Generating Graphs

A lot of learning tasks require dealing with graph data which contains rich relation information among elements. Modeling physical systems, learning molecular fingerprints, predicting protein interface, and classifying diseases require a model to learn from graph inputs. In general, there is no optimal method to generate graphs from raw data. In the image domain, some work utilizes convolution neural network (CNN) to obtain feature maps then upsamples them to form superpixels as nodes, while other ones directly leverage some object detection algorithms to get object nodes. Therefore, finding the best graph generation approach will offer a wider range of fields where GNN could make a contribution.

Finding a new algorithm to generate a graph from raw data. All in all, two significant approaches to enhance this domain are firstly to construct appropriate graphs in order to exhibit object relations. A second activity is the usage of GNN algorithms to find better embedding representation. We hope by the aforementioned, concise graphs with low dimensional node features and efficient graph structure will be constructed.

Close

Graph Representation Learning

Learning with graph-structured data requires an effective representation of their graph structure. Several well-known methods are known as learning node embeddings. The goal of these methods is to encode nodes as low- dimensional vectors that summarize their graph position and the structure of their local graph neighborhood. However, these shallow embedding approaches train unique embedding vectors for each node independently, which leads to a number of drawbacks. GNN paradigm was introduced to alleviate many of these shortcomings. The key idea is that we want to generate representations of nodes that actually depend on the structure of the graph, as well as any feature information we might have. GNNs broadly follow a recursive neighborhood aggregation scheme, where each node aggregates feature vectors of its neighbors to compute its new feature vector. After k iterations of aggregation, a node is represented by its transformed feature vector, which captures the structural information within the node’s k-hop neighborhood.

Traditional deep neural networks can stack hundreds of layers to get better performance because a deeper structure has more parameters, which improve the expressive power significantly. However, GNNs are always shallow, most of which are no more than three layers. However, Li et al. 2018 show that the performance of a GCN drops dramatically with an increase in the number of graph convolutional layers. As graph convolutions push representations of adjacent nodes closer to each other, in theory, with an infinite number of graph convolutional layers, all nodes' representations will converge to a single point. It seems that going deep is not still a good strategy for learning graph data.

Finding new algorithms to generate an effective representation for large graphs! How to apply embedding methods in web-scale conditions like social networks or recommendation systems has been a fatal problem for almost all graph embedding algorithms, and GNN is not an exception. Scaling up GNN is difficult because many of the core steps are computationally expensive in the big data environment. There are several examples of this phenomenon: First, graph data are not regular Euclidean, each node has its own neighborhood structure so batches can not be applied. Then, calculating graph Laplacian is also unfeasible when there are millions of nodes and edges.

Close

Molecules and Drug Discovery

Drug discovery and development are often fraught with problems such as high costs and time-consuming processes, e. g., the production of a drug by traditional methods costs an average of 2.6 billion USD and can take more than 12 years. Innovative approaches are essential to reduce the cost and accelerate the accurate development of a new drug. Computational methods can improve the drug discovery process due to their rapid and accurate prediction. The field of quantitative structure-activity relationships (QSARs) modeling was founded as a powerful computational approach to screen large libraries of molecules with desired properties. Artificial intelligence (AI) based methods are a kind of solution to the current problems of drug development, especially in predicting molecular properties, drug-drug interaction, and drug-target interaction. The performance of these methods relies heavily on finding the expressive representations of molecular structures. In the early works, molecular representations are based on fixed expert-designed molecular descriptors or fingerprints, such as Dragon descriptors or Morgan (ECFP) fingerprints. On the other hand, domain-specific expertise is used to select or engineer molecular descriptors and improve the performance of previous approaches.

One of the most significant advances in QSARs analysis has come from the use of deep learning because the highly flexible architecture of neural networks enables them to adapt to a variety of problems and achieve high-performance gains by extracting complex relationships. Instead of relying on features compiled by an expert, neural models can learn expressive representations directly from raw data.

GNNs, which have been developed as a powerful candidate for graph-structured data modeling, are among the hottest topics in recent years. A molecule can be represented naturally as a graph wherein atoms (nodes) connect by chemical bonds (edges). It has been shown in numerous studies that graph-based molecular property prediction can be more accurate than traditional descriptor-based methods. Also, recent publications have introduced various GNNs architectures that deliver promising performance in drug discovery over non-neural models.

The key idea of GNN and its variants is the use of a neural message passing process, in which message vectors of the nodes' neighbors (or edges' neighbors) are aggregated through neural networks to update node (and/or edge) representations. Therefore, a GNN layer combines the independent information of each node (and/or edge) with the structural information obtained from its neighbors. However, most learning problems require multiple GNN layers to exchange information between nodes (and/or edges) that are not directly connected.

Although existing GNN architectures have made good progress in various fields, such as drug discovery, they also face limitations that need further investigation.

Close

Scene Understanding

Computer Vision

The main goal of scene understanding is analyzing an image to recognize objects and their relationships. It is a high-level image processing field because it tries to reach information about an image beyond pixel level or ordinary image algorithms.

Recent progress of computer vision in deep learning has greatly impacted this area. Specifically, the detection and recognition of objects have achieved much better accuracy than in previous times. A better object detection algorithm, a better analyzer in scene understanding. Nevertheless, they suffer from defining a good relationship among objects. Some graphs have been introduced to depict semantic and spatial relationships among objects.

Close

Molecular Structure Analysis

Molecular structures have graph structures. By using GCNs, one can predict some physical, chemical, biochemical properties of a molecule by it's chemical formula. Also, it is possible to predict new formula and 3D structure for an unknown yet molecule or substance with certain desired properties.

Close