TWI502382B - A patent serch method applied with principal component analysis - Google Patents

Description

Patent search method using principal component analysis

The present invention relates to a patent search method and, in particular, to a patent search method using principal component analysis.

The patent contains a wealth of information, and the importance to the company has long been recognized by the world. According to the analysis of the World Intellectual Property Organization (WIPO), among the various technical publications, such as various professional journals, magazines, and encyclopedias, the only ones that can fully disclose the core of technology are patent specifications. Therefore, about 90%~95% of all research and development achievements in the world can be found in the patent specifications, 80% of which are not recorded in other magazines, which is a great way to absorb patent information and develop the technology trends. Help. According to further investigations by the World Intellectual Property Organization, the use of patent information can shorten the research and development time by 60% and save 40% of research and development expenses.

In the current booming technology industry, in the face of strong competitive pressures, the term "innovation" has become the representative source of global competitiveness. No matter whether domestic and foreign industries are actively investing a lot of time and human resources. innovating. However, before conducting R&D innovation, companies often use patent search to understand the patent layout of the technical field in which the R&D target is located. On the one hand, they can avoid their own research and development results infringing on others' patents, and on the other hand, they can also apply for patents in future enterprises. Reference.

However, in the process of patent search, it is often necessary to constantly adjust the search conditions, and to change the combination of various synonyms or their subordinate concepts, or to limit the conditions of specific words. In addition, most of the common professional terms used today are different from those used in the past, and if a person with a heart is writing a patent with special features that are not common and not commonly used to describe their technical characteristics, this deliberate circumvention method The barriers to search will be relatively increased. However, if the problem is to be solved by expanding the search range, the problem of the accuracy of the search results will also occur.

Therefore, the present invention will effectively improve the above problems, and as far as possible, without increasing the accuracy of patent search, effectively improve the breadth of patent search, so as to improve the credibility of patent search.

The present invention is expected to provide a patent search method using principal component analysis so as to increase the breadth of patent search without affecting the accuracy of patent search.

The invention provides a patent search method using principal component analysis, which comprises the following steps: (S1) searching a patent database according to a search condition to obtain N search patents; (S2) targeting the N according to a predetermined citation network rule The patent search patent searches for a patent for patent citation; (S3) performs a patent similarity calculation for the M-reference patent and the N-search patent to generate a patent similarity matrix; (S4) is similar according to the patent The principal matrix analysis performs a principal component analysis to screen and generate a Q pen similar patent, and merges the Q pen similar patent with the N pen search patent into a (Q+N) pen search result; (S5) the (Q+N) pen The search result replaces the N pen search patent of step (S2) and repeats steps (S2) to (S4) until Q is equal to zero, where N and M are both natural numbers greater than one, and Q is an integer greater than or equal to zero.

The patent similarity calculation of the step (S3) of the present invention comprises the following sub-steps: (S31) establishing a reference matrix of MxN; (S32) generating an associated intensity matrix of Mx1 according to the citation association matrix of the MxN, the association of the Mx1 The intensity matrix includes one of the associated strength values of each of the cited patents; (S33) normalizing the associated intensity matrix of the Mx1 to generate a normalized association strength value corresponding to each of the cited patents; (S34) from the M-citing The patent selects a patent of the P-referenced value whose normalized correlation strength value is greater than one, wherein P is a natural number greater than one; (S35) determines whether P is greater than N, and if so, establishes a PxN patent similarity matrix; if not, Then establish a patent similarity matrix of MxN. The patent similarity matrix is established by using the relationship between each cited patent and one of the predetermined patent classification codes of the N-search patents, and the predetermined patent classification code may be an international patent classification code or a US patent classification code. In addition, the predetermined patent classification code may also be a primary patent classification code, or a secondary patent classification code, or a combination of a primary patent classification code and a secondary patent classification code.

The principal component analysis of the step (S4) of the present invention comprises the following substeps: (41) calculating a common variance matrix according to the patent similarity matrix; (S42) performing feature decomposition according to the common variance matrix to generate a feature value and a Feature vector; (S43) calculating a principal component coefficient and a principal component interpretation variation Percentage; (S44) and then selecting the Q-like similar patent according to the principal component coefficient and the percentage variation of the principal component interpretation.

In addition, the predetermined citation network rule adopted by the present invention may include a Forward Citation and a Backward Citation, and the predetermined citation network rule may include a layer of patent citation network or two. Layer patent citation network or three-tier patent citation network.

Compared with the prior art, the present invention proposes a patent search method using principal component analysis, and combines principal component analysis and patent citation network to apply for patent search, without losing the accuracy of patent search. Effectively enhance the breadth of patent searches to improve the credibility of patent searches.

