Intelligent patent recommendation system for innovative design collaboration

(1)

Intelligent patent recommendation system for innovative

design collaboration

Amy J.C. Trappey

a,n

, Charles V. Trappey

b

, Chun-Yi Wu

a

, Chin Yuan Fan

c

, Yi-Liang Lin

a

Department of Industrial Engineering and Engineering Management, National Tsing Hua University, Taiwan b_{Department of Management Science, National Chiao Tung University, Taiwan}

c

Science & Technology Policy Research and Information Center, National Applied Research Laboratories, Taiwan

a r t i c l e i n f o

Article history: Received 11 August 2012 Received in revised form 27 December 2012 Accepted 8 February 2013 Available online 15 March 2013 Keywords:

Design collaboration Patent search Behavior records

Patent recommendation system Collaborativeﬁltering algorithm

a b s t r a c t

Patents' search is increasingly critical for a company's technological advancement and sustainable marketing strategy. When most innovative designs are created collaboratively by a diverse team of researchers and technologists, patent knowledge management becomes time consuming with repeated efforts creating additional task conflicts. This research develops an intelligent recommendation methodology and system to enable timely and effective patent search prior, during, and after design collaboration to prevent potential infringement of existing intellectual property rights (IPR) and to secure new IPR for market advantage. The research develops an algorithm to dynamically search related patents in global patent databases. The system clusters users with similar patent search behaviors and, subsequently, infers new patent recommendations based on inter-cluster group member behaviors and characteristics. First, the methodology evaluates the filtered information obtained from collaborative patent searches. Second, the system clusters existing users and identifies users' neighbors based on the collaborative filtering algorithm. Using the clusters of users and their behaviors, the system recommends related patents. When collaborative design teams are planning R&D policies or searching patents and prior art claims to create new IP and prevent or settles IP legal disputes, the intelligent recommendation system identifies and recommends patents with greater efficiency and accuracy than previous systems and methods described in the literature.

1. Introduction

Intangible assets such as Intellectual Property Rights (IPR) and trademarks are a significant part of a modern enterprise's net worth. In particular, the intellectual property registered interna-tionally as patents are used to legally protect the proprietary technology and insure the market advantage of thefirm, promote further commercialization, royalty, licensing, and sales. When new technology is developed and a patent is issued, official claims insure that the assignee(s) maintain their competitiveness by preventing others from using the patented technology without prior permission. As declared by US Patent Law (US Patent Act), patents are to be used to encourage and promote commercial development by providing legal protection. Whoever without authority makes, uses, imports, or sells any patented invention during the term of the patent stands in violation of the law and infringes upon the patent assignee. Certainly, companies with high quality patents hold a competitive and sustainable market position.

The World Intellectual Property Organization (WIPO) reports that the numbers of applications have reached a record high of a half million patents per year. When companies attempt to search, interpret, compare, and classify patent documents, they are over-whelmed by the difﬁculty of reviewing, analyzing, and synthesiz-ing the illustrations, information, claims, and technical knowledge. Thus, computer assisted patent information and knowledge man-agement systems are needed to facilitate the manual processing, organization, and knowledge management of relevant patents.

Companies and individual inventors can win or lose substantial profits and market advantage if their innovative designs unknow-ingly conflict or infringe upon existing technology or if others are misappropriating their pre-existing claims. There is a constant need to search, review, and interpret patents in various patent databases to understand inventions and prior arts (1) prior to creating new patent applications or new product commercializa-tion and (2) to maintain legal authority over existing IP. Therefore, patent search and recommendation methodologies are a critical function of a patent knowledge system and knowledge manage-ment program. When users lack specific domain knowledge, their patent search efforts are often impeded and result in useless searches. Patent documents use a plethora of domain concepts to describe the invention for comprehensive IP right protection. Thus, it is difficult for the majority of patent researchers (especially in a Contents lists available atScienceDirect

journal homepage:www.elsevier.com/locate/jnca

Journal of Network and Computer Applications

n_{Corresponding author. Tel.: þ886 933035375; fax: þ 886 35722204.}

E-mail addresses: [email protected] (A.J.C. Trappey),

[email protected] (C.V. Trappey), [email protected] (C.-Y. Wu), [email protected] (C.Y. Fan), [email protected] (Y.-L. Lin).

(2)

multi-lingual global IP environment) to accurately search for and retrieve related patent documents without using a computer based and intelligent recommendation system.

While the numbers of global patent applications continues to grow, the patent recommendation system must be useful for finding related patents rapidly and effectively. Thus, this research provides an intelligent recommendation methodology based on the records of the users' search behavior. The research implements a collaborative filtering algorithm and a clustering method to construct the intelligent patent recommendation system platform. When general users search patents, the platform automatically identifies and recommends related patents. This paper is orga-nized into several sections. InSection 2, the related literature is reviewed with a focus on recommendation systems and web search methods. In Section 3, the proposed recommendation methodology and algorithms are formulated and described.

Section 4describes the system framework, prototype implemen-tation, and the case study based on a solar cell technology patent search with recommendation results. Finally, theSection 5 sum-marizes the research outcomes, contributions, and provides direc-tions for future research.

2. Literature review

Patents protect the invention and intellectual property rights (IPR) claimed by the inventor. Inventors write their research and development (R&D) results or innovative achievements as patent documents following a required patent document format and specification guideline. The patent documents are then evaluated for qualification by the issuing patent office. During the issuance time, the patent owner (also called the patent assignee) can take legal action to prevent unauthorized manufacturing, selling or usage of the invention. This section describes related literature covering patent search and classification, patent analysis, and patent recommendation systems.

