Exploring Legal Patent Citations for Patent Valuation
`
`Shuting Wang
`Department of Computer
`Science and Engineering
`Pennsylvania State University
`sxw327@cse.psu.edu
`
`Zhen Lei
`Department of Energy and
`Mineral Engineering
`Pennsylvania State University
`zlei@psu.edu
`
`Wang-Chien Lee
`Department of Computer
`Science and Engineering
`Pennsylvania State University
`wlee@cse.psu.edu
`
`ABSTRACT
`Effective patent valuation is important for patent holders.
`Forward patent citations, widely used in assessing patent
`value, have been considered as reflecting knowledge flows,
`just like paper citations. However, patent citations also
`carry legal implication, which is important for patent val-
`uation. We argue that patent citations can either be tech-
`nological citations that indicate knowledge transfer or be
`legal citations that delimit the legal scope of citing patents.
`In this paper, we first develop citation-network based meth-
`ods to infer patent quality measures at either the legal or
`technological dimension. Then we propose a probabilistic
`mixture approach to incorporate both the legal and tech-
`nological dimensions in patent citations, and an iterative
`learning process that integrates a temporal decay function
`on legal citations, a probabilistic citation network based al-
`gorithm and a prediction model for patent valuation. We
`learn all the parameters together and use them for patent
`valuation. We demonstrate the effectiveness of our approach
`by using patent maintenance status as an indicator of patent
`value and discuss the insights we learned from this study.
`1.
`INTRODUCTION
`Patent valuation, i.e., assessing the value of patents, is
`an important but challenging task for firm technology and
`innovation management. Patent citations have been widely
`used in patent valuation [19, 8, 6, 7] on the ground that
`patent citations provide “paper trails” of knowledge flows
`among patents. The fact that a patent cites a large number
`of prior patents (hereafter, backward citations) suggests that
`the patented invention has built upon “the shoulders of gi-
`ants”, i.e., a significant amount of prior knowledge. This im-
`plies that the invention has great technological richness
`(defined as the amount of prior knowledge a patent builds
`upon) and likely high technological quality and economic
`value. Similarly, when a patent is cited by a large number
`of subsequent patents (hereafter, forward citations), this in-
`dicates that the patented invention has led to a number of
`successful lines of innovation. Thus, the invention is likely to
`
`Permission to make digital or hard copies of all or part of this work for personal or
`classroom use is granted without fee provided that copies are not made or distributed
`for profit or commercial advantage and that copies bear this notice and the full cita-
`tion on the first page. Copyrights for components of this work owned by others than
`ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
`publish, to post on servers or to redistribute to lists, requires prior specific permission
`and/or a fee. Request permissions from permissions@acm.org.
`CIKM’14, November 3–7, 2014, Shanghai, China.
`Copyright 2014 ACM 978-1-4503-2598-1/14/11 ...$15.00.
`http://dx.doi.org/10.1145/2661829.2662029 .
`
`be of high technological in(cid:13)uence (defined as the techno-
`logical impact that a patent has on subsequent inventions)
`and thus highly economically valuable.
`From such technological aspects, one might think that
`patent citations are similar to paper citations. However,
`patent citations are actually quite different from paper ci-
`tations in significant ways.
`In particular, patent citations
`could be interpreted in two dimensions: (i) a technological
`one that is related to knowledge flows, and (ii) a legal one
`that is related to delimitation of patent scope. Let’s con-
`sider the scenario where a patent application is under exam-
`ination. The patent examiner needs to search for relevant
`prior art (prior inventions) to determine whether the inven-
`tion is patentable based on its novelty and inventiveness in
`comparison to these prior inventions. Meanwhile, the exam-
`iner needs to determine the appropriate scope of the patent
`right by asking the applicant to modify, if necessary, the
`language of the claims. For example, if an inventor applies
`for a chemical compound that makes some novel structural
`modification to an existing drug. The examiner would grant
`a patent to the invention, but cite the prior patent on the ex-
`isting drug to: (i) show the knowledge link between the two
`inventions and (ii) to narrow down the scope of the newly
`granted patent so that it would cover only the modification,
`not the original chemical structure.