S1~S5‧‧‧Steps

S31~S35‧‧‧Steps

S351~S352‧‧‧Steps

S41~S44‧‧‧Steps

10‧‧‧citation association matrix

12‧‧‧Correlation strength matrix

14‧‧‧Normalized correlation strength matrix

FIG. 1 is a flow chart showing the patent search method of the present invention.

Figure 2 is a flow chart showing the calculation of the similarity of the patent of the present invention.

FIG. 3 is a schematic diagram showing the citation association matrix and the correlation strength normalization matrix of the present invention.

Figure 4 is a flow chart showing the principal component analysis of the present invention.

Figure 5 is a schematic diagram showing the first principal component coefficients of a particular embodiment of the present invention.

A detailed description of the patent search method using principal component analysis proposed by the present invention will be given below. Please refer to FIG. 1. FIG. 1 is a flow chart showing the patent search method of the present invention. The invention provides a patent search method using principal component analysis, which comprises the following steps: (S1) searching a patent database according to a search condition to obtain an N-search patent, wherein the patent database may be a patent material of the US Patent and Trademark Office. a library, or a related patent database of another country, etc., and the search condition may be a set of keywords or a set of patent numbers, etc.; (S2) performing a patent citation search for the N search patents according to a predetermined citation network rule The M-cited patent, wherein the predetermined citation network rule may be a patent citation for each search patent by Forward citation and Backward citation to obtain a M-cited patent, in addition to the citation The network specification may also include a patent citation network or a second-tier patent citation network or a three-layer patent citation network; (S3) performing a patent similarity calculation for the M-cited patent and the N-search patent to generate a patent similarity matrix; (S4) performing a principal component analysis according to the patent similarity matrix to screen and generate a similar patent for the Q pen, and Q T N T is similar to the retrieval Patent Patent Merging into (Q+N) pen search results; (S5) replacing the (Q+N) pen search result with the N pen search patent of step (S2) and repeating step (S2) to step (S4) until Q is equal to zero , wherein N and M are both natural numbers greater than one, and Q is an integer greater than or equal to zero.

Referring to FIG. 2 and FIG. 4, FIG. 2 is a flow chart showing the patent similarity calculation according to the present invention, and FIG. 4 is a schematic diagram showing the citation association matrix and the correlation strength normalization matrix of the present invention. The patent similarity calculation of the step (S3) of the patent search method of the present invention comprises the following sub-steps: (S31) establishing a MxN citation association matrix 10 (shown in FIG. 4), wherein A1 to An represent N-search patents, B1 to Bm represent the M-cited patent. If the A1 search patent has a citation relationship with the B3 citation patent (whether it is a forward citation or a backward citation), the citation value (Cij) corresponding to the citation association matrix 10 is 1 If the A1 search patent and the B2 citation patent do not have any citation relationship, the citation value (Cij) in the citation association matrix is 0, so the citation association matrix 10 is a matrix composed of 0 and 1; S32) generating, according to the reference matrix 10 of the MxN, an associated intensity matrix 12 of Mx1, where the associated intensity matrix 12 of the Mx1 includes one of the associated strength values of each of the cited patents ( i =1~ n ; j =1~ m ), where the relationship strength (RS) is defined as the sum of all the cited values (Cij) of any of the cited patents B1 to Bm (as shown in Figure 4) That is to say, when there is more citation relationship between any patent in the cited patent and the N-search patent, the sum of the citation value (Cij) is bound to increase, and it also represents the patent and N pen. Searching for the associated strength value (RSj) of the patent; (S33) normalizing the associated intensity matrix 12 of the Mx1 to generate a normalized correlation strength value corresponding to each of the cited patents and forming an Mx1 normalized correlation strength matrix 14, The normalized correlation strength value (RSnol_j) is defined as the correlation strength (RSj) divided by the average correlation strength (RSave); (S34) from the M-cited patent, the normalized association strength value is greater than one of the P-citing patents, wherein P is a natural number greater than one; (S35) determining whether P is greater than N, (S351) if yes, establishing a patent similarity matrix of PxN; (S352) if not, establishing a patent similarity matrix of MxN. In addition, the present invention uses the International Patent Classification Number (IPC) as a basis for calculating similarity, having a main section, a main class, a subclass, a group, and a subgroup. The five levels, the present invention uses this as the basis for the similarity score, defines the patent similarity as a value between 0 and 1, and sequentially determines the ranks in units of one hierarchy. The patent similarity matrix is established by using the relationship between each cited patent and one of the predetermined patent classification codes of the N-search patents, and the predetermined patent classification code may be an international patent classification code or a US patent classification code. The predetermined patent classification code may also be a primary patent classification code, or a secondary patent classification code, or a combination of a primary patent classification code and a secondary patent classification code.

