A patent quality analysis for innovative technology and product development

(1)

A patent quality analysis for innovative technology and product development

Amy J.C. Trappey

a

, Charles V. Trappey

b,⇑

, Chun-Yi Wu

a

, Chi-Wei Lin

a

Department of Industrial Engineering and Engineering Management, National Tsing Hua University, Taiwan

b

Department of Management Science, National Chiao Tung University, Taiwan

a r t i c l e

i n f o

Article history:

Available online 11 August 2011 Keywords:

Patent quality Patent indicator

Principal component analysis Back-propagation neural network

a b s t r a c t

Enterprises evaluate intellectual property rights and the quality of patent documents in order to develop innovative products and discover state-of-the-art technology trends. The product technologies covered by patent claims are protected by law, and the quality of the patent insures against infringement by com-petitors while increasing the worth of the invention. Thus, patent quality analysis provides a means by which companies determine whether or not to customize and manufacture innovative products. Since patents provide significant financial protection for businesses, the number of patents filed is increasing at a fast pace. Companies which cannot process patent information or fail to protect their innovations by filing patents lose market competitiveness. Current patent research is needed to estimate the quality of patent documents. The purpose of this research is to improve the analysis and ranking of patent qual-ity. The first step of the proposed methodology is to collect technology specific patents and to extract rel-evant patent quality performance indicators. The second step is to identify the key impact factors using principal component analysis. These factors are then used as the input parameters for a back-propagation neural network model. Patent transactions help judge patent quality and patents which are licensed or sold with intellectual property usage rights are considered high quality patents. This research collected 283 patents sold or licensed from the news of patent transactions and 116 patents which were unsold but belong to the technology specific domains of interest. After training the patent quality model, 36 his-torical patents are used to verify the performance of the trained model. The match between the analytical results and the actual trading status reached an 85% level of accuracy. Thus, the proposed patent quality methodology evaluates the quality of patents automatically and effectively as a preliminary screening solution. The approach saves domain experts valuable time targeting high value patents for R&D com-mercialization and mass customization of products.

1. Introduction

Patents play an important role in a knowledge based economy since companies use patents to protect innovation. Patents often establish a time period of protection for the intellectual property (IP) and the new products’ market domination. For the economics and management of mass customization, product planning consid-ers the customer’s needs and the sources of technology. Analyzing related patent documents help design engineers create detailed conceptual plans and understand the underlying component tech-nology. Patent quality analysis furthers the analysis by strategically identifying critical patents for mass customization. High quality patents contain wide claims, refer to few prior art designs, and are highly applicable. Owning high quality patents helps enterprise defend themselves against patent trolls and product infringement.

Mass customization products must consider the risk of patent infringement. Manufacturers must choose the highest quality pat-ents for mass customization to insure commercial production va-lue. High quality patents better enable companies to avoid costly litigation that impedes the production and sales of products in the marketplace. Thus, patent quality analysis facilitates mass cus-tomization and product personalization while minimizing the risk of infringing on the intellectual property rights (IPR) of others. For sustainable competitive growth, enterprises must claim and use intellectual property rights effectively.

Recent patent news shows that enterprises often purchase tech-nology speciﬁc patents to advance techtech-nology and create new products for timely commercialization. There are also incidences whereby manufactures are accused of IP infringement by compet-itors which impede new products from entering the market. The traditional process of patent trade includes three steps. First, the enterprise collects patents of interest from a patent trading plat-form. After collecting and organizing the patent collection, domain experts study and analyze the patents. Next, the enterprise evalu-ates and decides whether to purchase patents or invest in its own

⇑Corresponding author. Tel.: +886 3 5727686; fax: +886 3 5713796.

E-mail addresses: [email protected] (A.J.C. Trappey), trappey@faculty. nctu.edu.tw (C.V. Trappey), [email protected] (C.-Y. Wu), t7378020@ntut. edu.tw (C.-W. Lin).

Contents lists available atScienceDirect

Advanced Engineering Informatics

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / a e i

(2)

research and development (R&D) to create new intellectual prop-erty that can better protect or yield product development opportunities.

Traditional patent analysis requires significant costs, time, and manpower. Thus, the purpose of this research is to shorten the time required to determine and rank the quality of patents for new product R&D and innovation management. This research develops key patent indicators that are derived from principle component analysis. The patent quality models, created with the indicators as inputs, are trained using the back-propagation neural networks. The degree of patent quality is defined by different eval-uators in different situations including patent trade (i.e., sold pat-ents are of high quality), patent litigation (i.e., patpat-ents win lawsuits are of high quality), and patent assignment (i.e., patents have assignment processes are of high quality). The proposed ap-proach can be used to build different patent quality models based on the pre-defined situations.

In regards to the patent indicators, they are collected from patent corpuses, including the number of patent citations and the number of International Patent Classifications (IPC). The first step of the pro-cess is to extract the key quality impact factors using principal com-ponent analysis. The second step uses the key factors as input parameters for a back-propagation neural network (BPN) model. The BPN model is trained to identify technology specific patents. The system then processes the patent collection to rapidly identify patents matching the key identifying quality criteria. As a means to evaluate the methodology, patents that are successfully traded in the marketplace are compared to the model selection. The meth-odology helps experts rank and set values on patent quality. The key impact factors, when combined with a trained model to evaluate unknown patents’ quality, better enable engineers and product designers forecast patent potential for product development.

