Patent valuation method, electronic device and medium based on dynamic game attention and graph comparison learning

Abstract

This invention proposes a patent valuation method, electronic device, and medium based on dynamic game-theoretic attention and graph contrastive learning, belonging to the field of patent valuation technology. The method includes: dynamically weighting the multidimensional features of the transferor and transferee using an autoencoder and attention mechanism to generate subject feature representations; jointly modeling the patent's own features and its topological relationships in the citation network using a graph contrastive learning method to obtain a graph embedding representation of the patent; adaptively aggregating the three types of representations of the transferor, transferee, and patent through a cross-modal attention mechanism to generate a global representation integrating multi-source information; and inputting the global representation into a multilayer perceptron for patent price range prediction. This invention improves the accuracy, robustness, and interpretability of patent valuation prediction through the synergistic optimization of dual attention mechanisms and graph learning, providing intelligent decision support for patent transaction pricing.

Description

Technical Field

[0001] This invention belongs to the field of patent evaluation technology, and specifically relates to a patent value evaluation method, electronic device and medium based on dynamic game attention and graph comparison learning. Background Technology

[0002] With the rapid development of the global intellectual property market, patents, as carriers of core technologies, have become a research hotspot in terms of transaction price evaluation. Predicting patent prices has always been a complex and challenging problem. Existing patent price prediction methods typically rely on the structured information of the patent itself, such as in "Wan Xiaoli, Zhu Xuezhong. Evaluation Index System and Fuzzy Comprehensive Evaluation of Patent Value [J]. Scientific Research Management, 2008(02):185-191." However, these methods have significant limitations; relying solely on the structured information of the patent itself cannot fully reflect the complexity and multi-party interactions in patent transactions, and automation is not possible.

[0003] With the development of artificial intelligence technology, deep learning and graph learning technologies are increasingly being applied to patent valuation. Chinese patent application CN111861046A discloses an intelligent patent valuation system based on big data and deep learning, which uncovers important information hidden in patent text. However, patent value is not solely related to the technical information reflected in the text, and valuation has limitations. Chinese patent application CN108416535A discloses a deep learning-based patent valuation method that integrates textual information, attribute features, and patent citation relationships to assess patent value. Chinese patent application CN115983877A discloses a patent valuation method based on deep graphs and semantic learning. Unlike traditional text analysis, it constructs a semantic novelty index and combines patent citation relationships to assess patent value. Chinese patent application CN103679291A discloses a patent valuation method that mentions the impact of the patentee's company strength score and the technical competitiveness score of the patent being evaluated on patent value.

[0004] In practice, patent transactions are processes where technology suppliers and buyers agree to license or transfer patents through market prices. The final price of a patent depends not only on its inherent characteristics but also on the attributes of both the supplier and the buyer.

[0005] The aforementioned existing technologies lack feature analysis of both supply and demand sides, or have not yet utilized artificial intelligence technology, thus failing to automate the evaluation of large-scale data. Furthermore, patent valuation faces uncertainties influenced by technological, market, and legal factors. Traditional methods (such as expert evaluation or indicator-based methods) typically rely on manually setting feature weights when assessing the economic value of patents, lacking objectivity and flexibility.

[0006] Based on this, this invention proposes a new patent value prediction method by introducing a dual attention mechanism and graph learning method, combining information from both supply and demand sides with the technological relevance in the patent network. This method simulates the complex interactive relationships and market dynamics in patent transactions, thereby improving the accuracy and interpretability of patent price prediction. Summary of the Invention

[0007] The purpose of this invention is to provide a patent valuation method based on dynamic game attention and graph contrastive learning. Taking the patent technology value and market game relationship as the starting point, it employs an attention mechanism to dynamically assign weights to the characteristics of supply and demand entities, and combines graph representation learning methods to mine deep technological connections in the patent citation network. This establishes a valuation model that integrates market behavior and technology networks to predict patent transaction prices. This invention breaks through the reliance of traditional methods on single patent features, and significantly improves the accuracy of predictions through the deep integration of market dynamic game modeling and technology network connection mining.

[0008] The technical solution of this invention is as follows:

[0009] The first aspect of this invention provides a patent value assessment method based on dynamic game attention and graph contrast learning, comprising the following steps:

[0010] S1. Collect multi-source patent transaction data, including patent attribute features and citation relationships, transferor attribute features and transferee attribute features, and preprocess the data to obtain a dataset;

[0011] S2. An autoencoder is used to convert the multidimensional features of the transferor and the transferee into vector representations respectively. Then, based on the attention mechanism, the vector representations of the multidimensional features of the transferor and the transferee are weighted and aggregated to generate the transferor representation and the transferee representation.

[0012] S3. Based on the attribute features and citation relationship data of the patent, two heterogeneous views are generated using the GRACE algorithm. The two heterogeneous views are combined to train the encoder, and the trained encoder is used to generate the patent representation.

[0013] S4. Using an attention mechanism, the transferor representation, the transferee representation, and the patent representation are weighted and aggregated to obtain the final representation;

[0014] S5. Input the final representation into a multilayer perceptron to predict the patent value label.

