have been devoted to the automatic construction of CKGs. Previous approaches focus on generating concepts that have direct and obvious relationships with existing concepts and lack an capability to generate unobvious concepts. In this work, we aim to bridge this gap. We propose a general graph-to-paths pretraining framework that leverages high-order structures in CKGs to capture high-order relationships between concepts. We instantiate this general framework to four special cases: long path, path-to-path, router, and graph-node-path. Experiments on two datasets demonstrate the effectiveness of our methods. The code will be released via the public GitHub repository.