2. Background and literature review

This section highlights the core background knowledge and re-lated literature of the research, including patent analysis, patent characteristics, indicators of patent evaluation, the principal com-ponent analysis approach, and back propagation neural network models.

2.1. Patent analysis

Patent analysis is employed across organizations and is a re-search approach frequently used by R&D engineers, academics, and technology policy makers. The results of patent analysis are used to estimate trends, proﬁtability, and performance of technol-ogy[1]. The patent characteristics are in turn used for prior art searching and information extraction about patent history and activities[2]. The competition among companies is revealed using these characteristics, and careful analysis of this information often results in the discovery of mass customization development oppor-tunities. Yoon and Park[3]proposed a network based patent anal-ysis to show the relationship of domain speciﬁc patents within a virtual network, which is then used to evaluate a patent’s degrees of importance, degree of technique, and degree of similarity.

The creation of a patent portfolio is a combination process including a patent defense strategy. The process is analogous to creating a fence or patent cluster. A patent fence prevents or blocks competitors from registering related core technology. In order to create a defensive technology fence, the company must also devel-op non-core patent technologies. The fence makes it difﬁcult for competitors to incorporate similar technology without infringe-ment. On the other hand, a patent cluster brings together many patents of alternative technologies, and strategically makes

defin-ing the underlydefin-ing technology trends more difficult. The illustra-tions of patent fence and patent cluster are shown inFig. 1.

An advanced patent analysis must consider the characteristics of related patent documents. Most patent analysis techniques fo-cus on the classification of patents using the related prior art, the specific domain technology trend, and the patent defense strategy. Traditional patent quality assessment uses different patent indica-tors extracted from patent characteristics specified by the domain experts. However, the value or worthiness of patent indicators is changed based on many factors. Thus, the proposed patent quality methodology is flexible for building the patent quality model based on the domain of collected patents and their quality evaluation cri-teria, e.g., factors related to transaction, litigation, or maintenance.

2.2. Patent characteristics

Patent documentation and analysis use uniform format field, forward and backward analysis of citations, and the creation of patent portfolios. A patent document with a unified form consists of three parts [4]. The first part contains the patent publication number, the application date, the citation number and the interna-tional patent classification. The second part describes the back-ground, innovation content, and implementation methods. The third part defines the claims used by the assignees to protect the invention.

Most research focuses on the information contained in the pat-ent citation. Citations provide researchers with a historical trail about the development of technology and provide a means to as-sess and rank the importance of individual patents. Lai and Wu [5]employed patent co-citation analysis to establish a patent clas-sification system. The clasclas-sification system reveals the relationship of technologies and the evolution of a technology category. Com-mon technology trend analysis uses forward citation analysis, co-citation analysis, and backward co-citation analysis. The information contained within cited patents corresponds to specified fields, such as ‘‘US patent documents,’’ ‘‘foreign patent documents’’ and ‘‘other references’’. These metadata fields are also useful for evaluating patent quality.

2.3. Patent quality indicators

The primary patent quality indicators are related to investment, maintenance, and litigation, which form a basis for assessing pat-ent quality when the evaluation focuses on the patpat-ent’s potpat-ential for sale. These indicators are brieﬂy described as follows.

2.3.1. Indicators for investment

There are five indicators used by CHI Research to analyze patent portfolios for investment[6,7]. The first indicator represents the number of patent applications from a company and its subsidiaries in the previous year, the second indicator describes the percentage of patent growth in the previous year, and the third indicator is the current impact index. The fourth indicator, science linkage, is cal-culated using the average number of references which are cited from scientific papers. Finally, the technology cycle-time measures the median age of the cited patents.

(3)

2.3.2. Indicators for maintenance

Barney[8]proposes five indicators which demonstrate the com-petitiveness of patent documents. The five indicators are the number of independent claims, claim length, the length of the writ-ten specification, priority claims, and forward citations. The histor-ical patent metadata represents patents having more independent claims and are considered to be more valuable. The larger the num-ber of claims, the broader the scope of protection and the better the likelihood of surviving a validity attack. A longer specification pro-vides better support for patent claims and strengthens the patent against certain types of validity attacks. More priority claims means a patent is entitled to an earlier filing date in more countries, which can be beneficial in fending off patent validity attacks. Intuitively, a high forward citation rate indicates a high level of commercial interest or activity in the patented technology.

The patent maintenance rates generally increase with the num-ber of claims, the length of the written speciﬁcation, the numnum-ber of the recorded priority claims, and the forward citation rate. Further, the patent maintenance rates decrease with claim length. Statisti-cal data supports the hypothesis that patents having shorter claims (with fewer limitations and broader scope of protection) are more valuable[9].