Enterprises often use various patent databases to search patent documents, e.g., the United States Patent and Trademark Office (USPTO), the European Patent Of_{fice (EPO), and the World} Intel-lectual Property Organization (WIPO). In addition to these national and international open databases, there are integrated databases which require subscription fees or database purchases, such as Delphion and EPO PATSTAT. These databases generally use three functions in the patent search, i.e., a simple keyword search, an advanced metadata search (e.g., keywords, assignee, inventors, and year), and a patent number search. When users search patents without definite directions or specific keywords, they usually find the search results with many type-1 or type-2 errors where they either miss patents or retrieve the wrong patents. Therefore, many automatic classification methods have been proposed to help facilitate patent search. If patents are classified before searching, time and cost of search can be reduced. The EPO uses an auto-classification method based on the k-nearest neighboring and clustering algorithm (Krizer and Zacca, 2002). Researchers have depicted that the patent classification can be more accurate by considering both the patent metadata and the full-text of the patent (Richter and MacFarlane, 2005). Trappey et al. (2010)

propose a non-exhaustive clustering method to group patent documents into overlapping clusters. This approach determines whether a given patent can possibly be categorized into multiple clusters, which is consistent with the principle of a patent allowing multiple claims. Further,Chiang et al. (2011)developed an intelligent system for automated binary knowledge document classiﬁcation and content analysis. The system is constructed using a back-propagation artiﬁcial neural network, hierarchical ontology, and normalized term frequency methods to improve patent

classification in a binary and hierarchical fashion. Thus, the system iteratively identifies patents until a sufficiently reduced number of highly related patents are collected.

After patent search and classification, the collected patents are analyzed to extract detailed information. Patent analysis includes the identification of technology trends as well as the performance of legal due diligence. There are four general analytical approaches used including time series analysis, patent citation analysis, inter-national patent classifications (IPC) analysis, and the construction of patent maps. Time series analysis is used to analyze the change in the number of patents applied for overtime which in turn can be related to the development of the technology life cycle. Patent citations determine the relation between two different patents by using their citations and other references. For various technolo-gies, IPC analysis helps determine which technology is being or not being developed. If some IPC regions are limited to a small numbers of patents, then there may be a R&D bottleneck and business potential if a solution can be found. The result of IPC analysis provides information to support government or enterprise R&D strategy. Finally, patent maps are used to depict the potential technological relationships between two different patent groups. The patent map describes different patent groups that belong to technology groups or assigneesHuang and Li (2010).

User feedback is a type of rating or voting method for users to express their personal satisfaction with a search. This behavior is divided it into explicit or implicit feedback (Nichols, 1997). The explicit feedback represents information provided directly by the user such as personal information, common replies, survey results, and work experience. Implicit feedback is collected indirectly from user and is usually extracted from user's browsing records or query logs.Oard and Kim (1998) report that the user's implicit informa-tion can be extracted and classiﬁed as three behavioral types which in turn serve as the inference data for the recommendation system.

Researchers have also considered different techniques in de fin-ing user behaviors and methods to trace their queries and feed-back. For example, a service match maker plays an important role in ensuring the connectivity between the user and the service provider. However, the lack of relevant service domain knowledge and incorrect service queries prevents the semantic service match makers from identifying the service concepts that correctly repre-sent the service requests.Wang et al. (2007)demonstrate that the information coverage and update problems are a common bottle-neck for current web search engines. These problems limit the offers of service and make the resolution of complaints difficult to achieve. To solve these problems, a new search algorithm based on DNS is proposed in their research. This system adopts a layered distributed architecture, similar to DNS, which is different from current commercial search engines.Dong et al. (2011)present a novel semantic similarity model for describing the service ontol-ogy environment whereasBouras and Poulopoulos (2012) propose a web personalization mechanism based on dynamic creation and automatic updates of user profiles to better match users prefer-ences. This approach assumes that a user's profile is affected by other user's grouping details which are constructed with similar profiles. As a result, a real-time user-centric document grouping mechanism is implemented to support the web personalization system and provide data for experimental evaluation of the system.

Researchers (de la Torre-Diez et al., 2013) use generic and selective filters set up by the administrator from the module RSS_PROYECT installed in Joomla. The genericfilter allows a search of the words included in a series of sources indexed by the user. Thefilter categorizes all sources that contain the word without exception. Different languages such as PHP, MySQL, HTML, XML and the Application Program Interface (API) of Joomla were used

(3)

to evaluate the results. The results are favorable for the selective ﬁlter and strongly favorable for the generic ﬁlter. Better than average processing times were obtained for RSS_PROYECT with respect to other modules using Joomla.

Li et al. (2012) develops a novel framework which is presented and implemented for classifying patents according to the levels of invention as deﬁned by TRIZ theory.Liang et al. (2012) proposed an Issue, Solution and Artifact Layer (ISAL) model for design rationale representation. The research focuses on algorithm design to discover design rationale from design documents according to ISAL modeling. Moreover,Liang et al. (2012) use text mining as their primary method for analyzing issues, solutions and artifact layers. Text mining is useful for identifying keyword terminology, but requires additional analysis to correlate and organize the information.Li et al. (2012) use TRIZ as a method for identifying problem solutions. However, TRIZ must consider many different strategies, such as level of invention, which makes the system less practical for non-domain specialists.

Most recommendation approaches utilize implicit data extracted from user's behavior records. Kelly and Teevan (2003) integrate the research proposed by Oard and Kim (1998) and divide search behaviors intoﬁve different types including exam-ine, retain, reference, annotate, and create. For the above approach, it is necessary to record the user's operational history such as searching, browsing, markup, or editing using IT techniques and then transform and save the records in a standard format. From these records, a user's personal proﬁle is created.

Implicit data can also be utilized by a recommendation system.

Lee et al. (2008)use the opening time of commodities sold and the time that consumer's buy as a data source, analyze the consumer's preferences, and then make recommendations for further com-modities trading. The research considers that when buying times are closer together, the commodity betterfits consumer demand. Given the commodity's open selling time, commodities are assigned different weights, and inferences are made that there are other commodities that buyers might be in interested and recommendations are made. This type of data extraction helps avoid personal and subjective influence and is sufficient for model training.

