Patent transaction link prediction model fusing multi-dimensional proximity and network motif

Abstract

The invention discloses a patent transaction link prediction model fusing multi-dimensional proximity and a network motif, which comprises the following steps: constructing a multi-layer patent transaction network containing technical, geographical and social relationships, respectively extracting technical similarity, geographical proximity and social relationship proximity characteristics, and analyzing and capturing a local structure mode in combination with the network motif. And designing a multi-channel graph attention mechanism to realize adaptive fusion of multi-dimensional features and motif features, introducing a dynamic negative sampling strategy to optimize a model training process, and inputting the fused features into a trained model to carry out patent transaction link prediction. Through cross-layer correlation modeling of technology, geography and social dimensions and high-order structure representation fusion of a network motif, the problems of single-dimension analysis and insufficient structural information utilization are effectively solved, and the accuracy and generalization ability of patent transaction link prediction are remarkably improved; the method is suitable for potential transaction link mining of a patent transaction platform, an intellectual property service mechanism and other scenes, and provides decision support for efficient patent conversion.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to the field of patent transaction link prediction technology, specifically a patent transaction link prediction technology that integrates multidimensional proximity and network models. Background Technology

[0002] In the field of patent commercialization, patent transaction link prediction models play a crucial role in uncovering potential transaction relationships and promoting efficient patent commercialization. Existing patent commercialization prediction models are mainly divided into methods based on traditional machine learning and methods based on graph neural networks. Traditional machine learning methods typically utilize only basic patent attribute information, such as the technical field and applicant type, and construct prediction models using algorithms such as logistic regression and support vector machines. Graph neural network-based methods, on the other hand, treat the patent transaction network as a graph structure, using node embedding techniques to capture the relationships between nodes. For example, some models analyze the probability of a transaction only from the perspective of patent technical text similarity, or simply make predictions based on the geographical proximity of the applicants.

[0003] Existing patent transfer prediction models have significant shortcomings. On the one hand, existing models neglect the multidimensional interconnectedness of technology entities, analyzing node relationships from only a single dimension. This fails to comprehensively reflect the complex connections between technology entities in patent transactions across technological, geographical, and social aspects, leading to the omission of numerous potential transaction links. On the other hand, existing models fail to effectively utilize high-order network structure information, lacking sufficient mining of local structures such as network modules that can reflect complex node interaction patterns. This makes it difficult to accurately capture the deep-seated patterns in patent transaction networks, resulting in prediction accuracy that falls short of practical application requirements. Summary of the Invention

[0004] To address the issue that existing patent transfer prediction models neglect the multidimensional correlation characteristics of technology entities and the high-order structure of networks, this invention provides a link prediction model MTLP-MN that integrates technology / geographic / social proximity characteristics with network motifs. It aims to comprehensively capture the multidimensional correlations of technology entities in patent transactions, uncover local network structure patterns, and improve the accuracy and reliability of patent transaction link prediction.

[0005] A first aspect of the present invention provides a method for predicting patent transaction links that integrates multidimensional proximity and network models, comprising the steps of:

[0006] Construct a multi-layered patent transaction network, which includes a technology relationship network, a geographical relationship network, and a social relationship network;

[0007] Technical similarity features, geographical proximity features, and social proximity features are extracted separately. Technical similarity features are calculated based on the patent field and patent citation relationships. Geographical proximity features are calculated based on the geographical location information of the patent applicant or the right holder. Social proximity features are obtained based on the cooperation history and affiliation between the applicants.

[0008] Network motif analysis was performed on a multi-layered patent transaction network to capture local structural patterns.

[0009] A multi-channel graph attention mechanism is designed to fuse extracted technology similarity features, geographical proximity features, social relationship proximity features, and network motif features;

[0010] A dynamic negative sampling strategy is introduced to optimize model training, and the fused features are input into the trained model for patent transaction link prediction.

[0011] The extracted technological similarity features, geographical proximity features, social relational proximity features, and network motif features are fused together, specifically including:

[0012] The method for extracting motif features from network motif analysis includes the following steps: using graph mining algorithms, performing network motif analysis on the constructed multi-layer patent transaction network through subgraph enumeration, pattern matching, and frequency statistics, extracting network motifs of different orders and types, such as triangles and quadrilaterals, and statistically analyzing the frequency of occurrence and node composition of each motif in the network, capturing local structural patterns of the network, and forming an M-matrix.

[0013] In some implementations, network motif analysis can also use a fast algorithm based on subgraph matching to replace the traditional enumeration method. This algorithm is more efficient when dealing with large-scale networks, but may have the risk of missing some complex motif structures. Alternatively, it can focus only on specific types of high-order motifs (such as pentagonal motifs). For patent transaction networks in certain specific fields, this may better reflect their unique structural characteristics, but the general applicability will be reduced.

[0014] In a specific embodiment of the present invention, the fusion and prediction layer is used to perform cross-modal fusion of multidimensional proximity features, perform weighted summation according to feature weights, input the M matrix, and realize the concatenation with Motif features to achieve feature fusion. During the model training process, a dynamic negative sampling optimization strategy is adopted to input the fused features into the trained model, calculate the probability that there is a patent transaction link between the output node pairs, and realize patent transaction link prediction.