[0015] As a further improvement to this technical solution, the multi-source patent transaction data includes patent transaction record data, patent attribute feature data, patent citation relationship data, attribute feature data of the patent transferor, and attribute feature data of the patent transferee; wherein, each patent transaction record data contains one transferor, one transferee, and the patent; the dataset includes a patent transaction record dataset D. T Patent attribute feature dataset D px Patent citation relationship dataset D pa Patent assignor attribute feature dataset D t And the patent assignee attribute feature dataset D r The patent attribute feature dataset D px The total number of patents involved in transactions, including the patents cited and referenced by these patents, is q, of which the number of patents involved in transactions is q′.

[0016] As a further improvement to this technical solution, the preprocessing includes operations such as removing noisy data, filling missing values, and data normalization; wherein, numerical features are Max-Min standardized, text features are represented using SciBERT encoding, and classification features are converted into low-dimensional vectors using one-hot encoding.

[0017] As a further improvement to this technical solution, in S2, the autoencoder is used to convert the multidimensional features of the transferor and the transferee into vector representations, and the specific process is as follows:

[0018] Based on the patent assignor attribute feature dataset D t Extract multi-dimensional attributes of the patent holder (assignor), and base them on the patent assignee attribute feature dataset D. r Extract the multi-dimensional attributes of the transferee and construct the transferor feature matrix X respectively. t and the transferee's characteristic matrix X r :

[0019]

[0020] in, For the i-th transferor d t A feature vector of dimension i, i∈[1,m], where m is the number of transferors; For the j-th transferee, d r A feature vector of dimension, j∈[1,n], where n is the number of transferees;

[0021] Use separate encoders to transfer the transferor's d t Dimensional features and transferee's d rMap the 1D features to a k-dimensional space, and extract the hidden layer representation H of the transferor's features respectively. t Hidden layer representation H of transferee characteristics r :

[0022]

[0023] Among them, the output hidden layer matrix k is the dimension of the hidden layer; The weights are the encoder weights for the transferor, used to map the transferor features to the hidden layer space; The encoder bias is used to enhance the encoder's nonlinear expression capability; the output hidden layer matrix is ​​used. k is the dimension of the hidden layer; The weights are the encoder weights for the transferee, used to map the transferee's features to the hidden layer space; The encoder bias is applied to the transferee to enhance the encoder's nonlinear expression capability. To preserve the differences in characteristics between the transferor and the transferee;

[0024] Reconstruct the original features of the transferor and transferee, that is, obtain the hidden layer representation H from the decoder. t and hidden layer representation H r The original characteristics of the transferor, X, are restored respectively. t and the original features of the transferee X r Decoding features are obtained. and

[0025]

[0026] in, These are the weights for the transferor decoder, used to map the transferor's hidden layer representation back to the transferor's features; This is used as a decoder bias for the transferor to enhance the decoder's nonlinear expression capability; These are the transferee decoder weights, used to map the transferee's hidden layer representation back to the transferee's features; σ is the decoder bias for the transferee, used to enhance the decoder's nonlinear expressive power; σ is the Sigmoid function.

[0027] The autoencoder loss function is designed as follows:

[0028]

[0029] in, To reconstruct the loss, the autoencoder is constrained to achieve high-fidelity reconstruction of the original features; λ is the regularization term, used to prevent overfitting between the encoder and decoder; λ is the regularization coefficient, used to control the regularization strength of the encoder / decoder parameters. For the hidden layer alignment term, the trace of similarity between the transferor's representation and the transferee's representation is calculated to enhance the matching of dual-channel features; μ is an adjustment coefficient used to control the alignment strength between the transferor's and transferee's hidden layer representations.

[0030] As a further improvement to this technical solution, in S2, the specific process of calculating the multi-dimensional feature attention weights of the transferor and transferee based on the attention mechanism, and generating the representations of the transferor and transferee through weighted aggregation is as follows:

[0031] d for each transferor t Weights are assigned to the features respectively, and the following is used: and Calculate the importance of each feature in the transferor's hidden layer representation:

[0032]

[0033] in, This represents the weight of the l-th dimension feature of the i-th transferor; This represents the eigenvalue of the l-th dimension in the decoded feature of the i-th transferor; Let l represent the eigenvalue of the l-th dimension in the hidden layer representation of the i-th transferor; l∈[1,d] t ];

[0034] Similarly, calculate d for each transferee. r Weights of dimensional features:

[0035]

[0036] in, This represents the weight of the g-dimensional feature of the j-th transferee; This represents the feature value of the g-th dimension in the decoded feature of the j-th transferee; Let g represent the eigenvalue of the g-th dimension in the hidden layer representation of the j-th transferee; g∈[1,d] r ];

[0037] Hidden layer representation for each transferor By feature weights Weighted averages are applied to obtain the final representation from each transferor. and the final statement of all transferors

[0038]

[0039]

[0040] Similarly, the final representation of each transferee and the final statements of all transferees They are as follows:

[0041]

[0042] As a further improvement to this technical solution, in S3, the attribute features and citation relationship data based on the patent are used to generate two heterogeneous views using the GRACE algorithm, as follows:

[0043] Based on the patent feature dataset D px Extract multi-dimensional attributes of patents and construct a patent feature matrix X. p :

[0044]

[0045] Where q is the total number of patents, d p This refers to the total dimension of patent features;

[0046] Based on the patent citation relationship dataset D pa Construct a patent adjacency matrix A, if patent a u Patent a was cited v If u∈[1,q], then A uv =1, otherwise A uv =0:

[0047] A∈{0,1} q×q

[0048] Add self-loops to enhance the propagation of information within the node itself:

[0049]

[0050] For each edge Perform in-degree normalization:

[0051]

[0052] Among them, D in,uu The in-degree of a patent, i.e., the number of times it has been cited; I q An identity matrix; a normalized directed adjacency matrix containing self-loops.

