CN108665182B - Patent litigation risk prediction method - Google Patents
Patent litigation risk prediction method Download PDFInfo
- Publication number
- CN108665182B CN108665182B CN201810482063.3A CN201810482063A CN108665182B CN 108665182 B CN108665182 B CN 108665182B CN 201810482063 A CN201810482063 A CN 201810482063A CN 108665182 B CN108665182 B CN 108665182B
- Authority
- CN
- China
- Prior art keywords
- litigation
- company
- factors
- features
- meta
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 239000013598 vector Substances 0.000 claims abstract description 57
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 22
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 17
- 239000011159 matrix material Substances 0.000 claims description 20
- 238000011176 pooling Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 13
- 238000012512 characterization method Methods 0.000 claims description 9
- 238000012549 training Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 5
- 238000005516 engineering process Methods 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 4
- 238000012163 sequencing technique Methods 0.000 claims description 4
- 239000000126 substance Substances 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 230000009193 crawling Effects 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 108010014172 Factor V Proteins 0.000 claims description 2
- 230000004913 activation Effects 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 claims description 2
- 239000003795 chemical substances by application Substances 0.000 claims description 2
- 230000002860 competitive effect Effects 0.000 claims description 2
- OVBPIULPVIDEAO-LBPRGKRZSA-N folic acid Chemical compound C=1N=C2NC(N)=NC(=O)C2=NC=1CNC1=CC=C(C(=O)N[C@@H](CCC(O)=O)C(O)=O)C=C1 OVBPIULPVIDEAO-LBPRGKRZSA-N 0.000 claims description 2
- 230000009466 transformation Effects 0.000 claims description 2
- 230000001131 transforming effect Effects 0.000 claims description 2
- 239000000203 mixture Substances 0.000 claims 2
- 238000013475 authorization Methods 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000007418 data mining Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000000556 factor analysis Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/18—Legal services
- G06Q50/184—Intellectual property management
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Technology Law (AREA)
- General Physics & Mathematics (AREA)
- Operations Research (AREA)
- Physics & Mathematics (AREA)
- Primary Health Care (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Educational Administration (AREA)
- Game Theory and Decision Science (AREA)
- Quality & Reliability (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a patent litigation risk prediction method, which is characterized in that litigation factors serving as original reports, litigation factors serving as rejected and litigation factors of each company and litigation factors of each patent are obtained by a method of combining tensor decomposition and a convolutional neural network according to heterogeneous patent data (meta-features, text features and patent citation network) and patent litigation case records, and then the patent litigation risk prediction is carried out by utilizing the three litigation factors (all vector representations).
Description
Technical Field
The invention relates to the technical field of machine learning and patent data mining, in particular to a method for predicting risk of patent litigation.
Background
The patent is an important means for intellectual property protection. In recent years, with the development of science and technology and the innovation of technology, the number of patent applications and grants in various regions of the world is rapidly increasing, and the patent litigation cases caused by patent infringement are also remarkably increasing. Patent litigation cases are complex in cause, complex in procedure and high in cost, once the cases are greatly influenced by both litigation parties, early warning of patent litigation can provide more time for both litigation parties to make development strategies, potential litigation patents are negotiated and solved, and accordingly risks are dredged and avoided in time, and resources are saved.
In current research work and patents, the following methods are the main methods for patent litigation prediction:
1) patent litigation factor analysis based on statistics.
At present, the analysis of patent litigation factors based on statistics mainly focuses on analyzing the relationship between patent features and patent litigation, and patents meeting specific features are potential litigation patents. Through analysis by predecessors, factors that are known to affect patent litigation include patent forward citation, patent backward citation, patent family size, patent review process, patentee, and the like.
2) Corporate litigation risk prediction based on collaborative filtering.
The company litigation risk prediction based on collaborative filtering combines with a collaborative filtering (such as matrix decomposition) algorithm commonly used in the traditional recommendation system, and a learner uses the method to predict an industry or a company of which litigation is possible.
Both of the above approaches do not solve the problem of two companies regarding whether litigation can occur for a patent. In addition, the above methods do not utilize the heterogeneous data including text in patents, nor do they consider modeling relationships between companies, and between companies and litigation patents.
Disclosure of Invention
The invention aims to provide a method for predicting risk of patent litigation, which can improve the accuracy of a prediction result.
