Recommendations can be the guiding light for organizations seeking to increase customer engagement, satisfaction, and retainment—all culminating in greater revenue for the enterprise. Great recommendation systems, however, require great technology behind them, capable of instilling context and user research into the data itself.
Katie Roberts, PhD, data science solution architect at Neo4j, joined DBTA’s webinar, “Solving Data Challenges with Knowledge Graphs and Context-Aware Recommendation Systems,” to explore how building and leveraging the power of knowledge graphs can lead to better engagement with end customers through context-aware recommendation systems.
As knowledge graphs serve to identify relationships between data, Roberts emphasized that everything in a business is naturally connected—whether it’s networks of people (employees, customers, suppliers, etc.), transaction networks (risk management, supply chain, orders, payments, etc.), or knowledge networks (enterprise content, domain specific content, etc.).
Yet to make sense of these relationships with only traditional methods at its disposal, an enterprise might break up these networks in order to store its data. When these methods fail to answer relationship-defined data questions, it becomes an expensive task to reassemble the network structure to reveal those relationships.
Roberts posed an overarching question: “What if we could actually bring in all of that context into our decision making without having to apply all of that effort?”
As opposed to fighting the natural, interconnected network structure, Roberts implored viewers to take advantage of graphs to capture the network structure instead. Using graph queries to store and retrieve relationships, as well as graph algorithms to infer relationships, embraces the existing data order instead of paying the hefty price of changing it.
Graph data science, Roberts explained, is the nucleus of understanding relationships within data. Queries help find the patterns you know exist; ML uncovers trends and makes predictions; and visualizations help explore relationships, collaborate on its meaning, and provide explainability.
Knowledge graphs are the first step in the graph data science journey, where capturing customer interactions and the customer journey will reveal critical patterns in your connected data.
The building blocks that shape knowledge graphs exist as nodes, relationships, and properties. The node represents an entity in the graph; the relationship connects nodes to each other; and the property describes a node or relationship, as in name, age, etc.
Mapping relationships of user interactions—such as searches, browsing, and purchases—helps cultivate context-driven recommendations. With traditional methods, pre-processing and joining these different data points to increase recommendation relevancy would take hours; with knowledge graphs, traversing these relationships becomes a simplistic data query, as the relationship is embraced at the level of an organization’s data infrastructure.
Once the knowledge graph is adopted, the next step in the graph data science journey is leveraging your graphs to infer additional relationships with graph algorithms.
“They allow us to go beyond the data you collected initially, to start discovering and inferring these new connections, and flagging the most important or anomalous entities,” explained Roberts. “This lets you identify similar products and customers, find out which products are most important, or who the top ten social influencers within a social network are who might be good targets for marketing campaigns, or even infer relationships between nominal search terms and products.”
Personalized recommendations can be enhanced with graph algorithms to find customer communities based on common purchase behavior. Based on graph algorithms, you can begin to make a variety of connections between nodes based on, for example, similar products, what was purchased with that product, and who to recommend it to.
The final step of the graph data science journey, graph native ML, takes the graph features developed with your queries and graph algorithms and passes them into your ML models. This allows you to develop predictive pipelines, further personalizing recommendations based on your knowledge graph’s relationships, including product correlations, search queries, and historical purchases.
Though personalized recommendations are at the forefront of this discussion, Roberts reminded viewers that graph data science has a broad range of use cases, including fraud detection, disambiguation and segmentation, churn prediction, search and master data management, predictive maintenance, cybersecurity, and life sciences.
For an in-depth discussion of knowledge graphs and applying it to personalized recommendations, as well as examples and case studies, you can view an archived version of the webinar here.