Lafond and Kim, 2019

François Lafond and Daniel Kim. "Long-run dynamics of the U.S. patent classification system." Journal of Evolutionary Economics 2019. Full text online. 2018 version; SSRN version

Useful background for study of US patent classification system. Describes constantly changing nature of the USPC system. Section 2, "Why is studying classification systems important?" describes some recent research in this area and notes that not enough has been done to trace the historical evolution of patent classifications. Such research would of course provide more insight into the role of classifications in relation to contemporary technological development.

Classification systems may create a feedback on the system it describes, for instance by legitimizing the items that it classifies or more simply by biasing which items are found through search and reused in recombination to create other items.

Section 3 gives a history of the US classification system, which, aside from an abortive effort in 1790–1793, began in 1836. The number of classes increased from 14 to 36 by 1868. A new 'parallel' system was then established with 176 classes.

For the bulk of the paper, the authors quantitatively analyze the patent classification system in relation to number of patents and other data. They look at retroactive reclassification as an indication of technological innovation.

[A] patent related to a radical innovation is originally hard to classify. It is likely that it will have to be reclassified in the future, when a more appropriate set of concepts has been developed and institutionalized (that is, when the community of technologists have codified a novel understanding about the radical innovation).

In the conclusion, they contrast the fast-changing development of the American patent system to the relatively static French system at the turn of [which?] century.