The purpose of the invention is realized by the following technical scheme:
a method for predicting risk of patent litigation, comprising:
obtaining patent data in an authorized heterogeneous form, and crawling patent litigation case data;
transforming patent data in a heterogeneous form of each granted patent into comprehensive patent vector representation by using a convolutional neural network and a network embedding method;
establishing an original notice-defendant-patent third-order tensor and a tensor decomposition model by using patent litigation case data;
combining the comprehensive patent vector representation with a tensor decomposition model to obtain a mixed model;
training the mixed model by using a sequencing learning method;
calculating litigation factors of each company as original reports, litigation factors as defended reports and litigation factors of each patent by using the mixed model obtained by training;
the three litigation factors are used for predicting the litigation risk of a patent among companies.
According to the technical scheme provided by the invention, the litigation factors of each company as the original report, the litigation factors as the advised litigation factors and the litigation factors of each patent are obtained by combining tensor decomposition and a convolutional neural network according to heterogeneous patent data (meta features, text features and patent citation network) and patent litigation case records, and then the patent litigation risk prediction is carried out by utilizing the three litigation factors (all vector representations), so that the accuracy of the prediction result is greatly improved compared with the prior art.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
FIG. 1 is a flowchart of a method for predicting risk of patent litigation according to an embodiment of the present invention;
FIG. 2 is a diagram of a neural network structure combined with network characterization for a method for predicting risk of litigation according to an embodiment of the present invention;
FIG. 3 is a conceptual diagram of modeling a method for predicting risk of litigation according to an embodiment of the present invention;
FIG. 4 is a probability model diagram of a method for predicting risk of litigation according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention are clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.
The embodiment of the invention provides a method for predicting risk of patent litigation, which mainly comprises the following steps as shown in figure 1:
and step 11, obtaining authorized patent data in a heterogeneous form, and crawling patent litigation case data.
In embodiments of the present invention, authorized patent data in heterogeneous form and patent litigation case data may be obtained from the internet.
And step 12, converting patent data in a heterogeneous form of each granted patent into a patent vector representation by using a convolutional neural network and a network embedding method.
In the embodiment of the present invention, after obtaining the patent data in the authorized heterogeneous form, the patent data in the authorized heterogeneous form is represented in a unified mathematical form, where the patent data in the authorized heterogeneous form includes: meta-features, text features, and patent citation networks; using S p1, { k | k ═ 1,2,3, …, N } represents a patent set, and N is the number of patents; establishing a patent citation network G according to a patent citation relation, wherein each node of the patent citation network G represents a patent; each patent has certain characteristics, and thus the patent citation network G is an attribute network.
For patent k in the patent citation network G (i.e. the patent with the serial number k in the patent set), X is usedkRepresenting the characteristics of the characters, including meta characteristics and text characteristics; the meta-features refer to basic features that can be directly extracted from patent documents, and include: forward reference, backward reference, claim number, picture number, table number, patent classification information, grant gap, change in number of patents in patent category, locationThe number of patents in the company varies; text features refer to textual descriptions in patent documents, including: patent titles, patent abstract and patent claims.
The contents and descriptions of meta-features and text features are shown in table 1:
TABLE 1 contents and description of Meta-features and text features
In the embodiment of the invention, the patent data is processed by using a convolutional neural network and a network embedding method in the following way:
1. and processing the meta-characteristics and the patent citation network G in a network embedding learning mode. First, the meta-features of each patent node in the patent citation network G are spliced into a meta-feature vector (e.g., the meta-feature vector of patent k can be written as) Training the attributes as patent node attributes of network embedded learning; then, through network embedding learning of the patent citation network G, the high-dimensional patent meta-feature vector can be converted into the low-dimensional patent representation, and meanwhile, the patent citation relation is embedded into the patent representation, so that the patent features are accurately depicted. The treatment process is as follows:
firstly, the element characteristic vectors of all patents are spliced to form a characteristic matrix FN×QWherein Q is the dimension of the characteristic vector of the patent element; feature matrix FN×QIs denoted as fk(equivalent to) (ii) a Definition of input characterization of patent k as ek=ETfkIn which EIs a transformation matrix to be trained, obviously the dimension of E is Qxd1(d1Can be set by oneself);
secondly, for each node in the patent citation network G, taking it as a root node root, randomly sampling its neighbor nodes to generate different paths:
<root,neighborhood1,neighborhood2,…>;
wherein, the neighbor nodes are represented by neighbor 1 and 2;
for each path, a set of neighbor nodes (referred to as a scenario) for patent k is given:
context(k)={k-l,…,k+l}\{k};
that is, consider 2l neighbor nodes of patent k;
the following objective function, i.e. the probability of predicting the central node by the neighbor nodes, is maximized:
e 'in the formula'kAnd econtext(k)Respectively represents an output representation and a scene representation of a patent k, wherein the patent m is a neighbor node of the patent k, e'mAn output characterization representing patent m;
econtext(k)is defined as:
in the above formula emThe input characterization of the patent m;
finally, by approximating the objective function by negative sampling, an output representation is obtained, i.e. a vector representation of the meta-features, whose output representation is e 'for patent k'k。
2. The method adopts a convolutional neural network to process text features, and comprises the following steps:
firstly, words in the text features after the stop words are removed are converted into Word vectors with the dimension d through the Word2Vec technology0(ii) a Each sentence can be considered as a matrix and the sentences can form a tensor.