Patent specific search and analysis may be adapted to similar recommendation systems. After a patent search, a user may have difficulty in identifying patents of interest or related patents. Thus, this research proposes an intelligent recommendation methodology and system for patent search. A recommendation system is con-sidered to be an informationfiltering system that effectively reduces the cost and time of search activities. Resnick and Varian (1997) report that informationfiltering systems are widely applied in e-business and help users grasp information rapidly. The recommen-dation systemfilters and analyzes the feedback of users and helps with the specification and classification of items. The system is constructed as a dynamic model to collect information based on user_{'s requirements. Moreover, most recommendation systems are} based on two types of mechanisms (Ansari et al., 2000). One type uses collaborativefiltering and the other uses content based filter-ing. This research focuses on the collaborativefiltering since it relies on social or personal network recommendations. This mechanism clusters the target and similar behavior of other users in a closed group. By defining the group's common interests and similar behavior model, the mechanism infers the information or products (in this case, patents) that are of common interest.Bhavnani et al. (2008)propose a qualitative study of experienced patent searchers. The research assumes that the professional searchers will use well-formed search strategies that can rapidly and effectively search for the related patents and identify the novelty of invention. Many researchers have developed mechanisms to build recommendation systems for other applications. There is one for recommending

movies to viewers which merges combinations of features and attributes (Nazim-uddin et al., 2009). de Campos et al. (2010)

propose a hybrid system which uses a Bayesian network model to determine the weights of the target search and then provides a recommendation result.Barragáns-Martínez et al. (2010)propose a novel Web 2.0 TV program recommendation system. The hybrid approach combines contentfiltering techniques with collaborative filtering while providing the advantages of a social network. In order to eliminate the most serious limitations of collaborativefiltering, an item-based collaborative _{filtering algorithm was implemented to} improve performance. The resulting application simplifies the task of selecting programs to watch on TV.

3. Methodology

This research accumulates users_{' behavior records and the} related patent search information and applies collaborative filter-ing to recommend patents. The research methodology framework (Fig. 1) is divided into three parts which defines the behavior types and then records users' behavioral operations based on pre-defined types. The methodology summarizes the users' behavior records based on specific search conditions and when the target user searches patents, the methodologyfilters the results for the patent recommendation system. The following section describes the patent collaborative filtering process as the core of the dynamic patent recommendation methodology.

3.1. User behavior record analysis

In order to deﬁne the users' operational complexity under differing search conditions,ﬁve behavior types classify and record the user' behavior over time. The users' searching and viewing frequencies are accumulated and the analysis records and histor-ical bookmark records are logged. Finally, these data and records are exported to the users' behavior database. The behavior types,

Patents Collaborative Filtering Calculate OC Matrix Fixed Target User Selected Patents Recommended Results User Behavior Records Candidate Patents User Operating Save Users Clustering Patents Filtering Find Neighbor Users Target User Searching Save Evaluate Patents Correlations Select Recommend Patents Method Flow

User Operate and System Records Data Access

Behavior Records Extracting

Patents

(4)

their sub-behaviors, notations, and weights are presented in

Table 1.

The search is divided into four sub-behavior types by the search type and weights are assigned according to the search complexity. Bookmarks are divided into four sub-behavior types based on the patent search approach. The weights of the book-mark sub-types reflect the operation complexity. The algorithm distinguishes between single patent views and comparative patent views. The comparative patent view is used to select two patents of interest for simultaneous analysis. Thus, a lower weight is assigned to the behavior of single views and higher weights are assigned to the comparison views. The sub-behaviors, such as patent charting and patent quality analysis (Trappey et al., 2012) are an additional behavioral sub-type analysis. Patent charting analyzes the metadata attributes from patents and includes statistical patent trends based on the international patent classi-fication (IPC), assignees, and countries of applicants. The research further creates a two-dimension chart and a three-dimension chart based on the analyses. The more dimensions considered, the greater weight assigned to the behavior. Patent quality analysis is used to evaluate the patent value based on patent indicators. Thus, the patent quality analysis sub-type is given the highest weight among the analysis sub-types. Finally, export behavior is divided into general patent list exports and specific single patent exports. The former exports a series (a list) of all patent informa-tion and the latter only extract a single patent's information. Since selecting a single patent requires in-depth understanding and interest, a higher weight is assigned to the single patent export. The user behavior record is used by the system to evaluate each user's patent search process and grade the operational complexity. 3.2. Operation complexity function definition

The Operation Complexity (OC) value defined byEdwards and Barron (1994) is used to assign the weights of the pre-defined behavior types in the previous section (Table 1). The patent's attributes and patent search approach defines several data items such as patent number (PN), international patent classification (IPC), inventors (IN), assignees (AN), key phrase (KP), industry (INDT), and technology type (TECH). The OC function is shown in Eq.(1)and the variables are defined inTable 2.

OC ¼ ðWs S þWV V þWA AþWB BþWE EÞ ð1Þ

3.3. User cluster analysis

After collecting the users’ behavior records, the K-medoids clustering algorithm is used to group the target user and neighbors using the operation complexity (OC) matrix shown inTable 3. K-medoids minimizes outliers from over inﬂuencing the clustering result and better selects a real user as a cluster center (not a pseudo center as can occur with the K-means approach).

K-medoids helps decrease the calculation time. Furthermore, using an actual user as the cluster center is more appropriate when analyzing non-numerical data (Basumallick and Wong, 1996). The algorithm selects k objects as the initial centers with the target user (T) among them. Then the distances between the object and centers are calculated and the cluster with the shortest distance between its center and the object is selected. After finishing the first iteration, the algorithm randomly selects the object to replace its original center. If the new center produces better cluster results, then the clustering calculation continues If not, the clustering algorithm stops and provides thefinal result.

