Patent Landscaping is a valuable instrument for many stakeholders, such as patent examiners, company decision-makers, researchers, and policymakers. They use this method to analyze the state-of-the-art, compare organizations' patenting activities, assess entire industries, or identify gaps in internal R&D activities. However, analyzing vast amounts of patent documents and aggregating and visualizing information is cumbersome and complex. The paper presents an innovative approach to automated patent landscaping by combining natural language processing models with approximate nearest neighbor search, dimensionality reduction, and clustering methods. This entire approach only uses the textual content of the underlying patents and does not use any additional meta-data, such as technology classes or citations.
tecting potential infringements and assessing the validity The patent landscaping process involves analyzing and of existing patents by examining their legal status. This visualizing a collection of patents within a specific tech- provides stakeholders with a clear view of the competnological field, industry, or organization. The objective is itive landscape and potential barriers to entry for new to gain insights into the intellectual property landscape. innovations. For businesses and investors, this informaDepending on the unit of analysis, the results can be tion is critical for designing around existing technologies, utilized for various purposes. identifying licensing and merger and acquisition targets,
The process is used by diverse stakeholders, includ- and reducing legal risks associated with intellectual
proping researchers, patent examiners, decision-makers, R&D erty.
managers, investors, and policymakers [
One of the primary uses of patent landscaping is to as- to map scientific and technological trends [
apes/ (retrieved 22.04.2024)
processes and establishing a comprehensive and well- sures that resources are used more eficiently.
Moreinformed IP strategy [
Furthermore, patent landscaping aids in exploring com- opportunities through licensing. By identifying potenpetitor activity and enhancing competition and market tial licensing opportunities, companies can monetize intelligence. It provides insights into where competitors their patents beyond direct product sales, leveraging invest resources and predicts the next products they will their intellectual property for broader financial gains. likely release. Of course, they are also constantly mon- Additionally, patent landscaping provides researchers, itoring their competitors for possible infringements of policy-makers, companies, and investors with a better their patents. They also look out for potential targets for and clearer overview of the patent environment for speM&A activities [7]. cific technologies. This comprehensive understanding
By presenting complex patent data visually, stakehold- allows for strategic decision-making, ofering insights ers can better comprehend and derive actionable insights into product diferentiation and market positioning based based on empirical evidence. Visual representations such on intellectual property analysis. as charts, graphs, heat maps, network diagrams, and other visualizations are crucial in identifying patterns, relationships, and clusters within the patent dataset. These 2. Related Work tools make it easier for decision-makers to grasp the intricacies of the intellectual property landscape at a glance.2 There are simple of-the-shelf metadata visualization plat
Interpreting these visual findings involves drawing forms like PatentsView that are used for patent landconclusions about potential opportunities, gaps, and scaping. It was initiated in 2012 by the U.S. Patent and threats within the IP landscape. This analysis is vital Trademark Ofice and is a comprehensive platform for vifor guiding decision-making, strategic planning, and fos- sualizing, disseminating, and analyzing intellectual proptering innovation eforts. Through efective visualization, erty data. It is tailored to support a diverse user base, patent landscaping not only aids in understanding but including researchers, policymakers, and small business also significantly improves communication among stake- owners. It ofers several essential tools for patent landholders. This facilitates a more informed and collabora- scaping: visual analytics for exploring patent data, an tive approach to managing and capitalizing on intellec- API tool for data integration and advanced querying, and tual property. a data query builder for creating customized datasets.3
While patent landscaping is an invaluable tool, it is Trippe (2015), listed a vast array of analysis tool
essential to acknowledge that it is also a complex and providers. Many of these tools use the technologies
detime-consuming process [
Despite its complexities, the benetfis of making better- Lexical databases, such as WordNet, have also played informed decisions through patent landscaping are sig- a role in mining patents for key technological concepts, nificant and multifaceted. Firstly, eliminating redun- enhancing the extraction of meaningful data [9, 23]. dant research eforts helps reduce the costs associated More recently, models that focus on semantic simiwith research and development. At the same time, it larity leveraging word embeddings have been explored. shortens the time needed for commercialization and en- Skripnikova et al. (2021) applied a pre-trained word2vec
landscapes (retrieved 22.04.2024)
22.04.2024) model in combination with TF-IDF to create patent landscapes, utilizing dimensionality reduction and clustering techniques to analyze document relationships. Similarly, Abood and Feltenberger (2018) introduced an automated landscaping approach using patent metadata and word embeddings.