Secondly, dividing the patent titles, the abstract of the patent specification and the patent claims into a plurality of patent claims, wherein C-2 patent claims are taken before (0 is supplemented if the patent claims are not enough), and a patent title and the abstract of the patent specification are added to form C pieces; each piece is a word sequence consisting of word vectors, each piece takes the first H words (if the H words are not enough, 0 is complemented), and the text characteristics of the patent k are converted into tensorsWherein, the value of C can be set according to the actual need.
Then, a two-layered convolutional neural network pair tensor as shown in FIG. 2 is usedCarrying out treatment; the first layer is convolution and pooling of word level, and the second layer is convolution and pooling of sentence level; the convolution and pooling operations of the first layer are as follows:
and (3) convolution operation: the convolution kernel shape of the convolution operation is c' × d0Where c' is the dimension of the convolution kernel, the goal is to pass each slice s of patent k through a convolution operation as w1,w2,…,wHAll are converted into a new hidden layer sequence by ecAre shown hereEach item in the new hidden layer sequence(1. ltoreq. n. ltoreq.H) satisfies:wherein the content of the first and second substances,andis the parameter of convolution operation, d is the dimension of output, n-c' +1 < 0 means the boundary, needs to complement 0; relu (x) is a non-linear activation function, where relu (x) is max (0, x);is a join operation to join vectors; w is a word sequence, and subscripts are sequence numbers of the word sequence;
and (3) pooling operation: the step size of the pooling operation is u. The new hidden layer sequence e we obtained for the above convolution operationcPerforming u max pooling, converting into new global hidden layer sequence, and using ecuAre shown hereEach of the new global hidden layer sequencesAll satisfyWherein, r-u +1 < 0 means that the boundary is a boundary, and 0 needs to be supplemented:
the convolution and pooling operations of the second layer are the same as the first layer, except that the input is changed from a sequence of words to a sequence of sentences.
Vector representation of text features is obtained through a convolutional neural network, and vector representation of meta-features is obtained through a network embedding learning mode; and splicing the vector representation of the text features and the vector representation of the meta features, and then obtaining the comprehensive patent vector representation through the full-connection layer.
And step 13, establishing an original notice-defended-patent third-order tensor and establishing a tensor decomposition model by using patent litigation case data.
A patent litigation case includes three elements, namely a source company, a defendant company and related patents. Given M companies and N patents, assume the original company set is S U1,2,3, …, M, the advertised company set is S V1,2,3, …, M, patent setIs synthesized into SpAs shown in fig. 3, a third-order tensor is formed with the three axes as axes, and each value of the third-order tensor represents whether a litigation has occurred to a patent by a company.
Suppose the litigation case set is R ═ { Rijk|i∈SU,j∈SV,k∈SpGiven a grandfather i and a defendant j, called a company pair (i, j), if the grandfather i calls the defendant j about the patent k, a litigation record r is obtainedijk1 is ═ 1; if the company does not have litigation for k for (i, j) at present, the litigation record rijk=0;
Referring also to FIG. 3, prediction of risk of litigation for a patent refers to predicting the likelihood that a company will litigation for patent k for patent i, j given company pair (i, j) and patent k, with the prediction result being denoted as r'ijk(ii) a We break down litigation possibilities into the following forms:wherein, Ui、VjAnd PkThese three litigation factor solving methods will be described in detail later, each of which represents a litigation factor of the original reporting company i, a litigation factor of the reported company j, and a litigation factor of the patent k.
And 14, combining the comprehensive vector representation with a tensor decomposition model to obtain a mixed model.
From the viewpoint of data mining, companies have been classified into the following 4 causes of litigation on patent k for (i, j): 1) the value of patent k itself; 2) degree of importance of original company i to patent k; 3) degree of importance of the advertised company j to the patent k; 4) competitive relationship between the original company i and the defendant company j;
the comprehensive vector representation is obtained by splicing the vector representation of the text features obtained by the convolutional neural network and the vector representation of the meta-features obtained by a network embedding learning mode, the reason 1 can be described by utilizing the vector representation of the text features obtained by the convolutional neural network, and the reasons 2) -4 can be described by a tensor decomposition model); the combined vector representation and the patent litigation factors obtained from the tensor decomposition model are unified to form a hybrid model, and the details are shown in fig. 4 and in step 15.
