Measuring knowledge persistence: a genetic approach to patent citation networks

Abstract

The aim of this paper is to propose a new empirical method for identifying technologically important patents within a patent citation network and to apply it to the telecommunication switching industry. The method proposed is labelled the genetic approach, as it is inspired by population genetics: as geneticists are interested in studying patterns of migration and therefore the common origins of people, in innovation studies we are interested in tracing the origin and the evolution of today knowledge. In the context of patent and citation networks, this is done by calculating the patent’s persistence index, i.e., decomposing patent’s knowledge applying the Mendelian law of gene inheritance. This draws on the idea that the more a patent is related (through citations) to “descendent” patents, the more it affects future technological development and therefore its contribution persists in the technology. Results show that the method proposed is successful in reducing the number of both nodes and links considered. Furthermore, our method is indeed successful in identifying technological discontinuities where previous knowledge is not relevant for current technological development.

1 Introduction

The use of patents as an indicator of innovative activities is rather established in the innovation literature (Yoon and Kim 2011; Griliches 1991). However, one of the usual caveats pointed out by scholars is that patents are not informative about the importance (and the economic value) of the invention they disclose. The usual way to overcome this issue is to look at forward citations.¹ In fact, as the number of citations signals the importance of scientific articles, we have evidence that important patents receive as well more citations.

Recently, exploiting the similarity between scientific and patent citations a new stream of literature has adopted a network approach to the study of patent and citation data, and analytical tools previously used for publication networks have been used in identifying important patents (von Wartburg et al. 2005; Fontana et al. 2009). The aim of this paper is to contribute to this stream of research and to propose a new method for identifying technologically important patents in a patent citation network.

In order to make sense of the new indicator and method, we need to define our notion of importance. First of all, we need to stress that this work focuses on the technological importance of patents, as a patent citation network pinpoints only the technological relation between inventive steps (i.e. patents). In the next section, we will discuss at length the nature and the interpretation of such links. For the moment, it is enough to consider that being cited as prior art ² implies a technological relation between two inventions.

In this paper, the importance of a patent relates to the persistence of its technological contribution. This draws on the idea that the more a patent is related (through citations) to “descendent” patents, the more it affects future technological development and therefore its contribution persists in the technology. The choice of a genetic jargon is not accidental. In fact, from a purely conceptual perspective, the exercise carried out here is similar to the one made by population geneticists. As they trace our (geographical) origins by looking at genetic mutation in today’s population (Cavalli-Sforza et al. 1994), in this paper we propose a method to “decompose” the technological content of today’s patents into the technological contribution of the previous ones. We can therefore investigate the origin and development of today’s knowledge, identifying technological lineages representing cumulative chains of technological advances leading to today’s knowledge. For this reason, this new method is labelled the “genetic approach” (GA) to patent citation networks.

As we will see, differently from common approaches to patent and citation data, the GA (and, in particular, the knowledge persistence indicator) assesses patents’ citation structure from a broader perspective, considering not only direct citations but also all the direct and indirect “descendants”. Therefore, importance is not a “local” characteristic of the node (i.e. the number of citations a patent has) but a global characteristic.

In this article, we propose an application of the methodology in the context of the telecommunication switching industry. This represents an interesting case study because, in the last decades, this industry has gone through numerous incremental and radical technological changes. Results show that the method proposed is successful in reducing the number of both nodes and links considered. Furthermore, our method is indeed successful in identifying technological discontinuities where previous technological knowledge is not relevant for current technological development.

As regards the broader impact of this new methodology, we can expect two lines of generalization. The first (obvious) one is the possibility to apply the persistence index to other technological fields in order to characterize their cumulative and disruptive dynamics. The scientific importance of such an exercise is to foster quantitative research and appreciative theorizing defined as rigorous storytelling (Nelson 1989) in studies related to technology dynamics and its effect on industry dynamics.³ The second one regards the possibility of applying the persistence outside the context of patent citation networks. We postpone some speculations on this aspect until the last section.

The paper is structured as follows: Section 2 introduces patent and citation networks and provides the theoretical background for our notion of importance; Section 3 presents the details of the new method; Section 4 provides the comparison to existing alternative method; Section 5 introduces the industry here examined and the dataset; and Section 6 presents the empirical results. Conclusions follow.

2 Theoretical background

2.1 A network approach to patent and citation data

Patents and citations have been already used for building several types of knowledge networks the nodes of which are firms, inventors (Balconi et al. 2004), or technological classes (Bottazzi and Pirino 2010). The network analyzed in this work is composed of patents (the vertexes) and citations (the arcs). In particular, the forward citation between two patents establishes a directed link going from the cited to the citing patents. If we adopt a stylized (but realistic) view of patents, considering them as a collection of “technical problems and newly proposed solutions”, the advantage of the network approach is the possibility to appreciate the interrelated nature of technological developments.