Erhardt et al. (2022) employed the SPECTER [27] model to generate document embeddings in combination with a comprehensive system for semantic searches. The system is tailored to patents and scientific publications.
Since SPECTER was only trained on scientific publications, Ghosh et al. (2024) further enhanced this field by presenting dedicated patent similarity models.
This paper describes an approach to fuse these advanced methodologies by Erhardt et al. and Ghosh et al., building upon the frameworks established by Abood and Feltenberger and Skripnikova et al. to refine automated patent landscaping techniques.
36–50 This section introduces a new automated method for creating patent landscapes. This method significantly reduces many previously discussed constraints by utilizing human insights, semantic similarity, approximate nearest neighbor search, dimensionality reduction, and clustering. The most important aspect of this approach is that it only relies on text. No additional metadata is needed. and patent claims. PATSTAT 4, ofered by the European Patent Ofice, has established itself as a leading source of patent intelligence and statistics. It allows for advanced statistical analysis of patent data, including bibliographiLogic Mill This approach relies heavily on the capa- cal and legal events. This approach was used as a source bilities of Logic Mill [26], a software system wrapped to obtain the title and abstract. Furthermore, additional around an extensive vector database. It is designed to metadata is used as well. Another benefit of PATSTAT is ifnd semantic similar documents across single or multi- a simple aggregation of organization names. In addition ple domain-specific corpora. It employs state-of-the-art to the default metadata, Patstat provides aggregated and Natural Language Processing (NLP) techniques to create harmonized organization names linked by ownership renumerical representations of documents, utilizing a vast lations to the patent documents. That is especially helpful pre-trained language model for this purpose. The system in pinpointing patent documents with company names specializes in analyzing scientific and patent documents that are written in diferent styles or missing company and includes a database of over 200 million documents. designations.
Users can access Logic Mill through an Application Programming Interface (API) or a web interface. Logic Mill 3.2. Seed Documents is regularly updated and can be adapted to include text corpora from various other fields. It is envisioned as a Similar to the method introduced by Abood and Felversatile tool for future research in the social sciences and tenberger (2018), users must select what are known as beyond. The approach uses it to retrieve pre-computed seed patents. Any mistakes made in choosing these seed embeddings and identify the nearest neighbors of refer- patents will afect the entire resulting landscape. Thereence documents. fore, it is crucial to ensure that the patents included in the seed set are relevant to the subject matter. The seed 3.1. Patent Data set should only contain patents that are relevant to the desired landscape and accurately represent the theme.
cially patent text. Patent documents contain multiple text segments, such as a title, abstract, detailed description,
ml
documents of a Section, Class, Subclass, Group or Main Group of the Cooperative Patent Classification (CPC) system and analyze the patents within your unit of analysis.
Uncover diferences and similarities between diferent hierarchical siblings.
Depending on the use case, the seed selection can com- essary encoded documents are stored. According to a prise diferent documents. Based on the use case and the similarity metric (e.g., Cosine Similarity see Equation available metadata, end-users are able to gain insights in 1 or Euclidean Distance / L2 Distance see Equation 2), various contexts. these documents are indexed, and the database is able to provide the approximate nearest neighbors in highHand Picked Technology Assessment For a tech- dimensional space according to these metrics. After all nology assessment, one would start with the essential documents are in the database, users feed the encoded patents and then expand using related patents. seed query documents to the database and retrieve the closest semantically similar documents. They can specify Firm-Level For a patent portfolio analysis of an or- the number of nearest neighbors that should be included ganization’s patent portfolio, one would start with the for each reference document. Otherwise, they also can patent document of the focal organization. Along the specify a cut-of/threshold value. If the similarity score way, competitors of the focal organization can be identi- is below this value, the patent will not be added to the ifed using the approximate nearest neighbor search. landscaping process.