`In fact, the scope of
`the newly granted patent could be so narrowed down by the
`prior patent on the existing drug that the firm owning the
`newly issued patent may have to get a license from the paten-
`tee of the existing drug patent in order to market the new
`drug. In this case, the cited prior patent acts as a blocking
`patent to the newly granted patent. Similarly, an applicant,
`under the U.S. Patent Law, has the obligation to disclose rel-
`evant prior art that she knows during her research, though
`she has no obligation to identify all possible relevant prior
`art when filing an application. These applicant-inserted ci-
`tations could, on one hand, suggest knowledge flows, but on
`the other hand, be used to narrow down the patent scope.
`Therefore, a citation made by patent A to patent B could
`suggest that there are knowledge flows from the cited patent
`to the citing patent (this aspect of patent citations defines
`the technological citations), or that the cited patent puts le-
`gal constraints on the scope of the citing patent (this aspect
`defines the legal citations), or both. The legal interpretation
`of patent citations has quite different implication in terms
`of what patent citations mean for patent valuation, com-
`pared to the technological interpretation of patent citations.
`From the legal aspect, when a patent cites a large number of
`backward citations, it could suggest that many prior patents
`
`1379
`
`PMC Exhibit 2047
`Apple v. PMC
`IPR2016-00754
`Page 1
`
`

`
`might have been used to narrow down the scope of the citing
`patent or block the citing patent. Consequently the citing
`patent might have a very narrow legal patent scope (i.e.,
`the scope of patent right claimed), and thus likely small com-
`mercial value. On the other hand, when a patent receives
`a large number of forward citations, it might be blocking
`or putting constraints on these subsequent patents. In this
`case, a patent with a large number of forward citations im-
`plies a high level of legal blocking power, and thus it is a
`highly valuable patent.
`Furthermore, legal constraints in patent citations (from
`the legal aspect) are time sensitive. For a patent citation
`with two year lag between the citing and cited patents, the
`cited patent could block the citing patent for a long time.
`However, if a patent cites an expired prior patent, then the
`cited patent put no legal constraints on the citing patent.
`Our study aims to explore the insights about the tech-
`nological and legal dimensions of patent citations and pro-
`pose corresponding measures for patent valuation. Specif-
`ically, based on the technological and legal interpretations
`of patent citations, we propose to capture four quality mea-
`sures of patents, namely, technological richness, technological
`in(cid:13)uence, legal patent scope, and legal blocking power.
`More importantly, and quite intuitively, we propose that
`there exist mutual interdependence among these four mea-
`sures of patents; and the two measures in the technological
`dimension are related to each other in a different way than
`the other two measures in the legal dimension are related.
`Consider the technological richness and influence associated
`with a patent A.
`If patent A cites prior patents that are
`of greater technological influence, other things being equal,
`the technological richness of the citing patent A is greater,
`as the invention builds on a lot of influential prior inventions.
`Meanwhile, if patent A is cited by subsequent patents that
`are of high technological richness, the patent A’s techno-
`logical influence would be greater, as it leads to subsequent
`innovations of high quality.
`By an interesting contrast, if patent A cites prior patents
`that are of greater legal blocking power, other things being
`equal, the legal patent scope of the citing patent A is smaller,
`as it is narrowed down by prior patents with large blocking
`power. However, if patent A is cited by subsequent patents
`that are of large legal patent scope, the patent A’s legal
`blocking power is greater as it put constraints on subsequent
`patents with broad patent right that are highly valuable.
`We investigate different methods to quantify the four pro-
`posed measures. We first assume that a patent citation can
`be interpreted in the technological dimension or in the legal
`one (or both). We then consider the case where a patent
`citation represents a probabilistic mixture of both techno-
`logical and legal citations, with the significance of legal ci-
`tations decaying by time (i.e.,the grant lag between a cited
`and citing patents). Accordingly, we capture their mutual
`interactions and iteratively learn the four measures using
`the patent data. Technically, we adopt a parameter learning
`process that integrates multiple models (including a tempo-
`ral decay in legal citations, a probabilistic citation network
`based algorithm for quantifying the four proposed patent
`quality measures, and a prediction model for patent valua-
`tion).