And step 15, training the hybrid model by using a sequencing learning method.
Considering that litigation records are discrete data, and only litigation and non-litigation are the two cases, the tensors in the commonly used tensor decomposition store continuous values, and the direct use of the tensor decomposition is not feasible. The embodiment of the invention adopts a sequencing learning method to solve the problem. Mainly as follows:
the possibility of patent litigation between companies is similar to user preferences in traditional recommendation systems, namely: if it is not Then relative patent k-In other words, company prefers patent k to (i, j)+Then the problem translates into a predictionHere, theThe formula means that: company i about litigation patent k+Litigation of company j, company i regarding non-litigation patent k-No litigation company j. Wherein k is+Indicating that patent k is a litigation patent, k-Patent k is denoted as a non-litigation patent,represents a company pair (i, j) and is related to k+The occurrence of litigation is caused,represents a company pair (i, j) and is related to k-No litigation occurs.
For patent k, it contains k+And k is-In both forms, our target pairsA series of unknown patents are predicted by a probability ordering method, and both lawsuited patents and unlawned patents are required during training, so that both patents participate in the model.
Given a company pair (i, j), k is used+>i,j k-To indicate that the company is paired with (i, j) for patent k+And k-Bias-order relationship of preference degrees; then company pairs (i, j) in patent k+Litigation occurred in the above, but not in patent k-The probability of litigation occurring is expressed as:
From the perspective of probability theory, finding the best patent sequenceHerei,jAll the partial order relations representing possible patent disputes between the original company i and the reported company j are represented, so it is SP×SPThe problem can be achieved by maximizing the following a posteriori distribution: p (U, V, P, W | >)i,j)∝p(>i,jL U, V, P, W) P (U, V, P, W), where U, V, P, W represent the prosecutor matrix, the defendant litigation factor matrix, the patent litigation factor matrix, and all parameter matrices in the convolutional neural network, respectively.
Assuming that all company pairs are independent of each other, there are:
the conditional distribution of the above formula is:
wherein the content of the first and second substances,the partial order relation of all company litigation patents is stored;
as a generative model, assuming that litigation factors of the original agent all obey a 0-mean Gaussian distribution, the variance can be expressed asWherein deltaUI is an indication matrix, which is a diagonal matrix, and is 0 except for the diagonal.
The litigation factor matrix U of the original trailer also follows a 0-mean gaussian distribution and can be expressed as:
wherein N represents a gaussian distribution.
Similarly, suppose that the litigation factor matrix V of the defendant company is also subject to 0-mean Gaussian distribution with a variance of
Wherein, deltaVIs the standard deviation;
in order to better represent the patent hidden vector, on one hand, patent litigation record information is combined, and on the other hand, the meta-feature and the text feature of the patent are extracted. It is assumed that the patent hidden vector is determined by two factors, one is the weight in the convolutional neural network, and the other is the input patent feature. Assume the implicit vector of patent k is of the form:
wherein, PkIs the litigation factor, O, of the patent k that we finally expect to obtainkIs the comprehensive vector characterization of the final patent k in step 13, i.e., Ok=NCNN(W,Xk) Wherein X iskDenotes the characteristic of patent k, ∈kIs PkAnd OkDifference between, we assume εkObedience mean 0 and varianceOf a Gaussian distribution ofPIs the standard deviation.
The probability of a patent latent vector is thus expressed as:
wherein X is a characteristic of all patents;
for each weight W in WqAssuming they also obey a mean of 0 and a variance ofGaussian distribution of (a):
in the above formula, δwIs the standard deviation.
Through the method, the output (namely the comprehensive vector representation) of the NCNN can be used as the mean value of the Gaussian distribution of the hidden vector of the patent, and the mean value is connected with the convolutional neural network and tensor decomposition to play a role of a bridge.
Finally, the parameters U, V, P, W and the variances corresponding to these parameters in the hybrid model can be trained.
It will be appreciated by those skilled in the art that the above mentions various forms of standard deviation δU、δV、δP、δwWhich in turn correspond to the parameters U, V, P, W.
And step 16, calculating litigation factors of each company as original reports, as told litigation factors and as patent litigation factors by using the trained mixed model.
When calculating litigation factors as original reports, as rejected litigation factors and for each patent for each company using the trained hybrid model, the following objective functions are maximized:
taking the logarithm of the posterior distribution and taking the negative to obtain:
in the above formula, λU、λV、λW、λPAre all parameters that need to be adjusted to minimize the objective function.
In order to minimize the objective function, Adadelta optimizer is adopted to iteratively update parameters in the model, and the Adadelta optimizer can automatically derive parameters through Tensorflow so as to calculate litigation factor U of the original reporting company iiLitigation factor V for the defendant company jjAnd litigation factor P of patent kkAll three litigation factors are characterized by vectors.
And step 17, predicting the litigation risk of a patent among companies by using the three litigation factors.
The prediction result of the possibility that original carrier i and defendant j will litigation about patent k is denoted as r'ijkThe calculation formula is as follows:
wherein, Ui、VjAnd PkRespectively generation by generationLitigation factors for the indictor i, litigation factors for the indictor j, and litigation factors for the patent k.
According to the technical scheme of the embodiment of the invention, the litigation factors of each company as the original notice, the told litigation factors and the litigation factors of each patent are obtained by combining tensor decomposition and a convolutional neural network according to heterogeneous patent data (meta-features, text features and patent citation network) and patent litigation case records, and then the patent litigation risk prediction is carried out by using the three litigation factors (all vector representations), so that the accuracy of the prediction result is greatly improved compared with the prior art.
Through the above description of the embodiments, it is clear to those skilled in the art that the above embodiments can be implemented by software, and can also be implemented by software plus a necessary general hardware platform. With this understanding, the technical solutions of the embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a usb disk, a removable hard disk, etc.), and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods according to the embodiments of the present invention.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (8)
1. A method for predicting risk of patent litigation, comprising:
obtaining patent data in an authorized heterogeneous form, and crawling patent litigation case data; wherein the granted heterogeneous forms of patent data include: meta-features, text features, and patent citation networks; using SpWhere { k | k ═ 1,2,3, …, N } denotes a patent set, N isThe number of patents; establishing a patent citation network G according to a patent citation relation, wherein each node of the patent citation network G represents a patent; for patent k in patent citation network G, XkRepresenting the characteristics of the characters, including meta characteristics and text characteristics; the meta-features refer to basic features that can be directly extracted from patent documents, and include: forward reference, backward reference, claim quantity, picture quantity, table quantity, patent classification information, authorization gap, change of patent quantity in the patent category, change of patent quantity in the company; text features refer to textual descriptions in patent documents, including: patent titles, abstract of the specification of the patent and patent claims;
transforming patent data in a heterogeneous form of each granted patent into comprehensive patent vector representation by using a convolutional neural network and a network embedding learning method; processing the meta-features and the patent citation network G by adopting a network embedding learning mode, and processing the text features by adopting a convolutional neural network; vector representation of text features is obtained through a convolutional neural network, and vector representation of meta-features is obtained through a network embedding learning mode; splicing the vector representation of the text features and the vector representation of the meta features, and then obtaining comprehensive patent vector representation through a full connection layer;
establishing an original notice-defendant-patent third-order tensor and a tensor decomposition model by using patent litigation case data;
combining the comprehensive patent vector characterization with a tensor decomposition model to obtain a hybrid model, comprising: the reason why litigation occurs for patent k in (i, j) is divided into the following 4 aspects: 1) the value of patent k itself; 2) degree of importance of original company i to patent k; 3) degree of importance of the advertised company j to the patent k; 4) competitive relationship between the original company i and the defendant company j; the comprehensive vector representation is obtained by splicing the vector representation of the text features obtained by the convolutional neural network and the vector representation of the meta-features obtained by a network embedding learning mode, the aspect 1 can be described by utilizing the vector representation of the text features obtained by the convolutional neural network, and the aspects 2) to 4 can be described by a tensor decomposition model); unifying the patent litigation factors obtained by the comprehensive patent vector representation and the tensor decomposition model to form a mixed model;
training the mixed model by using a sequencing learning method;
calculating litigation factors of each company as original reports, litigation factors as defended reports and litigation factors of each patent by using the mixed model obtained by training;
the three litigation factors are used for predicting the litigation risk of a patent among companies.
2. The method of claim 1, wherein the patent data in the authorized heterogeneous form is represented in a unified mathematical form after being obtained.