Different clustering results are generated with different cluster numbers (k). To better select k, the minimizing root mean square standard deviation (RMSSTD) and maximizing R-squared (RS) value are used as recommended bySharma (1996). RMSSTD is used to calculate the data_{'s homogeneity within each cluster. A} smaller RMSSTD shows higher homogeneity within clusters. RS is used to measure the average divergence of two different clusters. The larger RS represents larger differences between clusters which is the desired outcome. The formulas for RMSSTD and RS are shown in Eqs.(2) and (3).

RMSSTD ¼ ∑i ¼ 1…nc j ¼ 1…v ∑nij k ¼ 1 ðxk−xkÞ2 ∑i ¼ 1…nc j ¼ 1…v ðnij−1Þ 2 6 6 4 3 7 7 5 ð2Þ nc: cluster numbers, v: data's dimension Table 1

Behavior type and weight.

Behavior Sub Behavior Notation Weight

Search Custom search S1 WS1

Patent number search S2 WS2

Industry patent search S3 WS3 Technology patent search S4 WS4

View Patent view V1 WV1

Patent comparison V2 WV2

Analysis Two-dimension chart A1 WA1

Three-dimension chart A2 WA2

Patent quality analysis A3 WA3

Bookmark Bookmark after B1 WB1

Custom search

Bookmark after B2 WB2

Patent number search

Bookmark after B3 WB3

Industry patent search

Bookmark after technology patent search B4 WB4

Export Patent list export E1 WE1

Single patent export E2 WE2

Table 2

OC function variables’ deﬁnition and weight.

Variable Deﬁnition Weight

OC Operation complexity –

S Search behavior Score WS

V View behavior Score WV

A Analysis behavior score WA

B Bookmark behavior score WB

E Export behavior score WE

Table 3

Operation complexity matrix.

User Search Condition (SC)

SC(1) SC(2) SC(j) … SC(s) User1 OC11 OC12 OC1j … OC1s User2 OC21 OC22 OC2j … OC2s Useri OCi1 OCi2 OCij OCis … … … … Userm OCm1 OCm2 OCmj … OCms where

OCTjis the target user's operation complexity for the jth search condition; Index T represents the target use;

Index j represents the jth search condition, and the total number of conditions is s, jo ¼s; and

(5)

nj: dimension j's data numbers

nij: dimension j's data numbers in cluster i

RS ¼SSb SSt ¼ SSt−SSw SSt ¼ f∑j ¼ 1…v½∑nk ¼ 1j ðxk−xkÞ2g−f∑i ¼ 1…c j ¼ 1…v ½∑nij k ¼ 1ðxk−xkÞ2g ∑j ¼ 1…v½∑nk ¼ 1j ðxk−xkÞ2 ð3Þ

SSb: sum of square between the clusters

SSt: total sum of square

SSw: sum of square within the cluster

3.4. Filtering recommended patents

Since the number of searched patents is quite large, the matrix is dif_{ficult to build effectively for system processing. After user} clustering, the patents' correlations within a given cluster are calculated as the basis for recommending related patents. Based on the clustering result, the target user and the target user's neighbors are identified as a group. The patents selected by the target user's neighbors are called the selected patents (SPs). Frequently appearing IPCs of selected patents are summarized and the other patents which contain the same IPCs from the database are chosen as the candidate patents (CPs). Thus, patents with little relationship arefiltered and eliminated to avoid com-plex calculations.

Afterﬁnding neighboring users and CPs, the target user and the neighbors' behavior records are summarized. The collaborative operation index is calculated (Table 4) using notations shown and explained inTable 5. The collaborative operation index ij repre-sents the operating count of ith patent and jth patent under different collaborative operations. The ith patent and jth patent can be the selected patents (S1, S2… Sm) or the candidate patents (C1, C2… Cn). Thus, the correlation score (called Co Score) is calculated individually by multiplying the pre-deﬁned weights and the count of each collaborative operation index (Formula (4)). After calculating all patent combination correlation scores, the

scores are normalized using the maximum value based on colla-borative operation indexes.Table 6shows the patent correlation matrix, where DCij represents the degree of correlation between ith patent and jth patent.

Co score ¼ WCoV1 CoV1þWCoV2 CoV2þ⋯þWCoE2 CoE2 ð4Þ

Through this process, the candidate patents which are highly related with the selected patents are identiﬁed. The recommended patents are sorted by the highest score calculated by the degree of the patents' correlation compared with each selected patent. The ﬁnal result is a method to automatically recommend patents to the user.

4. System construction and case study

The system integrates patent search and patent recommenda-tion funcrecommenda-tions. There are four modules within the patent recom-mendation unit. These four modules are the user's behavior record, related user clustering, patent_{ﬁltering, and system} para-meter management. The system framework describes the relation between user, system platform, and database (Fig. 2). The modules and the detailed actions of the patent recommendation function are shown inFig. 3.

The recommendation system was built for the solar energy alliance (http://www.wheeljet.com.tw/CIGS/) as a case study to validate the methodology. The system collects the related solar technology patents from the USPTO database and includes the key functions of patent searching, technology classiﬁcation, industry classiﬁcation, patent analysis, and patent bookmarking. The sys-tem also records the users' operation information for patent recommendation calculations. In the case study, the system collects twenty members_{' behavioral information which is used} to cluster the related users and to recommend the potential non-searched patents to a given user.

