Whether for modern Artificial Intelligence or traditional Business Intelligence, the insights derived from data for smart decision making and real-time action stem from analytics. Analyzing data to better achieve business objectives across domains is the main reason most organizations collect data to begin with.
Graph analytics is the foundation for delivering these insights in a comprehensive graph intelligence platform. Graph analytics gives meaning to the other facets of graph intelligence (which include graph query, graph mining, and graph AI) by oftentimes being the next logical step in workflows involving the first two while readying data for graph AI by determining things like machine learning model features.
Analytics itself ranges in several techniques and approaches that fulfill different needs. It involves descriptive analytics that tells users what’s already happened, diagnostic analytics that reveals what’s currently happening, predictive analytics that details what’s most likely to happen, and prescriptive analytics that pinpoint what to do when it does.
The horizontal need for analytics is intensified by the fact that only a fraction of the world’s data — most of which has been created in the past couple of years — has been analyzed. There are also temporal constraints about the value analytics produces. The time to insight usually includes a window of opportunity in which, if the analysis isn’t performed, the results become meaningless because the time for action has passed. Any recommendation engine use case proves this point.
Graph analytics is a revolutionary leap forward for analytics. It encompasses all the above analytics types and, with the High-Performance Computing (HPC) architecture of a graph intelligence platform, addresses the scale of modern analytics requirements. These HPC capabilities allow graph intelligence to satisfy the low latent speeds for creating timely action from analytics results — which is oftentimes the goal of any data-centric investment for competitive advantage.
Scalable, High-Speed Analytics
Most graph analytics techniques work by computing a global property on a full graph. Graphs, of course, contain nodes with properties. Graph analytics typically requires computing different aspects of desired properties. Graph algorithms perform the computations in many graph analytics deployments, which frequently involve relevant vertices in a graph reading the values from their neighboring nodes. Once they bring those values back to their own vertices, they perform the computations and update their own labeled properties.
PageRank provides a good use case for how graph algorithms work. This algorithm is widely used by Google to rank the results of web pages according to their relevance to a particular search term. Graph analytics uses this approach to rank nodes (instead of web pages) according to a formula depicting the importance of the graph’s node for a specific use case. This technique is ideal for ranking multiple matches from graph mining results. For instance, when looking for chemical compounds that fit the requirements for a certain graph query on a medical knowledge graph, PageRank can number the results in order of importance for meeting the specified query’s criteria, which is a good example of why it’s critical to have all elements of graph intelligence in a single platform. Most graph analytics techniques don’t work (certainly not well) outside of graph settings. It’s not possible to determine centrality, for example, in a relational warehouse. Therefore, graph analytics encompasses a suite of effective techniques that aren’t possible in other environments, in addition to those used in traditional settings.
Perfecting Graph Analytics
Graph intelligence offers several means of perfecting graph analytics. It includes several rapid domain-specific analytics designed for users in different industries to get the best results for common use cases. For medical knowledge graph users, for example, there are unstructured search capabilities useful for pharmaceutical research. It’s also possible to expedite domain-specific routines like compound similarity in this setting. The graph intelligence platform also contains a library of analytics algorithms with out-of-the-box functionality.
These algorithms (such as pathfinding, node ranking community detection, graph mining algorithms, centrality, and others) are easy to use to rapidly process large graphs and are also applicable to data science. Thus, users can fulfill horizontal needs like employing graph query to filter results before running the aforementioned algorithms to extract features for machine learning models. These filter data and feature data are locally available for local training of machine learning models, which minimizes data movement — another graph intelligence benefit. There’s also a Python integration that supports creating customized algorithms so users can orchestrate graphs in this popular data science tool.
High-Performance Computing Gains
The speed and scale of graph analytics are attributed to the HPC paradigm of graph intelligence. Users can scale out to hundreds of machines as needed, which is pivotal for removing bottlenecks pertaining to capacities for compute, storage, and memory. HPC enables firms to run graph analytics significantly faster than other approaches can without this functionality. These HPC advantages of graph analytics within the graph intelligence construct are primed for use cases involving complex networks of data.
In a cyber security network, graph analytics is optimal for identifying where in the network a security threat occurred. Graph analytics supports this use case by allowing users to see how many hops they can go while traversing the network to retrieve all the requisite information for identifying where such a security breach took place and how. Because this sophisticated technique is computationally intensive it requires HPC architecture for the sort of real-time responses that are integral for mitigating the damage inherent in this use case, which is why this architecture maximizes the efficiency of graph analytics.
Graph analytics provides similar desirable functionality for other complex network use cases like analyzing financial transactional data to combat financial crimes and fraud detection. Aided by its HPC architecture, graph analytics can swiftly denote how a fraud attack was perpetrated and where exactly in the network it came from. As in the foregoing example, the speed of the response is necessary to minimize any damage — which highlights the aforementioned necessity of responding to events within the window of opportunity for analytics endeavors.
Other compelling examples of the timeliness and scalability of graph analytics with HPC involve the telecommunications industry and electronic design automation. For the former, this approach is creditable for detecting instances of spam in real-time, which can also provide timely defenses for phishing scams. In the latter use case, this method is optimal for building circuit design tools. For this deployment, the circuits function as graphs, the pins are similar to nodes, and the wires are akin to the edges. Graph analytics approaches utilizing logic synthesis and static timing analysis are particularly effective for this use case, enabling users to run these algorithms in parallel with HPC.
Meaningful Action
Graph analytics represents the action within the graph intelligence framework. When coupled with graph query and graph mining, graph analytics is directly responsible for real-time, actionable responses and informed decision-making. Additionally, it sets the stage for workflows involving graph AI, which typically entails advanced analytics for predictive and prescriptive improvement over time for any use case.
This article was originally published by Gurbinder Gill, Senior Software Engineer at Katana Graph, on Medium.