As regards the nature of such relations, this depends on the interpretations of citations (e.g. knowledge spillovers, knowledge flows, etc.) (Marco 2007). If, on the one hand, we prefer to remain rather agnostic about the exact nature of this relation, on the other hand, such a technical relation is codified in the definition of prior art. The citation to an older patent means that the invention disclosed in the old patent can be relevant to the new patent’s claims of originality. This means that the definition of the prior art allows us to limit and to circumscribe the novelty of an invention. Following (Strandburg et al. 2009) “…A citation from one patent to another may indicate either that the later patent builds upon the technology of the earlier patent or simply that the earlier technology was closely enough related to be material to determining whether the later patent should be issued... (page 108)”. In this work, we adopt a “minimal” connotation of such links: if patent A cites patent B, we can think of a technological relation between the inventions disclosed by the two patents. In a conservative way we can claim that a patent citation network built as just described allows us to map the technological relations between all the inventions carried out within a specify technology.⁴ The analysis of its structure and topology (i.e. the layout of interconnections of the various elements) allows unfolding patterns of technical changes.

The challenge posed by patent citation networks is not only their size,⁵ but also other characteristics such as directionality and acyclicity. The former refers to the fact that ties are not symmetric. The latter means that no cycles are present in the network, as patents can only cite previous patents and therefore the time dimension is embedded into the ties’ direction. These characteristics make difficult to apply standard network analysis techniques, as these mainly deal with undirected networks and most of the usual indicators cannot be applied in a straightforward way to a directed network.⁶

For these reasons, the interest on patent citation networks is not confined to innovation studies. Researchers in complex systems began to be interested in the patent citation network as a whole, as it constitutes an example of a complex system emerging from human activities. They apply mechanical statistics in order to study network dynamics (i.e. the study of the organizing principle) and to model a system so as to replicate specific properties such as the power law distribution of the number of forward citations (Valverde et al. 2007).⁷ They explain the emergence of such a property by the combination of two (necessary) forces: network expansion and “preferential attachment” (Barabási and Albert 1999; Albert and Barabási 2002).⁸ In particular, they adapt the latter principle for the context of patent and citation network, making the probability of a patent to be cited dependent on the number of current citations and the age of the patent. These studies are rather recent and if, on the one hand, they shed some light on the structure and dynamics of patent networks, on the other hand, their approach is poorly grounded in innovation studies. For instance, they are not interested in estimating (and explaining) sectoral differences in the parameters fitting their models. For these scholars, the patent citation network is just an example of a complex system emerging from individual (innovative) behaviors and their aim is to find potential commonalities with natural complex systems.

In the innovation studies domain, there has been a recent⁹ interest in patent citation networks, and in a short time several empirical articles have been published. All of them use the bibliometric indicators developed by Hummon and Doreian (1989) for identifying the main flow of knowledge within a patent citation network (Mina et al. 2007; Verspagen 2007; Fontana et al. 2009; Barberá et al. 2010; Martinelli 2012).

The GA here proposed heavily differentiates from the studies just mentioned. Differently from the complex system approach, we are interested in grounding the indicators in the innovation literature and the persistence indicator is going to shed some light about technology evolution. Differently from the Hummon and Doreian approach (HDA), the GA is more flexible in its construction as it does not stress the importance of direct links. As the HDA is now increasingly used in the literature, we are going to discuss in detail their differences in Section 4.

2.2 Importance as knowledge persistence

The aim of this section is to provide a theoretical background for the notion of importance proposed in this paper. As already anticipated in the introduction, important patents are the ones the technological contribution of which is found in today’s patents. How is knowledge transmitted through the network and then found in recent patents? Endorsing a neutral outlook on the meaning of the links, we can think of citations as pipes through which pieces of knowledge are inherited from cited to citing patents. Therefore, as a population genetist compares the genotype of contemporary populations, we could study the “genetic” structure of today’s knowledge¹⁰ and rigorously trace their origin.

The genetic decomposition is explained in the next section. However, we can anticipate that it is operationalized using the Mendelian notion of genetic inheritance. We are, therefore, able to identify “knowledge lineages” depending on the citation structure between a patent and all its direct and indirect “descendants”. In fact, the contribution of each patent depends on the topological structure of the network composed of all the paths generated in it.¹¹

In this way, a patent citation network represents a system of knowledge generation where the inventive step is provided by the recombination of existing (as inherited) pieces of knowledge. Following the Mendelian decomposition mechanism, it appears that a building block of the evolutionary process, which is random mutation, is missing. In this respect, the genetic patent decomposition might look like a deterministic representation, where new inventions are simply a sum of proportion of previous knowledge and where nothing genuinely new is created. However, knowledge is recombined and therefore, by definition, is transformed into something different and therefore new. Undoubtedly, this concept of innovation used in this work is rather similar to the concept of “recombining knowledge” put forward by Weitzman (1998, 1996), according to which …new ideas arise out of existing ideas in some kind of cumulative interactive process” … (Weitzman 1996, page 209).