The patent recommendation system is an innovative platform which recommends speciﬁc and relevant patents based on the previous search results of clustered peers. Thus, the system considers different user situations and not only the experimental results of the domain experts. Moreover, in the case study,

Table 4

Collaborative operation indexes. Patents Indexes

Patent i Patent j CoV1 CoV2 CoA1 … CoE2 S1 S2 CoV1S1,S2 CoV2S1,S2 CoA1S1,S2 … CoE2S1,S2 S2 C1 CoV1S2,C1 CoV2S2,C1 CoA1S2,C1 … CoE2S2,C1 S1 C1 CoV1S1,C1 CoV2S1,C1 CoA1S1,C1 … CoE2S1,C1

… … … …

Sm Cn CoV1Sm,Cn CoV2Sm,Cn CoA1Sm,Cn … CoE2Sm,Cn

Table 5

Collaborative operation description and notation.

Operation Description Notation

Single patent view Select speciﬁc patent to view the detailed content CoV1

Patent comparison Choose any two patents to make comparisons CoV2

Two-dimension patent chart Choose some patents and two attributes to draw relationship chart CoA1 Three-dimension patent chart Choose some patents and three attributes to draw relationship chart CoA2 Patent quality analysis Select several patents for patent quality evaluation CoA3

Patent bookmark Bookmark the patents after search CoB1

Non-removed patent bookmark Do not remove the original bookmark patents after search CoB2 Patent list export Choose some patents and export their important information CoE1 Single patent export Select speciﬁc patent and export the complete patent information CoE2

Table 6

Patent correlation matrix.

SP(1) SP(2) … SP(m) CP(1) CP(2) … CP(n) SP(1) DCS1,S2 … DCS1,Sm DCS1,C1 DCS1,C2 … DCS1,Cn SP(2) DCS2,S1 DCS2,Sm DCS2,C1 DCS2,C2 … DCS2,Cn … … … … SP(m) DCSm,S1 DCSm,S2 … DCSm,C1 DCSm,C2 … DCSm,Cn CP(1) DCC1,S2 DCC1,S2 … DCC1,Sm DCC1,C2 … DCC1,Cn CP(2) DCC2,S2 DCC2,S2 … DCC2,Sm DCC2,C1 … DCC1,Cn … … … … CP(n) DCCn,S1 DCCn,S2 … DCCn,Sm DCCn,C1 DCCn,C2 …

(6)

sampling is based on the central limit theory. The research considers that users are normally distributed ranging from less experienced patent searchers to very experienced domain experts. Thus, the patent recommendation methodology is proved to work in a common collaborative R&D environment consisting of a wide range of users.

4.1. Calculating operation complexity

The patent search for the semiconductor material Copper Indium Gallium Selenide (CIGS) is used as the case study and methodology test. CIGS are thin-ﬁlm semiconductors made with copper, indium, gallium, and selenium which are elements used as light absorbing materials for thin-ﬁlm solar cell production. Nine hundred and eighty-seven related patents were used to build the patent database which includes the behavior records of the twenty researchers that conducted the patent search operations.

The algorithm selected user id“ieAC01” as the target user. The system classiﬁes the users' data according to the key phrases used for searching. After ﬁltering phrases, the system calculates the frequency of key phrases used by the patent search function. The

system clusters the users preliminarily into four groups as shown in Table 7 and makes comparisons using the clustering result. Next, the system summarizes the search conditions and counts the number of sub-behavior operations across all behavior types: search, view, analyze, bookmark and export. The system auto-matically calculates the scores for theﬁve behavior types as shown inTable 8based on the pre-deﬁned weights of sub-behaviors as listed inTable 1.

The operation complexity (OC) for the different search condi-tions across different search operacondi-tions are listed inTable 9. The system normalizes the scores based on the maximum score of each behavior type. Thus, all the scores fall between 0 and 1. Next, the system calculates the OC scores using Eq.(1)with the given parameter values as listed in Tables 10 and 11. Finally, the OC scores for the case are summarized into a matrix as shown in

Table 12.

4.2. User clustering

The OC matrix is the input data for the clustering analysis and shows the users' OC scores under different search conditions. The cluster groups are set to sizes of 2, 3 or 4.Table 13shows their RMSSTD and RS results with size 4 being the best cluster number. The size of 4 clusters best satisfies the statistical requirements for the most suitable number of groups (Table 14). User id“ieAC01” is the center of the cluster group, and the group contains_{“ieAC03,”} “ieAC09,” “ieAC10,” and “ieAC17.” Therefore, these four neighbors “ieAC01” and their records are the references for the recommen-dation process. Thefinal cluster result can be compared with the user's pre-classification (Table 7). The case study shows that they have the same group numbers, and except for the groups centered on “ieAC14” and “ieAC15” there are only slight differences (ieAC19). The other groups are the same. The clustering analysis performs consistently and the system automatically infers recom-mendations based on the users' clustering results.

The proposed recommendation approach extracts relevant patents by analyzing users’ behavior records and calculating each patent's collaborative contribution. Thus, the recommendation system clusters the target user with the members of the system according to the related patent search strategies and the

System Framework

System Platform Database Behavior Records Patent Documents User Interface Patent Search

Other Management Operation Searching Record Record Data Access User System Manager General User Search or Other Operation Maintain System Patent Recommend Data Access Search Results Results System Parameter Management User Clustering Recommended Patent Filtering Recommend Patent Behavior Record Information Transfer

Operation or System Execution Data Return

Information Transfer Fig. 2. System framework.

Recommend Patent Behavior Records Management Module Recommended Patent Filtering Module User Clustering Module System Parameter Management Module Personal Record Management Behavior Record Calculate Operation Complexity Cluster Users Evaluate Results Calculate Patent Correlations Filter Recommended Patents Set System Parameters

(7)

evaluation of RMSSTD. Following, the system recommends the related patents by analyzing the information of the patent search collaborative operations. Thus, the recommend patents are extracted by the users who use similar search strategies. Although

some previous researches are applying the concepts and principles of collaborative ﬁltering and collaborative operation evaluation (Nichols, 1997; Bhavnani et al., 2008; Nazim-uddin et al., 2009; de Campos et al., 2010; Barragáns-Martínez et al., 2010; Herlocker et al., 2004), this research has developed the unique algorithm speciﬁcally for patent recommendation application. This research focuses on recommending patents collected and prioritized based

Table 7

User pre-classiﬁcation.