3. The method for predicting risk of patent litigation according to claim 2, wherein the meta-features and the patent citation network G are processed by network embedding learning: splicing the meta-features of each patent node in the patent citation network G into a meta-feature vector, and training the meta-features as the attributes of the patent nodes for network embedding learning; then, by carrying out network embedding learning on the patent citation network G, the patent element feature vector can be converted into a patent representation, and meanwhile, the patent citation relation is embedded into the patent representation, so that the patent feature is described; the treatment process comprises the following steps:
firstly, the element characteristic vectors of all patents are spliced to form a characteristic matrix FN×QWherein Q is the dimension of the characteristic vector of the patent element; feature matrix FN×QIs denoted as fkRepresenting the meta feature vector of patent k; definition of input characterization of patent k as ek=ETfkWhere E is the transformation matrix to be trained;
secondly, for each node in the patent citation network G, taking it as a root node root, randomly sampling its neighbor nodes to generate different paths:
<root,neighborhood1,neighborhood2,…>;
wherein, the neighbor nodes are represented by neighbor 1 and 2;
for each path, the set of neighbor nodes for patent k is given:
context(k)={k-l,…,k+l}\{k};
the following objective function, i.e. the probability of predicting the central node by the neighbor nodes, is maximized:
e 'in the formula'kAnd econtext(k)Respectively representing the output representation and scene representation, e 'of patent k'mRepresenting the output representation of a patent m, wherein the patent m is a neighbor node of a patent k;
econtext(k)is defined as:
in the above formula emThe input characterization of the patent m;
finally, by approximating the objective function by negative sampling, an output representation is obtained, i.e. a vector representation of the meta-features, whose output representation is e 'for patent k'k。
4. A method for predicting risk of patent litigation according to claim 2 or 3, wherein a convolutional neural network is used to process the text features as follows:
firstly, words in the text features after the stop words are removed are converted into Word vectors with the dimension d through the Word2Vec technology0;
Secondly, dividing the patent titles, the abstract of the patent specification and the patent claims into a plurality of pieces, wherein C-2 pieces are taken before the patent claims, and a patent title and an abstract of the patent specification are added to form C pieces; each one from word to wordQuantity-formed word sequences, each of which takes the first H word sequences, then the text features of the patent k are converted into tensors
Then, a two-layer convolutional neural network pair tensor is usedCarrying out treatment; the first layer is convolution and pooling of word level, and the second layer is convolution and pooling of sentence level; the convolution and pooling operations of the first layer are as follows:
and (3) convolution operation: the convolution kernel shape of the convolution operation is c' × d0Where c' is the dimension of the convolution kernel, the goal is to pass each slice s of patent k through a convolution operation as w1,w2,…,wHAll are converted into a new hidden layer sequence ec:Wherein: andis a parameter of the convolution operation, d is the dimension of the output, relu (x) is a non-linear activation function,is a join operation to join vectors; w is a word sequence, and subscripts are sequence numbers of the word sequence;
and (3) pooling operation: step size of pooling operation is u, new hidden layer sequence e obtained for the above convolution operationcCarrying out uMax pooling and converting into a new global hidden layer sequence ecu:Wherein the content of the first and second substances,satisfy the requirement of
The convolution and pooling operations of the second layer are the same as the first layer, except that the input is changed from a sequence of words to a sequence of sentences.
5. The method of claim 1, wherein the step of establishing a source-defendant-patent third-order tensor using patent litigation case data to form a tensor resolution model comprises:
given M companies and N patents, assume the original company set is SU1,2,3, …, M, the advertised company set is SV1,2,3, …, M, patent set SpForming a third-order tensor by taking the three as coordinate axes, wherein each value of the third-order tensor represents whether litigation occurs to a certain patent by a company;
suppose the litigation case set is R ═ { Rijk|i∈SU,j∈SV,k∈SpGiven a grandfather i and a defendant j, called a company pair (i, j), if the grandfather i calls the defendant j about the patent k, a litigation record r is obtainedijk1 is ═ 1; if the company does not have litigation for k for (i, j) at present, the litigation record rijk=0;
The prediction of patent litigation risk refers to a prediction of the possibility that a company will litigation for patent k in relation to patent (i, j) given company pair (i, j) and patent k, and the prediction result is denoted as r'ijk(ii) a The litigation probability is decomposed into the following forms:wherein, Ui、VjAnd PkRepresenting litigation factors of the original reporting company i, litigation factors of the reported company j, and litigation factors of the patent k, respectively.