We expect that, in a dense network, few patents are successful in spreading their knowledge, displaying a high level of persistence and generating a large lineage of descendent patents. These lineages represent chains of technical change with a certain extent of cumulativeness. Therefore, the use of both the genetic decomposition and the persistence index allows us to answer questions related to patents technological importance and technology dynamics.

3 A genetic approach to patent citation networks: genetic decomposition and knowledge persistence

Note that directionality follows the direction of the knowledge flow between cited and citing patents.

Figure 1 represents a very simple patent citation network structure with five startpoints, two intermediates and two endpoints.Endpoints are generally recent patents and they represent the set of what in the previous section was labeled “today’s knowledge”. The genetic decomposition corresponds to decompose the knowledge content of the endpoints in function of the startpoints. After the decomposition is performed, it is possible to quantify the degree in which the startpoints’ prospective knowledge is retained in the endpoints and to look at how much knowledge the two endpoints actually share.¹³

Fig. 1

Simple patent citation network structure

Clearly, both these aspects depend on the structure of the forward citations and on the number of intermediates and their citations. This means that, for calculating the persistence index, we should not focus only on startpoints but we also need to account for the intermediates’ contribution that represents the “new” knowledge injected in the system. In fact, in the framework of the genetic decomposition in a patent citation network, both knowledge persistence and knowledge creation coexist. Summarizing, Fig. 1 represents a process of knowledge creation, transmission, and transformation.

The simple network displayed in Fig. 1 is composed of three layers of patents, indicated by TR0, TR1, and TR2. The genetic decomposition of the network in Fig. 1 is performed using the following heuristics:

• 1. Endpoints are identified, and working backwards, each patent is assigned to a layer;

• 2. For each startpoint belonging to the first layer (TR0 in Fig. 1), the persistence index is calculated. This is going to quantify how much of their knowledge is retained in the endpoints (TR2 in Fig. 1);

• 3. The startpoints are deleted (the network is truncated) and therefore a new layer of startpoints is created (TR1 in Fig. 1);

• 4. Calculation of the persistence index for the new group of startpoints;

• 5. Deletion of the layer and repetition of step 2 and 3 up to the last layer.

This procedure is repeated for each layer and the number of layers depends on the length of the largest geodesic distance in the network.¹⁴

Steps 2 and (recursively) 4 represent the core of the new method, which corresponds to the application of the Mendelian law of genetic inheritance to citations.

Looking at the first layer (TR0) in Fig. 1, we can see that the only patent cited by patent F is patent A; thus 100 % of the inherited knowledge embodied in patent F is the knowledge of patent A. Instead, patent G makes three citations to patents B, C, and D; thus the inherited knowledge embodied in patent G consists 33.3 % of patent B, 33.3 % of patent C and 33.3 % of patent D.

In the second layer (TR1), the endpoint I makes only one citation and that is directed to patent F. Since the embodied knowledge in patent F is 100 % that of patent A, the inherited knowledge embodied in patent I is again 100 % that of patent A. The endpoint H makes three citations. The first is to patent F that has 100 % patent A knowledge; thus $\frac {1}{3}100~\%=33.3~\%$ of the inherited knowledge embodied in patent H is the knowledge of patent A. The second citation of patent H is to patent G that embodies 33.3 % of each of the respective knowledge of patents B, C, and D. Since patent H inherits only $\frac {1}{3}$ of its knowledge from patent G, it inherits indirectly $\frac {1}{3}33.3~\%=11.1~\%$ of each of the knowledge of the startpoints B, C, and D. Finally 33.3 % of the inherited knowledge in patent H comes directly from startpoint E. These results are given in Table 1 in matrix form.

Table 1 Genetic decomposition and persistence of knowledge for Truncation 0

Focusing on the TR0 level, the genetic decomposition answers the question: How much knowledge of (startpoints) A , D , C , B , and E is retained in (endpoints) H and I ? The answer is displayed in Table 1, where each other row decomposes the inherited knowledge embodied in another endpatent down into the shares of the startpoints which have supplied the knowledge. For this reason, each row adds up to 1.0.

The persistence index can be found in the last row, the column sum. It supplies a fractional count that is the basis of the persistence index: effectively, 1.33 out of the 2 endpoints are the pure descendants of the startpoint A, 0.33 out of 2 are the pure descendants of the startpoint E, and each of the startpoints B, C, and D has 0.11 pure descendants. Clearly, startpoint A is the most important startpoint since it is $\frac {1.33}{0.33}=4$ times as important as patent E and $\frac {1.33}{0.11}=12$ times as important as patents B, C, and D. Patent E is the second most important startpoint and startpoints B, C, and D look not so important on their own.

