Patent co-recommendation method based on topic-enhanced neural network model

Abstract

This invention discloses a patent collaborative recommendation method based on a topic-enhanced neural network model, belonging to the field of patent recommendation technology based on deep learning models, and mainly involving a patent collaborative recommendation method based on a topic-enhanced neural network model. The specific method includes: S1 constructing a university-enterprise patent fusion dataset; S2 preprocessing the summary text in the fusion dataset to obtain preprocessed samples; S3 identifying R&D topics on the pre-trained samples based on the BERTopic model; S4 constructing feature datasets for universities and enterprises respectively based on the sample topics; S5 constructing a topic-enhanced neural network recommendation model; and S6 outputting a list of recommendation results. This invention effectively alleviates the problem of model overfitting caused by data sparsity and the low accuracy of traditional recommendation models.

Description

Technical Field

[0001] This invention relates to a patented collaborative recommendation method based on a topic-enhancing neural network model, belonging to the field of information recommendation technology based on deep learning models. Background Technology

[0002] In today's era, university-industry collaboration has become a crucial pathway for promoting technological innovation and facilitating the transformation of research results. This is especially true in technology-driven fields, where university-industry collaboration plays a vital role in the development and transformation of intellectual property. Patents, as significant technological achievements of university-industry collaboration, not only reflect technological innovation capabilities but also serve as core indicators for technology evaluation and intellectual property strategy. In the next-generation information technology industry, a field highly dependent on technological progress, university-industry patent collaboration is particularly important. The development of emerging technologies such as edge computing, 5G communication, and the Internet of Things requires close collaboration between universities and enterprises, combining cutting-edge theoretical research from universities with engineering practices from enterprises to generate high-value patent achievements.

[0003] However, accurately identifying key R&D technology themes in university-industry collaborative patented technologies and precisely recommending university patented technologies to enterprises has become a crucial issue in improving the efficiency of university-industry collaboration. On the one hand, identifying and matching R&D technology themes in university-industry collaborations is challenging. University research focuses more on cutting-edge theories, while enterprises prioritize market applications. These differing perspectives make it difficult to accurately pinpoint shared R&D themes and clarify the direction of cooperation. On the other hand, universities and enterprises often lack strong support from relevant patent resources when conducting collaborative R&D themes. The patent resources held by universities are difficult to accurately match with the actual needs of enterprises, resulting in patent outcomes that cannot effectively connect with enterprise production or other applications, failing to fully leverage the role of patents in promoting enterprise innovation. Furthermore, the process of using patent data for technology theme identification and matching often faces technical challenges due to the sparsity of patent data. The relatively limited number of university-industry patent collaborations, uneven distribution of textual information, and the professional and domain-specific nature of patent terminology lead to significantly sparse data for R&D theme identification. This makes traditional recommendation models prone to overfitting and insufficient generalization ability. Meanwhile, traditional matching models based on keywords or simple statistics are unable to deeply capture the semantic relationships between patented technologies, resulting in limited recommendation accuracy and making it difficult to fully support efficient and accurate patented technology matching and transformation between universities and enterprises. Summary of the Invention

[0004] Purpose of the invention: To address the problems and shortcomings of existing technologies, this invention provides a patented collaborative recommendation method based on a topic-enhanced neural network model, which solves the problems of model overfitting caused by data sparsity and the low accuracy of traditional recommendation models.

[0005] Technical solution: A patent collaborative recommendation method based on a topic-enhancing neural network model, comprising the following steps: S1. Construct a university-enterprise patent fusion dataset, which includes patents where the applicant is solely a university, patents where the applicant is solely an enterprise, and patents where the applicant is both an enterprise and a university.

[0006] S2. Preprocess the summary text in the fused dataset to obtain preprocessed samples.

[0007] S2 specifically includes: S21. Targets with missing values ​​in the patent abstract are deleted. S22. A stop word dictionary is set up based on the jieba library, and domain-specific dictionaries are introduced to retain professional terms in the text, improving the completeness and accuracy of keyword selection. The abstract text is segmented, and noise filtering is applied to the segmentation results. High-frequency irrelevant words and general stop words are manually selected to supplement the stop word list, thus constructing a composite stop word list.

[0008] S23. Output the word segmentation sequence based on the compound stop word list as a standardized preprocessing sample for subsequent research and development of topic recognition.