Commonly Used Key Phrase Members

coevaporation, evaporation, precursor, etc. ieAC01, ieAC03, ieAC09, ieAC10, ieAC17 printing, non-vacuum, precursor, etc. ieAC05, ieAC06, ieAC11, ieAC12, ieAC18 electrodeposition, RTP, annealing, etc. ieAC07, ieAC08, ieAC14, ieAC16, ieAC19 sputtering, vacuum, RTP, etc. ieAC02, ieAC04, ieAC13, ieAC15, ieAC20

Table 8

Sub-behavior count summary (partial).

User Field Key Term S1 S2 S3 S4 V1 V2 A1 A2 A3 B1 B2 B3 B4 E1 E2

ieAC01 Tech Coevaporation 0 0 1 12 14 0 5 2 0 0 0 0 24 4 0

ieAC01 AN Sotec Corp. 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0

ieAC01 KP Solar Cell 3 0 0 0 5 0 0 0 0 2 0 0 0 0 0

ieAC01 PN US4105471 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0

ieAC01 IPC H01C 2 0 0 0 1 0 0 0 0 0 0 0 0 0 0

ieAC01 KP Coevaporation 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0

ieAC02 Tech Sputtering 0 0 0 11 9 5 4 3 0 0 0 0 25 0 0

ieAC02 Tech Electrodepositon 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0

ieAC02 PN US4528082 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0

… … … …

Table 9

Operation complexity table (partial).

User Search Condition S V A B E

ieAC01 Tech Coevaporation 3.6 11.3333 1.5 7.2 1.3333

ieAC01 AN Sotec Corp. 0.1 0 0 0 0

ieAC01 KP Solar Cell 0.3 1.6667 0 0.2 0

ieAC01 PN US4105471 0.4 0 0 0 0

ieAC01 IPC H01C 0.2 0.3333 0 0 0

ieAC01 KP Coevaporation 0.1 0 0 0 0

ieAC02 Tech Sputtering 3.3 6.3333 1.6667 7.5 0 ieAC02 Tech Coevaporation 0.3 0 0 0 0 ieAC02 Tech Electrodepositon 0.3 0 0 0 0 ieAC02 PN US4528082 0.4 0.3333 0 0 0

… … … …

Max 4.2 11.3333 3.8333 13.2 2

Table 10

Sub-behavior weights.

Behavior Sub-behavior Weight (%)

Search WS1 10 WS2 40 WS3 20 WS4 30 View WV1 33 WV2 67 Analyze WA1 17 WA2 33 WA3 50 Bookmark WB1 10 WB2 40 WB3 20 WB4 30 Export WE1 33 WE2 67 Table 11

The behavior types weights.

Behavior Type Weight Behavior Type Weight (%)

WS 5.00% WB 25.67

WV 9.00% WE 45.66

WA 15.67%

Table 12

Operation complexity matrix (partial). User Tech: Coevaporation AN: SunPower Corp. Tech: Printing Tech: Sputtering IPC: H01L ieAC01 0.63 0 0 0 0 ieAC02 0.0029 0 0.0029 0.2957 0 ieAC03 0.1037 0 0 0 0 ieAC04 0.0029 0 0 0.0376 0 ieAC05 0 0 0.1524 0 0 ieAC06 0 0 0.0956 0 0 ieAC07 0.0213 0.0242 0 0 0 ieAC08 0 0 0 0 0 ieAC09 0.0196 0 0.0055 0 0.0189 ieAC10 0.2833 0 0 0 0 Table 13

Cluster result evaluation.

Cluster Numbers RMSSTD RS RMSSTDþ 1/RS

2 0.0503 0.2932 3.4609

3 0.0422 0.5298 1.9297

(8)

on the patent search characteristics and the search's collaborative operation records of the peers with the common domain interests.

4.3. Patent recommendation and inference

User id“ieAC01” searches patents for low-cost manufacturing solar cells and focuses on the assignee“Georgia Tech Research Corporation.” The search results include two selected patents (SPs), US5510271 and US5766964. The research collects another 985 related patents as candidate patents (CPs). The last step for the methodology is to infer the recommended patents from the CPs which are sufficiently related to SPs. The proposed collaborative filtering is based on neighbor's ratings which calculate the search conditions or user's operating status. The filtering analyzes users' records and selects appropriate patents to recommend. After user clustering, the system confirms the neighbor users' operating status for patent recommendation. The system focuses on the patents used by neighbors' collaborative operations and calculates the collaborative values between any two patents. The system

Table 14 Final cluster result.

Cluster center Cluster members

ieAC01 ieAC01, ieAC03, ieAC09, ieAC10, ieAC17 ieAC12 ieAC05, ieAC06, ieAC11, ieAC12, ieAC18 ieAC14 ieAC07, ieAC08, ieAC14, ieAC16

ieAC15 ieAC02, ieAC04, ieAC13, ieAC15, ieAC19, ieAC20

Table 15

Collaborative operation weight.

Collaborative Operation Weight Collaborative operation Weight

WCoB1 2 WCoV1 1.8 WCoB2 0.5 WCoV2 2.5 WCoE1 1.5 WCoA1 1 WCoE2 2 WCoA2 1.2 WCoA3 1.4 Table 16

Summary of patent collaborative operation counts (partial).