[0053] Structural disruption (RE) is performed on the edges in the patent adjacency matrix, deleting edges from adjacency matrix A with probability p1, i.e., if patent a u With patent a v If directed edges exist between patents, the directed references between patents are randomly removed, generating two topologically heterogeneous views:

[0054] See Figure 1 :

[0055]

[0056] View 2:

[0057]

[0058] Where ° represents the Hadamard product. It is an asymmetric mask matrix;

[0059] The in-degree matrix is ​​independently calculated and normalized for each generated view:

[0060] See Figure 1 :

[0061]

[0062] View 2:

[0063]

[0064] Feature masking (MF) is performed on the node features in the patent feature matrix, and the patent feature matrix X is randomly selected with probability p2. p By setting some dimensions of the node features to zero, i.e., randomly masking some feature dimensions of the patent, two heterogeneous views are generated:

[0065] See Figure 1 :

[0066]

[0067] View 2:

[0068]

[0069] in, For feature mask vectors;

[0070] The generated and Combine them to generate two heterogeneous views. and :

[0071]

[0072]

[0073] As a further improvement to this technical solution, in S3, the method for training the encoder is as follows:

[0074] A GCN is used as the encoder f(·), with the first layer employing the ReLU function for nonlinear activation and the second layer performing linear propagation; the input is the generated directed normalized adjacency matrix. With characteristic matrix Specifically:

[0075] 1) First-layer nonlinear activation:

[0076]

[0077] in, These are the weights of the first layer of GCN, which are trainable parameters;

[0078] 2) Second-layer nonlinear propagation:

[0079]

[0080] in, The weights for the second layer of GCN are trainable parameters; the output is the patent node embedding.

[0081] 3) Generate embeddings across views:

[0082] The two views generated from the input and To the encoder with shared weights, generate node embeddings:

[0083]

[0084] Where U and V are view embedding matrices, d e <<d p ;

[0085] Positive sample pairs are the embeddings of the same patent node in two different views (s u ,w u Negative samples are embeddings of other patent nodes in different views or the same view, and each positive sample pair (s) is evaluated based on cosine similarity. u ,w u Calculate the comparative loss :

[0086]

[0087] Among them, the discriminator θ(s) u ,w u ) is s u With w u The cosine similarity is calculated using the following formula:

[0088]

[0089] A nonlinear projection head g(s) = MLP(s) is introduced to enhance the discriminativeness of contrastive learning. τ is a temperature parameter (default τ = 0.5) used to adjust the sharpness of the similarity distribution.

[0090] The overall objective function is the average of all positive sample pairs, calculated as follows:

[0091]

[0092] The patent representation is generated using the trained GCN, specifically as follows:

[0093] The original patent map is generated using the constructed patent feature matrix and adjacency matrix. The original patent image is then input into the trained encoder f(·), which outputs the final representation of each patent. :

[0094]

[0095] Final representation of all patents Includes patent feature information and citation information, as follows:

[0096]

[0097] As a further improvement to this technical solution, in S4, the attention mechanism is used to perform weighted aggregation of the transferor representation, the transferee representation, and the patent representation to obtain the final representation, as follows:

[0098] The final statement of the transferor will be presented separately. The transferee ultimately stated Projecting onto the patent feature space yields the transferor representation, which has the same dimensions as the patent features. The transferee expressed The calculation formula is as follows:

[0099]

[0100] in, For the transferor's projection matrix, The projection matrix of the transferee;

[0101] The transferor's representation after unifying dimensions The transferee stated With patent representation By splicing the data, we obtain the spliced ​​feature z. i,j,k Used to capture the joint relationship between the three:

[0102]

[0103] splicing feature z i,j,k Mapped to attention score s i,j,k The calculation is as follows:

[0104]

[0105] in, All are trainable parameters, and W t′ W r′ W p′ It will be updated during the backpropagation of the MLP, that is, it is jointly optimized by the backpropagation algorithm and the classification loss; Attention scores are given to the transferor, transferee, and patentee, respectively.

[0106] The calculated transferor attention score Attention score of the transferee Patent attention score After normalization, their respective weights are obtained:

[0107]

[0108] Based on the normalized weights, aggregate the transferor representation after processing to a unified dimension. The transferee stated With patent representation The final representation is obtained :

[0109]

[0110] In the process of feature splicing and aggregation, including All intermediate variables, including the classification loss, maintain computational graph continuity to ensure that gradients can propagate back from the classification loss to the attention parameters.