Please refer to FIG. 3, which is a flow chart showing the principal component analysis of the present invention. The principal component analysis of the step (S4) of the patent search method of the present invention comprises the following substeps: (41) calculating a total variance matrix according to the patent similarity matrix, and the calculation method is if there are two random variables x _{1 i} And x _{2 i} , and i=1...N, then the covariation between the two variables is defined as: If the covariation number is generalized to a high dimensional space, a Covariance matrix can be obtained as follows: x _i = [ x _{1 i} ... x _Mi ] ^T ; (S42) performing eigen decomposition according to the covariance matrix to generate an eigenvalue and a eigenvector, wherein eigen decomposition is performed on the covariance matrix S to obtain a principal component (ie, a eigenvector) of the data and a weight thereof. (ie, the eigenvalue), using the characteristic equation | λI - S | = 0, the eigenvalues of the covariance matrix S can be obtained, and λ ₁ can be obtained according to the order of the eigenvalues. λ ₂ ... λ _p 0 total p eigenvalues, and then find the eigenvectors e ₁ , e ₂ ,..., e _p corresponding to λ _i , the meaning of the eigenvalue is the variance of the new variable, the total variance of the original data and the new total The variance number is the same, which is the sum of the eigenvalues; (S43) calculates a principal component coefficient and a principal component interpretation variation percentage, wherein the principal component coefficient is the coefficient between the new variable and the original variable transformation, representing the original variable for the new The influence and importance of variables, so the larger the coefficient, the greater the influence on the original variables. The coefficients are defined as follows: (I, j = 1,2, ... , p) where, l _ij is a j-th variable load in the i-th principal component, λ _i is the i-th eigenvalue of the principal component, e _ij is the eigenvectors e _i The jth value. After the main component extraction, the coefficient of each variable under each main component can be obtained, so that the main component after the main component conversion can be obtained. For example, X ₁ is y _{1 m} = l _m in the mth main component . ₁ x ₁₁ + l _{m 2} x ₁₂ +...+ l _{m 1} x _{1 p} , which can be expressed as: The percentage of variance of principal component interpretation indicates how much the principal component accounts for the whole data. The variables take more. Although most of the information of the original data can be retained, the purpose of dimension simplification cannot be achieved. Effectively reduce the dimensions for analysis, but may lose some useful information. [lambda] _i represents the i-th feature value, the percent explained variance of the i-th principal component and the cumulative percent of explained variance of the i-th principal component before were: (S44) selecting the Q-like similar patent according to the principal component coefficient and the percentage variation of the principal component interpretation.

Hereinafter, a specific embodiment of the present invention will be described. The patent search method using principal component analysis proposed by the present invention can be applied to search for "Through-Silicon Via" (TSV) as a search target, and the search database is an example of a US patent database. application.

In the step (S1) of the present invention, the patent is first announced in the United States according to a search condition of ABST/(("trough-silicon via$" or TSV or TSVs) and ("integrated circuit$") in 2013/4/16. An advanced search is performed in the database, and the search result is obtained by N search patents, where N is equal to 44.

In the step (S2) of the present invention, a patent citation search is performed on 44 search patents according to a predetermined citation network rule, and the citation network rule adopts Forward Citation and Backward Citation and uses A layer of citation network rules, through a patent citation search for 44 search patents to obtain a M-cited patent, where M equals 406.

In the step (S3) of the present invention, a patent similarity calculation is performed on the 406-stroke patent and the 44 search patents to generate a patent similarity matrix. The patent similarity calculation used in the present invention includes the following sub-steps: (S31) establishing a 406x44 citation association matrix; (S32) generating a 406x1 correlation strength matrix according to the 406x44 citation association matrix, the 406x1 correlation strength The matrix contains one of the associated strength values of each of the cited patents, that is, the correlation strength (RSj) of each of the cited patents in the 406 cited patents is calculated via the citation value (Cij) in the citation reference matrix, and then the 406 pens are The average correlation strength (RSave) of the cited patents (RSj) is 1.8705; (S33) normalizes the associated intensity matrix of 406x1 to generate a normalized correlation strength value corresponding to each of the cited patents; (S34) From the 406 cited patents, the patent of the P-referenced patent with the normalized correlation strength value greater than one is selected, wherein P is equal to 115; (S35) determining whether P is greater than N, (S351), if yes, establishing a PxN patent similarity matrix That is, a 115x44 patent similarity matrix is established. In practical applications, this embodiment uses the International Patent Classification Number (IPC) as the basis for the similarity score, and defines the patent similarity as a value between 0 and 1, and sequentially determines the ranks in units of one level. One order increases the similarity value of 0.2.