2.3.3. Litigation indicators

Allison et al.[10]noted that the identification of valuable pat-ents is needed to win lawsuits. Based on 6861 patpat-ents used in lit-igation, there are five common characteristics. First, the number of claims in a patent provides evidence of its breadth and therefore its value. A patent’s claims are the legal definition of the invention. In order to decide whether a patent has been infringed upon, a court must compare the claims of the patent to the defendant’s device. Second, the number of references cited (backward citations) in a patent is evidence of a patent’s validity. Third, the number of cita-tions received (forward citacita-tions) which represent references made by subsequent patents to the patent of interest is evidence of the importance other inventors accord the patent. Furthermore, cita-tions are positively related to patentee decisions to pay mainte-nance fees[11].

Fourth, Trajtenberg et al.[12]have identified a generality mea-sure which is a means of calculating the dispersion of citations re-ceived across different patent classes. They define a function of the sum of the percentages of citations received in each patent class. If a patent is cited by subsequent patents that belong to a wide range of fields, then the measure will be high. Finally, Lanjouw and Schankerman [13] have used the number of different IPCs into which an invention is categorized by the patent and trademark of-fice as evidence of both the breadth and originality of an invention, and hence as evidence of its value.

2.4. Principle component analysis

Principle component analysis (PCA) is a method ﬁrst proposed by Pearson[14] and later formalized by Hotelling[15]. Principal component analysis transforms several independent variables into a new set of variables which retain the most information. The goal of PCA describes the interrelationship among the variables and transforms the original variables into uncorrelated new variables. Moreover, PCA reduces the dimensions of multivariate data and can solve the colinear problems of linear regression[16].

According to the research of Lai and Chu[17], 17 patent indica-tors are defined which influence litigation. The report uses the PCA method to find the relationship between indicators and eliminate unimportant indicators. Tsai-Lin[18]used a questionnaire to ana-lyze the value of a patent, and then applied PCA to determine the four factors that influence patent value including the potential to lead technology, the potential for commercialization, the capacity

for market application, and ability to defend against litigation. Based on these factors, the study collects a set of valuable patents and clusters the patents into four strategic patent groups. Finally, the study suggests that the value of each patent also depends on its technology life cycle, commercial worth, and legal assessment.

2.5. Back-propagation neural (BPN) network

BPN network models are widely used for classification and fore-casting[19]. The BPN network is a fast learning pattern classifier based on a modified back propagation gradient descent algorithm. The BPN network uses a feed-forward and feed-backward flow reg-ulated by an error function. Each network contains an input layer (the neuron number correspond to the number of input vectors), a hidden layer (the weight calculated between the input and out-put layers), and an outout-put layer (the type of classifications corre-sponding to the number of output vectors). Using a non-linear transfer function, the BPN network builds the nonlinear relation-ship of weights between the hidden input layer and the hidden output layer and establishes the target model.

Trappey et al. [20] proposed a new document classification methodology based on a neural network approach. The result yields a significant improvement in document classification and R&D knowledge management. Trappey et al.[21]proposed a com-bined clustering and S curve approach for technology forecasting of RFID sub-technical groups. Chiang et al.[22]applied a back prop-agation artificial neural network, a hierarchical ontology, and nor-malized term frequencies for binary document classification and content analysis. Their approach reduces the effort needed to search and select patents for analysis.

3. Methodology

This research proposes an integrated methodology, combining the Kaiser–Meyer–Olkin (KMO) approach, PCA, and BPN, for deter-mining patent quality based on patent tradability potentials. The structure of the methodology is shown inFig. 2. First, patent data with IP usage rights are collected from the United States Patent and Trademark Ofﬁce (USPTO). Second, principal component anal-ysis is used to extract key patent indicators. Third, the key indica-tors are used as training input parameters for the back propagation neural network model. Fourth, after the BPN model is trained, the technology speciﬁc patent model is used to predict the quality of patents and forecast the IP market potential.

Patent data Collect patent trading information Extract indicators Extract key indicators using PCA

Over the KMO threshold ? No Apply BPN network model Yes Evaluate patent quality and forecast

(4)

Twelve indicators for patent quality are used based on the liter-ature review. The indicator data are used for principal component analysis and the back propagation neural network model. The first two indicators, the patent application and issue date, define the period of patent protection. Next, the international patent classifi-cation and US patent classificlassifi-cation define the technology specific domain. The indicators describe the technology source and appli-cation, including forward citations, foreign citations, and backward citations. The claims and independent claims represent the scope of the lawsuit. Finally, the patent family, the technology cycle time, the science linkage, and the length of detailed specifications are evaluated to quantify the value of patent. The above indicators are located in the patent document as highlighted inFig. 3.

3.1. Evaluation of patent indicators

After extracting the quality indicators, the Kaiser–Meyer–Olkin (KMO) approach is used for evaluating the strength of the relation-ship among variables [23]. The Kaiser–Meyer–Olkin measure of sampling adequacy tests whether the partial correlations among variables are small. The KMO measures the sampling adequacy which should be greater than 0.5 for a satisfactory factor analysis. Large values for the KMO measure indicate that a factor analysis of the variables is a good statistical ﬁt. A value of 0.6 is a suggested minimum threshold for the principal component analysis. The KMO function is shown in Formula(1)

KMO ¼ P i P jðijÞr2ij P i P jðijÞr2ijþ P i P jðijÞs2ij ð1Þ

where rijis the correlation coefﬁcients of indicator xiand indicator

xj. The sijis the offset correlation coefﬁcients of index xiand

indica-tor xj.