Patent I Patent II CoV1 CoV2 CoB1 CoB2 CoE1 CoE2 CoA1 C0A2 CoA3 Co Score

US5510271 US7842882 0 1 1 5 1 0 0 2 0 10.9 US5510271 US5928438 1 0 0 5 0 1 1 1 1 9.9 US5766964 US5871630 0 1 0 1 0 0 4 1 1 9.6 US5766964 US6518086 0 0 1 1 1 0 0 2 2 9.2 US5766964 US6897560 0 1 0 0 0 0 1 3 0 7.1 US5603778 US6897560 1 0 0 1 1 0 1 1 0 6 CO weights 1.8 2.5 2 0.5 1.5 2 1 1.2 1.4 Table 17

Simpliﬁed patent correlation matrix (partial).

Patent US7842882 US5928438 US5871630 US6518086 US6897560 …

US5510271 1 0.91 – – – …

US5766964 – – 0.88 0.84 0.65 …

Table 18

Patent recommendation result. Rank Recommend

result

Patent title Patent drawing

1 US7842882 Low cost and high throughput deposition methods and apparatus for high density semiconductorﬁlm growth

2 US5928438 Structure and fabrication process for self-aligned locally deep-diffused emitter (SALDE) solar cell

3 US5871630 Preparation of copper–indium–gallium–diselenide precursor ﬁlms by electro-deposition for fabricating high efﬁciency solar cells

(9)

analyzes the correlation scores and the weights of collaborative operations are deﬁned inTable 15.Table 16lists the summary for each patent combination and counts for different types of colla-borative operations.

Thefinal recommended patent is selected from the candidate patents which are highly related with the selected patents. The research simplifies and normalizes the matrix only for the selected patents versus the candidate patents as shown inTable 17. The normalized score is the final correlation of the patents. After setting a threshold, the system selects the highly related candidate patents for each selected patent.Table 18shows the recommenda-tion result of the case, and the system further analyzes the recommended patents by providing different ranking values. In summary, these patents are recommended to user“ieAC01” and the recommended patents describe low-cost and highly efficient solar cell manufacturing technology and disclose related technol-ogy that decrease manufacturing costs.

4.4. Recommendation and evaluation

Recommendations for patents are automatically generated by the proposed system and the results are judged byﬁeld experts.

Table 19 shows the overview of four recommended patents. Among four recommended patents, US7842882 and US5928438 are also the choice of experts. After content analysis, the top two selections are both in the area of thin-film improvement processes designed to decrease the manufacturing cost, to increase the production capacity, and to improve power capacity and efficiency. Nonetheless, the recommendation system also identifies US5871630 and US6518086, which are also relevant, yet, receive slightly lower Co Scores than the top two choices. The result indicates the system provides versatile andflexible recommenda-tions comparing to the human experts’ selection.

Many different websites are built for patent analysis and can be divided into two types including patent searching (e.g. Google Patent, PatTools, WIPS) and patent analysis (e.g. SOOPAT, GPSA, IPDSS). The patent recommendation system is different since it does not search directly from the ofﬁcial patent database. The system does not support patent text mining analysis, such as patent clustering or patent claim construction which is to be included in future research and development. However, the advantage of the system is that it automatically recommends patents by analyzing the users_{' operation behavior and search} rules. Moreover, the system continuously collects new domain patents ﬁled in the USPTO database by regular updating the

database. The system collects members' patent search behavior information to train the recommendation module. Then, the collected set of users’ behavior is used to analyze and cluster the members of the system and recommend the most relevant patents to users with demonstrated interest in the domain.

5. Conclusion

The intelligent recommendation methodology and system for patent search is based on the analysis of users' behavior records from the patent search platform. The research calculates each user's behavior records according to pre-deﬁned behavior types and analyzes the user clustering results according to decision rules. The recommendation system extracts the most appropriate patents based on the collaborativeﬁltering algorithm and helps users obtain related patents while saving time and costs. The intelligent patent search system recommends patents to users based on the user clustering result and the neighbors' behavior records.

The proposed system requires little time to infer recommenda-tions and works independently without human management. The system proposes the most relevant patents, in a shorter time, which provides users more options. Since the users' behavior records are an input source, personal factors which may influence search results are avoided. Since there are many users on the system, the recommendation system makes inferences with greater consensus and less bias. The proposed system updates the database quickly and recalculates collaborative information effectively. However, there still are restrictions in the analysis since the data only includes patent search result information and users_’ behavior records. Users’ behavior records also require sufficiently long periods of time to collect for accurate analysis. Researchers refer to this as the“cold-start” problem which is to say that the recommendation system will not perform well initially and only when there is sufficient data for analysis. However, the recom-mendation results’ accuracy improves with the long-term data collection (Schein et al., 2001; Herlocker et al., 2004). The recommendation system is trained by evaluating the search rules and analysis strategies of the patent engineers who work in the Green Energy and Environment Research Laboratories at the Industrial Technology Research Institute (ITRI). After collecting the search results of the domain's patent engineers, members in the solar energy alliance use the proposed system to search patents of interests and receive recommended patents. The system clusters the new members and the domain engineers by analyzing

Table 19

Recommended patents content overview. Patent

no.

Recommendation of proposed system Co Score Expert's

Choice Description

US 7842882

Discuss the precursor layer using new material, anticipating decrease the semiconductor thin-ﬁlm manufacturing cost and increase the battery charge capacity.

✓

US 5928438

The invention comprises a solar cell with reduced electron–hole recombination performance, relatively high efﬁciency including relatively low electrode resistance, which can be fabricated at relatively low cost using simpliﬁed fabrication techniques resulting in high yield.

✓

US 5871630

Fabricating theﬁlm by electrodepositing copper, indium, gallium, and selenium onto a glass/molybdenum substrate simultaneously, improving the battery energy conversing efﬁciency.

US 6518086

A two-stage production of thinﬁlm batteries, including non-heated substrate for the ﬁrst phase non-crystal deposition layer and the second phase of the pilot short-term treatment. This technique for optical correlation applications.

