Mapping patent classifications: portfolio and statistical analysis, and the comparison of strengths and weaknesses

Patent data provide a primary data source for scholars interested in the development of technological knowledge (e.g., Strumsky et al. 2012). In a comprehensive study of US patent data, Jaffe and Trajtenberg (2002), for example, argued that patents and patent citations provide “a window on the knowledge economy.” However, patents are indicators of inventions and not innovations (Archibugi and Pianta 1992; Pavitt 1985; Griliches 1990). Like publication data, patent data provide a wealth of information about knowledge claims (Hunt et al. 2007), references to prior art, forward citation, inventor and applicant addresses, and classifications. Note that patenting can also be strategic or defensive given a firm’s technical expertise and product portfolio (Alkemade et al. 2015).

Different from journal articles, patents are not refereed among peers, but by examiners at the patent office, among other things, on the criterion of whether the submitted application provides novelty. The purpose of the novelty requirement is to prevent prior art from being patented again. The examiner is entitled to add references to prior art,¹ but most importantly organizes the patent in a patent classification system. In the absence of an equivalent for journals, the classification system can be considered as the intellectual organization of the database of novel products and processes of economic value.

Jaffe (1986) was the first who proposed to use co-classifications of patents for characterizing the technological positions of firms with the objective of quantifying technological opportunities and research spillovers. For this purpose, each patent can be represented as a vector of classifications. Jaffe (1989) used the cosine between these vectors as a similarity measure (Salton and McGill 1983). Sets of patents (e.g., patent portfolios of firms) can be modeled as aggregates of these vectors into matrices. The resulting matrix can be used for statistical analysis and/or the visualization of the similarities as projections on a map (Schiffman et al. 1981; Waltman et al. 2010).

In a series of studies, we developed instruments for the mapping and analysis of patent data with a focus mainly on data of the US Patent and Trade Office (USPTO). US patenting is considered the most competitive and is therefore most relevant to technology and innovation studies (Granstrand 1999; Jaffe 1989; Jaffe and Trajtenberg 2002; Lee 2013). USPTO maintains a freely available database of patent data at http://patft.uspto.gov/netahtml/PTO/search-adv.htm. This database is copied at other places, such as Google Patents, the National Bureau of Economic Research (NBER) in the USA, and PatentsView (http://www.patentsview.com). PatentsView adds to the data by providing disambiguated inventor and assignee identifiers (Monath and McCallum 2015).

Building on approaches suggested in previous studies (e.g., Breschi et al. 2003; Leten et al. 2007; Verspagen 1997), we developed instruments for the mapping and analysis of patent data using International Patent Classifications (IPC). IPC was developed by the World International Property Organization (WIPO) and further elaborated by the European Patent Office (EPO) into the European Classification System ECLA. This classification uses up to fourteen characters for the indexing. As of 1 January 2013, however, USPTO and EPO use the new Cooperative Patent Classifications (CPC) which build on ECLA. In addition to the various patent offices of the EU member states, China, Korea, Russia, and Mexico, among others, are in the process of aligning their classifications with CPC. In this study, we use CPC for the mapping, review and integrate the results and routines from the previous studies, revise some of the methodological choices, and update the routines and analysis accordingly.

In addition to the choice of a similarity criterion such as the cosine, a number of parameter choices are relevant for the mapping of patent data: the choice of domains in terms of patent offices (USPTO, EPO, WIPO, etc.), the classification system (CPC, IPC, etc.), the clustering algorithm (e.g., co-classification, co-citation, bibliographic coupling), etc. The classifications can be used with different levels of detail by using more digits of the respective classes, i.e. main- and/or sub-classes.

Kogler et al. (2017b) used IPC at the four-digit level for portfolio analysis; Yan and Luo (2017) used IPC at the three-digit level (IPC-3). Leydesdorff et al. (2014) used both three and four digits of IPC for exploring a dynamic approach. With similar objectives, Kay et al. (2014) composed 466 IPC categories at different levels of depth after preprocessing their data from the comprehensive PATSTAT database.² Schoen et al. (2012) use a database of the Corporate Invention Board (CIB) for the construction of their own classification (Alkemade et al. 2015). Archibugi and Pianta (1992) combined IPC with the Standard Industrial Classification (SIC) using a concordance table. However, the objective of these last authors was not to provide a map or overlays.

In this study, we use CPC at the four digit level. CPC is similar to IPC in the first four digits, but the classification of individual patents can be changed in the process of reclassification; new classes have also been added. Furthermore, CPC contains new categories classified under “Y” that span different sections of CPC in order to indicate new technological developments such as nanotechnology and technologies for the mitigation of climate change (Scheu et al. 2006; Veefkind et al. 2012). Currently, there are nine of these Y-classes. Since these boundary-spanning classes are back-tracked into the database, they introduce new links between previously unrelated classes. As a result, one can expect changes in the networks among the classes. All values in this study are based on CPC at the 4-digit level (CPC-4) instead of IPC-4.