`To validate our idea of distinguishing legal citations from
`technological citations, we empirically apply the four pro-
`posed patent quality measures in patent valuation. We use
`
`patent renewal status (patent maintenance) as an indicator
`of patent value in our experimentation. Our results show
`that separating technological and legal dimension in patent
`citations achieves better accuracy in experiments for patent
`value prediction. And our proposed patent quality mea-
`sures based on legal citations show more important roles in
`predicting patent value than measures based on technolog-
`ical citations. Our study also confirms the mutual interde-
`pendence between technological influence and technological
`richness is different from that between blocking power and
`legal patent scope. Moreover, by applying a probabilistic
`model to quantify the proposed concepts, we validate that
`patent citation is a probabilistic mixture of technological
`and legal indications and the significance of a legal citation
`decays by time.
`To the best of the authors’ knowledge, this work repre-
`sents the first attempt to explore the insight that a patent
`citation could be a mixture of technological and legal ci-
`tations, to quantify the technological quality measures and
`legal quality measures corresponding to the two different di-
`mensions in patent citations, and to apply them in patent
`valuation.
`In summary, our work has made the following
`major contributions:
`(cid:15) This study aims to exploit the technological and legal
`interpretation of patent citations and apply them to
`patent valuation.
`(cid:15) Four different patent quality measures, namely, tech-
`nological richness, technological influence, legal patent
`scope and legal blocking power, and the interactions
`among them are proposed to capture the technological
`and legal information imbedded in patent citations.
`(cid:15) We propose a probabilistic model that considers a patent
`citation as a probabilistic mixture of technological and
`legal citations, with the relative weight on the legal di-
`mension decays by time. We develop an algorithm that
`captures the interdependence among the four proposed
`patent quality measures to iteratively derive these mea-
`sures and learn the model parameters, which are useful
`for analysis of patents in a firm or a field.
`(cid:15) Using patent renewals as an indicator of patent value,
`our experiments show that considering both the tech-
`nological and legal dimensions of patent citations and
`applying these four patent measures can significantly
`improve patent evaluation, compared to the current
`practice that only involves the technological interpre-
`tation in patent evaluation.
`The rest of this paper is organized as follows. We first as-
`sume a deterministic model of patent citations and introduce
`our algorithms to derive measures related to technological
`citations and legal citations in Section 2. Then we develop
`a probabilistic mixture model of patent citations to better
`capture those measures in Section 3. In Section 4, we intro-
`duce our evaluation methodology and conduct experiments
`on valuation of Drug&Medical patents in both firm-level and
`field-level. We finally review related works in Section 5 and
`draw conclusions in Section 6.
`2. DETERMINISTIC MODEL AND ALGO-
`RITHMS
`With the technological and legal interpretations for patent
`citations, an immediate question is how to model and quan-
`tify technological and legal citations. In this section, we first
`
`1380
`
`PMC Exhibit 2047
`Apple v. PMC
`IPR2016-00754
`Page 2
`
`

`
`logical influence score is the number of forward citations it
`receives, while its technological richness score is the number
`of backward citations it makes.
`
`∑ j
`
`(3)
`
`(4)
`
`∑ j
`
`BPi =
`
`EL(i; j)
`
`LSi = (cid:13) (cid:0)
`
`EL(j; i)
`
`where EL(i; j) refers to a legal citation to patent i made by
`patent j and (cid:13) is the global value of legal scope in all patents.
`Thus, for patent pi, its blocking power score is the number
`of legal citations to pi when it is not expired, while its legal
`scope score is dependent on the number of citations which
`it makes when the cited patents are not expired. With pi
`making more legal citations, its legal scope is likely further
`narrowed. Therefore, the resulting legal scope is reduced
`from the original legal scope of pi by its legal citations. Two
`issues arising here are (i) the setting of the initial (global)
`legal scope value and (ii) the amount of the legal scope to
`be deducted from this patent due to legal citations. We
`consider legal patent scope of patent i, as defined by Eq. (4),
`to be greater than 0. Thus, (cid:13) is inherently greater than
`the maximal number of legal citations/references made by
`patent pi. While in this paper all patents are assumed to
`have the same initial legal patent scope, we may empirically
`set appropriate (cid:13) and use it to analyze the initial legal scopes
`in different domains or patent sets.
`2.3 CiteNet Algorithm
`As CiteCount only counts on one-hop neighbors in the
`graph, the potential influence of neighbor patents located
`in multi-hop neighborhood is not considered. As a result,
`the relationship and interdependence between technological
`influence and richness as well as between blocking power
`and legal influence are not well considered. To address this
`issue, we derive CiteNet, a patent citation network based
`algorithm, to capture these mutual independence.