[0009] S3. Research and development topic identification is performed on preprocessed samples of the fused dataset based on the BERTopic model.

[0010] S3 specifically includes: S31. Use the BERTopic model to identify research and development topics in the preprocessed samples, and obtain topic clustering and topic support for each sample. S32. Select the top-5 topics as sample topics based on topic support.

[0011] S4. Based on the sample topics, construct patent feature datasets for universities and enterprises respectively.

[0012] S4 specifically includes: S41. Unifying pure university patent data and university-enterprise joint patent data, the constructed university patent feature dataset includes: (1) Topic support feature vector.

[0013] (2) Technical indicator vector: normalized indicators of the number of claims, the number of citations, the number of inventors and the number of transfers; cited patents refer to the patents that have been cited by subsequent patents during the application or examination process, that is, the current patent is cited by others as prior art, and the number of citations reflects the technical influence of the patent.

[0014] (3) Technical indicator interaction feature vector: The normalized interaction feature of the number of claims and the number of citations quantifies the synergistic effect of the two through the product of the normalized data of the number of claims and the number of citations, revealing the coupling relationship between technical value and technical influence, and more comprehensively reflecting the patent value.

[0015] S42. Prioritize and recommend companies with a good foundation for cooperation and strong technical needs under each theme; Patents with school as the transferor type are selected from enterprise patents to obtain a preliminary enterprise patent dataset, ensuring that enterprises have a good foundation for school-enterprise cooperation and a willingness to cooperate. In the preliminary enterprise patent dataset, enterprises that simultaneously meet the criteria of high number of patents and high theme support under the same theme are recommended as priority enterprises. The "double high" means that they simultaneously meet the preset threshold, ensuring that the enterprises have a high technical demand for the corresponding theme.

[0016] The constructed enterprise technology demand feature dataset includes: average topic support for enterprise technology demands and enterprise profile features; the enterprise profile features include enterprise size, geographical distribution, and R&D intensity. The total topic support for all eligible patents of the prioritized enterprise under the same R&D topic is calculated, and the average topic support for the enterprise is calculated after counting the number of eligible patents. The topic support for each patent is automatically calculated during topic modeling using BERTopic. The five topic support values ​​presented in this invention are obtained by calculating the cosine similarity between the patent text embedding vector and the five topic center vectors.

[0017] S5. Construct a neural network recommendation model based on topic enhancement.

[0018] S5 specifically includes: S51. Construct a feature input layer, which consists of two parts: patent features from universities and demand features from enterprises. The patent features from universities are represented by the topic support vector t. p Technical indicator vector m p Interaction feature vector h of technical indicators p Generated by fusion.

[0019] x p =Concat ( t p , m p , h p (1) Enterprise-side demand characteristics are derived from the average topic support vector t of enterprise technology demand. e and enterprise profile feature vector i e The composition is further derived by analyzing the average topic support vector t of enterprise technology needs.e Gaussian perturbation distribution is used to inject topic noise into the enterprise profile feature vector i. e Scaled Gaussian noise is injected into the portrait to reduce the risk of model overfitting. The two types of features are then fused to obtain the enterprise-side demand feature x. e .

[0020] (2) (3) x e =Concat ( t e , i e (4) in, and Let represent the features of the average topic support vector of enterprise technology requirements and the feature vector of enterprise profile before adding noise, respectively. and These are the average topic support vector noise for enterprise technology needs and the feature vector noise for enterprise profiles, respectively.

[0021] S52. Construct a dual-tower encoder layer to map multimodal patent features into low-dimensional embedding vectors. The multimodal patent features refer to the topic support vector of university patents, as well as technical indicator and interaction technical indicator vectors. By establishing a fully connected layer and combining layer normalization and ReLU activation functions, the patent features from the university side and the demand features from the enterprise side are fused and encoded respectively. e p = ReLU ( LayerNorm ( W p x p + b p (5) e e = ReLU ( LayerNorm ( W e x e + b e (6) Among them, e p and e e This includes patent embedding vectors from universities and demand embedding vectors from enterprises. p For the patent weight matrix on the university side, b p This is the patent bias vector for universities. We For the enterprise-side demand weight matrix, b e This is the enterprise-side demand bias vector.