As explained in the step list before, the decomposition algorithm puts only the startpoints into competition. In other words, the intermediate patents F and G do not show up in Table 1 as knowledge suppliers as they just transmit the knowledge from the startpoints to the endpoints. Therefore, after step 2, startpoints in layer TR0 are removed from the network (i.e., truncate the network from the left), and a new set of startpoints is created. In the example of Fig. 1, one left truncation (i.e., removal of patents A , B ,C, D, and E) brings patents F and G forward as startpoints. Table 2 shows the results for the genetic decomposition and the persistence index for the TR1 level.

Table 2 Genetic decomposition and persistence of knowledge for Truncation 1

How much knowledge of F and G is retained in H and I ? Patent F appears to have 1.5 pure descendants and patent G has 0.5 pure descendants; thus patent F is $\frac {1.5}{0.5}=3$ times as important as patent G.

In the simple network displayed in Fig. 1, only three layers are present since another step of truncation leaves us only with the endpoints. In real samples, the network is left truncated and analyzed as long as it is possible to truncate further. It follows that, for each layer, a matrix such as Table 1 and 2 is calculated and the persistence index is calculated as the sum of each contribution (the last row in Tables 1 and 2). Furthermore, the persistence index is then normalized using the maximum, meaning that for each truncation the persistence index takes a value between 0 and 1.¹⁵

4 The genetic approach vs. other approaches

Before moving to the empirical details, it is worth to spend few words to compare this new method to existing comparable methods. In particular, we are going to discuss differences with citation count, the originality indicator proposed by Trajtenberg et al. (1997), and the connectivity approach (Hummon and Doreian 1989). Furthermore, in the empirical section (Section 6.1), we will compare the persistence index to all these indicators in order to show their relations.

The network approach to patent and citation data represents a shift of perspective respect to citation counts. The two have already shown some complementarities: having a large number of citations is not a sufficient condition for becoming an important connection in the main flow of knowledge within the network (Fontana et al. 2009). Citation counts can be easily performed from a network perspective. In fact, it is always possible to count the direct number of ties a patent has. However, any network approach allows us to enlarge this local perspective and to evaluate the whole citation structure.

Trajtenberg et al. (1997) use a similar jargon for introducing some patent indicators of basicness and generality of invention. In particular, their measurement of patent importance does not consider only the forward citations but also the importance (i.e. the forward citation) of the citing patents. They compute patent importance as a sum of forward citation and a fraction of the forward citation of the citing patents. As they evaluate importance looking at two rounds of forward citations, they broaden the local perspective of citation counts. However, the choice of the weight is rather arbitrary (as well as the choice to stop at the second round) and it does not account for possible cross citations between all these subsequent patents. In this respect, the network approach here implemented allows us to account for all the patent citation structures.

Finally, as already anticipated in Section 2.1, recently scholars (among other see: Mina et al. (2007); Verspagen (2007); Fontana et al. (2009); and Barberá et al. (2010)) have used the bibliometric method proposed by Hummon and Doreian for identifying the main flow of knowledge in a patent citation network.

Ultimately, both the genetic approach (GA) and the Hummon and Doreian approach (HDA) have the same aim, which is to trace “important” technological advances. However, they deeply differ in the underlying rationale and definitions. An example from population genetics can be helpful. Summarizing, we can say that the work of geneticists is: given the observed population, genetic differences with ancestors¹⁶ highlight streams of migration. The application of the GA to endpoints follows exactly the same rationale and therefore it can be considered a backward mapping of successful (persistent) technologies. The application of the HDA to populations would work in a completely different way, that is: starting from the ancestors at each (population) bifurcation, follow the future development of the largest stream. This would correspond to tracing just the largest population and ignoring the remainders and their future development. Of course, the application of the HDA by geneticists would be nonsense. However, this example clarifies the basic differences between the two approaches that is, the direction of mapping: backwards for the GA and forward for the HDA. As a consequence, the HDA might be very sensible to “local” peaks and therefore discarding chains of innovations which from an ex-post perspective (the importance of the endpoints) is relevant. Furthermore, the HDA uses a search algorithm that selects highly valuable subsequent links, which introduces a certain bias towards a particular definition of cumulativeness and over-emphasizes the notion of incremental progress.¹⁷ By contrast, the implementation of the GA presented in this work is rather flexible, as the analysis of the persistence weighted network does not impose any specific structure.

5 Case study and dataset

5.1 The telecommunication switching industry

Before moving to the empirical section, we are going to discuss briefly the industry under examination. We do not want to summarize the milestones of technical change in the telecommunication switching industry¹⁸ but rather desire to convince the reader on the validity of this industry as a testbed. This consideration is based on two facts: first, the industry is an innovative one, and second, patents are representative of such advances. The study of technical change in the industry clearly show that the period between the 1975 and 2001 is characterized by both “normal” and “disruptive” technical change. The former refers to the development and the consolidation of digital switches, whereas the latter refers to the transition from circuit to packet switching.¹⁹ The presence of such waves of technical change makes this industry a suitable testbed for a new method for studying technology dynamics. Moving to the second point, interviews with engineers active in the period under examination support the idea that patents were used for protecting inventions and enhancing intense cross-licensing (Chapuis and AEj 1990; Martinelli 2010). Therefore, patents provide complete and comprehensive data about technological change in the telecommunication switching industry.