`In the
`algorithm, we make the following intuitive assumptions:
`
`1. The technological influence of a patent is determined
`by the number of technological citations it receives and
`the technological richness of these citing patent. The
`technological influence of a patent will be higher if it
`is cited by subsequent patents of higher technological
`richness, as it leads to follow-up innovations of high
`technological quality.
`
`2. The technological richness of a patent depends on the
`number of technological citations it makes and the
`technological influence of these cited patents. The
`technological richness of a patent will be higher if it
`cites patents of higher technological influence because
`it is based on prior innovations of high technological
`impacts.
`
`3. The blocking power of a patent depends on the number
`of legal citations it receives and the legal patent scope
`of these citing patents. The blocking power of patent
`will be greater, if it is cited by patents with larger legal
`patent scope because it blocks patents with broader
`scope.
`
`4. The legal patent scope of a patent is determined by the
`number of the legal citations it makes and the block-
`ing power of these cited patents in the legal citations.
`
`discuss a heuristic and deterministic model for interpreting
`the technological and legal dimensions of patent citations.
`Then we propose algorithms to quantify the four proposed
`patent quality measures: technological influence, technolog-
`ical richness, blocking power, and legal patent scope, for a
`focal patent.
`2.1 Modeling of Patent Citations
`Here we assume that a patent citation always reflects
`knowledge flow from the cited patent to the citing patent.
`Therefore, the technological dimension of a patent citation
`always exists. However, the legal dimension of a patent ci-
`tation only exists when the cited patent is not expired. In
`other words, if the cited patent is not expired when the cita-
`tion is made, the citation is of both a legal and technological
`citation. Otherwise, the citation only reflects the technolog-
`ical (knowledge) flow. Consider an example where patent
`A is granted in year 2000 and was maintained (renewed) at
`its 4th year renewal but not at the 8th year.
`If a patent
`B cites patent A in year 2006, we consider the citation to
`have both legal and technological interpretations. However,
`for a patent C citing patent A in year 2009, the citation is
`only a technological citation. Accordingly, it is fairly easy
`to determine whether a patent citation is a legal citation or
`not, since we know when a patent is expired (based on the
`data on patent maintenance at USPTO).
`Citation Graphs. Based on the discussion above, we de-
`rive two citation graphs. One represents the technological
`citation network, denoted as GT = (V; ET ) where GT is the
`same as the original patent citation network because here
`we assume that a patent citation always serves its techno-
`logical functionality, i.e., ET (i; j)=1, if pi is cited by pj.1
`On the other hand, the legal citation network, which cap-
`tures the legal implication between patents, is denoted as
`GL = (V; EL) where EL(i; j) = 1 if pi is not expired at the
`grant year of pj; and EL(i; j)=0 if pj is expired when pj is
`granted.2 Based on these two citation graphs, we propose
`to characterize a patent with four quality measures: techno-
`logical influence score, technological richness score, blocking
`power score and legal patent scope score. In the following,
`we describe two basic approaches in quantifying the four fea-
`tures: one is the CiteCount algorithm and the other is the
`CiteNet algorithm.
`2.2 CiteCount Algorithm
`The CiteCount algorithm, similar to the conventional ci-
`tation counting approach for assessing the quality of scien-
`tific literature, counts the number of citations of different
`types, based on the technological and legal citation graphs.
`Given the technological citation graph GT and the legal ci-
`tation graph GL, we formally define the Technological Influ-
`ence score (TI), Technological Richness score (TR), Block-
`ing Power score (BP) and Legal Patent Scope score (LS) of
`a patent pi as follow.
`
`∑ j
`
`∑ j
`
`ET (i; j)
`
`ET (j; i)
`
`(1)
`
`(2)
`
`T Ii =
`
`T Ri =
`
`where ET (i; j) refers to a technological citation to patent
`i made by patent j. Therefore, for patent pi, its techno-
`1V is the set of patents and ET is the set of edges in GT .
`2EL is the set of edges in GL.
`
`1381
`
`PMC Exhibit 2047
`Apple v. PMC
`IPR2016-00754
`Page 3
`
`

`
`years, given that they are all granted before the expiration
`of the cited patent. In other words, it is more realistic to
`assume that the legal implication of a patent citation decays
`as a function of the lagging years between the cited patent
`and the citing patent until the cited patent is expired.3
`Therefore, we propose to adopt a probabilistic approach
`to model patent citations. We assume that the technologi-
`cal and legal interpretation of a patent citation takes some
`weights, i.e., their total weight equals one. Meanwhile, we
`assume that the weight for the legal dimension decays as
`a function of the lagging years between the cited and the
`citing patents. Moreover, it becomes zero when the cited
`patent expires.
`With such a probabilistic/mixture model of patent cita-
`tions, we propose, similar to Section 3, probabilistic cita-
`tion count (ProbCiteCount) and probabilistic citation net-
`work (ProbCiteNet) to quantify the four quality measures of
`a patent. ProbCiteNet takes into consideration the citation
`network structure and the interdependence between the two
`technological measures and between the two legal measures,
`while ProbCiteCount does not.
`Once the decay behavior of legal citations, the constraint
`over total weight of legal and technological citations, and the
`relationships among legal and technological quality measures
`are properly modeled, we shall be able to model the corre-
`lations between the quality measures and the patent value,
`e.g., by formulating it as a classification problem. Putting
`all these together allows us to not only to derive quality
`measures but also learn the model parameters (such as the
`time dacay parameter) and analyze the importance of the
`patent quality measures in the classifiers for patent evalua-
`tion. Furthermore, learning the various model parameters
`enable us to study insights about valuation of patent cita-
`tions in a field or a firm.
`In the following, we first detail our approach to model
`the decay function for legal citations (see Section 3.1) as
`well as ProbCiteCount (see Section 3.2) and ProbCiteNet
`(see Section 3.3). Then, we introduce our prediction model
`for patent value and the learning process for deriving model
`parameters (see Section 3.4).
`3.1 Temporal Decay of Legal Citation Weights
`Here we discuss the selection of a decay function to model
`the temporal decay of legal power. Two candidate functions
`are exponential decay or linear decay functions. An expo-
`nential decay function assumes that the rate of decay is pro-
`portional to its current value, while a linear decay function
`assumes that the rate of decay is constant over time, which
`is less applicable to our case. Consider a case where patent
`A (granted in 2000 and to expire in 2016) is cited by patent
`B in 2002 and by patent C in 2003. Since these inventions
`are very close in time, they are likely close substitutes to
`each other in the market. Thus, the one year difference be-
`tween the citing patents B and C could mean significantly
`different market values. Consequently, the weights on the
`legal dimension of the two patent citations could be quite
`different. However, suppose that patent A is cited by patent
`D in 2014 and by patent E in 2015, which are very far away
`from patent A (which is about to expire). In this case, the
`weights for the legal dimension of the citations correspond-
`
`3Here the lagging years refers to the number of years the
`grant date of the citing patent is lagging behind the grant
`date of the cited patent.
`
`Intuitively, the legal patent scope of a patent will be
`smaller if it cites a lot of prior patents with stronger
`blocking power, because these cited patents would nar-
`row down its scope.
`
`Therefore, given a citation network graph G = (V; E) , the
`CiteNet algorithm computes the four quality measures for
`each patent iteratively until the derived measures converge.
`In each iteration, the measures are derived as follows.
`T Ii
`T Ri
`
`∑ j
`
`∑ j
`
`T Rj where ET (i; j) = 1
`
`T Ij where ET (j; i) = 1
`
`(5)
`
`(6)
`
`(7)
`
`(8)
`
`Note that Eq.(5) corresponds to the first assumption and
`Eq.(6) corresponds to the second assumption. Moreover,
`given the legal citation network GL and a patent pi, we
`have:
`BPi
`LSi (cid:13) (cid:0)
`
`∑ j
`
`∑ j
`
`LSj where EL(i; j) = 1
`
`BPj where EL(j; i) = 1
`
`Eq. (7) corresponds to the third assumption and Eq. (8)
`corresponds to the fourth assumption. Also, similar to the
`CiteNet method, (cid:13) is the initial value of legal scope and we
`subtract the blocking power of legal citations that pi cites
`from its original patent legal scope. Moreover, after each
`iteration, we normalize the calculated scores for the four
`quality measures using 2-norm normalization to guarantee
`the convergence of the algorithm.