Principal component analysis (PCA) is a mathematical proce-dure that transforms a number of possibly correlated variables into a smaller number of variables called principal components. The principal components analysis reduces the number of indicators

used to represent the entire sample. The correlation of components and indicators are shown in Formula(2)

Z1¼ a11x1þ a12x2þ þ a1pxp Z2¼ a21x1þ a22x2þ þ a2pxpþ

v

dotsZm¼ am1x1þ am2x2þ þ ampxp 8 > < > : ; Z1¼ ½ ai1 ai2 aip x1 x2 .. . xp 2 6 6 6 6 4 3 7 7 7 7 5 ð2Þ

where Z1is the subject’s value on principal component 1 (the ﬁrst

principal component extracted), xp represents the subject’s value

on observed indicator p, and a1pis the regression coefﬁcient for

ob-served indicator p, as used in creating principal component 1. On the other hand, the principal component analysis transfers p indica-tors to m components (Zi, i from 1 to m). In this research, the ﬁrst

step of PCA uses the matrix of the correlation coefﬁcients to calcu-late the indicators, as shown in Formula(3)

R ¼ 1 r12 r1p r21 1 r2p r31 r32 r3p .. . .. . .. . .. . .. . rp1 rp2 1 2 6 6 6 6 6 6 6 4 3 7 7 7 7 7 7 7 5 ð3Þ

where rijis the correlation coefﬁcient of indicator xiand indicator xj.

The second step calculates the principal component weights using a Lagrange equation. The objective function of the Lagrange equation and its constraints are show in Formula(4)

Max VarðZÞ ¼ Varða0_{xÞ ¼ a}0_Ra

st a0_{a ¼ 1} ð4Þ

The constraint of the Lagrange equation is a0_{a ¼ 1, which is the}

normalization of the principal component weights ½ a1 a2 ap.

The Lagrange equation is solved using the objective function minus

(5)

the product of the constraint and the Lagrange multiplier, as shown in Formula(5). The results of Lagrange multiplier are the eigenvalues which represent the principal components

L ¼ a0_{Ra kða}0_{a 1Þ} _ð5Þ

For the third step, the partial differentiation for the Lagrange equation (L) and the Lagrange multiplier ðkÞ are used to ﬁnd the weights of the principal components and eigenvalues, as shown in Formula(6). The maximum variance is calculated by Formula (7), and the derivation of principal components is calculated by Formula(8)

@L

@a¼ 2Ra 2ka ¼ 0 ) Ra ka ¼ 0 ) ðR kIÞa ¼ 0 @L

@k¼ a

0_{a 1 ¼ 0}

ð6Þ

VarðZ1Þ þ VarðZ2Þ þ VarðZ3Þ þ þ VarðZmÞ

¼ k1þ k2þ k3þ þ km ð7Þ

ki

k1þ k2þ k3þ þ km

ð8Þ

Based on above steps, the eigenvalues and variance of compo-nents are calculated, as shown inTable 1. In general, the cumula-tive variance should be above 70% for the colleccumula-tive practical principal components. In Table 1, ﬁrst three components will achieve the 70% cumulative variance level (i.e., 78.06%). Further, Hair et al.[24]deﬁned the indicator thresholds used for explaining different data sample sizes as shown inTable 2. For instance, when the data set example size is 85, the explanation values of the indi-cators for chosen components need to be above 0.6. Thus, Indicator I (0.656), Indicator II (0.704), Indicator III (0.713), and Indicator IV (0.895) are kept as key indicators to format the three valid princi-pal components.

3.2. Building patent quality model

After the principal components analysis, the key indicators are used as the input nodes for the back propagation neural network model as shown inFig. 4. The output nodes of the BPN network model represent the quality of patent documents[25,26].

The BPN network model is a supervised learning algorithm used to solve non-linear problems. The feed-forward processing of the BPN network is used to train the data model. The nodes of the hid-den layer and the result of output layer are calculated using the activation functions, as shown in Formula(9). Thus, the value of the BPN network from node i of the input layer to the node j of the hidden layer is calculated using Formula(10)

f ðxÞ ¼ 1 1 þ ex ð9Þ netn j ¼ X i wh ijXiþ bj ð10Þ where wh

ijis the weight from the input layer to the hidden layer, Xi

is the node i of the input layer, bjis the bias of the node j of the

hid-den layer. Thus, the value of node j representing the hidhid-den layer is calculated using Formula(11). The value of the BPN network from the node j of the hidden layer to the node k of the output layer is calculated by Formula(12) Hj¼ f nethj ð11Þ neto k¼ X j wo jkHj ð12Þ where wo

jk is the weight from the hidden layer to the output layer.

