Machine Learning and Deep Link Graph Analytics: A Powerful Combination

We investigate how graphs can help machine learning and how they are related to deep link graph analytics for Big Data.



By Yu Xu (founder and CEO, TigerGraph) and Gaurav Deshpande (VP of Marketing, TigerGraph)

Machine learning (ML) – an aspect of artificial intelligence (AI) that allows software to accurately identify patterns and predict outcomes – has become a hot industry topic. With ever-increasing advances in data analysis, storage, and computing power in the last few years, machine learning has been playing an increasingly important role in enterprise applications such as fraud prevention, personalized recommendation, predictive analytics, and so on.

Applying graph database capabilities to ML and AI apps is relatively new, however. That’s surprising in light of the fact that Google’s Knowledge Graph, which first popularized the concept of finding relationships within data to yield more relevant and precise information, dates back to 2012. Also, it’s a natural fit: Graphs are ideal for storing, connecting, and making inferences from complex data.

The main reason that graphs have not played an important role in ML is that legacy graph databases cannot deliver what is really needed for machine learning: deep link graph analytics for large datasets.

Let’s take a deeper dive into how graphs can help machine learning and how they are related to deep link graph analytics for Big Data.

Unsupervised Native Graph-Based Machine Learning Algorithms

First, graph analytics directly offers a unique set of unsupervised machine learning methods. A host of graph algorithms -- community detection, PageRank, label propagation, betweenness centrality, closeness centrality, and similarity of neighborhoods -- identify meaningful graph-oriented patterns which have wide applications. These include identifying fraud patterns, finding user groups, computing influential users or communities, and reporting weaknesses or bottlenecks in operations or a supply chain.

A common quality that these native graph algorithms share is that they all need to do deep link graph analytics -- traversing several hops or levels of connection within a graph to gather data and analyze it. Such traversal, especially for full-graph analyses like community detection, requires powerful graph computational capability. Only a native parallel graph database can provide that.

Graph Features for Training Supervised Machine Learning Algorithms

A second way that deep link graph analytics helps machine learning is by enriching the set of data features available for supervised machine learning. Consider the example of China Mobile, the world’s largest mobile service provider with more than 900 million subscribers. China Mobile’s network has over 2 billion voice calls every week, and their main challenge was finding phone-based fraudsters who used pre-paid SIM cards with little or no details about the subscriber. So it was hard to distinguish the good phone belonging to a regular subscriber from the bad phone belonging to a fraudster. Simple features such as the duration of the call or the percentage of rejected calls resulted in a lot of false positives.

China Mobile leveraged graph-based machine learning features such as stable group and in-group connections to improve the accuracy of the machine learning for fraud detection. A good phone also regularly calls a set of other phones – say, every week or month – and this group of phones is fairly stable over a period of time (“stable group”).

Another feature indicating good phone behavior is when a phone calls another that has been in the network for many months or years and receives calls back. We also see a high number of calls between the good phone, the long-term phone contact and other phones within a network calling both these numbers frequently. This indicates many “in-group connections” for the good phone. China Mobile has transformed its fraud detection process by generating 118 graph-based features for each phone, feeding tens of billions in new training data records to their machine learning solution.

Good Phone Bad Phone

Explainable Machine Learning/AI Models

Explainable models are becoming a much-discussed topic as machine learning adoption grows. A common criticism of methods like neural networks and their offspring, deep learning networks, is that the predictive models they produce often do not provide any insight into the causal factors.  They cannot give a simple answer to why they predicted a result. Explainable models, on the other hand, tend to highlight the key variables that lead to a decision.  A traditional example is a decision tree. Consider a decision tree to rate your risk of cancer: Do you smoke regularly? If yes, you have an above average risk. If no, ask the next question… Note that a decision tree is a specialized graph.

Similarly, when graph algorithms or graph features are used as part of an AI model, the natural semantics of graph relationships, such as "Customer --(bought)--> Product" lend themselves easily to interpretation.

They are many benefits of explainable AI models. Consumers have come to expect personalized recommendations, for example. If the recommendation comes with some explanation or evidence, the user is more likely to act on it. Graph analytics is well-suited to compute and show the evidence behind these personalized recommendations and explain with graph visualization as needed. For example, graph analytics may show that a percentage of your friends at two or three degrees of separation like this product or service, or that a percentage of other users who have similar interests to you like it, or that the product is similar to your previous purchases.

Graph-based machine learning and analytics is helpful not only to consumers, but also to enterprise users. Take fraud detection, for example. Many enterprises employ large teams of trained investigators to determine whether a transaction is likely to be fraudulent. Graphs excel at combining multiple data sources to connect the dots and allow investigators to see, visually and interactively, how the transaction in question is connected to those previously marked as fraudulent. This is more helpful than, say, just receiving a vague 0.7 fraud score from some machine learning models.

Another example: real-time phone scamming prevention. Each day, hundreds of millions of phone calls are made, but only a tiny fraction of these are malicious scams. Graph technology can quickly explore relationships among callers, phone numbers, and recipients, and be combined with ML to develop trained models to detect which are bogus.

What's more, explainable ML becomes a requirement when a regulatory agency or other auditing party is involved. Banks are required to have high-quality methods to detect possible money laundering.  More and more, banks are using machine learning to improve their detection accuracy, but at the same time, they must be able to show an auditor how and why their system is effective. An ML model that uses graph-based features provides the transparency that is needed.

Graph Powers Explainable AI

Native Parallel Graph Databases

Machine learning has always been computationally demanding, and graph-based machine learning is no exception. With every hop, or level of connected data, the size of data in the search expands exponentially, requiring massively parallel computation to traverse the data. This is computationally too expensive for key-value databases which require too many separate lookups or RDBMS that struggle with too many slow joins. Even a standard graph database may not be able to handle deep link analytics on large graphs.  A native graph database featuring massively parallel and distributed processing is needed.

In order to compute and explain the reasons behind personalized recommendations and fraud detection, the graph database needs a powerful query language that can not only traverse the connections in the graph but also support computation such as filtering and aggregation and complex data structures to remember the evidence.

Deep link graph analytics is powering the next advance in machine learning, through unsupervised learning of graph patterns, feature enrichment for supervised learning, and providing explainable models and results. Combined with AI and ML, it’s a potent combination that will serve enterprises well for years to come.

BioYu Xu is the founder and CEO of Tigergraph whos specialities include parallel computing, database management, Hadoop, MapReduce, large scale data analysis, data warehouse, information retrieval in semi-structured data, keyword search in XML documents, spatial databases, scalable web system development, search engine optimization, XML language and query engine, PigLatin, Cassadra, HBase and HIVE.

Gaurav Deshpande is a seasoned technology executive focused on building winning teams to create new products and industry solutions, establish new markets & dominate marketplace.

Resources:

Related: