top of page

Knowledge Graph Sweet Spots

At Reclassify AI, we are of the firm belief that knowledge graph technology can and should form the backbone of any organization's approach to data management and governance. In fact, it is the key for turning data systems into flexible platforms for generating new information and even knowledge. But beyond the great case that can be made for the *general utility* of a knowledge graph solution it is helpful to consider the kinds of use cases that particularly benefit from knowledge graphs.

Let's note three key features of knowledge graphs to help us understand the use cases to which they're most effectively applied:

1. Knowledge graphs link objects directly rather than requiring complex joins across tables.

2. Knowledge graphs leverage schemas that can encode meaning formally and provide a normalization template.

3. Partly because of (2) and partly because of the use of semantic web standards, data integration is far easier to accomplish.

This helps us to better understand what kinds of use cases might be particularly amenable to KG solutions.

1. Fraud Detection and Risk Analysis: The key challenge in these scenarios is usually that the relevant data for analysis is not in one place. More specifically, in fraud detection the perpetrators are deliberately trying to hide information to prevent themselves from being detected and the biggest risks in risk analysis, are by definition, the risks presented by situations about which we lack complete insight. So, if doing fraud detection or risk analysis we often need to adjust our datasets to include supplemental information that might shed light on the entities/scenarios under consideration. For example, we may need to understand the social network of potential perpetrators or their communication patterns; we may need to understand supply chains or weather patterns to understand long term risks. All of this entails quickly supplementing our existing datasets with additional data to facilitate analysis.

2. Multi-factor Analyses: Some of the great power of machine learning lies in its ability to analyze data that has heretofore remained hidden or for which we've failed to understand the full complexity of variable interplay. A drug analysis with a minimal dataset might identify a high risk in a particular age group, while failing to understand the locational or genetic factors combining to make it risk worthy. Hence, any attempt to extend our variable set to do deeper dives into prima facie conclusions benefits from a graph and its ability to draw new conclusions.

3. Clustering and Centrality Analysis: The data structures underlying graphs obviously make them particularly amenable to running algorithms to easily and effectively determine nodes and neighborhoods of particular importance but in addition to that, the semantic nature of knowledge graphs also allow us to update our graphs with new nodes and links as we learn new information and detect salient patterns. For example, if we want to run a centrality analysis based on adding links derived from the extent to which people are co-located, or share organization membership or more complex patterns, we can update our graphs with such data and iterate on the associated analysis.

4. Visualization-heavy Analysis: This is related to the above, but oftentimes we have use cases in which the ability to visualize clusters of information or the way in which entities are linked to each other is of particular utility in and of itself.

5. Recommenders: Very often, recommendation engines depend on the ability to do fast and effective one-hop, two-hop or three-hop queries based on some central entity. Graphs are far more performant for such use cases.

6. Simplified Query: If our use case requires a business user or analyst to be able to write and perform arbitrary queries, graph querying is a far more useful solution. The semantics in a graph allow users to write simpler queries that correspond far more directly to the natural language phrasing of the query than a query would in a more traditional data space. For example, a query of the sort "find all people reporting to Naomi who communicated with someone in Europe in the month of April" is easy to write in a semantic graph because it understands the semantics of "reports To", "communicates" (e.g., that it includes email and phone calls) and "Europe", e.g., that it includes Berlin and Zurich, etc. Similarly a knowledge graph allows for easily parameterizable queries that can aggregate along a variety of factors with ease.

These are but a few of a litany of example use case types to which knowledge graphs are particularly useful. In all likelihood, a knowledge graph will make addressing any data analysis use case easier, contact us to discuss how it might help with yours!


bottom of page