5.2 Data

The patent sample was retrieved from the USPTO website using technological subclasses that all belong to technological class 370 (“Multiplex Communication”).²⁰

The selection of the technological class and technological subclasses was driven by the reading of their descriptions and by the analysis of firms’ patent portfolio for companies highly specialized in switch production. In order to account for the complexity of a switch and to consider important technologies that might not have been captured by the first search, the first round of cited patents were added to the original set. These cited patents were retrieved from the NBER patent database (Hall et al. 2001); for patents granted before 1975, the citations were taken from the patent documents²¹ and manually added. The final sample includes 6214 patents covering the period 1924-2003.

Citations for patents issued before 1975 were manually collected. It is important to note that the citations considered are only the “internal” ones, meaning that once the patent sample is selected, only citations to patents included in the starting sample are considered.

6 Empirical analysis

For the results obtained using the HDA on the same data, see Martinelli (2010).

This section is going to present the results obtained using the GA introduced in the previous sections. In particular, two types of analyses are performed: (i) the persistence index is used to identify important patents and to assess firm’s patent portfolios and (ii) the persistence index is used to weight the citation network in order to shed some light about technology dynamics in the telecommunication switching industry.

6.1 Persistent patents

As mentioned in Section 3, the persistence index is calculated only for the startpoints generated after each truncation; therefore the persistence index can be calculated only for a subsample of the patents. Table 3 shows the number of startpoints evaluated in each of the 25 truncation levels present in our network.²³

Table 3 Number of startpoints per truncation

The persistence index calculated at each truncation presents a very skewed distribution displayed in Appendix A. Even if graphs are small, they clearly show the high left skewness, meaning that only a handful of patents are successful in spreading their knowledge, whereas the contribution of the high majority is diluted over time. This is consistent with the evidence on strongly skewed distribution of other indexes for patent importance and value (e.g. forward citations, license fees, etc. etc.) indicating that, generally, only few patents are important and valuable (Marsili and Salter 2005; Silverberg and Verspagen 2007).

As high persistence indicates high technological importance, we can rank patents and identify “important” ones. Given a rank, we can expect a certain extent of arbitrariness in deciding which patents are important, those corresponds to set a minimum threshold. However, in the case of persistence index, arbitrariness is very low because of its extremely skewed distribution. Using a rather conservative cut off point, which is 0.5, we extract 79 “important patents”.²⁴

The rationale about using a method for reducing network complexity is the possibility to infer some properties on the whole network just using that subsample. In order to test whether our subsample (representing 1.27 % of the full sample) is a representative one, we compare some patent indicators and test the persistent subsample against the full population. Table 4 reports some summary statistics for patents characteristics and citations indicators for the set of persistent patents and the remaining sample.²⁵ The table includes:

• 1. the patent citation count proposed by Jaffe and Trajtenberg (2005), which consists in the number of forward citations plus 1;

• 2. the average issue year of the patent;

• 3. the SPLC (Search Path Link Count) indicator introduced by Hummon and Doreian (1989) in their main path analysis. Without entering into the details, these indicators evaluate the connectivity of a citation by measuring how many downstream and upstream patents are connected throughout such citation;²⁶

• 4. the number of claims reported in the patent;

• 5. the generality index proposed by Trajtenberg et al. (1997) calculated on the IPC patent classes.

Table 4 Summary statistics for citation indicators

Table 4 shows that the patents extracted using genetic decomposition are older, receive more citations, have higher connectivity, and display more claims. Interestingly, figures related to the generality indicator are rather similar, suggesting some potential similarity between the two indicators (however see Fig. 2). These results hold both for the mean and the median.

Fig. 2

Persistence index and patent indicators

Table 5 shows the results of the comparison between the two samples. The non parametric Wilcoxon-Mann-Whitney used for accounting for the high skewness of the variables rejects the hypothesis of samples extracted from the same distribution for the citation count and the SPLC.

Table 5 Results of the Wilcoxon-Man2n-Whitney test on the median

As anticipated in Section 4, we now focus on the comparison between the persistence index and other patent indicators there exposed.

Figure 2 shows the scatterplots of the persistence index against the number of forward citations, the generality index (calculated on the IPC classes of backward citations), and the SPLC indicator. What clearly appears from the graphs is the lose relations between the new proposed indicator and the already existing ones. Indeed, this points to the fact that the former unfolds a different aspect of individual inventions covered by patents. In particular, we can see that high persistence is not systematically correlated to high frequency of forward citations, high level of generality, and high level of cumulativeness. In fact, going back to the discussion of the previous sections, persistence as explained and operationalized relates to both “long term” and widespread influence of a patent on subsequent innovations.