Finally, the value of the node k output layer is shown in Formula (13) Ok¼ f ðnetokÞ ¼ f X j wo jkHj ! ð13Þ

where f ðxÞ is the activation function of node k. The error function of the output layer of the BPN network model is described as Formula (14). And, Tkis the real value of the training data

E ¼1 2

X

k

ðTk OkÞ2 ð14Þ

Based on the error function of the output layer, the method which adjusts the weights from the hidden layer to the output layer is calculated by differentiation as shown in Formula (15). Moreover,

g

is the network learning rate. The error function of the hidden layer of the BPN network model is described by For-mula(16). Thus, the method which adjusts the weights from the input layer to the hidden layer can be calculated by the error func-tion of the hidden layer, as shown in Formula(17)

D

wo jk¼

g

@E @wo jk ¼

g

@E @Ok @Ok @wo jk ¼

g

ðTk OkÞf0ðnetkoÞHj¼

g

dokHj ð15Þ dok¼ f 0_ðneto kÞðTk OkÞ ð16Þ Table 1

Principal component analysis of patent indicators.

Key indicators Components matrix

Component 1 Component 2 Component 3 Component 4 Component 5

Indicator I 0.040 0.656 0.102 0.384 0.466 Indicator II _0.704 0.224 0.482 0.086 0.191 Indicator III 0.713 0.177 0.013 0.100 0.027 Indicator IV 0.194 0.068 0.895 0.024 0.152 Indicator V 0.262 0.165 0.541 0.796 0.238 Eigenvalues 3.75 1.62 0.80 0.54 0.06 Variance (%) 31.271 13.471 6.698 4.501 0.483 Cumulative variance (%) 31.27 62.78 78.06 92.96 100 Table 2

The indicator thresholds for different sample sizes.

Sample size Threshold Sample size Threshold

350 0.3 100 0.55

250 0.35 85 0.6

200 0.4 70 0.65

150 0.45 60 0.7

(6)

D

wh ij¼

g

@E @wh ij ¼

g

@E @hj @hj @wh ij ¼

g

@E @Ok @Ok @Hj @Hj @wh ij ¼

g

X k ðTk OkÞf0ðnetokÞwojkf0ðnethjÞxi ¼

g

X k dokwojkf0ðnetjhÞxi¼

g

dhjxi ð17Þ

After calculating these values using the formulas deﬁned by Beale and Jackson[27], the weights of the BPN network are trained. The trained model is then used to evaluate the testing data for clas-siﬁcation or forecast. Thus, the purpose of this research is to build the patent quality model for evaluating the tradability potential of patents. The patent indicators are extracted from patent docu-ments which have been sold and have changed IP usage rights. These transactional indicators are used to train the BPN patent model and the R&D and intellectual property engineers use the trained model to evaluate the quality of patents.

4. System implementation and case study

For the veriﬁcation of the patent transaction model, this re-search collects issued patents that have been sold or licensed to others to build the patent quality system. The proposed system automatically calculates the patent indicators, and uses principal component analysis to extract the transactional indicators. More-over, the patent transaction model is built for testing the unknown quality of patents using the derived key indicators. The quality of patents can be classiﬁed and evaluated after importing the col-lected patents. The case study collects 399 patents licensed or sold from news of patent transactions. Those patents are used to train the patent quality of the transaction model. The trained patent quality model focuses on patent transactions including digital

screens, light emitting diode (LED), information transfer, semicon-ductors, and cell phones. Each speciﬁc domain contains sold or li-censed patents and unsold patents for analyzing the patent indicators for different transaction types. Therefore, the well trained high-tech industry transaction model quickly evaluates the quality degree of the unknown patents. News about related patents is collected to verify the performance of the proposed pat-ent quality model.

The transaction model contains five technology specific do-mains including 78 digital screen patents, 90 LED patents, 52 infor-mation transfer patents, 71 semiconductor patents, and 108 cell phone patents. The first group consists of screen patents sold from Hitachi and Samsung. The usage rights of LED patents were trans-ferred from OSRAM to YaHsin Industrial Company (Taiwan). Third, news describes that the information transfer patents were sold by Agere System and GI. Fourth, Acer purchased semiconductors pat-ents developed by the Industrial Technology Research Institute (ITRI). Finally, the case study collects all current cell phone patents owned by Nokia and Sony. In total, the case study has 399 sample patents, containing 283 patents sold or licensed (i.e., considered in higher quality category) and 116 non-traded patents (i.e., in lower quality category) in the just described five technology domains.

After building the patent list from the news, the proposed sys-tem downloads patents from the USPTO patent database and auto-matically extracts patent indicators based on the proposed methodology. The extracted indicators are application length (be-tween application date and issue date using the unit month), the number of international patent classification (IPC), US patent clas-sifications (UPC), forward citations, foreign citations, backward citations, claims, independent claims, patent family (i.e., a set of patents in various countries taken to protect a single invention), technology cycle time[6], science linkage [7,28], and the length of detailed specification, as shown inTable 3.

Input Layer Output Layer Training weight Hidden Layer i X j H h ij w k O o jk w

Input nodes (patent indicators)

Output nodes (patent quality) Patent No. Indicator 1 Indicator 2 Indicator N Status 1 High 2 3 Low

Fig. 4. The BPN network model structure.