In the case of Logic Mill, these representations are stored in a vector search database called Elastic Search5. It uses the approximate nearest neighbor algorithm HNSW [29] to retrieve semantically similar documents. This approach uses the Logic Mill API to retrieve the closest neighbors based on the IDs of the seed documents.
Scientific Articles This text-based approach ofers several benefits. First, it can bridge the gap between scientific publications and patents by using scientific papers as starting points. Additionally, this method can be applied to scholarly articles to create a comprehensive research landscape. Combining scientific and patent documents could also generate a useful tool for prior art exploration.
patent documents generated by SPECTER / Pat-SPECTER is a vector of 768 dimensions or 1024 by PatSPECTER, the dimensions have to be reduced since they are impossible to visualize and interpret for humans. This can be done through a process called dimensionality reduction.
There are various methods, such as the linear variants Principal Component Analysis (PCA) [30] or t-Distributed Stochastic Neighbor Embedding (t-SNE) [31], as well as 3.3. Encoding the non-linear variants such as Uniform Manifold ApproxAll necessary documents need to be encoded before con- imation and Projection (UMAP) [32]. The objective of tinuing. This can be done with any machine-learning these methods is to increase interpretability while at the document encoder model that can encode text into a same time minimizing the loss of information [33, 34]. dense numerical vector. In the case of patents, a rea- For this approach, UMAP was chosen since it ofers sonable candidate would be SPECTER Cohan et al. (2020. multiple advantages, as seen in Table 1. On the one hand, This machine-learning model encodes the title and the ab- it can preserve the data’s global and local structures. In stract of the patent into a 768-dimensional vector. Since addition, it can work with non-linear data, and finally, it the model was trained by leveraging the citation graph is fast. of scientific documents, it has learned the semantic simi- PCA tSNE UMAP larities of related texts. [30] [31] [32]
Other state-of-the-art models, like SPECTER 2 [27], Non-linear data No No Yes PaECTER or Pat-SPECTER of [28]. Local structure Yes Yes Yes
As stated before, we rely on the functionality of Logic Global structure No Yes Yes Mill during this approach and make use of the pre- Speed Very Fast Slow Fast computed embeddings for our data. The initial version of Logic Mill uses SPECTER [27], while the current version Table 1 uses Pat-SPECTER [28]. Overview of dimensionality reduction algorithms.
various similarity measures can be selected to identify
the most similar neighbors in the high-dimensional space. eters that can be adjusted after the user reviews an initial
Cosine Similarity cos(a, b) =
a · b ‖a‖‖b‖ = √︀∑︀ ∑︀
=1 ab =1 (a)2√︀∑︀ =1 (b)2 Euclidean Distance / L2 Distance ⎯⎸ =1 2(a, b) = ‖b − a‖2 = ⎷⎸∑︁(b − a)2
from a 768/1024 dimensional representation into a 2 or
Using the resulting 2/3-dimensional data structure from the previous dimensionality reduction step, a clustering algorithm is applied. The HDBSCAN [35] algorithm was selected for this approach. As seen in Table 2, its parameter minimum amount of cluster members is very intuitive, and the modest speed is negligible. In comparison, K-Means [36] would not be feasible here since it is not clear in advance how many clusters are going to be needed. Since it requires some domain knowledge to form the neighborhood parameter of DBSCAN [37], it was not a feasible option for a generic and automated end-to-end approach. Furthermore, Grootendorst (2022) showed promising results by combining HDBSCAN [35] with UMAP [32]. It also generates better results if the clustering is done after the dimensionality reduction [38].
Type Speed Param.