Referring to FIG. 5, FIG. 5 is a schematic diagram showing a first principal component coefficient of a specific embodiment of the present invention. In the step (S4) of the present invention, principal component analysis is performed based on the patent similarity matrix of 115x44. The principal component analysis used in the present invention includes the following sub-steps: (41) calculating a common variance matrix according to the 115x44 patent similarity matrix; (S42) performing feature decomposition according to the common variance matrix to generate a eigenvalue and a feature vector; (S43) calculating a principal component coefficient and a principal component interpretation variation percentage, the first principal component in this embodiment is: Y ₁ =4.149506 X ₁ +3.432632 X ₂ +.....+3.720075 X ₁₁₅ , wherein each coefficient of the X term represents the weight value of the corresponding cited patent, and the greater the weight value, the higher the similarity between the cited patent and the 44 search patents, so the first principal component of the present invention The coefficient is presented in the form of a line graph, as shown in FIG. 5; (S44) according to the principal component coefficient and the percentage variation of the principal component interpretation, a Q-like patent is selected, wherein the first coefficient of the principal component coefficient corresponds to the peak of 4.590093. There are a total of 9 patents, a second coefficient peak of 3.807923 has 1 pen, a third coefficient peak of 3.798306 contains 3 pens, and a fourth coefficient peak of 3.641118 has 8 pens. After analysis and judgment, it is known that the fourth coefficient peak The eight patents corresponding to the value and one of the third coefficient peaks are not the search target of the present embodiment. Therefore, the present embodiment selects 12 similar patents for the cited patent corresponding to the first coefficient peak and the second coefficient peak. , that is, Q is equal to 12, and finally the 12 similar patents and the 44 search patents are combined into 56 search results; (S5) the 56 search results replace the 44 search patents of the step (S2) and repeated Step (S2) to step (S4) until the number of similar patents is equal to zero.

Compared with the prior art, the present invention proposes a patent search method using principal component analysis, and combines principal component analysis and patent citation network to apply for patent search, without losing the accuracy of patent search. Effectively enhance the breadth of patent searches to improve the credibility of patent searches.

The features and spirit of the present invention will be more apparent from the detailed description of the preferred embodiments. On the contrary, the intention is to cover various modifications and equivalents within the scope of the invention as claimed. Therefore, the scope of the patented scope of the invention should be construed in the broadest

S1~S5‧‧‧Steps

Claims (9)

A patent search method using principal component analysis includes the following steps: (S1) searching a patent database according to a search condition to obtain N search patents; (S2) searching for the N patents according to a predetermined citation network rule Performing a patent citation search to obtain a M-cited patent; (S3) performing a patent similarity calculation for the M-cited patent and the N-search patent to generate a patent similarity matrix; and (S4) according to the patent similarity matrix Performing a principal component analysis to screen and generate a similar patent for Q pen, and combining the similar patent of Q pen with the N pen search patent into a (Q+N) pen search result; wherein N and M are both natural numbers greater than one, Q For a maximum or equal to zero integer, the predetermined citation network rule may include a Forward Citation Network and a Backward Citation Network. The patent search method of claim 1, wherein the patent similarity calculation comprises the following sub-steps: (S31) establishing a MxN citation association matrix; (S32) generating an Mx1 according to the MxN citation association matrix Correlating the intensity matrix, the associated intensity matrix of the Mx1 includes one of the associated strength values of each of the cited patents; (S33) normalizing the associated intensity matrix of the Mx1 to generate a normalized correlation strength value corresponding to each of the cited patents; (S34) screening, from the M-cited patent, a patent of the P-referenced patent having a normalized correlation strength value greater than one, wherein P is a natural number greater than one; and (S35) determining whether P is greater than N, and if not, establishing one MxN patent similarity matrix. The patent search method according to claim 2, wherein in the sub-step (S35), it is determined whether P is greater than N, and if so, a patent similarity matrix of PxN is established. The patent search method according to claim 1, wherein the patent similarity matrix uses one of the cited patents to compare each of the cited patents with one of the N patents. The relationship is established. The patent search method of claim 4, wherein the predetermined patent classification code may be an international patent classification code or a US patent classification code. For example, the patent search method described in claim 4, wherein the predetermined patent classification code may be a primary patent classification code, or a secondary patent classification code, or a combination of a primary patent classification code and a secondary patent classification code. The patent search method of claim 1, wherein the principal component analysis comprises the following substeps: (S41) calculating a common variance matrix according to the patent similarity matrix; (S42) performing characteristics according to the common variance matrix Decomposing to generate a feature value and a feature vector; (S43) calculating a principal component coefficient and a principal component interpretation variation percentage; and (S44) selecting the Q pen similar patent according to the principal component coefficient and the principal component interpretation variation percentage. The patent search method according to claim 1, further comprising the steps of: (S5) replacing the (Q+N) pen search result with the N-search patent of the step (S2) and repeating the step (S2) to Step (S4) until Q is equal to zero. The patent search method as claimed in claim 1, wherein the predetermined citation network rule may further comprise a layer of patent citation network or a second layer patent citation network or a three-layer patent citation network.