The core of big data is data intelligence. The essence of data intelligence is to find and evaluate the correlation between several concepts in a large number of samples, then form mathematical expressions, and then use mathematical expressions for reasoning operations, thus completing the judgment and decision-making of unknown samples. This requires discovering the laws behind massive data and solving the problem of data representation. Data intelligence has gone through three stages: expert system, traditional machine learning and neural network. The input knowledge is more and more macroscopic from concrete to abstract, from rules to features to patterns, the efficiency of intelligent processing is getting higher and higher, and the explanation of the underlying perception and model is getting weaker and weaker. With the gradual fading out of expert system, traditional machine learning and neural network have become two commonly used technologies of data intelligence. Practice has proved that with the increase of data sets, the performance of traditional machine learning is not as good as that of neural network (see figure 1). This is mainly because the former is not as expressive as the latter. Goodfellow published a paper "MaxoutNetworks" at ICML (International Machine Learning Conference) 20 13. This paper proves that MaxoutNetworks can infinitely approximate any continuous function. In other words, neural network can fit any continuous function, and compared with traditional machine learning, neural network has outstanding expressive ability advantages.

? (Above): The horizontal axis represents the amount of data, and the vertical axis represents the accuracy of the algorithm?

We have seen several trends: the exponential growth of industry data, the growth of professional chip computing power represented by GPU, the endless emergence of new algorithms, the frontier research in academia, the capital investment in investment circles, and various industrial and commercial scenarios, which have promoted the rapid development of neural networks. The development of neural network has two directions: one is the vertical development represented by DNN deep connection and CNN convolutional neural network, that is, the vertical iteration with increasing layers, and the typical application is CV computer vision; The second is the lateral development represented by RNN recurrent neural network, that is, the lateral iteration between neurons. The typical application is sequence processing represented by NLP natural language understanding. Neural network technology presents two development forms at the same time, and has been widely used in many fields, indicating that this technology has entered a mature stage. Which direction is the next step? It is very possible to combine vertical development with horizontal development and penetrate into more application fields. This seems to be a logical conclusion. Facts have proved that this judgment is correct, and graph neural network is the combination of the two.

Throughout the development history of the technology circle, we can sum up a fact: whether a theoretical technology can be popularized in more fields depends on whether it can truly depict the substantive characteristics and relationships of the real world. The more realistic, the more application scenarios. For example, Markov chain theory truly depicts the characteristics and dependence of time series objects in the real world, so it is widely used in speech understanding, machine translation, national economy, event prediction and other fields; Another example is probability graph theory, which uses graphs to represent the dependence of event probability and truly depicts the entity relationship in the real world, so it is also widely used in anti-fraud, image understanding, event prediction and other fields. From the methodological point of view, in order to describe the entities in the real world, it is necessary to place nodes representing this entity in the model and design the transformation of dependencies between entities. However, Markov chain and probability graph both weaken the embedded representation, thus losing some hidden semantic information, which is flawed.

The appearance of graphic neural network (GNN) turned the situation around. In graphic neural networks, there are two kinds of networks. One is topological network, which usually describes many entities and their relationships; The other is feature transformation neural network, which is usually used for feature transformation of nodes, edges, graphs or subgraphs. The former completes the horizontal spread of information and realizes the transmission of topological relations of graphic signals, and its theoretical basis is graph theory; Based on deep learning, the latter completes the vertical dissemination of information and realizes the transformation from original features to embedded representation. Graphical neural network is a perfect combination of graph theory and deep learning, which considers both entity relationship and entity characteristics. Compared with traditional graph method and traditional deep learning, graph neural network has obvious advantages: it can model the source data more fully and better reflect the real relationship between entities in the real world. It can not only learn the semantic representation from the non-Euclidean spatial data represented by the graph structure, but also make the learned semantic representation conform to the entity relationship of the graph structure to the maximum extent.

More than 80% of the data in the real world are more suitable to be described by graph structure, such as traffic data, social data, molecular structure data, industry economic data and so on. Graphical neural network can adapt to this kind of data. Under the distributed learning architecture, the graphical neural network can process massive data, which is very suitable for processing industrial data of hundreds of millions of nodes. Therefore, the application scenarios of graph neural network are more extensive. In recent three years, various international summits have frequently published papers on graph neural networks. Many Internet technology companies (such as Ali, Baidu and ByteDance) have invested a lot of money in this field and made great progress, which are widely used in related search, real-time recommendation, risk prevention and control, anomaly detection, behavior prediction, pattern recognition and other fields. These phenomena undoubtedly show that graphic neural network is an important field direction of future technical development.

To sum up, under the background of industry data, algorithm theory, computing power support, market demand and capital influx, the rapid rise of graph neural network is inevitable in the era of big data.