As inventions covered by patents are developed within a firm, persistence can be used to a evaluate firm’s patent portfolio. In this respect, we use the persistence index for dividing the patent set into four equal groups and we look at the 10 top assignees for each quartile. Table 6 reports (for each quartile) the name of the assignee, the number and the percentage of its patents in that quartile, and the average issue year of these patents. The latter is displayed in order to give an idea of the “vintage” of the portion of the patent portfolio in each quartile.

Table 6 Assignees over patent persistence quartiles

Table 6 shows not great differences in the assignees’ names over quartiles. However, when we look at the concentration and the dispersions of the shares of patents, we can see that the upper and the top quartiles are more concentrated as the top four companies account for a larger share of (more persistent) patents. Furthermore, the increase in the HHI index over quartiles support the idea that persistent patents are developed by a lower number of companies. In a nutshell, companies developed both persistent and not-so persistent inventions, However, only a few companies are able to produce inventions the impact of which are found in subsequent inventions. In the final part of next session we will elaborate more on the relation between these companies and technological evolution.

6.2 The persistence weighted network

Following the previous section, we can see that the persistence index is associated with patents and, therefore, from a network perspective, it refers to a characteristic of the node. However, this indicator can be also used to weight the links of the patent citation network in order to build a persistence weighted network. In such a network, each link (i.e. the citations) is weighted using the product of the normalized (by the maximum) persistence index of the citing and cited patents to which the link connects. The persistence index is therefore used for transforming the binary patent citation network into a weighted one, where the values of the links inform the persistence level of the knowledge transferred by that citation. It is worth to point out that, if on the one hand, this use of the persistence index is rather straightforward, on the other hand, it is just one possibility, leaving many opportunities for alternative applications. In fact, we think the persistence index is a rather flexible indicator and it can be used in several ways both for network analysis and visualization. We leave for future research further explorations and methodological extensions.

Given the fact that the persistence index is available for a subsample of patents, it is not possible to weight all the citations. Therefore, the weighted network includes 16,747 citations, corresponding to 80.3 % of the full sample.

Figure 3 represents the distribution of the logarithm of the citations weight. It highlights the fact that very few links transmit persistently knowledge.²⁷ In fact, in this graph, the last column represents the number of citations with the highest weight, directly connecting the patents with the highest persistence index.²⁸

Fig. 3

Distribution of citation weights

In the following pages, the persistence weighted network is analyzed looking at how its structure evolves considering different cutoff points. Using a visual metaphor, we can conceive the persistent weighted network as the technological landscape, where patent height depends on the persistence index. Using a very high cut off point corresponds to deleting unimportant links and consider only citations with a (relative) high persistence measure. In this logic, the first step is to set a very high threshold such as 0.9 and to look at citations transmitting the most persistent knowledge within the network.

The resulting network structure (Fig. 4) presents two separated components indicating two disconnected areas of highly persistence knowledge unfolding from a network of more diluted knowledge. Undoubtedly, this fragmentation is dependent on the chosen threshold. However, this two component structure is stable down to 0.75 threshold.

Fig. 4

Network with cut off point 0.90

The existence of two separate components suggests the presence to two separate technologies that did not intensively interbreed. In fact, the lack of a bridge between the components indicates that the two components do not share any persistence knowledge. A way to validate this finding is by looking at some characteristics of the two components such as their vintage, their technological contents, and their assignees.

At first, components 1 and 2 differ in their vintage: component 1 comprises patents granted between 1949 and 1977, whereas components 2 are patents granted between 1980 and 1999. Beyond the years of grant, substantial differences emerge. Looking at their technical contents, component 1 deals with the development of a reliable digital telecommunication circuit switching, while component 2 is composed of three branches converging to patent 5953344 and finally to the endpoint 6272129. These patents address technical solutions responding to increasing demand for data communication and the development of packet switching and the use of Internet protocol.

According to the literature, the industry underwent a wave of disruptive technological change from circuit switching to packet switching, occurring in the period under examination. Following the innovation literature, this discontinuity represents a paradigmatic change (Dosi 1982) as it affects not only the design of telecommunication switches but also technological competences (both at firm and inventor level) needed for their development (Martinelli 2012). Therefore, patents in the two separated components in Fig. 4 disclose radically different inventions. The lack of highly persistent connections suggest that the later technology (i.e. the packet switching indicated with 2) is loosely built on previous technological development.

Given these two peaks represent such different technologies it is interesting to look at how they are connected and therefore to what extent knowledge from an early paradigm is retained in a subsequent one. This means to look at how the network structure changes, lowering the cutoff points. For instance, we can imagine two opposite scenarios: on the one hand, these two peaks may be connected by several links or they may be connected by unique link. Both examples would correspond to two different ways through which the technology evolves and a technological discontinuity emerges.