K-Means [36] Centroid Based Very Fast Number of Clusters
DBSCAN [37] Density Based Modest
Neighborhood the patent documents, the approach uses the c-TF-IDF [38] algorithm to generate cluster names. This is also in line with Grootendorst (2022). Here, the traditional document-level textitTF-IDF is transformed to function (2) efectively on a cluster-specific basis, which is essential for distinguishing the unique characteristics of each cluster. This approach, known as c-TF-IDF, modifies the standard method to reflect the diferences between clusters better.
aggregating the frequencies of each word within that cluster. This step forms the basis for a class-specific term frequency. An L1 normalization is applied to normalize these frequencies, ensuring that variations in cluster sizes do not skew the data.
puted by taking the logarithm of the quotient of the average word count per cluster and the occurrence of each word across all clusters, incremented by one to maintain non-negative values. This calculation yields the modified
c-TF-IDF
W, = ‖tf ,‖ × log 1 + (3) ︂(
A )︂ f where tf , is the frequency of word in class , f is the frequency of word across all classes, and A is the average number of words per class.
frequency and the adjusted inverse document frequency—are multiplied to derive a significance score for each word within a cluster. This method deviates from the conventional TF-IDF to provide a more tailored and efective topic representation [38].
common words per cluster and generates a fictional cluster name based on the, e.g., top 5 most common words.
(mRNA) vaccines have attracted considerable attention. BioNTech and Moderna vaccines were the first mRNA vaccines to be approved by a drug regulatory agency worldwide. Intensive research had been conducted on mRNA technology for many years prior to the development of the COVID-19 vaccines7.
The underlying technology used to develop such vaccines can be protected by patents. In an attempt to demonstrate the complexity involved in IP protections and licensing deals surrounding COVID-19 vaccine technology, Gaviria and Kilic (2021) developed a preliminary patent network analysis. Li et al. (2022) characterized the patent
7See https://www.pei.de/DE/newsroom/hp-meldungen/2022/22 0221-covid-19-pandemie-impfstofe-im-fokus.html?nn=169730 (retrieved 22.04.2024)
using the Plotly6 library. This library can be used by Python, R, and JavaScript. It allows for interactive, web- Objective This use case tries to generate a patent landbased visualizations, including line plots, scatter plots, scape analysis of mRNA patenting organizations. These bar charts, and more. Plotly uses a declarative syntax, findings are then compare to the studies of Gaviria and which makes it easy to create and customize 2D/3D plots. Kilic (2021) and Li et al. (2022).
It also supports a wide range of chart types and customization options. Additionally, Plotly allows for the creation Data BioNTech, CureVac, and Moderna have all deof interactive dashboards, which organize and present veloped mRNA-based vaccine candidates for COVID-19. multiple plots in one place. Furthermore, the plots can Gaviria and Kilic (2021) also display Acuitas and Arbube embedded into web pages and Jupyter notebooks. It tus. The identification of the companies was made using is an open-source library that can be used in commercial the han_name column in the tls206_person table of and non-commercial applications [39]. Patstat. For the search, the term BIONTECH, CUREVAC,
A 3D/2D scatterplot is used to visualize the landscapes. MODERNATX, ACUITAS, ARBUTUS were used. Only Here, each dot is a patent document. The x and y (or patents from the EPO were considered. No attention was z) coordinates result from the dimensionality reduction given to whether the patents were still active. All patent process. documents (A1, A2, B1, etc.) were used. These initial
For the visualization of clusters, the so-called convex documents represent the so-called seed documents in hull was computed. The convex hull is commonly used this approach. in computational geometry and computer graphics to create a shape that encloses a group of points. In other words, the smallest "convex" shape can be drawn around a set of points such that all the points are on or inside the shape [40].
This approach uses the Virtanen et al. (2020) implementation of the Quickhull algorithm [40]. The input to the algorithm is the set of points of each cluster. The output is a polygon of the outermost points.
were retrieved from Patstat by using the organization search terms. The initial step is to obtain the numerical representation of these documents generated by SPECTER [27]. These were retrieved using the API of the Logic Mill system [26]. The embeddings were then used for dimensionality reduction and clustering. Finally, the convex hull of the cluster was calculated. Cluster Names were generated using the C-TF-IDF algorithm. An interactive visualization can be seen in Figure 2. These visualizations make use of 3 dimensions, and the clusters are indicated as transparent hulls around the dots. Each dot represents a patent document. The color indicated the respective companies. Users can now interactively explore the patent landscape based on the patents of these organizations.
landscape of mRNA vaccines and analyzed 1852 patent families.
lected companies were retrieved to generate the patent landscape of the mRNA industry. The already encoded patent documents of the initial organizations (BIONTECH, CUREVAC, MODERNATX, ACUITAS, ARBUTUs) were used as seed documents. The embeddings of the documents were used in combination with the Logic Mill API to retrieve the closest neighbors. For each reference document, the closest ten documents were retrieved. The result set was then used to obtain the organizations. Finally, the results are aggregated and counted based on the organization’s name. The top 50 results can be found in Appendix 3. The han_names were not resolved to an organization level. There are multiple legal entities that belong to the same organization. This can be seen in the histogram in Figure 3. resented in the patent landscape.
On the other hand, important mRNA organizations were missing. Sanofi, Enanta Pharmaceuticals, Evelo Biosciences, the United States Department of Health and Human Services, and the Chinese Academy of Agricultural Sciences were not present. This could be due to the fact that only patent documents of the EPO were analyzed. The analysis did not include the USPTO, WIPO, and the Chinese Patent Ofice.
of the approach to automatically generate patent landscapes. Only initial reference organizations were selected. All other competitors have been identified automatically. This was done by only relying on the encoded patent text. With the help of an approximate nearest neighbor search in a vector search database.
The latest technological advancements have shed light
on the capabilities of quantum technology across
various fields. Among them are simulation, computation,
and communication. Quantum computers use
superconducting qubit-based programmable processors. They can
compute tasks in minutes, whereas state-of-the-art
classical supercomputers would take approximately tens of
thousands of years [
In their report, Aboy et al. (2022) produce a patent landscape of quantum technologies over the last 20 years. They evaluate patenting and strategies, key owners, dominant portfolios, and geographic distribution of patent activity, among other factors.
with the work of Gaviria and Kilic 2021, overlaps can be identified. UPenn (University of Pennsylvania) and UBC (University of British Columbia) could be identified automatically.
Comparing the results with Li et al. 2022, additional Analysis The initial response of Patstat was 23 patent overlaps could be found. Here, Merck & Co., Glaxo- documents. The numerical representation was obtained SmithKline, Boehringer, Roche, University of California, and TRON Translational Oncology Mainz were also repusing the Logic Mill System [26]. The approximate near- Interpretation It was shown that the approach is caest neighbors of the documents were retrieved using an- pable of identifying competitors for a patent landscape. other API endpoint. The documents were then resolved This was done only by using raw patent text, a docuback to the owner. Afterward, they are aggregated and ment encoder, and a search database. The approximated counted. nearest neighbor search returned competitors of Rigetti.
The top 20 results can be seen in Figure 4. The top 50 These could be matched with the organizations presented results can be found in Appendix 4. The results of Aboy by Aboy et al. 2022 in their patent landscape analysis. et al. (2022) can be found in the Appendix 4.
The overlap between the result generated by the approach and Aboy et al. (2022) is large. IBM, D- 5. Conclusions Wave, Northrop Grumman, Toshiba, MIT, Intel, Alphabet (Google), Honeywell, HP, Hitachi, Samsung, etc., were all in the top 50 results.
Furthermore, Lockheed Martin CORP (rank 80) QUBITEKK (rank 88) New South Innovations (rank 66) Raytheon Technologies (rank 52) were still within the top 100.
Only Bank of America, Seagate, and Michigan State University could not be found.
A reason could be that Patstat does not aggregate these
documents correctly. Companies and their subsidiaries
might not be aggregated. Furthermore, [
The patent landscaping method outlined in this paper provides an automated and scalable solution, significantly reducing the need for human intervention. This approach has various benefits for researchers, companies, policymakers, and investors who rely on patent landscape analyses but are often deterred by their associated costs and eforts. The approach is notably fast, adaptable, capable of integrating new advancements and accommodating future research.
However, the approach has clear limitations. As it is based on seed documents, a careful selection is necessary. It is also challenging to determine in advance the right amount of neighbors or a threshold value for the approximate nearest neighbors of the seed documents.
Additionally, the settings for dimensionality reduction and clustering might also depend on the use case and the number of documents.
Furthermore, this study is also limited in the evaluation of the approach. It is mostly episodic and limited in scope since the comparison relies only on EP data. A comparison with a gold-standard dataset is needed to compare the results of the automated approach with a human-generated one. It also lacks a comprehensive user study of the interface and visualization components.
Future research could explore how the latest generative AI models might be able to extend this approach further.
With this technology, cluster naming can be improved greatly. It might also be possible to start the approach without any seed documents and start with a prompt.
Please create a patent landscape of mRNA vaccines against COVID-19. [4] M. B. Heeley, S. F. Matusik, N. Jain, Innovation, g/stable/4134022, publisher: INFORMS.
Appropriability, And The Underpricing Of Initial [14] N. Bloom, M. Schankerman, J. Van Reenen, IdenPublic Oferings, Academy of Management Journal tifying Technology Spillovers and Product Market 50 (2007) 209–225. URL: https://journals.aom.o Rivalry, Econometrica 81 (2013) 1347–1393. URL: rg/doi/10.5465/amj.2007.24162388. doi:10.546 https://www.jstor.org/stable/23524180, publisher: 5/amj.2007.24162388, publisher: Academy of [Wiley, Econometric Society].
Management. [15] P. Thompson, M. Fox-Kean, Patent Citations and
[
Policy 30 (2001) 143–157. URL: https://www.scienc [18] S. Arts, N. Melluso, R. Veugelers, Beyond Citations: edirect.com/science/article/pii/S004873339900098 Text-Based Metrics for Assessing Novelty and its 0. doi:10.1016/S0048-7333(99)00098-0. Impact in Scientific Publications, 2023. URL: http: [8] M.-J. Shih, D.-R. Liu, M.-L. Hsu, Discovering com- //arxiv.org/abs/2309.16437. doi:10.48550/arXiv petitive intelligence by mining changes in patent .2309.16437, arXiv:2309.16437 [econ, q-fin]. trends, Expert Systems with Applications 37 (2010) [19] S. Lee, B. Yoon, Y. Park, An approach to dis2882–2890. URL: https://www.sciencedirect. covering new technology opportunities: Keywordcom/science/article/pii/S0957417409007829. based patent map approach, Technovation 29 doi:10.1016/j.eswa.2009.09.001. (2009) 481–497. URL: https://www.sciencedirec [9] H. Park, J. Yoon, K. Kim, Identifying patent in- t.com/science/article/pii/S0166497208001326. fringement using SAO based semantic technologi- doi:10.1016/j.technovation.2008.10.006. cal similarities, Scientometrics 90 (2012) 515–529. [20] I. Bergmann, D. Butzke, L. Walter, J. P. Fuerste, URL: https://doi.org/10.1007/s11192-011-0522-7. M. G. Moehrle, V. Erdmann, Evaluating the Risk of doi:10.1007/s11192-011-0522-7. Patent Infringement by Means of Semantic Patent [10] A. Abood, D. Feltenberger, Automated patent land- Analysis: The Case of DNA Chips, 2008. URL: scaping, Artificial Intelligence and Law 26 (2018) https://papers.ssrn.com/abstract=1288313. 103–125. URL: https://doi.org/10.1007/s10506-018 doi:10.1111/j.1467-9310.2008.00533.x. -9222-4. doi:10.1007/s10506-018-9222-4. [21] J. M. Gerken, M. G. Moehrle, A new instrument [11] T. Skripnikova, H. Aras, A. Weißhaar, S. Blank, H.-P. for technology monitoring: novelty in patents meaZorn, Semantic views — Interactive hierarchical sured by semantic patent analysis, Scientometrics exploration for patent landscaping, World Patent 91 (2012) 645–670. URL: https://doi.org/10.1007/s1 Information 65 (2021) 102043. URL: https://www.sc 1192-012-0635-7. doi:10.1007/s11192-012-0 iencedirect.com/science/article/pii/S01722190210 635-7.
00247. doi:10.1016/j.wpi.2021.102043. [22] S. Choi, H. Park, D. Kang, J. Y. Lee, K. Kim, An [12] A. B. Jafe, Characterizing the “technological po- SAO-based text mining approach to building a techsition” of firms, with application to quantifying nology tree for technology planning, Expert Systechnological opportunity and research spillovers, tems with Applications 39 (2012) 11443–11455. URL: Research Policy 18 (1989) 87–97. URL: https://www. https://www.sciencedirect.com/science/article/pii/ sciencedirect.com/science/article/pii/0048733389 S0957417412006112. doi:10.1016/j.eswa.2012. 900073. doi:10.1016/0048-7333(89)90007-3. 04.014. [13] L. Rosenkopf, P. Almeida, Overcoming Local Search [23] H. Park, J. J. Ree, K. Kim, Identification of promising through Alliances and Mobility, Management Sci- patents for technology transfers using TRIZ evoence 49 (2003) 751–766. URL: https://www.jstor.or lution trends, Expert Systems with Applications 40 (2013) 736–743. URL: https://www.sciencedir 20150202. URL: https://royalsocietypublishing.org ect.com/science/article/pii/S0957417412009694. /doi/10.1098/rsta.2015.0202. doi:10.1098/rsta.2 doi:10.1016/j.eswa.2012.08.008. 015.0202, publisher: Royal Society. [24] J. Yoon, H. Park, K. Kim, Identifying technolog- [34] R. Xiang, W. Wang, L. Yang, S. Wang, C. Xu, X. Chen, ical competition trends for R&D planning using A Comparison for Dimensionality Reduction Methdynamic patent maps: SAO-based content analy- ods of Single-Cell RNA-seq Data, Frontiers in Gesis, Scientometrics 94 (2013) 313–331. URL: https: netics 12 (2021). URL: https://www.frontiersin.org/ //doi.org/10.1007/s11192-012-0830-6. doi:10.100 articles/10.3389/fgene.2021.646936. 7/s11192-012-0830-6. [35] L. McInnes, J. Healy, S. Astels, hdbscan: Hi[25] H. Park, K. Kim, S. Choi, J. Yoon, A patent in- erarchical density based clustering, Journal of telligence system for strategic technology plan- Open Source Software 2 (2017) 205. URL: https: ning, Expert Systems with Applications 40 (2013) //joss.theoj.org/papers/10.21105/joss.00205. 2373–2390. URL: https://www.sciencedirect. doi:10.21105/joss.00205. com/science/article/pii/S0957417412012092. [36] J. MacQueen, Some methods for classification and doi:10.1016/j.eswa.2012.10.073. analysis of multivariate observations, Proceedings [26] S. Erhardt, M. Ghosh, E. Buunk, M. E. Rose, of the Fifth Berkeley Symposium on Mathematical D. Harhof, Logic Mill – A Knowledge Naviga- Statistics and Probability, Volume 1: Statistics 5.1 tion System, 2022. URL: http://arxiv.org/abs/23 (1967) 281–298. URL: https://projecteuclid.org/eboo 01.00200. doi:10.48550/arXiv.2301.00200, ks/berkeley-symposium-on-mathematical-statist arXiv:2301.00200 [cs]. ics-and-probability/Proceedings-of-the-Fifth-Ber [27] A. Cohan, S. Feldman, I. Beltagy, D. Downey, keley-Symposium-on-Mathematical-Statistics-a D. Weld, SPECTER: Document-level Representa- nd/chapter/Some-methods-for-classification-and tion Learning using Citation-informed Transform- -analysis-of-multivariate-observations/bsmsp/12 ers, in: Proceedings of the 58th Annual Meeting 00512992, publisher: University of California Press. of the Association for Computational Linguistics, [37] M. Ester, H.-P. Kriegel, J. Sander, X. Xu, A densityAssociation for Computational Linguistics, Online, based algorithm for discovering clusters in large 2020, pp. 2270–2282. URL: https://aclanthology.org spatial databases with noise, in: Proceedings of the /2020.acl-main.207. doi:10.18653/v1/2020.acl Second International Conference on Knowledge -main.207. Discovery and Data Mining, KDD’96, AAAI Press, [28] M. Ghosh, S. Erhardt, M. E. Rose, E. Buunk, Portland, Oregon, 1996, pp. 226–231.
D. Harhof, PaECTER: Patent-level Representation [38] M. Grootendorst, BERTopic: Neural topic modeling Learning using Citation-informed Transformers, with a class-based TF-IDF procedure, 2022. URL: 2024. URL: http://arxiv.org/abs/2402.19411. doi:10 http://arxiv.org/abs/2203.05794. doi:10.48550/a .48550/arXiv.2402.19411, arXiv:2402.19411 rXiv.2203.05794, arXiv:2203.05794 [cs]. [cs]. [39] I. Plotly Technologies, Collaborative data science, [29] Y. A. Malkov, D. A. Yashunin, Eficient and robust 2015. URL: https://plot.ly, place: Montreal, QC approximate nearest neighbor search using Hier- Publisher: Plotly Technologies Inc. archical Navigable Small World graphs, 2018. URL: [40] C. B. Barber, D. P. Dobkin, H. Huhdanpaa, The http://arxiv.org/abs/1603.09320. doi:10.48550/a quickhull algorithm for convex hulls, ACM TransrXiv.1603.09320, arXiv:1603.09320 [cs]. actions on Mathematical Software 22 (1996) 469– [30] I. T. Jollife, Principal Component Analysis, 483. URL: https://doi.org/10.1145/235815.235821.
Springer Series in Statistics, Springer-Verlag, New doi:10.1145/235815.235821.
York, 2002. URL: http://link.springer.com/10.1007/ [41] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberb98835. doi:10.1007/b98835. land, T. Reddy, D. Cournapeau, E. Burovski, P. Pe[31] L. van der Maaten, G. Hinton, Visualizing data terson, W. Weckesser, J. Bright, S. J. van der Walt, using t-SNE, Journal of Machine Learning Research M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. 9 (2008) 2579–2605. URL: http://jmlr.org/papers/v9 Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, I. Po/vandermaaten08a.html. lat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, [32] L. McInnes, J. Healy, UMAP: Uniform Manifold J. Perktold, R. Cimrman, I. Henriksen, E. A. QuinApproximation and Projection for Dimension Re- tero, C. R. Harris, A. M. Archibald, A. H. Ribeiro, duction, ArXiv (2018). F. Pedregosa, P. van Mulbregt, SciPy 1.0: fundamen[33] I. T. Jollife, J. Cadima, Principal component analy- tal algorithms for scientific computing in Python, sis: a review and recent developments, Philosophi- Nature Methods 17 (2020) 261–272. URL: https: cal Transactions of the Royal Society A: Mathemat- //www.nature.com/articles/s41592-019-0686-2. ical, Physical and Engineering Sciences 374 (2016) doi:10.1038/s41592-019-0686-2, number: 3 36–50
LUCENT TECH INC HONEYWELL INT INC SAMSUNG ELECT CO LTD MOTOROLA INC THALES 6 6 6 6 6 6
Organization ZYOMED HOLDINGS INC.
EVOQ NANO INC HONDA MOTOR CO
UNIV ILLINOIS
LIGHTMATTER INC
Matches: IBM, NORTHROP GRUMMAN, D WAVE SYSTEMS, MICROSOFT, INTEL, RIGETTI and CO, HITACHI, HONEYWELL INT, SAMSUNG ELECTRONICS, YALE UNIVERSITY, NEC, NOKIA, MITSUBISHI ELECTRIC.
Number of overlapping companies: 13