Understanding where an organization innovates—and how its research direction changes over time—is difficult when its activity spans hundreds of technology areas. Conventional charts can compress those dimensions into a simplified picture, obscuring meaningful differences. Topological data analysis offers another way to examine patent portfolios: it represents relationships among high-dimensional observations as a graph, helping analysts explore technological positions and trajectories without reducing the data to a handful of coordinates.
Why map technological innovation?
A firm’s patent portfolio can cover many technical classes, and large portfolios may extend across numerous fields. Each firm-year can therefore be represented as a vector of patent counts across technology categories. Comparing those vectors over time raises a practical challenge: the data are too high-dimensional to visualize directly, but aggressive dimensionality reduction can hide distinctions that matter.
Methods such as principal component analysis (PCA) are useful for summarizing variation, while clustering can group firms with similar portfolios. But a two-dimensional projection may make different firms appear close simply because information in the other dimensions has been discarded. Global clustering also imposes boundaries between groups, which can make a firm’s apparent movement depend on where those boundaries fall.
For technical decision-makers, the broader lesson is that a map should do more than display points. It should help preserve continuity and show how portfolios relate to one another as they evolve.
How the Mapper algorithm builds a shape graph
The approach applies the Mapper algorithm, a technique from topological data analysis, to patent data. In the study, each observation represents a firm’s patent activity over a five-year moving window, measured across 430 U.S. Patent and Trademark Office classes. Patent counts are transformed using a logarithm to reduce the influence of highly skewed portfolio sizes.
Mapper constructs a graph through a sequence of steps:
- Project the data through a filter. A lower-dimensional view, such as two-dimensional PCA, helps organize the observations into regions. This projection guides the process; it is not the final representation of all technological relationships.
- Cover the projected space with overlapping regions. Overlap allows observations near neighboring regions to contribute to more than one local group.
- Cluster within each region. A clustering method groups locally similar observations, using a selected distance measure. The analysis uses cosine distance in its baseline specification.
- Connect shared observations. Groups become nodes, and an edge connects two nodes when they share at least one observation.
The resulting shape graph uses nodes to represent local clusters and edges to indicate continuity between them. Although the graph can be drawn on a plane, its connections are based on relationships in the original high-dimensional data. In this way, it can expose branches, dense central regions, and separated components that a simple scatterplot may not make clear.
Reading branches and distinctive innovation trajectories
In the patent analysis, many engineering firms remain in a dense central region, reflecting portfolios that are relatively similar within the map. Other firms extend away from this central area along branches, or “flares.” These structures indicate that a firm’s patent activity has become distinct relative to other firms in the dataset.
The study’s examples include technology companies whose portfolios diverged from a shared electronics-centered area over time. A distinctive trajectory does not necessarily mean a firm has the largest portfolio: a firm can stand apart because of the direction of its patenting, even when its patent volume is not especially high. The map also captures convergence. Some firms’ trajectories move inward toward other groups, a pattern that can accompany changes in portfolio scope.
These patterns are relative, not absolute. A firm’s position depends on the full set of firms and patent portfolios being mapped. If other organizations enter a technological area or change their portfolios, the degree to which one firm appears unique may also change.
Turning graph structure into a measure
To quantify these patterns, the researchers define a firm’s interior in the graph as nodes containing that firm whose immediate neighbors also contain it. The boundary consists of its nodes that connect to nodes without the firm. A flare is a connected piece of the interior that remains linked to the broader graph; a separate connected component is treated as an island.
For each interior node, an exit distance measures the shortest graph path to the firm’s boundary. The flare index for a connected piece is its greatest exit distance, and the firm’s flare length is based on its longest finite flare. This graph-theoretic measure formalizes how far a firm’s distinctive path extends within the mapped structure. An island is handled separately because it has no path to the rest of the graph.
What the performance analysis indicates
The analysis reports a positive association between measured flare length and firms’ financial performance, including revenue, earnings before interest and taxes, and market value. The relationship remains in the reported regressions after accounting for factors such as patent portfolio size, firm survivorship, industry classification, and firm fixed effects.
This finding is an association, not proof that a distinctive patent trajectory causes stronger financial results. The map is a descriptive and measurement tool; it does not by itself establish why a firm’s performance changes. Still, the relationship suggests that information about the direction and distinctiveness of inventive activity may add perspective beyond patent counts alone.
Where topological mapping fits with existing methods
Mapper is best viewed as complementary to PCA and clustering rather than a universal replacement. PCA can provide a compact summary, and global clustering can be helpful when the goal is to sort firms into broad groups. Mapper’s local clustering and graph connections are more suited to examining continuity and firm-level movement across a complex portfolio landscape.
Results also depend on analytical choices, including the filter, distance measure, clustering approach, graph resolution, and overlap between regions. These parameters shape the graph and should be selected and tested with the analytical question in mind. The method’s value lies in making high-dimensional structure easier to inspect—not in removing the need for careful interpretation.
Conclusion: Use portfolio maps to explore direction, not just scale
For organizations assessing technology strategy, patent volume alone may not describe how an innovation portfolio is positioned or changing. Topological data analysis provides a graph-based way to explore relationships among firms’ patent portfolios, identify distinctive trajectories, and examine how those paths connect to broader technological clusters.
Used alongside conventional statistical tools, Mapper can help engineers, strategy teams, and decision-makers ask more precise questions about where inventive activity is concentrated, where it is becoming differentiated, and how technological positions evolve over time. The practical implication is to treat innovation mapping as an analysis of both what a firm patents and the direction its portfolio takes.