[0022] S53. Construct a Transformer topic feature interaction layer, employing a multi-head attention mechanism to dynamically capture the fine-grained relationship between enterprise needs and patent topics: (7) in , , , W q 、W k 、W v These are the weight matrices for Q, K, and V, respectively. , h The number of heads set for the multi-head attention mechanism. The result of the attention mechanism highlights the themes or technical indicators in the patent that are highly relevant to the company's needs, while downplaying irrelevant parts, thereby achieving fine-grained feature interaction.

[0023] A feedforward network (FNN) with the nonlinear activation function GELU is introduced. Based on a global attention interaction mechanism, it refines the combination relationships of local features such as patent technical indicators, dynamically adjusts the topic relevance score output by the self-attention layer, and suppresses patents with high topic similarity but low technical influence. Simultaneously, residual connections are used to avoid the attenuation of key dimensional information during the interaction process. Citation counts are used as a fundamental technical indicator to measure patent technical influence. Since the selected patents are based on data from the past five years, a citation count of 0 is set as the threshold to distinguish between high and low technical influence. This operation does not directly eliminate low-technical-influence patents, but rather imposes technical restrictions in the matching calculation, prioritizing the recommendation of high-technical-influence patents.

[0024] FFN ( x ) = W 2 ·GELU ( W 1 ·x in + b 1 ) + b 2 (8) x out =LayerNorm ( x in + Attention ( Q,K,V )) (9) x final =LayerNorm ( x out + FFN ( x out (10) Where W1 is the weight matrix that performs the linear transformation of increasing dimension, b1 is the relevant bias vector of W1, W2 is the weight matrix that performs the linear transformation of decreasing dimension, and b2 is the relevant bias vector of W2. x in The input vector for the Transformer layer. x out The feature vector is the result of processing by the self-attention sublayer. x final This is the final output feature vector of the Transformer layer.

[0025] Topic similarity regularization is introduced in the feature interaction layer to constrain the embedding vector distance of topic-similar patents, forcing the embedding vector distances of topic-similar patents to be close. The topic similarity regularization term... for: (11) in, This is the embedding vector for university patent i under the corresponding theme. Let be the embedding vector of university patent j under the corresponding theme. Let P be the topic cosine similarity between the embedding vectors of patent i and patent j from universities, and let P be the set of patent pairs with high topic similarity that satisfy a topic cosine similarity of not less than 0.8.

[0026] S54. Construct a prediction layer. Utilize a multilayer perceptron (MLP) to fuse the enterprise-side demand embedding vector with the attention result into a feature vector, and predict the matching probability through nonlinear transformation. Simultaneously, introduce a topic similarity regularization term combined with a binary cross-entropy loss function to construct a composite loss function, optimizing the model's ability to capture text topic-related features. Predict matching probability. The formula is as shown in formula (13): h concat = Concat ( e e , x final (12) (13) (14) in, A true indicator of university patents and corporate collaborations. This is used to predict the matching probability of university patents and enterprise collaborations. Here, W3 and W4 are the weight matrices, and b3 and b4 are the bias vectors. This is the feature vector resulting from the fusion of the enterprise-side demand embedding vector and the attention result. N is the total number of training samples in the same training batch, and (e,p) represents the enterprise-patent pairs in the training set. This is the regularization coefficient.

[0027] S6. Output the list of recommended results.

[0028] S61. Candidate Patent Determination: For priority recommended enterprises, a subset of pure university patents will be used as the candidate set. S62. Matching Probability Calculation: Based on the trained recommendation model and its unified feature representation system, the embedding vector of enterprise-side demand and the embedding vectors of all candidate patents in the pure university patent feature set are input into the model to calculate the matching probability of each patent. ; S63. Sort the candidate patents in the pure university patent feature set in descending order according to the matching probability, and output the top-5 patents with the highest ranking as the final recommendation list.

[0029] A patent collaborative recommendation system based on a topic-enhancing neural network model includes the following modules: Dataset Construction Module: Constructs a university-enterprise patent fusion dataset, which includes patents where the applicant is solely a university, patents where the applicant is solely an enterprise, and patents where the applicant is both an enterprise and a university. Preprocessing module: preprocesses the summary text in the fused dataset to obtain preprocessed samples; Develop a topic recognition module: Based on the BERTopic model, develop topic recognition for preprocessed samples of the fused dataset; Feature dataset construction module: Based on sample topics, construct patent feature datasets for universities and enterprises respectively; Recommendation model building module: Constructs a neural network recommendation model based on topic enhancement; Output module: Outputs a list of recommended results.

[0030] The implementation process and methods of the system are the same and will not be described again.

[0031] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the patent collaborative recommendation method based on a topic-enhanced neural network model as described above.

[0032] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the patented collaborative recommendation method based on a topic-enhancing neural network model as described above.

[0033] Beneficial Effects: Addressing bottlenecks in university-enterprise collaborative innovation, such as ambiguous R&D directions and inefficient resource matching, as well as existing university-enterprise patent recommendation models suffering from data cold start and semantic alignment issues between demand and patents, this invention constructs a methodological framework of "topic identification-feature mapping-intelligent recommendation." Specifically, the BERTopic model is first used to identify R&D topics for university-enterprise patents. Based on the identification results, a collaborative filtering recommendation method based on topic-enhanced neural networks is proposed. A dual-tower encoder for patents and enterprises is constructed to achieve heterogeneous fusion of static patent features and dynamic enterprise preferences, overcoming the limitation of traditional neural collaborative filtering that only processes ID-type features. Gaussian noise is injected into enterprise features to achieve dynamic enterprise feature enhancement, improving the model's robustness to small perturbations in enterprise preferences and avoiding overfitting to specific enterprise profiles. The Transformer's self-attention mechanism is used for cross-modal feature interaction between patents and enterprises, deeply capturing the high-order nonlinear correlation between technical topics and enterprise needs. Simultaneously, topic similarity constraints are introduced to supplement sparse enterprise-patent interaction data. Through transfer learning strategies, the R&D topic features of universities are mapped to the enterprise demand space, significantly improving the robustness and interpretability of the recommendation system. Attached Figure Description

[0034] Figure 1 This is a flowchart of the implementation method of the present invention; Figure 2 This is a detailed structural flowchart of the method according to an embodiment of the present invention; Figure 3 This is a block diagram of the overall structure of the model in an embodiment of the present invention. Detailed Implementation

[0035] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0036] like Figure 1 As shown, this embodiment of the invention provides a patent collaborative recommendation method based on a topic-enhanced neural network model, comprising the following steps: S1. Construct a university-enterprise patent fusion dataset; Specifically, based on the Incopat patent database, we collected Chinese authorized patents with application publication dates between 2020 and 2024, valid legal status, and Industry=1.1 patent codes for the next-generation information technology industry. The retrieved patent data was independently filtered according to applicant type: "school," "enterprise," and "enterprise, school," and the integrated patent data yielded the university-enterprise patent fusion dataset presented in this paper.

[0037] S2. Preprocess the summary text of the fused dataset; Specifically, target patent data with missing values ​​in the summary items of the fused dataset are deleted. A stop word dictionary is set up based on the jieba library, and a domain-specific dictionary is introduced to ensure the complete preservation of technical terms and improve the accuracy of topic keyword selection. The summary text is segmented, and high-frequency irrelevant words and general stop words are manually selected to supplement the stop word list, constructing a composite stop word list. Finally, the segmented sequence is output based on the composite stop word list as a preprocessing sample for subsequent topic identification in research and development.

[0038] S3. Based on the BERTopic model, semantic understanding and dimensionality reduction are performed on the preprocessed samples. HDBSCAN clustering is used to discover thematic clusters of patent data from university-enterprise integration. Keywords for each thematic cluster are selected using word frequency statistics.

[0039] Specifically, the BERTopic model was used to train the preprocessed patent text. 18,462 patents were identified as noisy patent data and could not be assigned to valid topics. The remaining 256,506 patents were clustered into 14 topics, and the top-5 topics were selected as sample topics based on topic consistency for subsequent patent recommendation. The R&D topic identification results are shown in Table 1. Keywords are high-frequency words within the same topic in the R&D topic identification results, displayed as 4-5 words. Topic summarization is performed using a large language model to summarize the topic based on the most frequent keywords.

[0040] Table 1. Results of R&D Theme Identification

[0041] S4. Construct feature datasets for universities and enterprises based on sample themes; Specifically, patent data from five sample topics were selected to construct feature datasets for universities and enterprises respectively.

[0042] This study unifies pure university patent data and university-enterprise joint patent data to construct a university patent feature dataset. The subject support vector of the patents in the dataset is input and normalized. The number of claims, the number of citations, the number of inventors, and the number of transfers are selected as patent technical indicators, and the indicator data are normalized. The interaction features of patent indicators are introduced, and the synergistic effect of the two is quantified in the form of the product of the normalized data of the number of claims and the number of citations, revealing the coupling relationship between technological value and technological influence.

[0043] To construct a corporate feature dataset, the process begins by filtering patents whose transferor type is a university. Companies with both high patent quantity and high theme support within the same theme are identified, and those with the most qualifying patents are prioritized to ensure strong collaboration foundations and technological needs. Next, the average theme support vector for companies within the same theme is input and normalized. Based on three fundamental corporate profile features—company size, geographic distribution, and R&D intensity—companies are categorized and coded. For training sample construction, positive samples consist of patents from university-enterprise joint patents within each technology theme that have collaboration relationships with the corresponding prioritized companies, with a true label of 1. Negative samples employ a negative sampling strategy: for each theme's prioritized company, several patents that the company has never collaborated on are randomly selected from a pure university patent feature database, their features are paired with the company's features, and a true label of 0 is assigned. A stratified sampling strategy is used to partition the dataset, allocating 80% of the total samples to the training set and the remaining 20% ​​equally to the validation and test sets.

[0044] S5. Construct a neural network recommendation model based on topic enhancement; Specifically, the topic-enhanced neural network recommendation model includes a feature input layer, a dual-tower encoder layer, a Transformer topic feature interaction layer, and a prediction layer; The specific model structure is as follows Figure 3 As shown, it specifically includes: The feature input layer includes patent features from universities and demand features from enterprises. University-side patent features x p From topic support vector t p Technical indicator vector m p Interaction feature vector h of technical indicators p Fusion generation; Enterprise-side demand characteristics x e The average topic support vector t of enterprise technology needs e and enterprise profile feature vector i e Composition, where t e and i e Gaussian perturbation distribution and scaled Gaussian noise are used to inject topic noise to reduce the risk of model overfitting.

[0045] The dual-tower encoder layer includes a patent encoder and an enterprise encoder. Both are fully connected layers that combine Layer Normalization optimization and ReLU activation function to fuse and encode the input features of university patents and the demand features of enterprises, respectively.

[0046] The Transformer topic feature interaction layer incorporates a feedforward network FNN with a nonlinear activation function GELU. Based on a global attention interaction mechanism, it dynamically captures the fine-grained correlation between enterprise needs and patent topics. Topic similarity regularization is introduced to force the vector distance of patents with similar topics to be close. The evaluation metrics of the best training model under each topic are shown in Table 2.

[0047] Table 2 Evaluation metrics for the best training models for each topic

[0048] The prediction layer first loads the best model saved during training to ensure the accuracy of the prediction results, and uses a multilayer perceptron to fuse the feature vector after fusing the enterprise-side demand embedding vector with the attention result, and predicts the matching probability through nonlinear transformation; it introduces a topic similarity regularization term and combines it with the binary cross-entropy loss function to optimize the model's ability to capture text topic association features.

[0049] S6. Output the list of recommended results; Specifically, for the target enterprise, the feature vectors of all patents in the pure university patent feature set and the enterprise-side demand embedding vector are input into the pre-trained recommendation model to calculate the matching probability of each patent. The patents are sorted in descending order of matching probability from high to low, and the Top-5 pure university patents are output as the final recommendation results, as shown in Table 3.

[0050] Table 3 Recommendation Results

[0051] The above embodiments are merely examples to clearly illustrate the technical solutions of the present invention and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, improvements, and arbitrary combinations of technical features made within the spirit and principle framework of the claims of the present invention should be included within the protection scope of the claims of the present invention. The above description uses preferred embodiments as examples, and while striving for specific detail, it should not be construed as limiting the scope of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these also fall within the protection scope of the present invention. The final protection scope of this patent should be determined by the appended claims.

Claims

1. A patent collaborative recommendation method based on a topic-enhancing neural network model, characterized in that, Includes the following steps: S1. Construct a university-enterprise patent fusion dataset, which includes patents where the applicant is solely a university, patents where the applicant is solely an enterprise, and patents where the applicant is both an enterprise and a university. S2. Preprocess the summary text in the fused dataset to obtain preprocessed samples; S3. Research and development topic identification based on preprocessed samples of the fused dataset using the BERTopic model; S4. Based on the sample topics, construct patent feature datasets for universities and enterprises respectively; S5. Construct a neural network recommendation model based on topic enhancement; S6. Output the list of recommended results.

2. The patent collaborative recommendation method based on a topic-enhanced neural network model according to claim 1, characterized in that, S2 specifically includes: S21. Targets with missing values ​​in the patent abstract are deleted. S22. A stop word dictionary is set up based on the jieba library, and domain-related specialized dictionaries are introduced to retain professional terms in the text and improve the completeness and accuracy of keyword selection. The abstract text is segmented, and noise filtering is performed on the segmentation results. High-frequency irrelevant words and general stop words are manually selected to supplement the stop word list, and a composite stop word list is constructed. S23. Output the word segmentation sequence based on the compound stop word list as a standardized preprocessing sample for subsequent research and development of topic recognition.

3. The patent collaborative recommendation method based on a topic-enhancing neural network model according to claim 1, characterized in that, S3 specifically includes: S31. Use the BERTopic model to identify research and development topics in the preprocessed samples, and obtain topic clustering and topic support for each sample. S32. Select the top-5 topics as sample topics based on topic support.

4. The patent collaborative recommendation method based on a topic-enhancing neural network model according to claim 1, characterized in that, S4 specifically includes: S41. Unifying pure university patent data and university-enterprise joint patent data, the constructed university patent feature dataset includes: (1) Topic support feature vector; (2) Technical indicator vector: normalized indicators of the number of claims, the number of citations, the number of inventors, and the number of transfers; (3) Technical indicator interaction feature vector: the normalized interaction feature of the number of claims and the number of citations. By multiplying the normalized data of the number of claims and the number of citations, the synergistic effect of the two is quantified, revealing the coupling relationship between technical value and technical influence. S42. Prioritize and recommend companies with a good foundation for cooperation and strong technical needs under each theme; Patents with school as the transferor type are selected from enterprise patents to obtain the initial enterprise patent dataset; in the initial enterprise patent dataset, enterprises that simultaneously meet the criteria of high number of patents and high theme support under the same theme are selected as priority recommended enterprises; The constructed enterprise technology demand feature dataset includes: average topic support of enterprise technology demand and enterprise profile features; the enterprise profile features include enterprise size, geographical distribution and R&D intensity; calculate the total topic support of all eligible patents of the enterprise under the same R&D topic, and calculate the average topic support of the enterprise after counting the number of eligible patents.

5. The patent collaborative recommendation method based on a topic-enhanced neural network model according to claim 1, characterized in that, S5 specifically includes: S51. Construct a feature input layer, which consists of two parts: patent features from universities and demand features from enterprises. The patent features from universities are composed of the topic support vector t. p Technical indicator vector m p Interaction feature vector h of technical indicators p Fusion generation; x p =Concat ( t p , m p , h p ) (1) Enterprise-side demand characteristics are derived from the average topic support vector t of enterprise technology demand. e and enterprise profile feature vector i e The composition is further derived by analyzing the average topic support vector t of enterprise technology needs. e Gaussian perturbation distribution is used to inject topic noise into the enterprise profile feature vector i. e Scaled Gaussian noise is injected into the portrait to reduce the risk of model overfitting; the two types of features are fused to obtain the enterprise-side demand feature x. e ; (2) (3) x e =Concat ( t e , i e ) (4) in, and Let represent the features of the average topic support vector of enterprise technology requirements and the feature vector of enterprise profile before adding noise, respectively. and These are the average topic support vector noise for enterprise technology needs and the feature vector noise for enterprise profiles, respectively. S52. Construct a dual-tower encoder layer to map multimodal patent features into low-dimensional embedding vectors. The multimodal patent features refer to the topic support vector of university patents, as well as technical indicator and interaction technical indicator vectors. By establishing a fully connected layer and combining layer normalization and ReLU activation functions, the patent features from the university side and the demand features from the enterprise side are fused and encoded respectively. e p = ReLU ( LayerNorm ( W p x p + b p )) (5) e e = ReLU ( LayerNorm ( W e x e + b e )) (6) Among them, e p and e e For university-side patent embedding vectors and enterprise-side demand embedding vectors; W p For the patent weight matrix on the university side, b p For the patent bias vector on the university side; W e For the enterprise-side demand weight matrix, b e This represents the enterprise-side demand bias vector. S53. Construct a Transformer topic feature interaction layer, employing a multi-head attention mechanism to dynamically capture the fine-grained relationship between enterprise needs and patent topics: (7) in , , , W q 、W k 、W v The weight matrices for Q, K, and V are respectively. , h The number of heads set under the multi-head attention mechanism; A feedforward network (FNN) with the nonlinear activation function GELU is introduced. Based on the global attention interaction mechanism, local features are extracted and the topic relevance score output by the self-attention layer is dynamically adjusted to suppress patents with high topic similarity but low technical influence. At the same time, residual connections are used to avoid the attenuation of key dimension information during the interaction process. The number of citations is used as a technical indicator to measure the technical influence and fundamental technical indicators of the patent. FFN ( x out ) = W 2 ·GELU ( W 1 ·x out + b 1 ) + b 2 (8) x out =LayerNorm ( x in + Attention ( Q,K,V )) (9) x final =LayerNorm ( x out + FFN ( x out )) (10) Where W1 is the weight matrix that performs the linear transformation of increasing dimension, b1 is the relevant bias vector of W1, W2 is the weight matrix that performs the linear transformation of decreasing dimension, and b2 is the relevant bias vector of W2. x in The input vector for the Transformer layer. x out The feature vector after processing by the self-attention sublayer. x final This is the final output feature vector of the Transformer layer; Topic similarity regularization is introduced in the feature interaction layer to constrain the embedding vector distance of topic-similar patents, forcing the embedding vector distances of topic-similar patents to be close. The topic similarity regularization term... for: (11) in, This is the embedding vector for university patent i under the corresponding theme. Let be the embedding vector of university patent j under the corresponding theme. Let P be the topic cosine similarity between the embedding vectors of patent i and patent j from universities, and let P be the set of patent pairs with high topic similarity that satisfy a topic cosine similarity of not less than 0.

8. S54. Construct a prediction layer. Utilize a multilayer perceptron to fuse the enterprise-side demand embedding vector with the attention result's feature vector, and predict the matching probability through nonlinear transformation. Simultaneously, introduce a topic similarity regularization term combined with a binary cross-entropy loss function to construct a composite loss function, optimizing the model's ability to capture text topic-related features; predict the matching probability. The formula is as shown in formula (13): h concat = Concat ( e e , x final ) (12) (13) (14) in, A true indicator of university patents and corporate collaborations. This is used to predict the matching probability of university patents and enterprise collaborations. Here, W3 and W4 are the weight matrices, and b3 and b4 are the bias vectors. The feature vector is the result of fusing the enterprise-side demand embedding vector with the attention result; N is the total number of training samples in the same training batch, and (e,p) represents the enterprise-patent pairs in the training set. is the regularization coefficient.

6. The patent collaborative recommendation method based on a topic-enhanced neural network model according to claim 1, characterized in that, S6 includes: S61. Candidate Patent Determination: For priority recommended enterprises, a subset of pure university patents will be used as the candidate set. S62. Matching Probability Calculation: Based on the trained recommendation model and its unified feature representation system, the embedding vector of enterprise-side demand and the embedding vectors of all candidate patents in the pure university patent feature set are input into the model to calculate the matching probability of each patent. ; S63. Sort the candidate patents in the pure university patent feature set in descending order according to the matching probability, and output the top-5 patents with the highest ranking as the final recommendation list.

7. A patent collaborative recommendation system based on a topic-enhancing neural network model, characterized in that, Includes the following modules: Dataset Construction Module: Constructs a university-enterprise patent fusion dataset, which includes patents where the applicant is solely a university, patents where the applicant is solely an enterprise, and patents where the applicant is both an enterprise and a university. Preprocessing module: preprocesses the summary text in the fused dataset to obtain preprocessed samples; Develop a topic recognition module: Based on the BERTopic model, develop topic recognition for preprocessed samples of the fused dataset; Feature dataset construction module: Based on sample topics, construct patent feature datasets for universities and enterprises respectively; Recommendation model building module: Constructs a neural network recommendation model based on topic enhancement; Output module: Outputs a list of recommended results.

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the patent collaborative recommendation method based on the topic-enhanced neural network model as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the patent collaborative recommendation method based on the topic-enhanced neural network model as described in any one of claims 1-6.

Images