Skip to main content


Zuse Analytics 

Zuse Analytics leverages the power of artificial intelligence, predictive analytics, and sophisticated data management techniques to help answer the most fundamental questions about a patent.  

Zuse Analytics traces its history to early pioneering work of Prof. George Karypis, the Distinguished McKinght  University Professor in the Department of Computer Science and Engineering at the University of Minnesota.  

In 2007, Professor Karypis, Dan Bork, Erich Spangenberg, and a team of data and computer scientists turned their attention to building the best possible patent analytic tool. It helped generate hundreds of millions of dollars in hundreds of patent licensing, acquisition, and disposition transactions.

After spending 10 years and over $20 million on developing the system,  IPwe acquired the exclusive rights to what we believe is the world's premier patent analytic tool.  

Most users will be able to quickly find answers to fundamental questions like whether a patent exists and what it covers. Patent offices and patent examiners can use additional features that aide validity and prior art searches.   

While we're very proud of Zuse Analytics and think it should become the patent analytics tool of choice, IPwe recognizes there are other patent analytic tools that users may prefer. We happily will integrate free and pay versions of other patent analytic tools into the IPwe Platform.  

Zuse Analytics scores patent attributes like quality and validity, based on years of experience. 

We're continuing to invest heavily in Zuse Analytics and have a product roadmap that we'll update regularly.  For now, Zuse Analytics functionality is limited to basic search. We are continually adding increasingly sophisticated functions to Zuse. Features to be released in 2018 include analysis of entire portfolios and downloading results in different formats.  

If you want more technical information on Zuse Analytics, you can get it here

If you want additional information on how to use Zuse Analytics or have questions on how to use the system, you can contact us here.

If you want to contact us about Zuse Analytics with some other idea (including any open source ideas), we would love to hear from you and you can reach us here.

Zuse Analytics


A Brief Overview of How Zuse Analytics Evaluates Prior Art

Zuse Analytics identifies relevant prior art by taking into account the limitations of the claim under consideration (query claim of query patent), the text of the art, the link structure of the citation network, and the patent classification. The overall process proceeds generally as follows:

Zuse Analytics constructs a network that consists of two types of nodes: (i) the art (patents and non-patent literature) and the (ii) classes of the patent classification (currently IPCR). Each art node is linked to all the art nodes that it cites and is linked to all the classification nodes that it belongs to. The weights of the first set of edges are determined based on the content similarity of the corresponding arts, whereas the weights of the second set of edges are determined based on the classification strength. If the art corresponds to patents and/or patent applications, their known classifications are used (the primary class has higher weight than secondary classes). If the art corresponds to non-patent literature (NPL), their weights are determined by machine-learning classifiers that were estimated from the patents. In addition to the above edges, each classification node is also connected to its parent within the classification scheme being used (the patent classification has a hierarchical structure). We refer to this network as the “CAN” or Art and Classification Network.

Given a query patent, Zuse performs a random walk with restart on ACN from the node in the ACN that corresponds to the query. Upon convergence, this process determines the steady-state probability of visiting each of the nodes in ACN starting from the query patent. The nodes in ACN that correspond to art with a priority date that is earlier than that of the query patent and for which we have access to their full text are ranked based on their steady-state probabilities and the N highest-ranked nodes are selected. The art corresponding to this set of nodes defines the candidate prior art subset (CPAS) that is being further analyzed by Zuse in order to determine if it is relevant with respect to the query claim. By leveraging the citation network, the classification hierarchy, and the connections between art and classifications, the CPAS contains prior art that is highly related to the query patent because there are many short paths within ACN passing through related art and/or classifications that connect them.

For each limitation in the query claim, Zuse identifies the text segments in CPAS that have a similarity that is above a threshold (referred to as sufficiently similar segments (SSS)). The similarity is determined by comparing the corresponding words while taking into account variations in word-forms, thesaurus-derived synonyms, and user-supplied synonyms. In addition, when provided, Zuse takes into account user-supplied information related that relate to “must-have”, “important”, and “forbidden” terms. Upon completion, Zuse eliminates from CPAS any art that does not contain at least one SSS to any of the limitations of the query claim. As a result, what remains in CPAS is potential prior art that is now also related to the query claim in terms of its content (each contains at least one SSS).