Figure 5 shows citations with a weight larger than 0.75, which allows us to include in the structure more patents and citations. The additionally included patents are indicated in white. Looking at the structure, we can observe that most of the citations newly included are in the area of the two separate components (where some triads are closed). This means that, in this case, persistent (and therefore important) knowledge links tend to cluster around the peaks rather than connecting them. Second, the citation network is by no means broken anywhere between the earlier and the latest patents and a single semipath connect the two components. More on this point: it is interesting to notice that the connection between the two components takes place through one single patent (patent 4245341).

Fig. 5

Network with cut off point 0.75

As the structure is dependent on the choice of cut off points, Fig. 6 shows the resulting network when the threshold is lowered to 0.5.

Fig. 6

Network with cut off point 0.5

The two structures are different but still comparable, despite the increase in the number of nodes and edges.

Looking at the whole structure, we can notice: (i) the emergence of a few short paths (indicated with A in Fig. 6), and (ii) the emergence of a shortest path connecting the two isolated components of Fig. 4 (indicated with B in Fig. 6). Following the analysis of the previous figures, it is interesting to look at the technical contents of these patents.

Despite the different vintage of the newly added patents indicated in A and B, they all disclose similar technological developments. They relate to the design of an hybrid switch (the ATM), including features of both circuit and packet switching. The early patents in A put forward the idea of packets, but still in a “connection-oriented” framework, whereas the patent in group B includes later patents related to the early development of ATM. ATM switches data are still divided into packets but the path is conceived at the outset, and all the packets are then sent through the same circuit. Despite its potential in mitigating some problems of packet switching (such as the decrease in the Quality of Service), it was still not optimal for quick demand increase for data communication, and therefore quickly dismissed by manufacturers .

The technical content of the patents included by lowering the threshold sheds some light on chains of less-and-less persistent innovation, detecting ways through which the technological space was explored. In the case of the telecommunication switching industry, this corresponds to abandoned technologies. In a dynamic perspective, this method highlights patents that contained relatively persistent knowledge but that are made obsolete with the emergence of new knowledge.

In order further to validate our finding. we can look at the assignees for these highly persistent patents in the network. The lower-right pane of Table 6 displays the top 10 assignees of the patents in the top quartile of the persistence index. All these companies have been top-players in telecommunication switches, however, specialized in different market segments and therefore on slightly different technologies. The average issue year of their persistent patents help to place such patents (and therefore firm’s technological choices) in the technological evolution timeline. Bell Laboratories was the most advanced institution as regarding telecommunication switching and in particular in the use of circuit switching. Not surprisingly, its share of persistent patents are rather old and reflect the commercial introduction of digital switches based on circuit switching. As regarding digital switches, Nortel also represents a case of success being able to internationalize rapidly in the early 1990s (Sutton 1998). By contrast, IBM has never active in the telephone switching market, as its main interest was in computer networking and data transmission. In this respect, its patent portfolio is slightly younger than that of Bell Laboratories. Consistently with what emerges from the history of telecommunication switching industry (Fransman 1995; Martinelli 2010), only a limited number of companies (among which the Japanese NEC and Fujitsu, and Nortel) were involved in developing of ATM switches (the branches A and B in Fig. 6) in the late 1980s. Finally, Motorola has been active in a specific and later-developed submarket designing and selling wireless network infrastructure equipment (for instance: cellular transmission base stations and signal amplifiers).

Going back to Section 4, we can now return on one of the difference between the GA and the HDA. In the empirical exercise just carried out, there is no reason to assume that technology evolution developed only along one path, especially in case of a technological discontinuity where lot of “search around” is performed. In the HDA, the search algorithm sequentially searches for a path between each startpoint edges with the highest connectivity measure are sequentially selected up to an endpoints. It follows that, in the case of the HDA approach, the single ridge connecting the two peaks is endogenous in the search algorithm that does not allow for the emergence of multiple paths. On the contrary, in the analysis so far performed using the GA approach, no “greedy” algorithm is used and multiple connecting semipaths might emerge. From these different algorithms, different considerations about the time structure emerge. If, on the one hand, in both cases we have the representation of technology evolution, on the other hand, in the GA, the time dimension is less constraining, but is still present in the direction of the arcs and in the numbering of patents.

7 Conclusion

The availability of patents and citations data has increased their use as innovation indicators. Recently, we observe the emergence of a complementary approach to patent and citation counts, which is to consider them from a network perspective. In fact, it is possible to exploit citations to map the technological relation between inventions (i.e. patents). In this setting, a patent citation network represents the space of the “technologically possible solutions”, the structure and dynamics of which characterize technical change in a specific technology.

The aim of this paper is to propose a new empirical method for identifying technologically important patents within a patent citation network and to apply it to the telecommunication switching industry.

