Start this year. Graph Neural Networks has become the focus of discussion among researchers. As a researcher in this field, I am very pleased. I remember that about three years ago, the graph neural network was still in the cold. When I communicated with colleagues who studied GAN and Transformers, they thought that my research direction was extremely niche. Up to now, this field is finally well guarded. Therefore, I will introduce to you the recent hot spots of GNNs applications in this article.
Recommended system
In an e-commerce platform, the interaction between users and products constitutes a graph structure, so many companies use graph neural networks for product recommendation. The typical approach is to model the interaction between users and products, and then learn node embedding through some negative sampling loss, and recommend similar products to users in real time through kNN. Uber Eats has adopted this method for product recommendations for a long time. Specifically, they use the graph neural network GraphSage to recommend food and restaurants to users.
In food recommendation, due to geographical restrictions, the graph structure used is relatively small. On some large-scale graphs containing billions of nodes, graph neural networks can also be used. It is very difficult to use traditional methods to process such large-scale graphs. Alibaba is researching graph embedding and GNN on a network containing billions of users and products . Recently, they proposed Aligraph , which can be built with 400M nodes in only five minutes. Of the figure. Very powerful! In addition, Aligraph also supports efficient distributed graph storage, optimizes the sampling process, and integrates many GNN models internally. The framework has been successfully used in the company's multiple product recommendations and personalized search tasks.
Alibaba , Amazon and many other e-commerce platforms use GNN to build recommendation systems
Pinterest proposed the PinSage model, which uses personalized PageRank to efficiently sample neighborhoods and update node embeddings by aggregating neighborhoods. The follow-up model PinnerSage further extends the framework to deal with the multi-embedding problem of different users. Due to limited space, this article only lists some of the applications of GNN in recommendation systems (other cases include: Amazon uses GNN in the knowledge graph or Fabula AI uses GNN to detect fake news, etc.), but these are enough to show that if the user interacts with enough information Rich, then GNNs will significantly promote the further development of the recommendation system.
Portfolio Optimization
Many fields such as finance, logistics, energy, life sciences, and hardware design are facing the problem of combinatorial optimization (CO). Most of these problems can be modeled through graph structures. Therefore, the research work in the past century has been devoted to solving the CO problem from the algorithm level. However, the development of machine learning provides another possibility for solving the CO problem.
The Google Brain team has successfully used GNN for hardware design, such as the optimization of the power consumption, area and performance of the Google TPU chip block. You can think of a computer chip as a graph composed of memory and logic components, and each graph is represented by the coordinates and types of its components. The job of an electrical engineer is to determine the location of each component while complying with density and wiring congestion limits. The Google Brain team combined GNN and strategy/value RL to achieve the design and optimization of the circuit chip layout, which performed better than the artificially designed hardware layout.
Comparing the chip layout with the complexity of chess and Go ( source )
In addition, machine learning (ML) models can be integrated into existing solvers. Gasse et al. proposed a graph network for learning the branch and bound variable selection strategy (the key to the MILP solver of the mixed integer linear program), by which the running time of the solver can be minimized. At the same time, the paper shows that this method can not only guarantee the reasoning time, but also guarantee the quality of decision-making.
In the latest work of DeepMind and Google, Graphnet is used for two key subtasks of the MILP solver: joint variable assignment and target value delimitation. On the large-scale data sets of Google's production package and planning system, their neural network method is 2-10 times faster than existing solvers. For more information, please refer to the relevant review .
Computer vision
Since all objects in the world are closely related, GNN can be used for object images. The image can be perceived through the scene graph , that is, a group of objects appear in the same scene, then there is a connection between them. Scene graphs have been used in tasks such as image inspection, understanding and reasoning, caption generation, visual question and answer, and image generation, which greatly improves the performance of the model.
A Facebook study showed that a scene graph was created based on the shape, position, and size of objects in the classic CV data set COCO, and then GNN was used to embed the objects in the graph, and then combined with CNN to generate the mask, border and exterior of the object. Finally, GNN/CNN can generate object images at user-specified nodes (determine the relative position and size of the nodes).