For each piece of art Zuse identifies, it calculates a V Score, which is a blended score for the reference. The V Score is comprised of five scores: (i) the aggregate similarity of the most similar SSS to each of the limitations, (ii) the textual similarity between the query claim and the art’s abstract, (iii) the similarity between the classifications of the query patent and art (that is informed by the hierarchical nature of the classification scheme), (iv) the random-walk distance between the query patent and the art in the ACN, and (v) the similarity between the ancestors of the query patent and the art within the art citation network (i.e., the subset of ACN involving only the art nodes). These five scores are then combined to produce an overall score for each art element in CPAS, and the top-scoring prior art documents are returned.

Zuse Analytics is a patent pending solution developed by Professor George Karypis, Dan Bork and a team of dedicated artificial intelligence, predictive analytics, database management and machine learning experts over the last 10-years.   The Zuse Team welcomes thought and input and encourages you to contact us with any questions or thoughts on how to improve our systems and methods or if you would like to join our team.  Please contact us at


Q Score Overview

Zuse Analytics presents for users a single score that is designed to measure the overall relative quality of a patent in a large patent collection by taking into account various attributes, including the citation network of the collection and a measure of value that the patent owners assign to it. We call this score the “Qscore.”  

The Qscore is computed by performing a random walk with restart, in which the destination of the random walk’s restart node is proportional to the values assigned to the patents by their owners. The steady-state probabilities of ending up at any given node in the network of that random walk are used to determine the Qscore of each node. Nodes with higher steady-state probabilities will in general have higher Qscores than nodes with smaller steady-state probabilities. Extensive research in using random walks on networks have shown that the nodes with high steady-state probabilities correspond to central nodes in the network and depending on the underlying domain, they correspond to topical authorities (e.g., performing a random walk on the citation network of scientific articles), important web-pages (e.g., performing a random walk on the network corresponding to the hyperlink structure of the web), proteins involved in many biological processes (e.g., performing a random walk in a protein-protein interaction network), and influencers (e.g., performing a random walk in a social network or a follower-followee network).

The Qscore is computed using the following process. A directed weighted network is constructed from the collections’ citations. In that network, each patent in the collection forms a node, and every time a patent x cites patent y, a pair of weighted edges are created from x to y and from y to x. Each node is assigned an initial weight that corresponds to the number of times that a patent has been active. This initial weight corresponds to patent owner’s assigned measure of value as it correlates to the costs associated with prosecuting and maintaining the patent. The initial node weights are scaled to form a probability distribution, and the weighted directed network is converted into a row-stochastic matrix. Random-walk with restart is performed on this network to compute the steady-state probabilities that a random walked will end up at any given node. The steady-state probability of each patent is then divided by the average steady-state probability of the patents that were published in the prior two years. A standard logistic function is then used to convert these ratios into a probability value, which is then scaled to a number between 0 and 100. 

To address the issue that in a citation network, older sources tend to accumulate more citations, the weighted network used for Qscore reduces the weight of the source-destination links in which the publication date of the source is more recent than that of the destination based on the publication time difference of the two nodes. For similar reasons, it increases the weight of the source-destination links when the date of the source is older than that of the destination based on the publication time difference. This reduces the tendency of older patents to have higher Qscore values just because they are older. However, since it usually takes a few years for newly published patents to be well-connected in the network, the reliability of the Qscore values for recently published patents is expected to be lower.

Zuse Analytics is a patent pending solution developed by Professor George Karypis, Dan Bork and a team of dedicated artificial intelligence, predictive analytics, database management and machine learning experts over the last 10-years.   The Zuse Team welcomes thought and input and encourages you to contact us with any questions or thoughts on how to improve our systems and methods or if you would like to join our team.  Please contact us at


IPWe Inc.

  • #

    Address 160 Greentree Dr, STE 101, Dover, DE 19904, USA 

  • #

    Phone  USA: 1.214.438.0820

  • Email

send us a message