`In summary, CiteNet captures the independence between
`technological influence and technological richness, and block-
`ing power and legal scope in each iteration as shown in
`Eqs. (5)-(8). The proposed CiteNet algorithm derives the
`patent technological influence and richness in a way similar
`to the HITS algorithm that captures the mutually reinforce-
`ment between authoritative and hub web pages [13]. On the
`other hand, the derivation of the patent legal blocking power
`and patent scope is totally different. As Eqs. (7) and (8)
`show, they are based on different rules.
`3. PROBABILISTIC MODELING
`In the previous section, we assume that a citation is always
`a technological citation. Moreover, depending on whether
`the cited patent is expired at the time of being cited, the
`citation could have legal implication on the citing patent.
`Note that some potential issues may arise with such deter-
`ministic heuristics. For example, some patent citations are
`counted twice (as a technological citation and as a legal ci-
`tation), whereas others are counted only once (only as a
`technological citation). This seems to be ad hoc. Is there
`a better way to model the two dimensions of a patent cita-
`tion coherently? In particular, as explained earlier, a patent
`citation is likely to be a mixture of a technological citation
`and a legal citation, with different weights.
`Additionally, the legal measures of a cited patent, corre-
`sponding to a patent citation, are assumed to remain con-
`stant, whether it is cited by a citing patent granted just a
`few years later or by another patent granted many years
`later. We argue that intuitively it may be more reasonable
`to assume that the legal power of a cited patent varies cor-
`responding to different citing patents granted at different
`
`1382
`
`PMC Exhibit 2047
`Apple v. PMC
`IPR2016-00754
`Page 4
`
`

`
`where AT
`L is the transpose of AL and (cid:13) is the global legal
`patent scope for each patent, defined in Section 2.
`3.3 Probabilistic Citation Network Algorithm
`Given the weighted technological citation matrix AT and
`legal citation matrix AL defined above, the ProCiteNet al-
`gorithm iteratively derives the four quality measures, based
`on the mutual interdependence among them, as discussed in
`Section 2. Formally, we define technological influence (TI)
`and technological richness (TR) for the given patent pi as
`follows.
`T Ii
`AT (cid:1) T Rj
`
`∑ j
`
`∑ j
`
`T Ri
`
`(15)
`
`T (cid:1) T Ij
`AT
`
`(16)
`
`In Eq. (15) and Eq. (16), j refers to the patents citing patent
`pi and the patents cited by patent pi, respectively.
`Similarly, we define blocking power score (BP) and legal
`(9)
`patent scope score (LS) for the given patent pi as follows.
`AL (cid:1) LSj
`BPi
`
`(17)
`
`∑ j
`
`LSi ((cid:13) (cid:0)
`
`L (cid:1) BPj)
`AT
`
`(18)
`
`where (cid:13) is the initial value of legal scope for each patent,
`and j refers to the patents citing patent pi and the patents
`cited by patent pi, respectively.
`ProbCiteNet, similar to HITS algorithm [13], considers
`mutual reinforcement between the technological influence
`and richness, as well as the blocking power and legal patent
`scope. However, it is different from HITS in that it oper-
`ates on weighted technological and legal citation graphs that
`take into account the two dimensions of patent citations and
`exponential decaying in weights. The weights on the cita-
`tion networks are to be learned in an integrated learning
`process, which uses different rules for updating different fea-
`tures based on mutual interdependence among features.
`3.4 Prediction Model for Patent Valuation
`We argue that the four legal and technological measures
`can be used for patent valuation and thus derive a model to
`predict patent value using the proposed quality measures as
`features. In this section, we introduce a prediction model,
`which we combine with the probabilistic modeling of patent
`citations to learn model parameters and to derive classifiers
`for patent evaluation in an integrated learning process. We
`use logistic regression model for predicting patent value, and
`maximize the object function with gradient ascent method
`to predict patent value. Accordingly, each patent pi in the
`training sample can be represented as (x; y) where x is the
`feature set of pi. i.e., the four features/quality measures, and
`y is the predictive label used in regression. For example, the
`patent maintenance status may serve as a label/indicator of
`patent value (as discussed in detail later in Section 4). For
`logistic regression, we define our hypothesis function h(cid:18)(x)
`as follows.
`
`h!;(cid:21)(x) =
`
`1
`1 + exp((cid:0)!T x((cid:21)))
`where the model parameter (cid:18) consists of (i) ! – the weights
`of the features and (ii) (cid:21) = f(cid:21)1; (cid:21)2; :::; (cid:21)ng – the parame-
`ters of the exponential decay functions corresponding to US
`
`(19)
`
`∑ j
`
`(13)
`
`(14)
`
`ing to A to D and A to E should be no much different,
`because inventions in patent A is fading out of the market
`when it is cited by patent D and E.
`Moreover, the legal weight of a patent citation depends
`on the power of the cited patent in narrowing down and/or
`blocking the citing patent. Therefore, the rate at which the
`weight on a legal citation decays should be correlated with
`the turnover or product life cycle for the technology field
`of the cited patent. Hence, in this model, we assume that
`the technology domain of the cited patent, which reflected
`by the U.S. patent class of the cited patent, determines the
`temporal decay pattern, i.e., patents in the same U.S. class
`share the same temporal decay pattern.
`Formally, let the parameter of the decay function for a
`given U.S. class u be (cid:21)u. Given two patents pi and pj where
`pi is cited by pj, we define the weight for the legal dimension
`of this citation as follows.
`−(cid:21)u|tj−ti|
`
`(cid:21)ue
`
`0
`
`PL(pi; pj) =
`
`pi is not expired when cited
`by pj
`pi is expired when cited by pj
`
`where ti and tj are the grant dates of pi and pj respectively,
`and u denotes the U.S. Class of the cited patent pi.
`As we consider any citation to be a mixture of the techno-
`logical and legal dimensions (with their total weight equals
`1), the technological weight for patent pj citing patent pi is
`defined as follows.
`PT (pi; pj) = 1 (cid:0) PL(pi; pj) = 1 (cid:0) (cid:21)ue
`−(cid:21)u|tj−ti|
`As such, Eq. (9) and Eq. (10) govern the weighted techno-
`logical citation network and legal citation network, respec-
`tively. Next, we discuss the modeling of interdependency
`among the four quality measures with ProbCiteCount and
`ProbCiteNet.
`3.2 Probabilistic Citation Count Algorithm
`To present the probabilistic citation count (ProbCiteCount)
`algorithm, we first introduce an adjacency matrix defined for
`deriving the four quality measures, based on the weighted
`technological and legal citation networks.
`Adjacency Matrix for Technological and Legal Cita-
`tions. We use an adjacency matrix AT to denote techno-
`logical citations and AL to denote the legal ones. AT (i; j) =
`PT (pi; pj) if pi is cited by pj, otherwise AT (i; j) = 0. On
`the other hand, AL(i; j) = PL(pi; pj) if pi is cited by pj,
`otherwise AL(i; j) = 0.
`Based on AT and AL, we define technological influence
`(TI) and technological richness (TR) for a given patent pi
`as follows.
`T Ii
`T Ri
`
`(10)
`
`(11)
`
`(12)
`
`AT (i; j)
`
`AT
`T (i; j)
`
`∑ j
`
`∑ j
`
`where j is bounded to the set of patents in the corpus citing
`pi and AT
`T is the transpose matrix of AT . Next, we define
`blocking power (BP) and legal patent scope (LS) for a given
`patent pi, based on citation count as follows.
`BPi
`LSi (cid:13) (cid:0)
`
`∑ j
`
`∑ j
`
`AL(i; j)
`
`AT
`L(i; j)
`
`1383
`
`PMC Exhibit 2047
`Apple v. PMC
`IPR2016-00754
`Page 5
`
`

`
`∑ n
`
`we perform n iterations of updates for the scores, Z(cid:21) is:
`T AT )N−1 +
`@AT
`!ti[(AT
`T
`@(cid:21)j
`
`
`
`(ATT AT )n](
`
`AT +
`
`@AT
`@(cid:21)j
`
`AT
`T )
`
`∑
`
`∑ n
`
`where !ti, !tr, !bp and !ls are weights for technological
`influence, technological richness, blocking power and patent
`legal scope, respectively.
`4. EVALUATION USING PATENT VALUA-
`TION
`We conduct experiments to validate our proposition that
`patent citations have legal and technological dimensions in
`interpretation and such distinctions can be useful for patent
`valuation. Specifically, we test whether the proposed patent
`quality measures (i.e., technological influence and richness,
`and blocking power and legal patent scope) can improve
`performance in predicting patent value. In this section, we
`first discuss our evaluation methodology (i.e., using patent
`renewal status as a proxy for patent valuation), then conduct
`experiments to evaluate the effectiveness of the propose