[0015] As a further technical solution of the present invention, the construction of the multi-layered patent transaction network specifically involves: using patents and patent transaction participants as nodes, constructing a technology relationship network based on patent technology association, constructing a geographical relationship network based on the geographical location of the participants, and constructing a social relationship network based on the social relationship of the participants.

[0016] Specifically, the construction of the multi-layered patent transaction network involves using patents and patent transaction participants (such as enterprises, universities, and research institutions) as nodes to build a multi-layered patent transaction network. The technology relationship network is based on the technological connections between patents, such as those in the same technical field or those involving technology citations, to construct edges. The geographical relationship network connects entities that are geographically close by edges based on the geographical location information of patent applicants or rights holders. The social relationship network constructs connecting edges based on the history of cooperation and affiliation between applicants, thus forming a multi-layered network structure encompassing technological, geographical, and social dimensions.

[0017] As a further technical solution of the present invention, the method for calculating the technical similarity features is as follows: the patent technical field is encoded, a technical similarity matrix is ​​constructed by combining the patent citation relationship, and the technical similarity features are obtained through matrix operations.

[0018] Specifically, the patent technology field is coded, for example, using the International Patent Classification (IPC) for One-Hot encoding. A technology similarity matrix is ​​constructed by combining patent citation relationships, and technology similarity features are obtained through matrix operations. For two patent nodes, technology similarity features are obtained through matrix operations, such as calculating the similarity score of corresponding elements in the matrix.

[0019] As a further technical solution of the present invention, the method for calculating the geographic proximity feature is as follows: obtain the latitude and longitude coordinates of the patent applicant or the right holder, calculate the distance between the coordinates, and generate the geographic proximity feature by combining it with a distance decay function. Specifically, the distance between nodes is calculated using a geographic distance calculation formula (such as the Haversine formula), and then combined with a distance decay function, such as setting a higher weight for closer distances, to generate the geographic proximity feature.

[0020] As a further technical solution of the present invention, the method for obtaining the social relationship proximity feature is as follows: analyze information such as the number of cooperative projects, the time span of cooperation, and the affiliation between applicants, and obtain the social relationship proximity feature through quantitative scoring.

[0021] The method for obtaining the social relationship proximity characteristics is as follows: analyze the cooperation history (including the number of cooperation projects and the time span of cooperation), affiliation and other information between applicants, and assign corresponding scores to different social relationships through a quantitative scoring system, thereby obtaining the social relationship proximity characteristics.

[0022] As a further technical solution of the present invention, the network motif analysis includes: using graph mining algorithms to perform network motif analysis on the constructed multi-layer patent transaction network, identifying and extracting network motifs of different orders and types, such as triangles and quadrilaterals. Information such as the frequency of occurrence of each motif in the network and node composition is statistically analyzed to capture local structural patterns in the network, which reflect the complex interactive relationships between nodes.

[0023] As a further technical solution of the present invention, the multi-channel graph attention mechanism includes multiple attention channels, each channel performs attention calculation on different types of features, and then the results of each channel are fused together.

[0024] Specifically, a multi-channel graph attention mechanism is designed, comprising multiple attention channels. Each channel performs attention calculations on technology similarity features, geographical proximity features, social relationship proximity features, and network motif features. Through this attention mechanism, the model can automatically learn the importance of different features in patent transaction link prediction, highlighting key features and suppressing secondary features. Finally, the calculation results from each channel are fused to obtain a comprehensive feature representation.

[0025] As a further technical solution of the present invention, the dynamic negative sampling strategy dynamically adjusts the sampling ratio and sampling method of negative samples according to the sample distribution during model training. Specifically, a dynamic negative sampling strategy is introduced during model training. This strategy dynamically adjusts the sampling ratio and sampling method of negative samples according to the distribution of positive and negative samples during training. For example, in the early stage of training, the sampling ratio of negative samples is increased to expand the learning range of the model; as training progresses, the selection strategy of negative samples is adjusted according to the convergence of the model, prioritizing the selection of difficult negative samples for training, thereby improving the generalization ability and training efficiency of the model.

[0026] As a further technical solution of the present invention, the fused features are input into the trained model, and the probability of a patent transaction link between node pairs is output through the model's calculation, thereby realizing patent transaction link prediction.

[0027] Secondly, the present invention provides a patent transaction link prediction model that integrates multidimensional proximity and network models, comprising:

[0028] The input layer is used to obtain the data source;

[0029] The processing layer is used to construct a multi-layered patent transaction network and extract technical similarity features, geographical proximity features, and social relationship proximity features; it is also used to perform network phantom analysis on the multi-layered patent transaction network.

[0030] The fusion and prediction layer is used to design a multi-channel graph attention mechanism that fuses technology similarity features, geographical proximity features, social relationship proximity features, and network motif features. It is also used to introduce a dynamic negative sampling strategy to optimize model training. The fused features are input into the trained model to predict patent transaction links and calculate the probability that there is a patent transaction link between the output node pairs.

[0031] Thirdly, the present invention provides a patent transaction link prediction device that integrates multidimensional proximity and network models, including a memory and a processor. The memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to perform the steps of the patent transaction link prediction method that integrates multidimensional proximity and network models as claimed above.

[0032] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to perform the steps of the patent transaction link prediction method that integrates multidimensional proximity and network models.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] Comprehensive Capture of Multidimensional Relationships: This invention comprehensively considers the multidimensional relationships between technology entities in patent transactions by extracting features of technological similarity, geographical proximity, and social proximity. Technological similarity features can identify potential trading partners with similar technological backgrounds, promoting collaborative technological innovation; geographical proximity features help uncover the possibility of transactions between entities located close to each other, reducing transaction costs; and social proximity features leverage existing cooperative and affiliation relationships to strengthen the trust foundation for transactions. These complementary features allow the model to more comprehensively describe the potential relationships between technology entities, uncovering more potential transaction links compared to existing single-dimensional analysis models.

[0035] Effectively utilizing higher-order network structures: Network motif analysis can uncover local structural patterns in patent transaction networks, revealing complex interaction patterns between nodes. For example, a triangular motif might represent a stable network of technological collaborations where potential patent transaction opportunities exist among three nodes. By extracting and utilizing these network motif features, the model can gain a deeper understanding of the structural characteristics of patent transaction networks, capturing deep connections that existing models often overlook, thereby improving the accuracy of predicting patent transaction chains.

[0036] Optimized feature fusion and training: The multi-channel graph attention mechanism can adaptively learn the importance of different features, achieving effective feature fusion and avoiding information redundancy and the burying of key information caused by simple feature concatenation. The dynamic negative sampling strategy dynamically adjusts negative samples according to the training process, enabling the model to learn the distribution differences between positive and negative samples more efficiently, improving the model's training efficiency and generalization ability.

[0037] Compared with existing models, this invention has significant improvements in prediction accuracy and recall, and can better meet the demand for accurate prediction in the patent transaction market. It is of great significance for promoting patent transformation and optimizing resource allocation. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0039] Figure 1 This is a schematic diagram of the patent transaction link prediction model framework that integrates multidimensional proximity and network models according to the present invention.

[0040] Figure 2 This is a schematic diagram of the multidimensional proximity processing of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0042] In an embodiment of the present invention, a method for predicting patent transaction links that integrates multidimensional proximity and network models is provided, comprising the following steps:

[0043] Construct a multi-layered patent transaction network, which includes a technology relationship network, a geographical relationship network, and a social relationship network;

[0044] Technical similarity features, geographical proximity features, and social proximity features are extracted separately. Technical similarity features are calculated based on the patent field and patent citation relationships. Geographical proximity features are calculated based on the geographical location information of the patent applicant or the right holder. Social proximity features are obtained based on the cooperation history and affiliation between the applicants.

[0045] Network motif analysis was performed on a multi-layered patent transaction network to capture local structural patterns.

[0046] A multi-channel graph attention mechanism is designed to fuse extracted technology similarity features, geographical proximity features, social relationship proximity features, and network motif features;

[0047] A dynamic negative sampling strategy is introduced to optimize model training, and the fused features are input into the trained model for patent transaction link prediction.

[0048] In some embodiments, a multi-layered patent transaction network is constructed using patents and patent transaction participants (such as enterprises, universities, and research institutions) as nodes. The technology relationship network is based on the technological connections between patents, such as those in the same technical field or those involving technology citations, forming edges. The geographical relationship network connects entities that are geographically close by edges based on the geographical location information of patent applicants or rights holders. The social relationship network forms connecting edges based on the history of cooperation and affiliation between applicants, thus forming a multi-layered network structure encompassing technological, geographical, and social dimensions.

[0049] In some embodiments, the patent technology field is encoded, for example, using the International Patent Classification (IPC) for One-Hot encoding, and a technology similarity matrix is ​​constructed by combining patent citation relationships. For two patent nodes, technology similarity features are obtained through matrix operations, such as calculating the similarity score of corresponding elements in the matrix. Technology similarity feature calculation can also employ topic models (such as LDA) to extract topics from the patent text, and measure technology similarity through topic similarity. Compared to methods based on IPC and citation relationships, this approach can delve deeper into the patent technology's connotation, but it has higher computational complexity.

[0050] In some embodiments, the latitude and longitude coordinates of the patent applicant or patentee are obtained, and the distance between nodes is calculated using a geographic distance calculation formula (such as the Haversine formula). This is then combined with a distance decay function, where closer distances are assigned higher weights, to generate geographic proximity features. Geographic proximity feature calculation can also refer to administrative division information, assigning higher proximity weights to entities within the same administrative region. This method is simple and intuitive, but its accuracy may be lower than that based on coordinate distance.

[0051] In some embodiments, information such as the number of collaborative projects, the time span of collaboration, and affiliation among applicants is analyzed. Through a quantitative scoring system, corresponding scores are assigned to different social relationships, thereby obtaining the proximity characteristics of social relationships.

[0052] This invention employs graph mining algorithms to perform network motif analysis on a constructed multi-layered patent transaction network, identifying and extracting network motifs of different orders and types, such as triangular and quadrilateral motifs. It statistically analyzes the frequency of occurrence of each motif in the network, node composition, and other information to capture local structural patterns that reflect complex interactions between nodes. A multi-channel graph attention mechanism with multiple attention channels is designed, with each channel calculating attention to technical similarity features, geographical proximity features, social relationship proximity features, and network motif features. Through this attention mechanism, the model can automatically learn the importance of different features in patent transaction link prediction, highlighting key features and suppressing secondary features. Finally, the calculation results from each channel are fused to obtain a comprehensive feature representation. During model training, a dynamic negative sampling strategy is introduced. This strategy dynamically adjusts the sampling ratio and sampling method of negative samples based on the distribution of positive and negative samples during training. For example, in the early stages of training, the sampling ratio of negative samples is increased to expand the model's learning range; as training progresses, the negative sample selection strategy is adjusted based on the model's convergence, prioritizing the selection of difficult negative samples for training, thereby improving the model's generalization ability and training efficiency. The fused features are input into the trained model, and the model calculates and outputs the probability that there is a patent transaction link between node pairs, thereby realizing the prediction of patent transaction links.

[0053] In embodiments of the present invention, a patent transaction link prediction model integrating multidimensional proximity and network models is also provided. Figure 1 This is a schematic diagram of the patent transaction link prediction model framework that integrates multidimensional proximity and network models according to the present invention. The prediction model of the present invention includes:

[0054] The input layer is used to obtain data sources, which include patent IPC classification (International Patent Classification), TF-IDF text vectors, institutional latitude and longitude coordinates, cooperation records and transaction history, Motif pattern library, etc.

[0055] The processing layer is used to construct a multi-layered patent transaction network, perform multi-dimensional proximity processing, calculate and extract multi-dimensional proximity features including technological similarity features, geographical proximity features, and social relationship proximity features; perform network motif analysis on the multi-layered patent transaction network, extract motif features (network motif features), and output the M matrix;

[0056] The fusion and prediction layer is used to design a multi-channel graph attention mechanism that fuses technology similarity features, geographical proximity features, social relationship proximity features, and network motif features. It is also used to introduce a dynamic negative sampling strategy to optimize model training. The fused features are input into the trained model to predict patent transaction links and output the prediction results.

[0057] In an embodiment of the present invention, the method for extracting motif features through network motif analysis includes the following steps: using graph mining algorithms, performing network motif analysis on the constructed multi-layer patent transaction network through subgraph enumeration, pattern matching, and frequency statistics, extracting network motifs of different orders and types, such as triangles and quadrilaterals, and statistically analyzing the frequency of occurrence and node composition of each motif in the network, capturing local structural patterns of the network, and forming an M matrix.

[0058] In some implementations, network motif analysis can also use a fast algorithm based on subgraph matching to replace the traditional enumeration method. This algorithm is more efficient when dealing with large-scale networks, but may have the risk of missing some complex motif structures. Alternatively, it can focus only on specific types of high-order motifs (such as pentagonal motifs). For patent transaction networks in certain specific fields, this may better reflect their unique structural characteristics, but the general applicability will be reduced.

[0059] In a specific embodiment of the present invention, the fusion and prediction layer is used to perform cross-modal fusion of multidimensional proximity features, perform weighted summation according to feature weights, input the M matrix, and realize the concatenation with Motif features to achieve feature fusion. During the model training process, a dynamic negative sampling optimization strategy is adopted to input the fused features into the trained model, calculate the probability that there is a patent transaction link between the output node pairs, and realize patent transaction link prediction.

[0060] In some implementations, the model architecture can also use a graph recurrent neural network (GRN) instead of a graph convolutional neural network or a graph attention network. GRNs can better handle patent transaction data with time-series characteristics and have a stronger ability to model the dynamic changes in the patent transaction chain, but they may be slightly inferior in capturing the static structure of the network. Alternatively, graph Transformer networks can be used, which utilize the self-attention mechanism of the Transformer architecture and have advantages in feature fusion and long-distance dependency modeling, but the training cost is high.

[0061] In an embodiment of the present invention, a patent transaction link prediction method integrating multidimensional proximity and network models is specifically as follows: Figure 1-2 As shown, it includes:

[0062] First, construct the GAT layer and perform technology proximity calculation to obtain the technology similarity matrix, which is used to represent technology similarity features.

[0063] Among them, technology proximity calculation includes:

[0064] 1. Patent Text Vectorization: For all patent texts of organization u, extract IPC subclasses (such as H01L) and expand their semantics to construct bag-of-words vectors.

[0065]

[0066] in, It is the weight value (bag-of-words vector component) of institution u in the technical field c. The meaning is the number of patents held by organization u in the technology field c. It is the total number of patents in the entire dataset that involve technology c; It represents the total number of all patents in the dataset.

[0067] 2. Similarity Calculation: Calculate the inter-institutional technology similarity matrix.

[0068]

[0069] in, It is the technical similarity score between institutions u and v; , These are the patent technology vectors of institutions u and v (composed of bag-of-words vectors). constitute); , It is the set of technical fields of institutions u and v (such as the set of IPC classification numbers). , It represents the set of real numbers.

[0070] 2. Construct a GCN layer with distance constraints to perform geographic proximity modeling and obtain map background to represent geographic proximity features.

[0071] Geographic proximity modeling includes:

[0072] 1. Spatial distance calculation: Converting the longitude dimension of the institution's address to three-dimensional Cartesian coordinates.

[0073]

[0074] in, It is the dimension of the registered address of organization u; It is the longitude of the registered address of organization u; , , ) represents the three-dimensional Cartesian coordinates of mechanism u (used for accurate calculation of spherical distance).

[0075] The spatial distance between computing units u and v (Calculated using three-dimensional coordinates):

[0076]

[0077] R is the Earth's radius (a scale factor for converting the central angle to the distance of a great circle on a sphere), commonly used value: R = 6371 km. Indicates enterprise Latitude and longitude (in radians), similarly. .

[0078] 2. Distance decay function: Introducing a double exponential kernel function to characterize the geographical proximity effect (rapid decay from nearest neighbors + smooth decay from medium to long distances)

[0079]

[0080] in, It is the geographical proximity score (decayed similarity) of institutions u and v. It is the weighting coefficient (balancing the double exponential term, typical value) ∈[0,1]); , It is the attenuation radius parameter (controls the attenuation rate over distance, and the unit is consistent with the distance).

[0081] If the three types of proximity need to be unified into "primary similarity × reduction / enhancement factor", then geographical proximity can be rewritten as:

[0082]

[0083] Maintain numerical monotonicity and hyperparameter interpretability, and be equivalent or approximately equivalent to the current additive expression (fine-tuned according to the normalization method).

[0084] Third, construct the GraphSAGE layer to model social proximity and form a social network graph to represent social proximity features.

[0085] Social proximity modeling includes:

[0086] 1. Construction of historical interaction network:

[0087]

[0088] in, : Time window length (e.g., τ=3 years, only the most recent years are counted) Events within the year). Indicates the intensity of unnormalized historical interactions; Indicates to The result after min–max normalization is in the range [0,1]. The time interval is expressed as (current moment - moment when the kth interaction event occurs), and the unit can be "year" or "day". Indicates the time decay coefficient ( The larger the value, the faster the weight of long-term events decays. This indicates the number of historical interaction events between u and v (such as the number of joint patent applications).

[0089] 2. Enhanced Structural Similarity: Fusing local (Common Neighbors) and global (SimRank) similarity:

[0090]

[0091]

[0092]

[0093] in, Geometric mean normalized version of common neighbor (CN) similarity, with values ​​ranging from... ; , These are respectively represented in the social relationship diagram. The neighbor set of v (and) v (nodes with social connections such as collaboration / co-authorship / affiliation with the same institution); : degree (number of neighbors) Represents the cardinality of a set; This represents the self-similarity boundary condition of SimRank, where the similarity between two objects is 1; C is the SimRank decay parameter (usually C∈(0,1), which controls the decay of global similarity propagation; the smaller C is, the weaker the influence of long paths); a and b are... and Any neighboring node in; The similarity of social relationships between institutions a and b is calculated recursively. It is the similarity score of the social relationship structure of institutions u and v (a weighted fusion of CN and SimRank); It is the fusion weight of local and global similarity ( ), controlling the ratio of CN to SimRank.

[0094] 3. Social Proximity Integration:

[0095]

[0096] in, It is the score of the social relationship similarity between institutions u and v; The fusion weights representing the strength of historical interactions and structural similarity ( The larger the value, the stronger the influence of historical interactions.

[0097] If the three types of proximity need to be unified into "primary similarity × reduction / enhancement factor", then social proximity can be rewritten as:

[0098]

[0099] Maintain numerical monotonicity and hyperparameter interpretability, and be equivalent or approximately equivalent to the current additive expression (fine-tuned according to the normalization method).

[0100] IV. Multidimensional Proximity Feature Fusion (Unified Encoding and Channel Attention Fusion)

[0101] Scalars of three proximity classes are mapped to the same dimension using independent linear / MLP methods. :

[0102]

[0103] in, It is a linear layer or a small MLP used to map proximity features from different sources and with different dimensions to a unified dimension. The aligned channel representation is obtained. .

[0104] To adaptively aggregate the importance of different channels, channel attention (normalized with Softmax) is introduced:

[0105]

[0106] Meaning: Project the activated r-dimensional representation onto a scalar energy. Meaning: Learnable affine transformation of attention (can be shared across three channels, or made independent for each channel) ); Indicates the current attention weight Specific channel index (e.g.) (This indicates a technical channel). It's a dummy variable (placeholder) in the summation operator, meaning "sum across all channels," and doesn't point to a specific channel; it only refers to... It is valid within its scope. For example, calculating... Time: The molecule uses The denominator is ,in, , Take in sequence .

[0107] The nearest neighbor feature fusion is obtained by weighted summation based on the feature weights, as shown in the following formula:

[0108]

[0109] V. Feature Fusion of Multidimensional Proximity and Network Motif Features

[0110] set up To Extracted and compressed phantom features (such as three / four node phantom counts or normalized intensity, after Log / standardization and linear dimensionality reduction).

[0111] The multidimensional proximity and network motif features are concatenated and projected:

[0112]

[0113] Represent a A real vector space of dimension 1, that is The dimension and numerical domain; Proj means: the concatenated... Dimensional vector mapping / projection to Dimensional (usually using a single linear or small MLP, with learnable parameters, trained end-to-end).

[0114] VI. Predictive Decoder

[0115] A lightweight decoder is used to output transaction probabilities:

[0116]

[0117] in, The predicted probability represents the predicted probability that there is a transaction link (a transaction occurs) between institutions u and v, and its range is (0,1). The learned parameters are typically a vector that is a linear mapping connected to the predictor and decoder. The weight matrix represents the proximity representation of the input (e.g., ...). ); This represents the proximity representation after feature fusion. This represents the bias term, used to adjust the model's output; This represents an additional bias term, typically connected to the final output layer of the network; Activation functions provide non-linearity. The Sigmoid activation function maps the output to the range (0,1).

[0118] VII. Training Objectives and Dynamic Negative Sampling

[0119] Given the observed positive sample set Dynamic negative sample sets generated by degree or time window Using binary cross-entropy:

[0120]

[0121] in, The loss function is represented, including cross-entropy loss and L2 regularization; This represents the positive sample set, which includes observed trading pairs (e.g., known cooperative trades). This represents the negative sample set, typically sampled from institutions that have never had a transaction. The L2 regularization coefficient is used to prevent overfitting by adjusting the parameters. Impose punishment; Includes all learnable parameters of the encoder-decoder, attention, and decoder, such as wait; This represents the square of the L2 norm, which is the sum of the squares of all model parameters, and is used to perform regularization.

[0122] In summary, the prediction algorithm process includes:

[0123] (1) Feature preparation: calculation ;Statistical analysis and compression of phantom features ;

[0124] (2) Uniform coding: ;

[0125] (3) Channel attention fusion: ;

[0126] (4) Attach and project the model: ;

[0127] (5) Decoding prediction: Output ;

[0128] (6) Training: Dynamic negative sampling + BCE loss optimization .

[0129] It is important to note that:

[0130] (1) The three types of proximity are edge-level scalars, which have low computation and storage costs and are convenient for large graphs.

[0131] (2) Motif statistics can be performed by sampling / approximate counting and parallelization; then PCA / linear layer dimensionality reduction can be used.

[0132] (3) The channel attention parameters are small and the gradient is stable, making it suitable for joint training with the GNN module or independent backend calls.

[0133] (4) Can be flexibly set according to task Hyperparameters; if industry / regional stratification is required, independent parameter groups can be set for different subplots.

[0134] In specific examples, the hyperparameters can be set as follows:

[0135] Geographical core: Take the local city scale, Take the regional / national scale; .

[0136] Social similarities: SimRank .

[0137] Encoding Dimensions: Projected dimensions after stitching .

[0138] Positive-negative ratio: And perform negative sampling aligned with degree distribution.

[0139] Example 1

[0140] Data Preparation and Network Construction: Patent transaction data was collected from a province, covering 5,000 patents from 100 enterprises, 20 universities, and 10 research institutions. Using these patents and entities as nodes, a multi-layered patent transaction network was constructed. In the technology relationship network, edges were constructed based on the similarity or identicality of patent IPC classification numbers and patent citation relationships. In the geographical relationship network, the latitude and longitude coordinates of the registered addresses of each entity were obtained, and a distance threshold of 100 kilometers was set; edges were established between entities with distances less than this threshold. In the social relationship network, edges were established between entities with cooperative projects, based on their records of such projects.

[0141] Multidimensional proximity feature extraction:

[0142] Technology similarity features: One-Hot encoding is performed on patent IPC classification numbers to construct patent technology vectors. In conjunction with patent citation relationships, if patent A cites patent B, the technology association weight between the two is increased. By calculating the cosine similarity of the patent technology vectors, a technology similarity matrix is ​​obtained, thereby extracting technology similarity features.

[0143] Geographic proximity feature: The Haversine formula is used to calculate the geographic distance between subjects. Subjects with a distance of less than 100 kilometers are assigned different weights according to their distance, with the closer the distance, the higher the weight, thus generating the geographic proximity feature.

[0144] Social proximity characteristics: Analyze information such as the number of collaborative projects and the duration of collaboration among subjects to develop quantitative scoring standards. For example, 1 point is awarded for each additional collaborative project, and 2 points are awarded for each full year of collaboration. The social proximity characteristics are obtained based on the scoring results.

[0145] Network motif analysis: A graph mining algorithm was used to analyze the constructed multi-layer network, focusing on extracting triangular motifs. Statistical analysis revealed 800 triangular motifs in the network. Information such as the number of times each node participated in a triangular motif was recorded as motif features.

[0146] Model Training and Prediction: A multi-channel graph attention mechanism was designed, comprising four attention channels corresponding to technology similarity features, geographical proximity features, social relationship proximity features, and network motif features, respectively. During training, a dynamic negative sampling strategy was introduced, with an initial negative sampling ratio of 3:1, which was dynamically adjusted based on changes in the model's loss value as training progressed. A graph convolutional neural network was used as the basic model architecture, with three convolutional layers: 32 kernels in the first layer, 64 in the second, and 128 in the third, using ReLU as the activation function. The Adam optimizer was used with a learning rate of 0.001, a batch size of 64, and training for 200 epochs. Ultimately, the model achieved an accuracy of 88% and a recall of 85% on the test set.

[0147] Example 2

[0148] Data Preparation and Network Construction: Patent transaction data from across the country was selected, involving 20,000 patents from 500 enterprises, 100 universities, and 50 research institutions. A multi-layered patent transaction network was constructed. When constructing the technology relationship network, in addition to considering IPC classification numbers and citation relationships, patent keyword matching was also introduced. In the geographical relationship network, the distance threshold was adjusted to 300 kilometers. The social relationship network incorporated factors such as enterprise investment relationships to construct edges.

[0149] Multidimensional proximity feature extraction:

[0150] Technical similarity features: Based on Example 1, the patent keywords are vectorized and combined with IPC classification numbers and citation relationships to calculate technical similarity features using a weighted fusion method.

[0151] Geographic proximity features: Geographic information system (GIS) technology is used to calculate geographic distance more accurately and adjust the weights by taking into account factors such as transportation convenience to generate geographic proximity features.

[0152] Social proximity characteristics: In addition to cooperation projects and time, the scoring system incorporates factors such as corporate investment amount and equity relationship to obtain a more comprehensive social proximity characteristic.

[0153] Network motif analysis: Not only triangular motifs were extracted, but also quadrilateral motifs were extracted. The frequency of occurrence of different types of motifs and the position of nodes in the motifs were statistically analyzed, resulting in a total of 3,000 triangular motifs and 1,200 quadrilateral motifs.

[0154] Model Training and Prediction: The multi-channel graph attention mechanism was optimized by increasing the number of attention heads to four per channel. In the dynamic negative sampling strategy, a hard negative sample mining algorithm was employed, prioritizing negative samples whose predicted probabilities are close to those of positive samples for training. The model architecture uses a graph attention network (GAT) with four attention layers, each with eight attention heads, and the activation function is LeakyReLU. Training parameters were set to a learning rate of 0.0005, a batch size of 128, and a training duration of 300 epochs. Ultimately, the model achieved an accuracy of 90% and a recall of 87% on the test set.

[0155] Comparative Example 1

[0156] This study uses only technology similarity and geographic proximity features, ignoring social proximity and network motif features. On the same dataset as in Example 1, a technology relationship network and a geographic relationship network are constructed, and their corresponding features are extracted and input into a graph convolutional neural network model. The parameter settings during training are the same as in Example 1. The final model achieves an accuracy of 82% and a recall of 80% on the test set. Compared to Example 1, both accuracy and recall are significantly lower, indicating that social proximity and network motif features play a crucial role in improving prediction performance.

[0157] Comparative Example 2

[0158] This invention employs a traditional graph neural network model without using multi-channel graph attention mechanisms and dynamic negative sampling strategies. On the dataset of Example 2, only simple node embedding is performed on the multi-layer network, and the features are input into the traditional graph neural network model. The training parameters are the same as in Example 2, and the model achieves an accuracy of 85% and a recall of 82% on the test set. Compared to Example 2, this invention significantly improves the predictive performance of the model through multi-channel graph attention mechanisms and dynamic negative sampling strategies.

[0159] Comparative Example 3

[0160] This comparative example only uses technology similarity features, geographical proximity features, and social relationship proximity features, without introducing network motif features; the rest of the dataset, time-rolling partitioning, and training process are consistent with Example 2 (multi-channel attention fusion and dynamic negative sampling strategy are retained; the base network is a 4-layer GAT with 8 attention heads per layer; learning rate 0.0005, batch size 128, training for 300 epochs). In the fusion and prediction layers, the motif channels are removed, and the original four-channel attention is changed to three-channel attention. Weight learning and fusion are performed only on the three proximity categories of "technology-geography-society", which are then input into the decoder to output the link transformation probability.

[0161] Under the same testing conditions, this comparative example achieved an accuracy of 83% and a recall of 82% on the test set. The results indicate that the three-class proximity model can still achieve high recognition performance without introducing higher-order structural information. Meanwhile, compared to Example 2 (which includes phantom features), the overall performance is slightly lower, suggesting that higher-order local structural patterns have a positive effect on improving recognition and coverage in certain contexts. To ensure comparability, this comparative example kept all hyperparameters and training procedures unchanged except for the removal of phantom features.

[0162] In summary, this invention discloses a patent transaction link prediction model that integrates multidimensional proximity and network motifs. Addressing the issue that existing patent conversion prediction models neglect the multidimensional correlation characteristics of technological entities and the high-order structure of networks, this invention constructs a multi-layered patent transaction network encompassing technological, geographical, and social relationships. It extracts technological similarity, geographical proximity, and social proximity features respectively, and combines this with network motif analysis to capture local structural patterns. A multi-channel graph attention mechanism is designed to achieve adaptive fusion of multidimensional features and motif features, while a dynamic negative sampling strategy is introduced to optimize the model training process. This invention effectively solves the problems of single-dimensional analysis and insufficient utilization of structural information by fusing cross-layer correlation modeling across technological, geographical, and social dimensions with the high-order structural representation of network motifs. This significantly improves the accuracy and generalization ability of patent transaction link prediction, making it suitable for potential transaction link mining in scenarios such as patent trading platforms and intellectual property service agencies, providing decision support for efficient patent conversion.

[0163] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the device in the above device embodiments by referring to the specific technical solutions in other method embodiments and combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, on the basis of the above system embodiments, and on the premise of ensuring the practicality of the technical solutions, they can obtain corresponding device-type embodiments for implementing the methods in other method-type embodiments.

[0164] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a patent transaction link prediction device that integrates multidimensional proximity and network models, including a memory and a processor. The memory stores program instructions that are executed by the processor, and the processor calls the program instructions to execute the steps of the above-described patent transaction link prediction method that integrates multidimensional proximity and network models.

[0165] In embodiments of the present invention, the memory can be non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or it can be volatile memory, such as random access memory (RAM). Memory is any other medium capable of carrying or storing desired program code having an instruction or data structure form and accessible by a computer, but is not limited thereto. The memory in embodiments of the present invention can also be a circuit or any other device capable of implementing a storage function for storing program instructions and / or data.

[0166] In this embodiment of the invention, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in this embodiment of the invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in this embodiment of the invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0167] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the steps of the patent transaction link prediction method that integrates multidimensional proximity and network models.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting patent transaction links that integrates multidimensional proximity and network models, characterized in that, Includes the following steps: Construct a multi-layered patent transaction network, which includes a technology relationship network, a geographical relationship network, and a social relationship network; Technical similarity features, geographical proximity features, and social proximity features are extracted separately. Technical similarity features are calculated based on the patent field and patent citation relationships. Geographical proximity features are calculated based on the geographical location information of the patent applicant or the right holder. Social proximity features are obtained based on the cooperation history and affiliation between the applicants. Network motif analysis was performed on a multi-layered patent transaction network to capture local structural patterns. A multi-channel graph attention mechanism is designed to fuse extracted technology similarity features, geographical proximity features, social relationship proximity features, and network motif features; A dynamic negative sampling strategy is introduced to optimize model training, and the fused features are input into the trained model for patent transaction link prediction.

2. The patent transaction link prediction method integrating multidimensional proximity and network models according to claim 1, characterized in that, The construction of the multi-layered patent transaction network specifically involves: using patents and patent transaction participants as nodes, constructing a technology relationship network based on patent technology associations, a geographical relationship network based on the geographical location of the participants, and a social relationship network based on the social relationships of the participants.

3. The patent transaction link prediction method integrating multidimensional proximity and network models according to claim 1, characterized in that, The method for calculating the technical similarity features is as follows: the patent technology field is encoded, a technical similarity matrix is ​​constructed by combining the patent citation relationship, and the technical similarity features are obtained through matrix operations; The inter-institutional similarity matrix is ​​as follows: ； in, It is the technical similarity score between institutions u and v; , These are the patent technology vectors of institutions u and v, derived from bag-of-words vectors. constitute; , It is the collection of technical fields of institutions u and v.

4. The patent transaction link prediction method integrating multidimensional proximity and network models according to claim 1, characterized in that, The method for calculating the geographic proximity feature is as follows: obtaining the latitude and longitude coordinates of the patent applicant or the right holder, calculating the spatial distance between the coordinates, and generating the geographic proximity feature by combining a distance decay function, including: Convert the longitude dimension of the organization's address to three-dimensional Cartesian coordinates; The spatial distance between computing units u and v ; The distance decay function: ； in, It is the geographical proximity score of institutions u and v; These are weighting coefficients; , It is the attenuation radius parameter.

5. The patent transaction link prediction method integrating multidimensional proximity and network models according to claim 1, characterized in that, The method for obtaining the social relationship proximity characteristics is as follows: analyze the number of collaborative projects, the time span of collaboration, and the affiliation information among applicants, and obtain the social relationship proximity characteristics through quantitative scoring.

6. The patent transaction link prediction method integrating multidimensional proximity and network models according to claim 5, characterized in that, Modeling of social proximity includes: (1) Construction of historical interaction network: ； in, : Length of time window; Indicates the intensity of unnormalized historical interactions; Indicates to The result after min–max normalization; Indicates a time interval; Indicates the time decay coefficient; This indicates the number of historical interaction events between u and v, such as the number of joint patent applications; (2) Enhanced structural similarity: fusing local and global similarity: ； in, It is the fusion score of the social relationship structure similarity between institutions u and v; It is a fusion weight of local and global similarity; , These are respectively represented in the social relationship diagram. The set of neighbors of v; It is the SimRank decay parameter; It is the recursive calculation of the social relationship similarity between institutions a and b; a and b are and Any neighboring node in; (3) Social Proximity Integration ； in, It is the score of the social relationship similarity between institutions u and v; This represents the fusion weight between historical interaction strength and structural similarity.

7. The patent transaction link prediction method integrating multidimensional proximity and network models according to claim 1, characterized in that, The multi-channel graph attention mechanism includes multiple attention channels, each channel performs attention calculations on different types of features, and then the results of each channel are fused together.

8. A patent transaction link prediction model integrating multidimensional proximity and network models, characterized in that, include: The input layer is used to obtain the data source; The processing layer is used to construct a multi-layered patent transaction network and extract features such as technological similarity, geographical proximity, and social relationship proximity. Used for network schema analysis of multi-layered patent transaction networks; The fusion and prediction layer is used to design a multi-channel graph attention mechanism that fuses technology similarity features, geographical proximity features, social relationship proximity features, and network motif features. This is used to introduce a dynamic negative sampling strategy to optimize model training, and the fused features are input into the trained model for patent transaction link prediction.

9. A patent transaction link prediction device integrating multidimensional proximity and network models, characterized in that, The method includes a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to perform the steps of the patent transaction link prediction method integrating multidimensional proximity and network modalities as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the steps of the patent transaction link prediction method that integrates multidimensional proximity and network models as described in any one of claims 1 to 7.

Images