Table 3

Sample values of indicators from ﬁve patents.

Patent indicators US Patent No.

5075742 5101478 5131006 5146465 5151920 1 Application length 11.6 44.5 18.9 19.5 12.7 2 Number of IPC 8 2 3 9 7 3 Number of UPC 5 1 4 3 2 4 Forward citations 3 25 3 4 6 5 Foreign citations 1 2 7 0 0 6 Backward citations 1 2 1 4 1 7 Number of claims 4 24 14 14 10 8 Independent claims 1 7 4 1 7 9 Patent family 5 7 9 2 1

10 Technology cycle time 3 8 1 1 2

11 Science linkage 1 0 1 0 0

(7)

4.1. Evaluation of patent indicators

The Kaiser–Meyer–Olkin approach is used to evaluate the appropriateness of the principal component analysis. The KMO analysis yields a value of 0.721 from the extracted 12 patent indi-cators. The eigenvalues, variance and cumulative variance of com-ponents are shown inTable 4. This research assumes that the value of cumulative variance should be above 70% to sufficient collection of principal components. The cumulative variance represents the explanation of components. Therefore, the top five principal com-ponents are sufficiently chosen for building the patent quality model. The five components explain and represent the patent

indi-cators respectively in various strength as calculated and listed in Table 5.

As described in Section 3, the thresholds of indicators for explaining different sample sizes are shown inTable 2. The case study collects 399 patents. Thus, the threshold of indicator of the case study should be above 0.3. Thus, the ﬁrst component selects ‘‘Forward Citation’’, ‘‘Claims’’ and ‘‘Independent Claims’’ (as under-lined in column one,Table 5). The second component selects ‘‘IPC’’, ‘‘UPC’’, and ‘‘Backward Citation’’. The third component selects ‘‘Technology Life Cycle’’, and ‘‘Science Linkage’’. Moreover, the fourth component selects ‘‘Foreign Citation’’ and ‘‘Patent Families’’. Finally, the ﬁfth component only selects ‘‘Application length’’. In

Table 4

The principal component analysis of extracted patent indicators.

Components ₁ ₂ ₃ ₄ ₅ 6 7 8 9 10 11 12

Eigenvalues 3.47 1.73 1.42 1.18 0.90 0.88 0.71 0.61 0.48 0.39 0.12 0.07

Variance (%) 28.91 14.42 11.83 9.84 7.56 7.33 5.95 5.14 4.06 3.29 1.02 0.60

Cumulative variance (%) _28.91 _43.33 _55.16 _65.01 _72.58 79.92 85.87 91.01 95.07 98.37 99.39 100

Table 5

Five principal components collectively represent the indicators in various degrees.

Patent Indicator Component Matrix

Component 1 Component 2 Component 3 Component 4 Component 5

Application Length 0.053 0.058 0.091 0.036 _0.933 IPCs 0.232 0.672 0.339 0.168 0.272 UPCs 0.161 _0.738 0.197 0.376 0.101 Foreign Citations 0.142 0.257 0.072 _0.505 0.113 Forward Citations 0.309 0.343 0.006 0.150 0.101 Backward Citations 0.232 0.516 0.372 0.498 0.176 Claims _0.814 0.108 0.525 0.148 0.57 Independent Claims 0.863 0.185 0.405 0.449 0.196 Patent Families 0.115 0.131 0.253 0.678 0.349

Technology Cycle Time 0.145 0.104 0.913 0.103 0.552

Science Linkage 0.019 0.085 0.414 0.269 0.114

The length of speciﬁcation 0.162 0.049 0.075 0.035 0.254

Training Pass

Learning Rate

Momentum

Iteration

Error Rate

Train 0.17

1 0.1 0.8

8000

Test 0.42

Train 0.18

2 0.2 0.8

8000

Test 0.35

Train 0.16

3 0.2 0.8

10000

Test 0.29

The Training

Error Rate

Iteration

(8)

summary, this case study selects eleven key indicators (excluding indicator ‘‘Length of speciﬁcation’’). Thereafter, values of these ele-ven indictors are extracted from training patent documents to train the patent quality model using BPN approach.

4.2. Training the patent transaction model

The extracted indicators are reduced to 11 signiﬁcant indica-tors. These indicators are then used as the input layer for training the BPN network model. The nodes of the output layer contain the trading quality of the patents and the non-trading quality of pat-ents. The case study prepares 260 training patents (182 sold or licensed and 78 unsold patent documents) to train the patent transaction model. The test patents include 101 sold or licensed patents and 38 unsold patent documents. The training parameters include learning rate, momentum, and iteration. The results of the patent transaction model are shown inFig. 5.