[0111] As a further improvement to this technical solution, in S5, the final representation is input into a multilayer perceptron to predict the patent value label, as follows:

[0112] The patent value is discretized into C categories, labeled as follows:

[0113] The final representation obtained from input S4 The MLP outputs the class probability distribution, specifically as follows:

[0114] 1) Hidden layer 1:

[0115]

[0116] in, The weights of the first layer of the MLP For the first layer bias of the MLP, d h For the hidden layer dimension;

[0117] 2) Hidden layer 2:

[0118]

[0119] in, For the second layer weights of the MLP, This is the output of the first layer of the MLP. For the second layer bias of the MLP, d h For the hidden layer dimension;

[0120] 3) Output layer:

[0121]

[0122] in, For class probability vectors, For the third layer weights of the MLP, This is the output of the second layer of the MLP. This is the third layer bias of the MLP, where C is the number of classes;

[0123] Cross-entropy loss is used:

[0124]

[0125] In the formula, ||(·) is an indicator function. The labels are the real labels; the gradient is backpropagated from the output layer of the MLP to... Then, the attention weights are calculated and passed to W. t′ W r′ W p′ This allows for the updating of attention weights, and further, through backpropagation, updates to the representations of the patent, the assignor, and the assignee, as well as the attention mechanism parameters involved, thereby enabling accurate prediction of patent value.

[0126] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the electronic device executes the patent value assessment method based on dynamic game attention and graph contrast learning.

[0127] A third aspect of the present invention provides a storage medium comprising a computer program that, when run on an electronic device, causes the electronic device to execute the patent value assessment method based on dynamic game attention and graph contrast learning.

[0128] The beneficial effects of this invention are as follows: By integrating supply and demand game dynamics, patent network correlation, and multi-source heterogeneous data modeling, this invention significantly improves the accuracy, robustness, and interpretability of patent value prediction; specifically, this is reflected in the following aspects:

[0129] (1) Achieve adaptive discrimination of multi-source features

[0130] Based on subject feature attention, key influencing factors in supply and demand sides and patent features are automatically identified, and differentiated weights are assigned to improve model accuracy and decision interpretability.

[0131] (2) Deeply explore the relevance of patented technologies

[0132] The patented method models the topology of the network by contrastive learning. Compared with existing graph learning algorithms, it can better capture the complex relationships between nodes through contrastive learning, and generate a graph embedding representation that integrates its own attributes and network relationships, thus breaking through the limitations of traditional isolated analysis.

[0133] (3) Improve the alignment between forecasts and actual pricing.

[0134] By dynamically empowering the multidimensional features of the transferor, transferee, and patent through cross-modal attention, we model market game behavior, breaking through the limitations of traditional methods that rely on static patent attributes, and making the prediction results more in line with the dynamic pricing logic in actual transaction scenarios.

[0135] In summary, this invention achieves breakthroughs in prediction accuracy, scenario adaptability, and intelligent decision-making through the synergistic innovation of multi-source feature dynamic fusion, deep patent network correlation, and market game modeling. It provides the patent transaction market with an efficient, transparent, and interpretable evaluation tool, facilitating the efficient transformation of intellectual property value. Attached Figure Description

[0136] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0137] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0138] This embodiment uses 2356 patents transferred between universities and enterprises as examples to construct a patent valuation model and verify its effectiveness. Specifically, 1579 patent samples were selected for constructing the patent valuation model, and 777 patent samples were used for patent valuation and model effectiveness verification. The implementation steps of this invention are as follows:

[0139] S1. Collect multi-source patent transaction data and preprocess it to obtain the required dataset; the specific process is as follows:

[0140] This process involves acquiring authentic patent transfer transaction records between universities and enterprises, patent feature data, patent citation relationship data, attribute data of universities (transferors), and attribute data of enterprises (transferees). The data undergoes data cleaning, including noise removal, missing value imputation, and data normalization. Numerical features are standardized using Max-Min encoding, text features are represented using SciBERT encoding, and categorical features are converted into low-dimensional vectors using one-hot encoding. The final result is patent transaction record data D for the target domain. T Patent attribute feature dataset D px Patent citation relationship dataset D pa University Attribute Feature Dataset D t And Enterprise Attribute Feature Dataset D r The patent attribute feature dataset D px The dataset contains transaction patents and the patents cited and referenced by these patents, totaling q, of which q′ are transaction patents. This dataset is obtained to subsequently construct the university feature matrix, enterprise feature matrix, patent feature matrix, and patent adjacency matrix, thereby obtaining vector representations of these three matrices.

[0141] S2. An autoencoder is used to convert the multidimensional features of the transferor and the transferee into vector representations respectively. Based on the attention mechanism, the attention weights of the multidimensional features of the transferor and the transferee are calculated respectively. The vector representations of the multidimensional features of the transferor and the transferee are weighted and aggregated respectively to generate the transferor representation and the transferee representation.

[0142] The generation of transferor and transferee representations is based on an autoencoder, an unsupervised feature learning model based on neural networks. An autoencoder, through a symmetrical structure of encoder and decoder, compresses input data into a latent space and reconstructs it, capturing the inherent distribution patterns of the data in the process. The autoencoder transforms the feature information of the transferor and transferee into vector representations. Specifically:

[0143] S2.1, Construct the transferor's feature matrix and the transferee's feature matrix.

[0144] Based on the university attribute feature dataset D obtained from S1 t Extract multi-dimensional attributes of universities and construct a university feature matrix X. t :

[0145]

[0146] in, Let d be the i-th university t A feature vector of dimension d, i∈[1,m], where m is the number of universities; the universities include 10 feature dimensions such as university address, university ranking, number of teachers, etc., therefore d t It is 10.

[0147] Based on the enterprise attribute feature dataset D obtained from S1 r Extract multi-dimensional attributes of enterprises and construct enterprise feature matrix X. r :

[0148]

[0149] in, For the j-th enterprise d r The feature vector is d, j∈[1,n], where n is the number of enterprises; the enterprises include 19 feature dimensions such as enterprise address, industry, and number of past patent transactions, therefore d r It is 19.

[0150] S2.2, Feature encoding of the transferor and transferee.

[0151] Use an independent encoder to convert the d of the university t Mapping 1D features to a k-dimensional latent space, and extracting the hidden layer representation H of the university features. t :

[0152]

[0153] Among them, the output hidden layer matrix k is the dimension of the hidden layer, which is set to 128 dimensions; These are the encoder weights for universities, used to map university features to the hidden layer space; This is used as a bias for the encoder in high-performance applications to enhance its nonlinear expression capabilities.

[0154] Using an independent encoder to d of the enterprise r The 1D feature is mapped to a k-dimensional latent space, and the hidden layer representation H of the enterprise features is extracted. r And ensure that parameters are not shared:

[0155]

[0156] Among them, the output hidden layer matrix k is the dimension of the hidden layer; These are the enterprise encoder weights, used to map enterprise features to the hidden space; For enterprise encoder bias, used to enhance the encoder's nonlinear expression capability;

[0157] S2.3 Decode and reconstruct the characteristics of the transferor and transferee.

[0158] Reconstructing the original characteristics of universities (d) t (dimensional), that is, representing H from the hidden layer through the decoder. t The original characteristics of Chinese universities Xt This makes the hidden layer representation H t Retain sufficient information to reconstruct the original features of the university with high accuracy:

[0159]

[0160] in, These are the weights for the university decoder, used to map the hidden layer representation of universities back to university features; σ is the bias of the high-performance decoder, used to enhance the nonlinear expression capability of the decoder; σ is the Sigmoid function.

[0161] Reconstructing the original characteristics of the enterprise (d) r (dimensional), that is, representing H from the hidden layer through the decoder. r Restore the original characteristics of the enterprise X r This makes the hidden layer representation H r Retain sufficient information to reconstruct the original characteristics of the enterprise with high accuracy:

[0162]

[0163] in, These are the enterprise decoder weights, used to map the enterprise hidden layer representation back to enterprise features; This is used as a bias for the enterprise decoder to enhance its nonlinear expressive capabilities.

[0164] S2.4, Joint Optimization Objective

[0165] The loss function of the autoencoder is:

[0166]

[0167] in, To reconstruct the loss, the autoencoder is constrained to achieve high-fidelity reconstruction of the original features; This is a regularization term to prevent overfitting between the encoder and decoder; the regularization coefficient λ is set to 0.01. For the hidden layer alignment term, the trace of similarity between the transferor's representation and the transferee's representation is calculated to enhance the matching of dual-channel features; the mutual information coefficient μ is set to 0.1.

[0168] The calculation of attention weights for the transferor and transferee features is based on an attention mechanism. This mechanism is used to assign different weights to the features of universities or enterprises. Features with higher weights indicate less information loss during the compression-decompression process of the autoencoder, and are likely to be features with high information content or low noise. Specifically:

[0169] Different characteristics of each university (d) t (Dimension) assign weights, and use the obtained and Calculate the importance of each feature in the hidden layer representation of this university:

[0170]

[0171] in, This represents the weight of the l-th dimension feature of the i-th university; This represents the eigenvalue of the l-th dimension in the decoded feature of the i-th university; Let l represent the eigenvalue of the l-th dimension in the i-th hidden layer representation of the university; l∈[1,d] t ].

[0172] Similarly, calculate enterprise characteristics (d) r Dimensional weights:

[0173]

[0174] in, This represents the weight of the g-dimensional feature of the j-th transferee; This represents the feature value of the g-th dimension in the decoded feature of the j-th transferee; Let g represent the eigenvalue of the g-th dimension in the hidden layer representation of the j-th transferee; g∈[1,d] r ].

[0175] Hidden layer representation for each university By feature weights Weighted summaries are applied to obtain the final representation for each university. and the final statement from all universities

[0176]

[0177] Similarly, we obtain the final representation from each company. and the final statement of all enterprises

[0178]

[0179] S3. Based on the attribute features and citation relationship data of patents, two heterogeneous views are generated using the GRACE algorithm. These two heterogeneous views are then combined to train an encoder, which is used to generate patent representations. The GRACE algorithm is a graph representation learning algorithm based on contrastive learning. It constructs two different graph views by performing structure disruption (RE) on edges in the patent adjacency matrix and feature masking (MF) on node features in the patent feature matrix, maximizing the representation consistency of the same node in both views to train the GCN. Details are as follows:

[0180] S3.1 Constructing the patent feature matrix and patent adjacency matrix

[0181] Based on the patent attribute feature dataset D obtained from S1 px Extract multi-dimensional attributes of patents and construct a patent feature matrix X. p :

[0182]

[0183] Where q is the total number of patents; d p This refers to the total dimensions of patent features; a patent includes 16 feature dimensions such as patent title and abstract, application date, and number of inventors, therefore d p It is 16.

[0184] Based on the patent citation relationship dataset D obtained from S1 pa Construct a patent adjacency matrix A, if patent a u Patent a was cited u If u∈[1,q], then A uv =1, otherwise A uv =0:

[0185] A∈{0,1} q×q

[0186] Add self-loops to enhance the propagation of information within the node itself:

[0187]

[0188] For each edge Perform in-degree normalization:

[0189]

[0190] Among them, D in,uu The in-degree of a patent, i.e., the number of times it has been cited; I q An identity matrix; a normalized directed adjacency matrix containing self-loops.

[0191] S3.2, Use the GRACE algorithm to generate two views.

[0192] S3.2.1, Perform structural disruption (RE) on the edges in the patent adjacency matrix.

[0193] 1) Generate two topologically heterogeneous views

[0194] Delete edges in the constructed adjacency matrix A with probability p1 (edge ​​deletion probability p1 is set to 0.4), that is, if patent a u With patent a v If directed edges exist between patents, randomly remove directed references between them; generate two topologically heterogeneous views respectively:

[0195] See Figure 1 :

[0196]

[0197] View 2:

[0198]

[0199] Where ° represents the Hadamard product. It is an asymmetric mask matrix.

[0200] 2) Normalized update:

[0201] The in-degree matrix is ​​independently calculated and normalized for each generated view:

[0202] See Figure 1 :

[0203]

[0204] View 2:

[0205]

[0206] S3.2.2, Perform feature masking (MF) on the node features in the patent feature matrix.

[0207] The patent feature matrix X is randomly selected according to probability p2 (feature mask probability p2 is set to 0.3). p Some dimensions of the features in the middle nodes are set to zero, i.e., some feature dimensions of the patent are randomly masked; two heterogeneous views are generated respectively:

[0208] See Figure 1 :

[0209]

[0210] View 2:

[0211]

[0212] in, This is the feature mask vector.

[0213] S3.2.3, Combine the views generated after structural destruction and feature masking.

[0214] The generated and Combine them to generate two heterogeneous views. and :

[0215]

[0216] S3.3, Training the graphical encoder

[0217] A GCN is used as the encoder f(·), with the first layer employing the ReLU function for nonlinear activation and the second layer performing linear propagation; the input is the generated directed normalized adjacency matrix. With characteristic matrix Specifically:

[0218] 1) First-layer nonlinear activation:

[0219]

[0220] In the above formula, The first layer weights of the GCN are trainable parameters; the hidden layer dimension h is set to 256 layers.

[0221] 2) Second-layer nonlinear propagation:

[0222]

[0223] In the above formula, The weights for the second layer of GCN are trainable parameters; the output is the patent node embedding. Output embedding dimension d e Set to 64 dimensions.

[0224] 3) Generate embeddings across views:

[0225] The two views generated from the input and To the encoder with shared weights, generate node embeddings:

[0226]

[0227] Where U and V are view embedding matrices. d e <<d p .

[0228] S3.4, Comparison Loss Calculation

[0229] Positive sample pairs are the embeddings of the same patent node in two different views (s u ,w u Negative samples are embeddings of other patent nodes in different views or the same view, and each positive sample pair (s) is evaluated based on cosine similarity. u ,w u Calculate the comparative loss :

[0230]

[0231] Among them, the discriminator θ(s) u ,w u ) is s u With w u The cosine similarity is calculated using the following formula:

[0232]

[0233] A nonlinear projection head g(s) = MLP(s) is introduced to enhance the discriminativeness of contrastive learning. τ is a temperature parameter (default τ = 0.5) used to adjust the sharpness of the similarity distribution.

[0234] The overall objective function is the average of all positive sample pairs, calculated as follows:

[0235]

[0236] S3.5, Generate patent feature representation

[0237] Generating the original patent map using the patent feature matrix and adjacency matrix The original patent image is then input into the trained encoder f(·), which outputs the final representation of each patent. and the final representation of all patents :

[0238]

[0239]

[0240] S4. An attention mechanism is used to weight and aggregate the transferor's representation, transferee's representation, and patent representation. By reusing the attention mechanism, vector weights can be assigned to the transferor, transferee, and patent to simulate real-world market dynamics. If the transferor (transferee) has a significant weight among the three, it indicates that the patent transaction price depends more on the transferor's (transferee's) information, and also suggests that the transferor (transferee) has strong bargaining power in this transaction. If the patent has a significant weight among the three, it indicates that the patent transaction depends more on the patent's own information. Specifically:

[0241] S4.1, Unifying the representation of the transferor, transferee, and patent.

[0242] The final representation of the universities obtained from S2 The company ultimately stated Projecting onto the patent feature space yields a high-level representation with the same dimensions as the patent features. And the company said The calculation formula is as follows:

[0243]

[0244] in, For university projection matrix, For enterprise projection matrix.

[0245] S4.2, Features of splicing the three

[0246] The university's representation after splicing together a unified dimension The company stated With patent representation Obtain the splicing feature z i,j,k Used to capture the joint relationship between the three:

[0247]

[0248] S4.3, Calculate the attention scores for each item.

[0249] Through trainable parameters W t′ W r′ W p′ Using the LeakyReLU activation function, the concatenated features z i,j,k Mapped to attention score s i,j,k The calculation is as follows:

[0250]

[0251] in, All are trainable parameters, and W t′ W r′ W p′ It will be updated during the backpropagation of the MLP, that is, it is jointly optimized by the backpropagation algorithm and the classification loss; Attention scores were given to universities, enterprises, and patents, respectively.

[0252] S4.4, Normalized Weights

[0253] The calculated attention scores of college students Enterprise Attention Score Patent attention score After normalization, their respective weights are obtained:

[0254]

[0255] S4.5, weighted aggregation yields the final representation.

[0256] Based on the normalized weights, aggregate the university representations after unifying the dimensions. The company stated With patent representation The final representation is obtained :

[0257]

[0258] In the process of feature splicing and aggregation, including All intermediate variables, including the classification loss, maintain computational graph continuity to ensure that gradients can propagate back from the classification loss to the attention parameters.

[0259] S5. Input the final representation into the MLP to predict the patent value label; MLP is a feedforward deep learning model based on artificial neural networks, which models complex data patterns through multi-layer nonlinear transformations of input, hidden, and output layers. In the patent valuation scenario, after inputting the final representation into the MLP, the model can automatically learn the nonlinear mapping relationship between the final representation and the value, and finally output a label representing the patent value; as follows:

[0260] S5.1, Classification Label Definition

[0261] The patent value is discretized into 7 categories, labeled as follows:

[0262] S5.2, MLP Prediction Layer Design

[0263] The final representation obtained from input S4 The MLP outputs the class probability distribution, specifically as follows:

[0264] 1) Hidden layer 1:

[0265]

[0266] in, The weights of the first layer of the MLP For the first layer bias of the MLP, d h The hidden layer dimension is set to 128.

[0267] 2) Hidden layer 2:

[0268]

[0269] in, For the second layer weights of the MLP, This is the output of the first layer of the MLP. For the second layer bias of the MLP, d h For the hidden layer dimension.

[0270] 3) Output layer:

[0271]

[0272] in, For class probability vectors, For the third layer weights of the MLP, This is the output of the second layer of the MLP. This is the bias of the third layer of the MLP, where C is the number of categories.

[0273] The cross-entropy loss method is used here:

[0274]

[0275] In the formula, ||(·) is an indicator function. The labels are the real labels; the gradient is backpropagated from the output layer of the MLP to... Then, the attention weights are calculated and passed to W. t′ W r′ W p′ This allows for the updating of attention weights, and further, through backpropagation, updates to the representations of the patent, the assignor, and the assignee, as well as the attention mechanism parameters involved, thereby enabling accurate prediction of patent value.

[0276] The above description represents a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A patent value assessment method based on dynamic game attention and graph contrastive learning, characterized in that, include: S1. Collect multi-source patent transaction data, including patent attribute features and citation relationships, transferor attribute features and transferee attribute features, and preprocess the data to obtain a dataset; S2. An autoencoder is used to convert the multidimensional features of the transferor and the transferee into vector representations respectively. Then, based on the attention mechanism, the vector representations of the multidimensional features of the transferor and the transferee are weighted and aggregated to generate the transferor representation and the transferee representation. S3. Based on the attribute features and citation relationship data of the patent, two heterogeneous views are generated using the GRACE algorithm. The two heterogeneous views are combined to train the encoder, and the trained encoder is used to generate the patent representation. S4. Using an attention mechanism, the transferor representation, the transferee representation, and the patent representation are weighted and aggregated to obtain the final representation; S5. Input the final representation into a multilayer perceptron to predict the patent value label.

2. The patent value assessment method based on dynamic game attention and graph contrastive learning according to claim 1, characterized in that, The multi-source patent transaction data includes patent transaction record data, patent attribute feature data, patent citation relationship data, attribute feature data of patent transferors, and attribute feature data of patent transferees; wherein, each patent transaction record data contains one transferor, one transferee, and the patent; the dataset includes a patent transaction record dataset D. T Patent attribute feature dataset D px Patent citation relationship dataset D pa Patent assignor attribute feature dataset D t And the patent assignee attribute feature dataset D r The patent attribute feature dataset D px The total number of patents involved in transactions, including the patents cited and referenced by these patents, is q, of which the number of patents involved in transactions is q′.

3. The patent value assessment method based on dynamic game attention and graph contrast learning according to claim 2, characterized in that, The method of using an autoencoder to convert the multidimensional features of the transferor and transferee into vector representations is as follows: Based on the patent assignor attribute feature dataset D t Extract the multi-dimensional attributes of the transferor, based on the patent transferee attribute feature dataset D r Extract the multi-dimensional attributes of the transferee and construct the transferor feature matrix X respectively. t and the transferee's characteristic matrix X r ;in, d t Let m be the total dimension of the characteristics of the transferors; d r is the total dimension of the transferee's characteristics, and n is the number of transferees; Use separate encoders to transfer the transferor's d t Dimensional features and the transferee's d r Map the 1D features to a k-dimensional space, and extract the hidden layer representation H of the transferor's features respectively. t Hidden layer representation H of transferee characteristics r : in, k is the dimension of the hidden layer; The encoder weight for the transferor; For the encoder bias of the transferor; k is the dimension of the hidden layer; The encoder weights for the transferee; For the encoder bias of the transferee; To preserve the differences in characteristics between the transferor and the transferee; The hidden layer representation H is obtained from the decoder. t and hidden layer representation H r The original features of the transferor and the transferee are recovered respectively to obtain the decoded features. and in, For the transferor's decoder weight; For the transferor's decoder bias; For the decoder weights of the transferee; The receiver's decoder bias is σ; σ is the Sigmoid function. The autoencoder loss function is designed as follows: in, To reconstruct the loss, the autoencoder is constrained to achieve high-fidelity reconstruction of the original features; λ is the regularization term to prevent overfitting between the encoder and decoder; λ is the regularization coefficient. is the hidden layer alignment term, which calculates the trace of cross-modal similarity to enhance the matching of dual-channel features; μ is the adjustment coefficient.