• They say there are about 10 million patents in the US system, presumably counting all of history together.
• The first US classification was created in 1829-30
• "we find that historically the growth of the number of distinct classes has been more or less linear, with about two and a half classes added per year"
• "using the historical classes we find that the relationship between the number of classes and the number of patents is compatible with Heaps’ law, a power law scaling of the number of categories with the number of items, originally observed between the number of different words and the total number of words in a text (Heaps 1978)."
• "the size distribution of classes .... is exponential, confirming a result of Carnabuci (2013) on a much more restricted sub-sample. We also find that there is no clear relationship between the size and the age of classes, which rules out an explanation of the exponential distribution in terms of simple stochastic growth models in which classes are created once and for all."
• "we hypothesize that new technology fields and radical innovations tend to be associated with a higher reclassification activity. This suggests that the history of reclassification contains interesting information on the most transformative innovations. Our work here is related to Wang et al. (2016) who study how a range of metrics (claims, references, extensions, etc.) correlate with reclassification for 3 million utility patents since 1994. We used the data since 1976, for which we observe the class of origin and the citations statistics. It appears that reclassified patents are more cited than non-reclassified patents. We also construct a reclassification flow diagram, with aggregation at the level of NBER patent categories (Hall et al. 2001). This reveals that a non-negligible share of patents are reclassified across NBER categories. We find that patents in “Computers” and in “Electronics” are often reclassified in other NBER categories, which is not the case of other categories such as Drugs." We examine three example classes: Fabric, Combinatorial Chemistry and AI.
• using Heaps's law they have a model that predicts reclassification frequency, if I understand correctly
• "patented technologies are classified according to their function, not their industry of use or origin"
• Discusses mappings of patent techs to industries
• "Kutz (2004) studied the growth and distribution of patent classes since 1976; Leydesdorff (2008), Antonelli et al. (2010), Strumsky et al. (2012) and Youn et al. (2015) studied co-classification patterns; and Caminati & Stabile (2010) and Acemoglu et al. (2016) studied the patterns of citations across USPCS or NBER technology classes."
• "classification systems are used to estimate technological distance, typically between firms or inventors in the “technology space” based on the classification of their patent portfolio (Breschi et al. 2003, Nooteboom et al. 2007, Aharonson & Schilling 2016, Alstott et al. 2016). Additional methodological contributions include Benner & Waldfogel (2008), who have pointed out that using all the codes listed on patents increases the sample size and thus reduces bias in measuring proximity, and McNamee (2013) who argues for using the hierarchical structure of the classification system"
• counts of the number of distinct classes by Bailey (1946) and Stafford (1952), which we update here.
• Youn et al. (2015) studied the subclasses (“technology codes”) [and] found that the number of subclasses used up to a given year is proportional to the cumulative number of patents until about 1870, but grew less and less fast afterwards.
• Examining "how professional diversity scales with city size ... Bettencourt et al. (2014) and Youn et al. (2016) exploited the different layers of industry and occupation classifications systems to identify resolution-independent quantities"
• Per "Pavitt (1985) and Hicks (2011), a new technology which fits perfectly in the existing classification scheme may be considered an incremental innovation, as opposed to a radical innovation which challenges existing schemes. " (PBM: relatedly, macroinvention)
• "Patent officers are generally ... highly qualified (often PhDs) . . .. Rotkin et al. (1999) mention that classification work was not particularly attractive and that the Classification division had difficulties attracting volunteers. . . Paradise (2012) [refers] to “high turnover, less than ideal wages and heavy workloads”. There is an emerging literature on patent officers’ biases and incentives (Cockburn et al. 2003, Schuett 2013) but it is focused on the decision to grant the patent. Little is known about biases in classification"
• "Xiang (2005) finds that new goods, as measured by changes to the SIC system, have a higher skill intensity than existing goods."
• "Among the patent classification systems, the USPCS is the oldest still in use (as of couple of years ago) (Wolter 2012)."
• "The U.S. patent system was established on 31st July 1790." It had an examination requirement which was quickly abolished in 1793, so they didn't need to search prior art and didn't badly need a classification system. An examination requirement was reestablished in 1836.
• "The earliest known official subject matter classification appeared in 1823 as an appendix to the Secretary of State’s report to the Congress for that year (Rotkin et al. 1999). It classified 635 patents models in 29 categories such as “Bridges and Locks”, 1184 in a category named “For various purposes”, and omitted those which were not “deemed of sufficient importance to merit preservations”."
• The authors give exact numbers for the classification systems, which I've put in my spreadsheet. Cite this.
• Principles of classification: 19th century classifications were based on the industry or profession using the invention, e.g. Bee culture (449) or Butchering (452).
• "A useful example "(Falasco 2002a, USPTO 2005, Strumsky et al. 2012) is that of cooling devices which were classified separately if they were used to cool different things, such as beer or milk. Today’s system would classify both as cooling devices into the class “Heat exchange” (165), which is the utility or function of the invention. Another revealing example (Schmookler 1966, Griliches 1990) is that a subclass dealing with the dispensing of liquids contains both a patent for a water pistol and one for a holy water dispenser. This change in the fundamental principles of classification took place at the turn of the century with the establishment of the Classification division (Falasco 2002a, Rotkin et al. 1999)."
• "The division [redesigned] the classification system so that inventions would be classified according their utility. The fundamental principle which emerged is that of “utility classification by proximate function” (Falasco 2002a) where the emphasis on “proximate” means that it is the fundamental function of the invention, not some example application in a particular device or industry. For instance “Agitating” (366) is the relevant class for inventions which perform agitation, whether this is to wash clothes, churn butter, or mix paint (Simmons 2014). Another classification by utility is the classification by effect or product, where the result may be tangible (e.g. Semiconductors device and manufacture, 438) or intangible (e.g. Audio signal system, 381). Finally, the classification by structure (“arrangement of components”) is sometimes used for simple subject matter having general function. This rationale is the most often used for chemical compounds and stock material. It is rarely used for classes and more often used at the subclass level"
• three classification rationales (by industry, utility and structure) coexist. Each class “reflects the theories of classification that existed at the time it was reclassified” (USPTO 2005). In addition, the system keeps evolving as classes (and even more so subclasses) are created, merged and split. New categories emerge when the need is felt by an examiner and approved by the appropriate Technology Center; in this case the USPCS is revised through a “Classification order” and all patents that need to are reclassified (Strumsky et al. 2012). An example of how subclasses are created is through alpha subclasses. Alpha subclasses were originally informal collections created by patent examiners themselves to help their work, but were later incorporated into the USPC. They are now created and used as temporary subclasses until they become formalized (Falasco 2002b, USPTO 2005). When a classification project is completed, a classification order is issued, summarising the changes officially." Past patents may then be reclassified.
• an 1868 source said: "division became a necessity to secure a proper apportionment of work among the corps of examiners."
• "our aim is to build a database13 of 1) the evolution of the USPCS primary classes, and 2) the reclassification of patents from one class to the other. . . we relied on several sources. First, our most original data collection effort concerns the historical number of classes. For the early years our main sources are Bailey (1946) and Rotkin et al. (1999), complemented by Reingold (1960) and the “Manual of Classification” for the 5 years within the period 1908–1923."
• Lafond and Kim's data is available at https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/ZJCDCE !!
  • This table shows the first year various categories appear: https://dataverse.harvard.edu/file.xhtml?persistentId=doi:10.7910/DVN/ZJCDCE/JEEUCS&version=1.1. However 244 shows up as appearing first in 1936, when we think it appeared in 1912.
• [1] <== this is a table of how many classes in a RECENT USPC would be needed to classify the patents in each year.
• "In many systems, it has been found that the number of categories grows as a power law of the number of items that they classify, a result known as Heaps’ law (for an example based on a classification system –the medical subject headings– instead of a language, see Petersen et al. (2016)). Here we find that using the 2015 classification, Heaps’ law is clearly violated.Footnote23 Using the historical data, Heaps’ law appears as a reasonable approximation. We estimate the Heaps exponent to be 0.378 with standard error of 0.010 and R2 = 0.95. "

https://heinonline.org/HOL/Page?collection=journals&handle=hein.journals/jpatos28&id=497&men_tab=srchresults

stopped around Fig 1 from these notes in order to shut down and restart.

Original title	Long-run dynamics of the U.S. patent classification system
Simple title	Long-run dynamics of the U.S. patent classification system
Authors	François Lafond, Daniel Kim
Date	2019-01
Countries	US
Languages	en
Keywords	patents, history, United States Patent Office, USPC
Journal	Journal of Evolutionary Economics
Related to aircraft?	0
Page count	34
Word count
Wikidata id