4.3. Veriﬁcation of patent quality model using historical patent cases

After analyzing the principle components and building the pat-ent transaction model, this research collects total of 36 historical patents as test data to verify the proposed methodology. These pat-ents include 22 sold or licensed patpat-ents and 14 unsold patpat-ents. Ele-ven network communication patents were sold from patent owners to Stragent LLC – a company focusing on development, acquisition and licensing of patented technology[29]. Further, ele-ven computer technology patents were licensed from Taiwan ITRI to ACER Computer using for litigations[30]. These patents, consid-ered as high quality patents, are traded by organizations (e.g., Stragent, ACER, and ITRI) to pursuit or defend legal cases. The test-ing sample also includes 9 light switch patents owned by Lutron Electronics Company[31]and unsold LED patents owned by Lite-panels, LLC[32]. The trading statuses (licensed, sold or unsold) of these 36 patents are depicted inTable 6.

The trained patent transaction model is used to analyze the potentials of these 36 historical patents (purchased, licensed or un-sold). The results of the tradability-based patent quality analysis are shown inTable 7.

The predictions for the 36 patents include 17 in the strong/high, 4 medium, and 15 low quality ranges. Eighteen out of 22 patents sold or licensed are classified by the model as high quality, while 11 out of 14 patents unsold are classified as low quality and low potential for patent transactions. The matching between the ana-lytical prediction and the actual trading status reached 85% accu-racy. Therefore, the proposed patent quality analysis can be effectively used to evaluate the quality of unknown patents for pre-evaluation and preliminary screening. After extracting the high quality patents, the R&D engineers can confirm the patent claims and advances of inventions.

5. Conclusion

Patent news shows that enterprises often purchase technology specific patents to advance technology or make new products. Enterprises use patents to protect their innovations and to estab-lish a time period of protection. However, traditional patent anal-ysis requires significant costs, time, and manpower to evaluate the quality of patents. Thus, the purpose of this research is to shorten the time required to determine and rank the quality of pat-ents with respect to their potential values in IPR marketplace. The proposed patent quality analysis uses principal component analy-sis to identify critical patent indicators. These indicators in turn are used to build the patent quality model using BPN network ap-proach. The research has made contribution in applying the patent quality model in real application. The case study uses data set of 399 patents in five high-tech domains to identify the most suitable 11 key indicators. Then, 260 patents’ data set are used for training the BPN model and, finally, 36 historical patents to test the model. The case has yield 85% accuracy in patent quality prediction, which is considered a valid result for automatically pre-evaluating huge

Table 6

The information of the patents sold and unsold.

Current assignee Domain Type Patent no.

Stragent LLC Network

communication

Purchased from inventors

US6848972; US7028244; US7320102; US6832226; US6665722; US6393352; US6285945; US6604043; US7289524; US7543077; US7095753

ACER Computer

technology

Licensed from ITRI US5977626; US6188132; US6280021; US6788257; US5101478; US6075686; US5903765; US5870613; US5410713; US5214761; US5581122

Lutron Electronics Company

Light switches Unsold US4893062; US5248919; US5637930; US5905442; US5949200; US7190125; US5982103;

US4797599; US5736965 Litepanels, LLC Light emitting

diode (LED)

Unsold US6948823; US7163302; US7510290; US7429117; US6749310

Table 7

The results of the tradability-based patent quality analysis.

Patent no. Assignee Strength Score

US6188132 ACER Strong 98.5

US6285945 Stragent LLC Strong 98.4

US5949200 Lutron Electronics Company Strong 92.3

US5736965 Lutron Electronics Company Strong 88

US6832226 Stragent LLC Medium 67.2

US5410713 ACER Medium 66.3

US6665722 Stragent LLC Medium 66.2

US7163302 Litepanels, LLC Medium 64.9

US7289524 Stragent LLC Low 36.5

US5101478 ACER Low 33

US5214761 ACER Low 32.2

US4893062 Lutron Electronics Company Low 32

US5982103 Lutron Electronics Company Low 29.1

US7510290 Litepanels, LLC Low 28.8

US6788257 ACER Low 25

(9)

number of patents for commercialization. The patent quality pre-diction and methodology can be further reﬁned by adding other criteria and factors for patent quality evaluation. More case studies should be conducted for the practical applications when many companies are increasingly concern about patent infringement and IPR litigations, particularly for rapid product development and mass customization.

Acknowledgements

The authors thank the referees for their thorough review and valuable suggestions that helped improve the quality of the paper. This research was partially supported by the National Science Council research grants.

References

[1] R. Paci, A. Sassu, S. Usai, International patenting and national technological specialization, Technovation 17 (1) (1997) 25–38.

[2] V.K. Gupta, Technological trends in the area of fullerenes using bibliometric analysis of patents, Scientometrics 44 (1) (1999) 17–31.

[3] B. Yoon, Y. Park, A text-mining-based patent network: analysis tool for high technology trend, Journal of High Technology Management Research 15 (2004) 37–50.

[4] A.J.C. Trappey, C.V. Trappey, C.Y. Wu, Automatic patent document summarization for collaborative knowledge systems and services, Journal of Systems Science and Systems Engineering 18 (1) (2009) 71–94.

[5] K.K. Lai, S.J. Wu, Using the patent co-citation approach to establish a new patent classiﬁcation system, Information Processing and Management 41 (2) (2005) 313–330.