4. The patent value assessment method based on dynamic game attention and graph contrastive learning according to claim 3, characterized in that, The specific process of weighted aggregation of the vector representations of the multi-dimensional features of the transferor and transferee based on the attention mechanism to generate the transferor representation and the transferee representation is as follows: Calculate d for each transferor separately. t The weights of the dimensional features and each transferee d r Weights of dimensional features: in, This represents the weight of the l-th dimension feature of the i-th transferor; This represents the eigenvalue of the l-th dimension in the decoded feature of the i-th transferor; Let l represent the eigenvalue of the l-th dimension in the hidden layer representation of the i-th transferor; l∈[1,d] t ]; This represents the weight of the g-dimensional feature of the j-th transferee; This represents the feature value of the g-th dimension in the decoded feature of the j-th transferee; Let g represent the eigenvalue of the g-th dimension in the hidden layer representation of the j-th transferee; g∈[1,d] r ]; Final representation of each transferor and the final representation of each transferee They are as follows:

5. The patent value assessment method based on dynamic game attention and graph contrast learning according to claim 4, characterized in that, The patent-based attribute features and citation relationship data are used to generate two heterogeneous views using the GRACE algorithm, as follows: Based on the patent attribute feature dataset D px Extract multi-dimensional attributes of patents and construct a patent feature matrix X. p , q represents the total number of patents, d p This refers to the total dimension of patent features; Based on the patent citation relationship dataset D pa Construct a patent adjacency matrix A, if patent a u Patent a was cited v If u∈[1,q], then A uv =1, otherwise A uv =0: A∈{0,1} q×q Add self-loops to enhance the propagation of information within the node itself: For each edge Perform in-degree normalization: Among them, D in,uu The in-degree of a patent, i.e., the number of times it has been cited; I q An identity matrix; a normalized directed adjacency matrix containing self-loops. By structurally disrupting the edges in the patent adjacency matrix and deleting edges from adjacency matrix A with probability p1, two topologically heterogeneous views are generated: View 1: View 2: Where ° represents the Hadamard product. It is an asymmetric mask matrix; The in-degree matrix is ​​independently calculated and normalized for each generated view: View 1: View 2: Feature masking is performed on the node features in the patent feature matrix, and the patent feature matrix X is randomly selected with probability p2. p By setting some dimensions of the node features to zero, two heterogeneous views are generated: View 1: View 2: in, For feature mask vectors; The generated and Combine them to generate two heterogeneous views. and 6. The patent value assessment method based on dynamic game attention and graph contrastive learning according to claim 5, characterized in that, The trained encoder is specifically as follows: Using GCN as the encoder f(·), the first layer utilizes the ReLU function for nonlinear activation: in, First-level authority for GCN; The second layer performs linear propagation: in, The second layer weights of GCN are used; the output is the patent node embedding. The two views generated from the input and To the encoder with shared weights, generate node embeddings: Where U and V are view embedding matrices, d e <<d p ; Positive sample pairs are the embeddings of the same patent node in two different views (s u ,w u Negative samples are embeddings of other patent nodes in different views or the same view, and each positive sample pair (s) is evaluated based on cosine similarity. u ,w u Calculate the comparative loss Among them, the discriminator θ(s) u ,w u ) is s u with w u The cosine similarity is calculated using the following formula: Furthermore, a nonlinear projection head g(s) = MLP(s) is introduced to enhance the discriminativeness of contrastive learning, where τ is a temperature parameter; The overall objective function is the average of all positive sample pairs, calculated as follows: The original patent map is generated using the patent feature matrix and the adjacency matrix. And Input the trained encoder and output the final representation of each patent.

7. The patent value assessment method based on dynamic game attention and graph contrastive learning according to claim 6, characterized in that, The transferor representation, the transferee representation, and the patent representation are weighted and aggregated using an attention mechanism to obtain the final representation, as follows: The final statement of the transferor The transferee ultimately stated Projecting onto the patent feature space yields the transferor representation, which has the same dimensions as the patent features. The transferee expressed The calculation formula is as follows: in, For the transferor's projection matrix, The projection matrix of the transferee; The transferor's representation after unifying dimensions The transferee stated With patent representation By splicing the pieces together, we can obtain the splicing features. splicing feature z i,j,k Mapped to attention score s i,j,k The calculation is as follows: in, All are trainable parameters, and W t′ W r′ W p′ Update during MLP backpropagation; Attention scores are given to the transferor, transferee, and patentee, respectively. right and Normalization is performed to obtain their respective weights. The transferor of the aggregated transferor stated The transferee stated With patent representation The final representation is obtained in, 8. The patent value assessment method based on dynamic game attention and graph contrastive learning according to claim 7, characterized in that, The final representation is then input into a multilayer perceptron to predict the patent value label, as follows: The patent value is discretized into C categories, labeled as follows: Input final representation Output category probability distribution through MLP The loss function used is cross-entropy loss: In the formula, ||(·) is the indicator function. This is a real label.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor; characterized in that, When the processor executes the computer program, it causes the electronic device to perform the patent valuation method based on dynamic game attention and graph contrast learning as described in any one of claims 1-8.

10. A storage medium comprising a computer program that, when run on an electronic device, causes the electronic device to perform the patent valuation method based on dynamic game attention and graph contrast learning as described in any one of claims 1-8.

Images