Use in AI

Graphs have always been closely related to the field of artificial intelligence, especially machine learning. In the field of pattern recognition, graph algorithms are used (Figure 1), for example, for unsupervised machine learning, neural networks, and deep learning processes. Typical use cases include fraud detection, personalized recommendations, identification of target groups and influential users, and identifying weaknesses and bottlenecks in operations and in the supply chain.

Figure 1: Machine learning algorithms use comprehensive datasets to perform differentiated classification, forecasting, and processing tasks.

Machine learning is a challenging field, and graph-based models are no exception. With every hop (i.e., every level of networked data), the size of the data pool to be searched increases exponentially, and from a computational point of view, this approach is simply too expensive for other database architectures. For data connections in relational databases to be evaluated meaningfully, uneconomic table joins are necessary. Additionally, classic NoSQL databases can recreate graphs at the application level only, are too cost-intensive, and do not provide support for complex queries.

Machine learning algorithms in graphs are used, among other things, to detect spam calls on mobile phones by analyzing the behavior of the source phone in relation to the target phone. However, simply determining whether the source phone is known to the target phone and how many times calls from the source phone have been rejected by other callers can only determine whether the call is spam with a rough probability.

Mobile phone provider China Mobile has developed a more sophisticated procedure for this: With the help of a graph database, the behavior of the 900 million mobile phones registered with China Mobile can be checked, which means that about 2 billion calls per week can be analyzed. An algorithm that analyzes 118 different characteristics for each investigated phone, combined with a native graph database featuring massively parallel processing capacities, enables calls to be examined in real time and allows the called person the option of rejecting an incoming call while the phone is still ringing.

Conclusions

Companies face the growing challenge of having to process vast amounts of information. This problem is not only about storing and retrieving data. Analysts also need to be able to perform in-depth analysis to obtain meaningful results. Graph databases are already helping in many areas, including supply chain management, customer analysis, and social media management. Together with complex algorithms based on graph query languages, this type of data storage enables the development of new applications. In much the same way that relational databases revolutionized data collection in the 1970s and 1980s, graphs will make it possible to break new ground in data processing over the next few decades.