[6] F. Narin, Patent bibliometrics, Scientometrics 30 (1) (1994) 147–155. [7] F. Narin, K. Hamilton, D. Olivastro, The increasing linkage between US

technology and public science, Research Policy 26 (1997) 317–330. [8] J.A. Barney, Statistic measurement of patent quality, Ocean Tomo, LLC.<http://

www.oceantomo.com>, 2010 (retrieved 25.12.2010).

[9] P. Klemperer, How broad should the scope of patent protection be?, The RAND Journal of Economics 21 (1) (1990) 113–130

[10] J.R. Allison, M.A. Lemley, K.A. Moore, R.D. Trunkey, Valuable patents, Georgetown Law Journal 92 (2004).

[11] D. Hegde, B. Sampat, Examiner citations, applicant citations, and the private value of patents, Economics Letters 105 (3) (2009) 287–289.

[12] M. Trajtenberg, R. Henderson, A. Jaffe, University versus corporate patents: a window on the basicness of inventions, Economics of Innovation and New Technology 5 (1) (1997) 19–50.

[13] J.O. Lanjouw, M. Schankerman, Characteristics of patent litigation: a window on competition, RAND Journal of Economics 32 (1) (2001) 129–151.

[14] K. Pearson, On lines and planes of closest ﬁt to systems of points in space, Philosophical Magazine 2 (6) (1901) 559–572.

[15] H. Hotelling, Analysis of a complex of statistical variables into principal components, Journal of Educational Psychology 24 (1933) 417–441. [16] S.Y. Chen, Multivariate Analysis, 4th ed., Hwa-Tai Publishing Co., Taipei,

Taiwan, 2005, ISBN 957-412-759-1.

[17] Y.H. Lai, H.C. Chu, Modeling patent legal value by extension neural network, Expert Systems with Applications 36 (7) (2009).

[18] T.F. Tsai-Lin (Advisor: Prof. Y.C. Chang), Patent evaluation and clustering strategy, M.S. Thesis, Department of Institute of Technology Management, National Tsing Hua University, Hsinchu, Taiwan, 2009.

[19] H.C. Che, Y.H. Lai, S.Y. Wang, Assessment of patent legal value by ⁄⁄regression and back-propagation neural network, International Journal of Systematic Innovation 1 (2009) 31–47.

[20] A.J.C. Trappey, F.C. Hsu, C.V. Trappey, C.I. Lin, Development of a patent document classiﬁcation and search platform using a back-propagation network, Expert Systems with Applications 31 (2006) 755–765.

[21] C.V. Trappey, H.-Y. Wu, F. Taghaboni-Dutta, A.J.C. Trappey, Using patent data for technology forecasting: China RFID patent analysis, Advanced Engineering Informatics 25 (1) (2011) 53–64.

[22] T.A. Chiang, C.Y. Wu, A.J.C. Trappey, C.V. Trappey, Applying BPANN and hierarchical ontology to develop a methodology for binary knowledge document classiﬁcation and content analysis, in: The 14th International Conference on Computer Supported Cooperative Work in Design (CSCWD), Shanghai, China, 2010.

[23] H.F. Kaiser, The application of electronic computers to factor analysis, Educational and Psychological Measurement 20 (1) (1960) 141–151. [24] J.F. Hair, R.E. Anderson, R.L. Tathan, W.C. Black, Multivariate Data Analysis, 5th

ed., Prentice Hall, Upper Saddle River, New Jersey, 1998, ISBN 013-894-858-5.

[25] G. Cybenko, Approximation by superpositions of a sigmoidal function, Mathematical Control, Signals, and Systems (MCSS) 2 (4) (1989) 303–314. [26] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with artiﬁcial neural networks: the

state of the art, International Journal of Forecasting 14 (1) (1998) 35–62. [27] R. Beale, T. Jackson, Neural Computing: An Introduction, Adam Hilger, IOP

Publishing Ltd, Bristol, 1990, ISBN 085-274-262-2.

[28] D. Hicks, T. Breitzman, D. Olivastro, K. Hamilton, The changing composition of innovative activity in the US a portrait based on patent analysis, Research Policy 30 (2001) 681–703.

[29] Science and Technology Policy Research and Information Center (STPI), Stragent LLC patent litigation. <http://cdnet.stpi.org.tw/techroom/pclass/ 2010/pclass_10_A075.htm>, 2010 (retrieved 25.12.10).

[30] Science and Technology Policy Research and Information Center (STPI), Acer purchases patent to litigate HP. <http://cdnet.stpi.org.tw/techroom/pclass/ 2008/pclass_08_A043.htm>, 2010 (retrieved 25.12.2010).

[31] Science and Technology Policy Research and Information Center (STPI), Lutron Electronic Co. patent litigation. <http://cdnet.stpi.org.tw/techroom/pclass/ 2009/pclass_09_A021.htm>, 2010 (retrieved 25.12.2010).

[32] Science and Technology Policy Research and Information Center (STPI), Litepanels LLC LED patent litigation. <http://cdnet.stpi.org.tw/techroom/ pclass/2009/pclass_09_A024.htm>, 2010 (retrieved 25.12.2010).