Use scene graphs to generate images. The user can place the object anywhere on the canvas, and the image will change accordingly (for example, if the red "river" is moved from the middle to the lower right corner, the river generated in the image will also be moved to the lower right corner)
For the classic task of CV-the matching of two related images, the previous method can only be achieved manually. But now, 3D graphics company Magic Leap has developed a GNN architecture called SuperGlue , which can perform graphics matching in real-time video to complete tasks such as 3D reconstruction, location recognition, localization, and mapping (SLAM). SuperGlue consists of an attention-based GNN. GNN learns the representation of the key points of the image, and then matches these key point representations at the best transmission layer. The model can be matched in real time on the GPU and can be easily integrated into the existing SLAM system. For more research and applications of graphics and computer vision, see the following review articles .
Physical Chemistry
Constructing a graph based on the interaction between particles or molecules, and then using GNN to predict system properties has gradually become an important research method in life sciences. The Open Catalyst project between Facebook and CMU is dedicated to finding new ways to store renewable energy, such as solar or wind energy. One of the possible solutions is to convert this energy into other fuels, such as hydrogen, through a chemical reaction. However, this requires the discovery of new and more efficient catalysts to accelerate chemical reactions, and the known DFT method is extremely expensive. The Open Catalyst project open sourced large-scale catalyst data sets, DFT relaxation and GNN benchmark methods, hoping to find new, efficient, and low-cost catalyst molecules.
The initial state and relaxed state of the adsorbate (small linking molecule) and the surface of the catalyst. In order to find the relaxed state of the adsorbate-catalyst pair, expensive DFT simulations are required, and it takes several days. Zitnick et al. 2020
DeepMind researchers also use GNN to simulate the dynamics of complex particle systems such as water or sand. By gradually predicting the relative motion of each particle, the dynamics of the entire system can be reconstructed reasonably and the basic laws of controlling motion can be understood. This can be used to understand the glass transition , one of the most interesting unanswered questions in solid-state theory. In addition, using GNN can not only simulate the dynamics of the transition, but also better understand how particles interact with each other based on distance and time.
In addition, Fermilab of the Physics Laboratory in the United States is committed to applying GNNs to analyze the results of the CERN Large Hadron Collider (LHC) , hoping to process millions of images, discover and select images related to new particles. Their task is to deploy GNN on FPGA and integrate it with data collector , so that GNN can be remotely run on a global scale. For more applications of GNNs in particle physics, see the following review article .
Drug Discovery
The pharmaceutical industry is highly competitive, with leading companies investing billions of dollars each year to develop new drugs. In biology, graphs can represent interactions on different scales. For example, at the molecular level, the edges of the graph can be the bonds between atoms in a molecule or the interactions between amino acid residues in a protein; on a larger scale , The diagram can represent the interaction between more complex structures (such as proteins, mRNA or metabolites). Graphs in different levels of scale can be used for target recognition, molecular property prediction, high-throughput screening, new drug design, protein engineering and drug reuse, etc.
The time flow of applying GNN to drug development, Gaudelet et al., 2020
MIT researchers and their collaborators published an article on Cell (2020) that GNN helps in drug development. They trained a deep GNN model called Chemprop to predict whether the molecule has antibiotic properties, that is, its growth inhibitory effect on E. coli. After training it with about 2500 molecules in the FDA-approved drug library, Chemprop was applied to a larger data set, the Drug Repurposing Hub containing Halicin molecules, and based on the movie "2001: A Space Odyssey" Rename it in HAL9000 .
It should be noted that the molecular structure of Halicin is very different from that of known antibiotics, so previous work only studied this molecule. However, in vivo and in vitro clinical trials have shown that Halicin is a broad-spectrum antibiotic. Compared with the extensive benchmark tests conducted by the NN model, the application of GNN found that Halicin has more demonstrated the powerful learning and characterization capabilities of GNN. In addition, the Chemprop architecture is also worthy of attention: Unlike most GNN models, Chemprop has 5 layers and 1600 hidden layer dimensions, far exceeding other GNN parameters.