6. The method for predicting risk of patent litigation according to claim 5, wherein the training of the mixture model by using the method of ranking learning is as follows:
the possibility of patent litigation between companies is similar to user preferences in traditional recommendation systems, namely: if patent k is litigation patent, it is marked as k+If patent k is a non-litigation patent, it is recorded as k-(ii) a If it is notThen relative patent k-In other words, company prefers patent k to (i, j)+Then the problem translates into a predictionHere, theWherein k is+Indicating that patent k is a litigation patent, k-Patent k is denoted as a non-litigant patent;represents a company pair (i, j) and is related to k+The occurrence of litigation is caused,represents a company pair (i, j) and is related to k-No litigation occurs;
given a company pair (i, j), k is used+>i,jk-To indicate that the company is paired with (i, j) for patent k+And k-Bias-order relationship of preference degrees; then company pairs (i, j) in patent k+Litigation occurred in the above, but not in patent k-The probability of litigation occurring is expressed as:
wherein σ (x) is a sigmoid function;
from the perspective of probability theory, finding the best patent sequenceThe problem can be solved by maximizing the following posterior distribution:
p(U,V,P,W|>i,j)∝p(>i,j|U,V,P,W)p(U,V,P,W);
wherein >i,jAll partial order relations which represent possible patent disputes between the original company i and the reported company j; u, V, P and W respectively represent a litigation factor matrix of the original reporting company, a litigation factor matrix of the reported company, a patent litigation factor matrix and all parameter matrices in the convolutional neural network;
assuming that all company pairs are independent of each other, there are:
the conditional distribution of the above formula is:
wherein the content of the first and second substances,the partial order relation of all company litigation patents is stored;
as a generative model, suppose the litigation factors of the original agent obey a 0-mean Gaussian distribution with a variance ofThe litigation factor matrix U of the original trailer also follows a 0-mean gaussian distribution, expressed as:
wherein N represents a Gaussian distribution;
suppose that litigation factor matrix V of the defendant company also obeys a 0-mean gaussian distribution with a variance of
Wherein, deltaVIs the standard deviation;
assume the implicit vector of patent k is of the form:
in the above formula, PkIs litigation factor, O, of patent kkIs a comprehensive vector characterization of patent k, i.e. Ok=NCNN(W,Xk) Wherein X iskDenotes the characteristic of patent k, ∈kIs PkAnd OkDifference between, let us assume ∈kObedience mean 0 and varianceOf a Gaussian distribution ofPIs the standard deviation;
the probability of a patent latent vector is thus expressed as:
in the above formula, X is the characteristic of all patents;
for each weight W in WqAssuming they also obey a mean of 0 and a variance ofThe mean gaussian distribution of (a), as follows:
in the above formula, δwIs the standard deviation;
finally, the parameters U, V, P, W and the variances corresponding to these parameters in the hybrid model can be trained.
7. The method of claim 6, wherein the following objective functions are maximized when the litigation factors as a source, the litigation factors as a destination, and the litigation factors for each patent are calculated using the trained mixture model:
taking the logarithm of the posterior distribution and taking the negative to obtain:
in the above formula, λU、λV、λW、λPAll parameters need to be adjusted when the objective function is minimized;
iterative updating of parameters in the model by using Adadelta optimizer to calculate litigation factor U of original trailer company iiLitigation factor V for the defendant company jjAnd litigation factor P of patent kk。
8. The method for predicting risk of litigation for a patent according to claim 1, wherein the prediction of risk of litigation for a patent between companies using the three litigation factors comprises:
the prediction result of the possibility that original carrier i and defendant j will litigation about patent k is denoted as r'ijkThe calculation formula is as follows:
wherein, Ui、VjAnd PkRepresenting litigation factors of the original reporting company i, litigation factors of the reported company j, and litigation factors of the patent k, respectively.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810482063.3A CN108665182B (en) | 2018-05-18 | 2018-05-18 | Patent litigation risk prediction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810482063.3A CN108665182B (en) | 2018-05-18 | 2018-05-18 | Patent litigation risk prediction method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108665182A CN108665182A (en) | 2018-10-16 |
CN108665182B true CN108665182B (en) | 2021-12-14 |
Family
ID=63776982
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810482063.3A Active CN108665182B (en) | 2018-05-18 | 2018-05-18 | Patent litigation risk prediction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108665182B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112949665A (en) * | 2019-11-26 | 2021-06-11 | 阿里巴巴集团控股有限公司 | Data processing method, device, equipment and storage medium |
CN111309900B (en) * | 2020-01-17 | 2022-09-06 | 中国科学技术大学 | Legal class similarity judging and pushing method |