Co-classification is a binary measure at the level of individual patents. Co-citation of classes or bibliographic couplings among them can be considered as more refined measures of the strength of the relationships. Yan and Luo (2017) compared twelve techniques for the mapping of patent data in terms of citation relations, inventive activities, etc., among 121 IPC classes at the 3-digit level. The conclusion was that maps based on Jaccard-normalized bibliographic coupling at the (disaggregated) patent level—“A1” in the classification of methods provided by these authors—significantly outperformed maps based on the cosine values among of aggregated citations among IPC classes (Leydesdorff et al. 2014, 2015). Actually, Yan and Luo (2017, at p. 435) conclude that “class-to-class cosine similarity among aggregated citations (A2) performed the worst in various analyses, although it is popularly used for constructing network maps of patents.” This critique prompted us to reconsider on previous choices of parameters.

In this study, we combine our two approaches: we use bibliographic coupling among CPC-4 classes at the level of patents as individual documents. We first construct the asymmetrical (2-mode) matrix of patents (cited) in the rows versus citing patents aggregated into the 654 CPC classes at the 4-digit level in the columns. The Jaccard index and cosine values are then computed over the 654 columns (Leydesdorff and Vaughan 2006). The resulting (1-mode) matrices can be utilized as input into the mapping exercise.

For the clustering and subsequent coloring of the resulting map we use a methodology recently developed by Leydesdorff et al. (2017) for journal mapping. VOSviewer (v1.6.5; 28 September 2016) provides both a community-finding algorithm and visualization (Waltman et al. 2010). We use this algorithm for generating a hierarchically decomposed set of maps of patent classes. Although the resulting maps are statistical and cannot claim semantic authority, they can serve the heuristics by offering a baseline. Both the Jaccard-based map and the cosine-based one were decomposed into clusters using these statistics. VOSviewer enables us to visualize 654 data points without cluttering the labels on the screen by foregrounding and backgrounding strong and weak presences, respectively. The CPC classes are fractionally counted so that each patent contributes with a value of one.

Indicators for measuring the distance or relatedness between distinct sets of technological knowledge categories are of interest in a number of disciplines; for example, management studies (Almeida 1996; Makri et al. 2010), economics (Jaffe 1986; Teece et al. 1994), and regional science (Fischer et al. 2006; Quatraro 2010; Boschma et al. 2015). The strength of the relationships between unique sets of technological expertise can inform us about the level of specialization and coherence amongst firms, regions, and/or countries. This in turn has been linked to the level of productivity of these units (Nesta and Saviotti 2005; Kogler et al. 2013), and can also serve as an indicator for potential diversification and innovation opportunities as well as limitations (Colombelli et al. 2014; Feldman et al. 2015; Kogler et al. 2017a).

One of our objectives has been to show how sensitive the maps and overlays are to choices of parameters. The maps are not natural, like geographic maps, but remain constructs that originate from discursive reasoning. Consequently, one cannot infer on the basis of visual inspection that one representation is better than another; but one needs analytical arguments for the choices. For this reason, it was urgent to take Yan and Luo’s (2017) critique seriously and redo the maps. As we have shown, the results based on CPC-4 are roughly similar, but different in important respects from IPC-based maps. In our opinion, the ostensable similarity is to a large extent due to the use of similar methods; the empirical results are in important respects significantly different.

We did not follow Yan and Luo’s preference for the Jaccard index, but used the cosine. The Jaccard index is relational and binary, whereas the cosine is a numerical and a positional measure (Burt 1982; Leydesdorff 2014). Portfolios projected on a global map can be considered as positions. We suggested to study relations using a local map and provided a tool for this purpose.

The global maps were not interpretable when using the Jaccard index (Fig. 1), but they were interpretable on the basis of the cosine (Fig. 2). However, we agree on using the basic (2-mode) matrix of cited patents versus citing aggregates into CPC classes. Using this disaggregated level can be considered as an important improvement on previous maps. The input data is more finely-grained by orders of magnitude.

The multi-variate analysis confirmed the improvement. The prediction based on discriminant analysis was improved from 85 to 95%. The (varimax) rotation of the eigenvectors provided us with additional and meaningful insights in the comparisons among cities. The positions of the major Chinese cities among cities in the global domain (factor 1) and cities with strong engineering (factor 2) are suggested in the cosine-based map.

Visualization is a strong instrument because maps can be provided with an interpretation more easily than numerical statistics. However, the latter are needed in the background for making analytical arguments. A direct comparison between two units (e.g., competing firms) using a map, however, can provide a first orientation to their differences in terms of strengths and weaknesses.