The method proposed is inspired by population genetics: as geneticists are interested in studying patterns of migration and therefore the common origins of people, in innovation studies we are interested in tracing the origin and the evolution of today knowledge. In this respect, the genetic parallel is rather clear: as patent A cites patent B, we can say that patent A inherits some knowledge from patent B. At the heart of this genetic method for analyzing patent and citation data is the persistence index that measures how much knowledge from older patents is retained (and therefore persists) into the recent patents. Accordingly, to the genetic parallel, the persistence index is calculated decomposing patent’s knowledge applying the Mendelian law of genes inheritance. In this framework, the novelty of a patent (i.e. its inventive step) derives from the recombination of its “inherited” knowledge.

The empirical exercise of this paper consist in the use of this persistence index for identifying important patents and citations within the network. In particular, the analysis of the structure and evolution of the persistence weighted network allows us to unfold specific patterns of technological change in the case of the telecommunication switching industry.

If a network approach on patent data is not new, this paper suggests an alternative method overcoming some limitations in the existing approaches. In particular, the genetic approach is a flexible method that places very little assumptions on the possible outcomes of the emerging patterns of technical change. First of all, this approach is designed to account for differences in technology evolution, meaning that new technologies characterized by different levels of cumulativeness or radicalness can display different network structures. In this respect, this method allows for testing hypothesis that are not confined to structural network properties but that are rooted in innovation studies. Second, the genetic approach does not make any assumptions about optimality or efficiency structure in the pattern of network evolution. Again, this means that it accommodates for differences in patterns of technical change. Finally, as regards the Hummond and Doreian approach, the method here presented has a different rationale with less emphasis on direct links and uniqueness of the emerging pattern of technical change.

Summarizing, we can conclude that the method proposed is successful in reducing the number of both nodes and links considered. This reduction might look arbitrary as it is based on cut-off points. However, the persistence index displays a highly left-skewed distributions that mimics a scale-free distribution. This makes easier to justify cut off choices and confirm the idea that few patents are “important” both in terms of economic value and in terms of knowledge contribution.

Furthermore, our method is indeed successful in identifying technological discontinuities where previous knowledge is not relevant for current technological development. Indeed, we can show that our methodology can be successfully used for identifying different technological paradigms are defined by Dosi (1982).

As regarding the broader impact of this new methodology, we can expect two lines of generalization. The first obvious one is the application of the persistence index to other technological fields in order to characterize their cumulative and disruptive dynamics. The scientific importance of such exercise is to foster quantitative research and appreciative theorizing defined as a rigorous storytelling (Nelson 1989) in studies related to technology dynamics and its effect on industry dynamics.

The second one concerns the relevance of the methodology to other types of networks. The GA method could be used to analyze any graph representation of dynamic evolutionary processes (that follows the arrow of calendar time) where entities evolve into different/other entities by mutation and/or recombination while retaining various properties of their predecessors. In fact, all these processes can, in principle, be represented as a directed and acyclical graph on which it is possible to apply our method. Nature provides other examples of networks with these characteristics, such as: family trees, phylogenetic networks, food webs, feed-forward neural networks, and software call (Karrer and Newman 2009). Any such graph is a system that is potentially analyzable by our GA, which is essentially a fractional system of compound inheritence accounting. Related to the field of innovation studies, publication citation networks are another acyclical network “commonly” studied; the application and the interpretation of the persistence index in such context would be rather straightforward. In particular, our indicator could be used as a scientometrics tool to evaluate the long-run impact of individual papers and indirectly their authors/affiliated organizations. In historical perspective, the persistence approach differentiates between fashion of the day and long lasting influence, also refereed in network analysis as the difference between popularity and prestige. Finally, an other interesting application related to innovation could be for studying hierarchical organizations of learning (i.e., successions of master(s) and apprentices or supervisors and PhD. students) to find out the most (i.e., persistently) influential/prolific individual or different lineages of styles/paradigm.

Looking at the limitations of this method, we can identify two main caveats. If, on the one hand, this method can potentially be applied for comparative studies about technology dynamics in different technology, on the other hand, this is only possible for sectors in which patents are used for innovation appropriability. Limitations on the methodology itself relate to the (necessary) assumptions of the use of the Mendelian laws for the operationalization of the persistence indexes. However, in our connotation of the links between patents, the network under examination represents a system of both knowledge creation and retention. Therefore, it sounds plausible to assume that the innovative step in each patent is in the recombination of existing knowledge (i.e. inherited by the previous patents).

Appendices

Appendix A: Distribution of persistence index

ᅟ

Histogram of the frequency distribution of the persistence index for the first 6 truncations

ᅟ

Histogram of the frequency distribution of the persistence index for the second 6 truncations

ᅟ

Histogram of the frequency distribution of the persistence index for the last 3 truncations

Appendix B: List of persistent patents

Table 7 ᅟ