(10)

the users’ operation behaviors and search strategies. One of the system's advantages is actually preventing following gurus by recommending relevant patents through the collaborativeﬁltering of all clustered users, which consist of new, less experienced, seasoned, and domain-expert engineers.

Acknowledgments

This research was partially supported by the Taiwan National Science Council research Grant (NSC100-2221-E-007-034-MY3 and NSC100-2410-H009-011-MY2), as well as the Ministry of Education Grant (100N2074E1/101N2074E1) for the Advanced Manufacturing and Service Management Research Center at National Tsing Hua University.

References

US Patent Act, Part III. Patents and protection of patent rights. Infringement of patents, Section 271. Infringement of patent. Available from:〈http://www.law. cornell.edu/patent/35uscs271.html〉[accessed 21.12.12, chapter 28].

World Intellectual Property Organization (WIPO). Available from: 〈http://www. wipo.int/portal/index.html.en〉[accessed 21.12.12].

Ansari A, Essegaier S, Kohli R. Internet recommendation systems. Journal of Marketing Research 2000;37(3):363–75.

Barragáns-Martínez AB, Costa-Montenegro E, Burguillo JC, Rey-López M, Mikic-Fonte FA, Peleteiro A. A hybrid content-based and item-based collaborative ﬁltering approach to recommend TV programs enhanced with singular value decomposition. Information Sciences 2010;180(22):4290–311.

Basumallick S, Wong JSK. Design and implementation of a distributed database system. Journal of System Software 1996;34(4):21–9.

Bhavnani C, Clarkson G, Scholl M. Collaborative search and sensemaking of patents. In: Proceedings of the CHI EA‘08 CHI ‘08 extended abstracts on human factors in computing systems, New York, USA; 2008. p. 2799–804, ISBN: 978-1-60558-012-8.

Bouras C, Poulopoulos V. Enhancing meta-portals using dynamic user context personalization techniques. Journal of Network and Computer Applications 2012;35(5):1446–53.

Chiang TA, Wu CY, Trappey AJC, Trappey CV. An intelligent system for automated binary knowledge document classiﬁcation and content analysis. Journal of Universal Computer Science 2011;17(14):1991–2008.

Dong H, Hussain F, Chang E. A service concept recommendation system for enhancing the dependability of semantic service matchmakers in the service ecosystem environment. Journal of Network and Computer Applications 2011;34(2):619–31.

Edwards W, Barron FH. SMARTS and SMARTER: improved simple method for multiattribute utility measurement. Organizational Behavior and Human Deci-sion Processes 1994;60:306–25.

Herlocker JL, Konstan JA, Terveen LG, Ridel JT. Evaluating collaborativeﬁltering recommender systems. ACM Transactions on Information Systems 2004;22 (1):5–53.

Huang LC, Li Y. Research on technology trend based on patent information. In: Proceedings of IEEE management of innovation and technology (ICMIT) inter-national conference; 2010. p. 209–13.

Kelly D, Teevan J. Implicit feedback for inferring user preference: a bibliography. SIGIR Forum 2003;37(2):18–28.

Krizer M, Zacca F. Automatic categorisation applications at European Patent Ofﬁce. World Patent Information 2002;24(1):187–96.

Lee TQ, Park Y, Park YT. Time-based approach to effective recommender system using implicit feedback. Expert Systems with Applications 2008;34 (4):3055–62.

Li Z, Tate D, Lane C, Adams C. A framework for automatic TRIZ level of invention estimation of patents using natural language processing, knowledge-transfer and patent citation metrics. Computer-Aided Design 2012;44(10):987–1010. Liang Y, Liu Y, Kwong CK, Lee WB. Learning the“Whys”: discovering design

rationale using text mining—an algorithm perspective. Computer-Aided Design 2012;44(10):916–30.

Nazim-uddin, M, Shrestha J, Jo GS. Enhanced contented-basedﬁltering using diverse collaborative prediction for movie recommendation. In: Proceeding of 2009ﬁrst asian conference on intelligent information and database systems; 2009. p. 132–7.

Nichols DM. Implicit rating and filtering. In: Proceedings of the 5th DELOS workshop onfiltering and collaborative filtering; 1997. p. 31–6.

Oard DW, Kim J. Implicit feedback for recommender systems. In: Proceedings of the AAAI workshop on recommender systems; 1998. p. 81–3.

Resnick P, Varian HR. Recommender systems. Communications of ACM 1997;40 (3):56–8.

Richter G, MacFarlane A. The impact of metadata on the accuracy of automated patent classiﬁcation. World Patent Information 2005;27(3):13–26.

Schein AI, Popescul A, Ungar LH, Pennock DM. Methods and metrics for cold-start recommendations. In: Proceedings of the 25th annual international ACM SIGIR conference on research and development in information retrieval; 2001. p. 253–60.

Sharma S. Applied multivariate techniques. New York: John Wiley & Sons, Inc; 1996. de la Torre-Diez I, Álvaro-Muñoz S, López-Coronado M, Rodrigues J. Development and performance evaluation of a new RSS tool for a web-based system: RSS_PROYECT. Journal of Network and Computer Applications 2013;36 (1):255–61.

Trappey AJC, Trappey CV, Wu CY, Lin CW. A patent quality analysis for innovative technology and product development. Advanced Engineering Informatics 2012;26(1):26–34.

Trappey CV, Trappey AJC, Wu CY. Clustering patents using non-exhaustive overlaps. Journal of Systems Science and Systems Engineering 2010;19(2):162–81. Wang L, Guo YP, Fang M. Web search engine based on DNS. Journal of Network and

Computer Applications 2007;30(2):466–78.

de Campos LM, Fernández-Luna JM, Huete JF, Rueda-Morales MA. Combining content-based and collaborative recommendations: a hybrid approach based on bayesian networks. International Journal of Approximate Reasoning 2010;51 (7):785–99.