CN115796285B (en) * | 2023-02-13 | 2023-05-09 | 上海百事通法务信息技术有限公司浙江分公司 | Litigation case pre-judging method and device based on engineering model and electronic equipment |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104850905A (en) * | 2015-05-15 | 2015-08-19 | 上海集本信息技术服务有限公司 | Machine-learning-based legal risk quantitative evaluation system and method |
CN105631783A (en) * | 2015-12-25 | 2016-06-01 | 深圳联合产权交易所股份有限公司 | Objective and quantitative Chinese patent evaluation system and method |
CN105677726A (en) * | 2015-12-29 | 2016-06-15 | 上海律巢网络科技有限公司 | Data search and result presenting method and system |
CN105677727A (en) * | 2015-12-29 | 2016-06-15 | 上海律巢网络科技有限公司 | Data search and search results presenting method and system |
CN107015981A (en) * | 2016-01-27 | 2017-08-04 | 广州知商互联网科技有限公司 | A kind of risk control of patent operation and evaluation system |
CN107122444A (en) * | 2017-04-24 | 2017-09-01 | 北京科技大学 | A kind of legal knowledge collection of illustrative plates method for auto constructing |
CN107784603A (en) * | 2017-09-27 | 2018-03-09 | 国家电网公司信息通信分公司 | A kind of intellectual property operation system and method |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20030028406A1 (en) * | 2001-07-24 | 2003-02-06 | Herz Frederick S. M. | Database for pre-screening potentially litigious patients |
US8355932B2 (en) * | 2005-04-08 | 2013-01-15 | Ocean Tomo Llc | System and method for managing intellectual property-based risks |
-
2018
- 2018-05-18 CN CN201810482063.3A patent/CN108665182B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104850905A (en) * | 2015-05-15 | 2015-08-19 | 上海集本信息技术服务有限公司 | Machine-learning-based legal risk quantitative evaluation system and method |
CN105631783A (en) * | 2015-12-25 | 2016-06-01 | 深圳联合产权交易所股份有限公司 | Objective and quantitative Chinese patent evaluation system and method |
CN105677726A (en) * | 2015-12-29 | 2016-06-15 | 上海律巢网络科技有限公司 | Data search and result presenting method and system |
CN105677727A (en) * | 2015-12-29 | 2016-06-15 | 上海律巢网络科技有限公司 | Data search and search results presenting method and system |
CN107015981A (en) * | 2016-01-27 | 2017-08-04 | 广州知商互联网科技有限公司 | A kind of risk control of patent operation and evaluation system |
CN107122444A (en) * | 2017-04-24 | 2017-09-01 | 北京科技大学 | A kind of legal knowledge collection of illustrative plates method for auto constructing |
CN107784603A (en) * | 2017-09-27 | 2018-03-09 | 国家电网公司信息通信分公司 | A kind of intellectual property operation system and method |
Non-Patent Citations (1)
Title |
---|
Minimizing Legal Exposure of High-Tech Companies through Collaborative Filtering Methods;Bo Jin等;《Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining》;20160813;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN108665182A (en) | 2018-10-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Darko et al. | Artificial intelligence in the AEC industry: Scientometric analysis and visualization of research activities | |
Pan et al. | Transfer learning to predict missing ratings via heterogeneous user feedbacks | |
Pavlichenko et al. | Best prompts for text-to-image models and how to find them | |
CN110674850A (en) | Image description generation method based on attention mechanism | |
CN108665182B (en) | Patent litigation risk prediction method | |
Miller et al. | Preference construction, sequential decision making, and trade space exploration | |
Yu et al. | Personalized item recommendation for second-hand trading platform | |
CN113761359B (en) | Data packet recommendation method, device, electronic equipment and storage medium | |
Ghafouri et al. | A survey on exponential random graph models: an application perspective | |
Wang et al. | Adversarial training-based mean Bayesian personalized ranking for recommender system | |
CN115221413B (en) | Sequence recommendation method and system based on interactive graph attention network | |
Wang et al. | Combined approach for government E-tendering using GA and TOPSIS with intuitionistic fuzzy information | |
Wu et al. | Inferring emotional tags from social images with user demographics | |
Eide et al. | Deep neural network marketplace recommenders in online experiments | |
Kozhevnikov et al. | Digital trust as a basis for the digital transformation of the enterprise and economy | |
Zhang et al. | User multi-modal emotional intelligence analysis method based on deep learning in social network big data environment | |
CN112905906A (en) | Recommendation method and system fusing local collaboration and feature intersection | |
Sima et al. | Lsh-smile: Locality sensitive hashing accelerated simulation and learning | |
Lan et al. | Learning from the past: uncovering design process models using an enriched process mining | |
CN115935067A (en) | Article recommendation method integrating semantics and structural view for socialized recommendation | |
Martinez et al. | Towards estimating and predicting user perception on software product variants | |
Stephens et al. | Bayesian analysis of discrete time warranty data | |
Barbon Jr et al. | Trace encoding in process mining: a survey and benchmarking | |
Zhao et al. | A DEMATEL and consensus based MCGDM approach for with multi-granularity hesitant fuzzy linguistic term set | |
Zhang et al. | Improvement and optimization of slicing algorithm for rapid prototyping manufacturing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |