University patent personalized recommendation method and system

Abstract

The invention relates to the technical field of big data analysis and processing, in particular to a university patent personalized recommendation method and system. Constructing a citation network based on the preprocessed college patent data; clustering the quotation network, and dividing all patents into a plurality of theme clusters; calculating the text similarity between each patent and each enterprise demand based on the text of each patent and the text of each enterprise demand; constructing a time difference matrix based on the time difference between each patent and other patents cited for the last time; summing elements of corresponding rows of each patent in the time difference matrix to obtain a reference span score of each patent; according to the reference span score and the second score of each patent, determining the weights of the reference span score and the second score, and performing weighted calculation to obtain a comprehensive score of each patent; and generating a patent recommendation list meeting the requirements of each enterprise. The accuracy and scientificity of patent screening are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis and processing, and in particular to a method and system for personalized recommendation of university patents. Background Art

[0002] At present, the patent transfer and commercialization technology of universities has developed rapidly under the impetus of multidisciplinary cross-border. With the deepening of scientific and technological innovation and industrial transformation, the innovative achievements of universities in the field of scientific research have gradually become an important source of technology transfer and industrialization. However, unlike the technology transformation needs and methods of enterprises or other research institutions, the technology transformation of universities has some unique challenges and characteristics. First, due to the strong basic, innovative and relatively complex technical background of university patent technology, its marketization and transformation path is often not as direct and clear as that of corporate patents. Therefore, the screening and promotion of university patents needs to focus on identifying projects with potential market application prospects from university patents, and should be able to evaluate the convertibility and commercialization potential of these technologies. Secondly, the technology transformation path of universities is relatively complex, involving different intellectual property protection, policy support, market analysis and docking channels from enterprises. The resources and experience of universities in intellectual property management, patent layout, etc. are relatively limited. It is necessary to consider the complexity of the university transformation model to help university TTOs efficiently identify and dock with suitable industry needs. In addition, universities often achieve knowledge spillovers through cooperative networks during technology transfer, and cooperate with multiple companies or research institutions. The diversity and dynamism of such cooperative networks make the potential partners of university technology transfer more complex, so the screening and promotion of university patents requires the ability to dynamically identify and match, in order to support cross-border cooperation between universities and different industries. In addition, the patent transformation of universities not only pursues commercial profits, but also needs to consider social value orientation, such as promoting innovation ecology, talent cultivation and promoting social progress. Therefore, the screening and promotion of university patents needs to take into account both economic benefits and social value.

[0003] Traditional patent conversion usually relies on expert judgment and limited economic indicators, and these methods are inefficient in the modern market with a huge number of patents and diversified demands. With the advancement of artificial intelligence (AI), natural language processing (NLP), and data mining, new technical support has been provided for the intellectual property conversion of universities. At present, the technical means of patent identification are mainly concentrated in two aspects: on the one hand, methods based on statistical and historical data analysis, which mainly evaluate the conversion potential based on the historical performance of patents (such as the number of citations, the number of collaborations, etc.). On the other hand, methods based on text mining and similarity calculation, commonly used methods include text similarity measurement and semantic analysis, which analyze the similarity between patent texts and market demands to identify patents that may have conversion potential, and to a certain extent improve the efficiency of patent identification in universities. These emerging technologies make it possible to promote the personalized promotion of university patents and build a more efficient bridge for cooperation between universities and enterprises.

[0004] However, in the existing technology, the evaluation of patent value and transformation potential usually only considers single or a small number of indicators such as the number of citations and life cycle, while ignoring the time when the patent was last cited. Although some patents have been cited many times, they may no longer represent the current technological frontier or market demand with the development of technology. The evaluation of the life cycle is usually based on the application and authorization time of the patent. In many cases, a technology may be abandoned or replaced by a new technology at a certain stage of the life cycle, resulting in a sharp decline in its market demand. On the contrary, some patents may usher in new applications in the later stages of the life cycle, especially under the change of technology iteration or market demand. Therefore, only considering the life cycle may cause misjudgment of development potential, which makes it difficult for existing evaluation methods to identify patents that are truly forward-looking and have high innovation value. Especially in the rapidly developing technology field, relying solely on conventional indicators such as the number of citations cannot effectively evaluate whether a patent has a "leading" position in technology, which may miss the cutting-edge technology with high potential, resulting in the promotion of patents that may be more outdated or less influential, and lacking technological leadership and innovation. In addition, the existing technology evaluates the conversion potential by analyzing the similarity between patent text and market demand, and usually directly converts the patent text into a vector representation for similarity calculation. However, the patent text is usually long and contains a lot of redundant information and legal terms. Direct processing may cause information noise, affecting the accuracy of similarity calculation, resulting in insufficient precision of recommended patents, and in the case of a large number of patents, it will affect the response speed of the system. In addition, the scoring system of the existing technology usually lacks comprehensiveness, focusing only on the technical indicators of the patent, such as innovation and citation frequency, but ignoring economic and legal factors. This limitation leads to the one-sidedness of the evaluation system, affecting the accuracy and scientificity of patent screening. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the defects that the prior art ignores the foresight of patents in technological innovation and its impact on subsequent technologies, and the prior art directly converts the patent text into a vector representation, which reduces the accuracy of similarity calculation. In addition, the scoring system of the prior art usually lacks comprehensiveness, resulting in reduced accuracy and scientificity of patent screening.

[0006] In order to solve the above technical problems, the present invention provides a method for personalized recommendation of university patents, comprising the following steps:

[0007] Obtain university patent data and preprocess it to obtain preprocessed university patent data; wherein the university patent data includes: citation information of different patents, the number of citations, the time of citation, text, patent owner information, and the number of IPC classification numbers;

[0008] Based on the preprocessed university patent data, each patent is taken as a node and the citation relationship between patents as an edge to construct a citation network;

[0009] Cluster the citation network and divide all patents into multiple subject clusters, each of which represents a technical direction;

[0010] Based on the text of each patent and the text of each enterprise demand, calculate the text similarity between each patent and each enterprise demand;

[0011] Construct a time difference matrix based on the time difference between each patent and other patents when it was last cited;

[0012] Sum the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent;

[0013] After text vectorization of the texts of each patent and other patents, the text similarity between each patent and other patents is calculated;

[0014] Calculate the first score for each patent based on the number of citations, texts and IPC classification codes of each patent;

[0015] Calculate the second score of each patent based on the maximum value of the text similarity between each patent and other patents and the first score of each patent;

[0016] According to the citation span score and the second score of each patent, the weights of the citation span score and the second score are determined, and the comprehensive score of each patent is obtained by weighted calculation;

[0017] Based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent, a patent recommendation list that meets the needs of each enterprise is generated.

[0018] Preferably, the calculation formula for each element in the time difference matrix is:

[0019]

[0020] Among them, t ij is the element in the i-th row and j-th column of the time difference matrix, d ij For the i-th patent P i and patent P j The time difference of the last citation, d ij =t j -t i , t j is the time when the jth patent was last cited, t i is the time when the i-th patent was last cited, e is a natural constant, α is the time decay factor, q ij is the quality weighting factor of the i-th patent and the j-th patent, t current is the current time.

[0021] Preferably, the row elements corresponding to each patent in the time difference matrix are summed to obtain the citation span score of each patent, and the calculation formula is:

[0022]

[0023] Among them, Y i is the citation span score of the i-th patent, n is the total number of patents, t ij is the element in the i-th row and j-th column of the time difference matrix.

[0024] Preferably, the first score of each patent is calculated based on the number of citations, texts and IPC classification numbers of each patent, and the calculation formula is:

[0025] S1=ω1·C norm +ω2·N(K)+ω3·S IPC +ω4·Sim,

[0026] Among them, S1 is the first score of the current patent, ω1, ω2, ω3, and ω4 are the weights of the number of citations, the weight of keyword scarcity, the weight of technology branch concentration, and the weight of feature deviation, respectively, and C norm is the normalized value of the number of citations of the current patent, N(K) is the keyword scarcity score of the current patent, μ is the frequency of the keywords of the current patent appearing in other patents, S IPCis the normalized value of the number of IPC classification numbers to which the current patent belongs, Sim = 1-CosineSimilarity(F,G), Sim is the feature deviation between the current patent and the patent that cites it, CosineSimilarity(.) is the cosine similarity, F is the vector representation of the current patent text, and G is the vector representation of the patent text that cites it.

[0027] Preferably, the second score of each patent is calculated based on the maximum value of the text similarity between each patent and other patents and the first score of each patent, and the calculation formula is:

[0028]

[0029] Among them, S2 is the second score of the current patent, max(Similarity) is the maximum value of the text similarity between the current patent and other patents, Similarity is the similarity between the vector representation of the current patent text and the vector representation of other patent texts, is the first scoring weight, and S1 is the first scoring of the current patent.

[0030] Preferably, the university patent data is obtained and preprocessed to obtain the preprocessed university patent data, including:

[0031] After word segmentation, stop word removal and word form restoration, the university patent data is formatted to obtain the preprocessed university patent data.

[0032] Preferably, the calculating of the text similarity between each patent and each enterprise requirement based on the text of each patent and the text of each enterprise requirement includes:

[0033] Decompose each module of each patent to obtain decomposition information of each module in each patent; the modules include: invention title, abstract, claims, and specification;

[0034] Structuring the decomposition information of each module in each patent to generate structured input data for each module;

[0035] The structured input data of all modules of each patent are passed through the set prompt word template to generate the prompt words of each patent; the prompt words of each patent are input into the large language model respectively, and the characteristic summary of each patent is output;

[0036] The feature summary of each patent and the text of each enterprise requirement are converted into their corresponding vector representations, and the text similarity between each patent and each enterprise requirement is calculated.

[0037] Preferably, the method further includes: obtaining scores of each dimension, determining weights of the citation span score, the second score and each dimension according to the citation span score, the second score and the score of each dimension of each patent, and performing weighted calculation to obtain an optimized comprehensive score of each patent;

[0038] Among them, the scoring indicators of different dimensions include: technical indicators, economic indicators and legal indicators; technical indicators include: the number of citations, patent life cycle and the number of patents in the technical field to which the patent belongs; economic indicators include: R&D funds, patent owner type and the market size of the technical field to which the patent belongs; legal indicators include: patent authorization status, patent family size and patent protection history.

[0039] Preferably, the weights of the citation span score and the second score are determined according to the citation span score and the second score of each patent, and the comprehensive score of each patent is obtained by weighted calculation, and the calculation formula is:

[0040] S i =w Y ·Y i +w I I i ,

[0041] Among them, S i is the comprehensive score of the i-th patent, w Y With w I are the citation span scoring weight and the second scoring weight, respectively, i Score the citation span of the i-th patent, w I is the second score of the i-th patent.

[0042] The present invention also provides a personalized recommendation system for university patents, including:

[0043] The data collection and preprocessing module is used to obtain and preprocess university patent data to obtain preprocessed university patent data; the university patent data includes: citation information of different patents, the number of citations, the time of citation, text, patent owner information, and the number of IPC classification numbers;

[0044] The citation network construction module is used to construct a citation network based on the preprocessed university patent data, taking each patent as a node and the citation relationship between patents as an edge;

[0045] Patent theme identification module, which is used to cluster the citation network and divide all patents into multiple theme clusters, each of which represents a technical direction;

[0046] An enterprise text similarity calculation module is used to calculate the text similarity between each patent and each enterprise requirement based on the text of each patent and the text of each enterprise requirement;

[0047] A time difference matrix building module, used to build a time difference matrix based on the time difference between each patent and other patents when it was last cited;

[0048] The citation span score calculation module is used to sum the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent;

[0049] Patent text similarity calculation module, used to calculate the text similarity between each patent and other patents after text vectorization;

[0050] A first score calculation module, used to calculate the first score of each patent based on the number of citations, texts and IPC classification numbers of each patent;

[0051] A second score calculation module, used for calculating a second score of each patent based on the maximum value of text similarities between each patent and other patents and the first score of each patent;

[0052] A comprehensive scoring module is used to determine the weights of the citation span score and the second score of each patent according to the citation span score and the second score, and to obtain the comprehensive score of each patent through weighted calculation;

[0053] The recommendation module is used to generate a patent recommendation list that meets the needs of each enterprise based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent.

[0054] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0055] The method and system for personalized recommendation of university patents described in the present invention construct a time difference matrix through the time difference between each patent and other patents when it was last cited; the elements of the corresponding row of each patent in the time difference matrix are summed to obtain the citation span score of each patent; by analyzing the time when each patent was last cited and the relative citation relationship with other patents, the time difference matrix can quantify the patent's leading position in technological innovation, and in calculating each element of the time difference matrix, the weighted factor of the citation quality and the dynamic time decay factor are added. The weighted factor of the quality can be assigned different weights according to the technical maturity of the patent, so that the calculation of the leading position is more in line with the actual market value of the patent; due to the rapid development of some technical fields, early citations may no longer have the current innovation, and the time decay factor is introduced to gradually reduce the citations over time. The weight can reflect the dynamics of the technology field. By calculating the citation span score of each patent, it can reflect the foresight of the patent in technological innovation and its impact on subsequent technologies. It can accurately evaluate the innovative leadership of a patent, making the personalized promotion of university patents more scientific and accurate. In addition, by determining the citation span score and the weight of each dimension based on the citation span score of each patent and the score in each dimension, the weighted calculation is used to obtain the comprehensive score of each patent. Taking into account multi-dimensional indicators such as technology, economy and law, the comprehensive score of each patent can help universities more accurately identify patents with high transformation potential. This comprehensive evaluation can not only provide data support for the efficient screening of university scientific research results, but also help decision makers formulate more scientific transformation strategies, making the personalized promotion of universities more comprehensive and improving the matching accuracy of patents and corporate needs.

[0056] In addition, the present invention generates accurate input data by decomposing each patent into multiple modules (such as title, abstract, claims, specification, etc.) and performing structured processing on each module. After the information of each module is processed, a feature summary is generated through a set prompt word template. These summaries are more focused on the core technical content of the patent, and these streamlined feature summaries are compared with the enterprise demand text. The matching degree between the patent and the enterprise demand is evaluated by calculating the similarity. Compared with directly converting the patent text into a vector representation, this method can more accurately extract key information related to the enterprise demand, and improve the matching accuracy and calculation efficiency between the patent and the enterprise demand. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0058] Figure 1 It is a flow chart of the steps of a personalized promotion method for university patents of the present invention. DETAILED DESCRIPTION

[0059] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0060] Reference Figure 1 As shown, the first embodiment of the present invention provides a method for personalized recommendation of university patents, including the following steps:

[0061] Step S1: Acquire university patent data and preprocess it to obtain preprocessed university patent data;

[0062] In this embodiment, preferably, the university patent data is mainly obtained from patent databases (such as Zhihuiya and the National Intellectual Property Administration database) that have been updated in the past five years. The university patent data includes: citation information of different patents, the number of citations, the time of citation, the text of the patent, and the patent owner information. The data is screened to ensure the integrity and timeliness of the data, and expired or revoked patents are eliminated.

[0063] In this embodiment, preferably, the university patent data is obtained and preprocessed to obtain the preprocessed university patent data, including:

[0064] After the university patent data is segmented, stop words are removed and morphological restoration is performed in sequence, the data format is formatted to obtain the preprocessed university patent data. The TF-IDF and inverse document frequency methods can be used to ensure the discrimination of the retained keywords in patent matching.

[0065] The TF-IDF method measures the frequency of words in a document by using the term frequency (TF), and the inverse document frequency (IDF) by penalizing common words that appear in most documents, thereby highlighting those representative and uncommon words. This method can assign higher weights to keywords with important information in the text, while suppressing common, non-discriminative words.

[0066] The formula for calculating TF-IDF weight is:

[0067] TF-IDF i,l =TF i,l ×IDF l

[0068]

[0069] Among them, TF i,l is the frequency of the lth word in the i-th patent, f i,l is the number of occurrences of the lth word in the i-th patent, f r,lis the number of occurrences of the rth word in the i-th patent, IDF l is the inverse document frequency of the lth word, R l is the number of patents containing the lth word, TF-IDF i,l is the TF-IDF weight of the lth word in the i-th patent.

[0070] Step S2: Based on the preprocessed university patent data, each patent is used as a node and the citation relationship between patents is used as an edge to construct a citation network;

[0071] Step S3: Cluster the citation network and divide all patents into multiple subject clusters, each of which represents a technical direction;

[0072] In this embodiment, preferably, the citation network is clustered, and the clustering method is any one of a K-means clustering algorithm, a Louvain algorithm, and a hierarchical clustering algorithm.

[0073] Step S4: Based on the text of each patent and the text of each enterprise demand, calculate the text similarity between each patent and each enterprise demand;

[0074] The text similarity between each patent and each enterprise requirement is calculated based on the text of each patent and the text of each enterprise requirement, including:

[0075] Decompose each module of each patent to obtain decomposition information of each module in each patent; the modules include: invention title, abstract, claims, and specification;

[0076] Structuring the decomposition information of each module in each patent to generate structured input data for each module;

[0077] For example:

[0078] Invention Name Module: Question 1: What is the core technical field of the patent?

[0079] Specification Abstract Module: Question 2: Please summarize the technical features in the patent abstract and output it in dot format.

[0080] Claims Module: Question 3: What is the core protection scope involved in the patent claims?

[0081] The structured input data of all modules of each patent are passed through the set prompt word template to generate the prompt words of each patent; the prompt words of each patent are input into the large language model respectively, and the characteristic summary of each patent is output;

[0082] For the patent text module, a step-by-step prompt word template is designed to guide the large model to generate a summary of the patent's technical features and an analysis of its innovation points. For example:

[0083] Tip 1: "Based on the following patent titles, extract the core areas of the patents and summarize the key keywords in point format."

[0084] Prompt 2: "Combining the following technical features, list the three main differences from the existing technology."

[0085] Use multi-round conversational prompts to optimize output quality. According to the semantic characteristics of the patent field, adjust the terminology and expression of prompt words to adapt them to the language habits in the patent text. At the same time, it can guide the big model to generate improvement suggestions based on the parts with insufficient innovation or unclear protection scope, such as:

[0086] Prompt: "Based on the analysis of the innovation points of the following patents, list the possible technical weaknesses of the patents and provide directions for improvement."

[0087] The feature summary of each patent and the text of each enterprise requirement are converted into their corresponding vector representations, and the text similarity between each patent and each enterprise requirement is calculated.

[0088] Through the modular design of structured prompt words, the big model is guided to conduct in-depth analysis of the key elements of the patent text, avoiding the shortcomings of the traditional big model's output results being lengthy or off-topic; with the help of semantic similarity calculation and prompt word optimization, the system can more accurately match university patents with corporate needs and improve the recommendation effect. Combined with prompt word design, it automatically generates patent improvement suggestions and provides directional guidance for university patent transformation.

[0089] In the present invention, specifically, the large language model is any one of GPT-3, PaLM, Galactica and LLaMA.

[0090] In this embodiment, preferably, the text similarity between each patent and each enterprise demand is calculated, and the similarity is any one of cosine similarity, Euclidean distance, and Pearson correlation coefficient.

[0091] The formula for cosine similarity is:

[0092]

[0093] Among them, cos(θ) is the text cosine similarity between the i-th patent and the u-th enterprise demand, A i is the vector representation of the i-th patent text, B u is the vector representation of the u-th enterprise demand text, A i,k is the kth element in the vector representation of the i-th patent text, B u,kis the kth element in the vector representation of the uth enterprise demand text, n represents the length of the vector representation of the ith patent or the uth enterprise demand text, ‖A i ‖ is the modulus of the vector representation of the i-th patent text, ‖B u ‖ is the modulus of the vector representation of the u-th enterprise demand text.

[0094] Step S5: construct a time difference matrix based on the time difference between each patent and other patents when it was last cited;

[0095] In this embodiment, preferably, the calculation formula for each element in the time difference matrix is:

[0096]

[0097] Among them, t ij is the element in the i-th row and j-th column of the time difference matrix, d ij P i With P j The time difference of the last citation, d ij =t j -t i , t j is the time when the jth patent was last cited, t i is the time when the i-th patent was last cited, e is a natural constant, α is the time decay factor, q ij is the quality weighting factor of the i-th patent and the j-th patent, t current is the current time.

[0098] The present invention not only considers the number of times a patent is cited, but also introduces the quality of the cited patents into the calculation. Specifically, if the literature citing a patent itself has a high number of citations or comes from an authoritative source (such as a large enterprise or scientific research institution), the citation should be given a higher weight. Some citations may occur in the early stages of patent application, while others may occur after the technology is more mature. To this end, different weights can be assigned according to the technical maturity of the patent (for example, the citation weight is higher after the patent is granted), so that the leadership calculation is more in line with the actual market value of the patent. q ij Values ​​can be assigned based on factors such as the influence of cited patents and the type of patent holder to enhance the accuracy of leadership calculations.

[0099] By introducing a time decay factor α, the dynamic nature of the technology field is reflected. Some technologies develop rapidly, and early citations may no longer be as innovative as they are today. Therefore, a time decay factor α can be set based on the technology field to which the patent belongs, gradually reducing the weight of the citation over time. For example, for fields with rapid technological updates, a higher decay rate can be set so that citations with a longer citation time contribute less to the leading degree.

[0100] Step S6: summing the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent;

[0101] In this embodiment, preferably, the row elements corresponding to each patent in the time difference matrix are summed to obtain the citation span score of each patent, and the calculation formula is:

[0102]

[0103] Among them, Y i Score the citation span of the i-th patent, t ij is the element in the i-th row and j-th column of the time difference matrix, and n is the total number of patents.

[0104] Step S7: After text vectorization of the texts of each patent and other patents, the text similarity between each patent and other patents is calculated;

[0105] Step S8: Calculate the first score of each patent based on the number of citations, texts and IPC classification codes of each patent;

[0106] S1=ω1·C norm +ω2·N(K)+ω3·S IPC +ω4·Sim,

[0107] Among them, S1 is the first score of the current patent, ω1, ω2, ω3, and ω4 are the weights of the number of citations, the weight of keyword scarcity, the weight of technology branch concentration, and the weight of feature deviation, respectively, and C norm is the normalized value of the number of citations of the current patent, N(K) is the keyword scarcity score of the current patent, μ is the frequency of the keywords of the current patent appearing in other patents, S IPC is the normalized value of the number of IPC classification numbers to which the current patent belongs, Sim = 1-CosineSimilarity(F,G), Sim is the feature deviation between the current patent and the patent that cites it, CosineSimilarity(.) is the cosine similarity, F is the vector representation of the current patent text, and G is the vector representation of the patent text that cites it.

[0108] The first score is used to measure the influence and innovation of the current patent. It reflects the importance of the patent in the academic and technological fields through a weighted evaluation of multiple factors. Specifically, the score combines the normalized value of the number of times the patent is cited (C_norm), the scarcity of keywords (N(K)), the concentration of technical branches (S_IPC), and the characteristic deviation (Sim) from other patents that cite the patent. Among them, the number of citations reflects the influence of the patent, the scarcity of keywords reflects the uniqueness of the patent, the concentration of technical branches indicates the depth of the patent in a specific technical field, and the characteristic deviation reflects the relevance of the patent in related patents. Through the weighted combination of these indicators, the first score provides a quantitative basis for evaluating the innovation, market potential and academic value of the patent, which helps to support decision-making in multiple fields such as technical review, patent screening, and intellectual property management.

[0109] Step S9: Calculate the second score of each patent based on the maximum value of the text similarity between each patent and other patents and the first score of each patent;

[0110]

[0111] Among them, S2 is the second score of the current patent, max(Similarity) is the maximum value of the text similarity between the current patent and other patents, Similarity is the similarity between the vector representation of the text of the current patent and the vector representation of the text of other patents, is the first scoring weight, and S1 is the first scoring of the current patent.

[0112] The second score ensures that the scoring system can effectively identify patents with high differences by maximizing the difference in text similarity between the current patent and other patents.

[0113] Step S10: Determine the weights of the citation span score and the second score according to the citation span score and the second score of each patent, and perform weighted calculation to obtain a comprehensive score for each patent;

[0114] S i =w Y ·Y i +w I I i ,

[0115] Among them, S i is the comprehensive score of the i-th patent, w Y With w I are the citation span scoring weight and the second scoring weight, respectively, i Score the citation span of the i-th patent, w I is the second score of the i-th patent.

[0116] This embodiment preferably further includes: obtaining scores of each dimension, determining weights of the citation span score, the second score and each dimension according to the citation span score, the second score and the score of each dimension of each patent, and performing weighted calculation to obtain an optimized comprehensive score of each patent;

[0117] Among them, the scoring indicators of different dimensions include: technical indicators, economic indicators and legal indicators; technical indicators include: the number of citations, patent life cycle and the number of patents in the technical field to which the patent belongs; economic indicators include: R&D funds, patent owner type and the market size of the technical field to which the patent belongs; legal indicators include: patent authorization status, patent family size and patent protection history.

[0118] Through objective data analysis methods, a comprehensive assessment of university patents is conducted from the three dimensions of technology, economy and law to scientifically measure the transformation potential and innovation value of patents.

[0119] The technical indicators are as follows:

[0120] Number of citations: This indicator reflects the technological influence of a patent. A higher number of citations generally indicates that the technology has greater academic recognition and industry influence.

[0121] Patent life cycle: The market durability of a patent is judged by the application period and maintenance time of the patent. Patents with longer life cycles tend to have continuous market demand.

[0122] The number of patents in the technical field to which the patent belongs: This indicator means that innovation activities in this technical field are active and the market demand is high, or that the technology is advancing rapidly and the competition is fierce.

[0123] The economic indicators are as follows:

[0124] R&D funding: Higher R&D investment usually means more technological innovation and patent output. These patents not only represent technological breakthroughs, but may also bring commercialization opportunities, thereby promoting the company's market competitiveness and economic benefits.

[0125] Patent holder type: Different types of patent holders (such as enterprises, research institutions) indicate the market orientation of the patent. Patents held by enterprises tend to be more commercially applicable.

[0126] Market size of the technical field to which the patent belongs: Analyze the size of the market in which the technology belongs. A larger market size means greater demand potential, more commercialization opportunities for the patent, and more sources of revenue to determine the market appeal of the patent.

[0127] The legal indicators are as follows:

[0128] Patent authorization status: Authorized patents have stronger legal protection capabilities and have lower risks of market transformation than patents that are pending or rejected.

[0129] Patent family size: The size of a patent family shows the layout of patents in different regions and markets. Patents with larger family sizes usually have a wider scope of application.

[0130] Patent protection history: The success of patent protection can be used to evaluate the strength of its legal protection. Patents with a successful protection record usually show higher technical content and market recognition.

[0131] According to the citation span score, the second score and the score in each dimension of each patent, the weights of the citation span score, the second score and each dimension are determined, and the weighted calculation is used to obtain the optimized comprehensive score of each patent. The calculation formula is:

[0132] S i ′ =w T ·T i +w E ·E i +w L ·L i +w Y ·Y i +w I ·Y I ,

[0133] Among them, S i ′ is the comprehensive score of the optimized patent, w Y With w I are the citation span scoring weight and the second scoring weight, respectively, i Score the citation span of the i-th patent, w I is the second score of the i-th patent, T i Score the technical indicators of the i-th patent, E i Score the economic indicators of the i-th patent, L i Score the legal indicators of the i-th patent.

[0134] In this embodiment, any one of the analytic hierarchy process (AHP) and the entropy weight method is used to calculate the citation span score, the second score and the weight of each dimension.

[0135] The entropy weight method determines the weight by calculating the degree of dispersion of each indicator value. The higher the dispersion, the greater the amount of information in the indicator, and its weight increases accordingly. The entropy weight method provides a weight allocation based on the objectivity of the data for each indicator, so that the weight does not depend on subjective judgment.

[0136] The specific steps of using the entropy weight method to calculate the citation span score and the weight of each dimension are as follows:

[0137] The data were standardized so that the score of each patent in each dimension was between 0 and 1 to eliminate the effect of dimensional blocking.

[0138] Calculate the proportion P of the i-th patent in the v-th dimension iv , the calculation formula is:

[0139]

[0140] Among them, P iv is the weight of the i-th patent in the v-th dimension, y iv is the score of the vth dimension of the ith patent, and n is the total number of patents.

[0141] The entropy value of each dimension is calculated according to the entropy formula. The calculation formula is:

[0142]

[0143] Among them, e v is the entropy value of the vth dimension.

[0144] According to the entropy value of each dimension, the weight of each dimension is calculated. The smaller the entropy value, the greater the amount of information, the greater the impact on the comprehensive score, and the higher the weight. The calculation formula is:

[0145] d v =1-e v

[0146]

[0147] Among them, d v is the credit utility value of the vth dimension, w v is the weight of the vth dimension, and V represents the total number of dimensions.

[0148] Step S11: Generate a patent recommendation list that meets the needs of each enterprise based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent.

[0149] In this embodiment, based on the text similarity between each patent and each enterprise demand, a patent-enterprise similarity matrix is ​​constructed for dynamic adjustment. According to real-time changing enterprise needs, market trends or user feedback, the matching between patents and needs is dynamically updated to improve the flexibility and accuracy of the recommendation system.

[0150] Initialize a similarity threshold to filter patent pairs in the similarity matrix that are highly matched with the requirements.

[0151] For example, if the initial similarity threshold is set to 0.8, only pairs with a similarity greater than or equal to 0.8 are retained. The pairs initially screened out are used as the initial results of the recommendation list.

[0152] The system dynamically adjusts based on the following conditions:

[0153] Changes in enterprise requirements: When an enterprise updates its requirements or new requirements enter the system, the similarity between the new requirements and existing patents is recalculated and added to the similarity matrix.

[0154] Changes in market trends: When changes occur in popular technology areas in the market, the vector representation of the patent can be adjusted by increasing or decreasing the weight of specific keywords, thereby affecting the similarity calculation.

[0155] User feedback: Based on user clicks, likes, and detailed information viewing behaviors, the most influential patents are identified and the keyword weights are adjusted accordingly.

[0156] In response to the above demand changes and feedback, the following strategies are adopted to dynamically adjust the similarity matrix:

[0157] Adjust keyword weights: Increase the weights of keywords whose user attention is higher than the set threshold, and recalculate the similarity of related patents that include these keywords.

[0158] Update similarity threshold: Dynamically increase or decrease the threshold based on the overall accuracy and recall requirements of the recommendation system. If there are not enough high-quality patents in the system, the similarity threshold can be appropriately lowered to expand the recommendation scope; if there are too many recalled patents, the similarity threshold can be increased to improve accuracy.

[0159] Real-time update of recommendation list: With the dynamic adjustment of the similarity matrix, a new recommendation list is generated in real time and pushed to the enterprise interface. Patents with higher similarity scores are displayed first to ensure the accuracy and relevance of the recommendation results.

[0160] First, based on the subject clusters corresponding to each patent and the text similarity between each patent and each enterprise demand, patents related to the current enterprise needs are screened out to ensure that the recommendation results are concentrated in relevant technical fields; each patent in the recommendation result set is sorted according to the comprehensive score to generate a personalized recommendation list for the current enterprise.

[0161] This embodiment screens out high-potential patents based on a multi-dimensional scoring system and establishes a hierarchical recommendation mechanism. Patents with a comprehensive score higher than 0.8 are included in the "priority recommendation" list, while patents with a comprehensive score between 0.5-0.8 are included in the "secondary recommendation" list, which facilitates efficient management by university technology transfer departments. By tracking corporate needs and market trends, the system automatically updates the recommendation results. In particular, by introducing the mining and analysis of leading user comments, their latest needs and industry trends are used to provide supplementary data for patent recommendations.

[0162] "Priority recommendation" list (comprehensive score higher than 0.8): usually shows excellent performance in terms of technological innovation, market demand and legal protection. It is recommended to give priority to these patents for transformation and promotion. For such patents, further market research can be combined to verify their commercialization needs, or discuss their transformation potential with potential investors.

[0163] "Sub-recommendation" list (comprehensive score between 0.5-0.8): It is considered that the patents have advantages in one dimension but average performance in other dimensions. For these patents, it is recommended to conduct in-depth analysis to evaluate whether their transformation potential can be enhanced through technical improvement or market promotion.

[0164] For patents with a comprehensive score lower than 0.5, it is believed that these patents have not shown obvious advantages in the three dimensions of technology, economy and law. It is recommended to temporarily shelve the transformation plan of these patents, but the possibility of re-evaluation based on future market or technological changes is not ruled out.

[0165] The present invention not only improves the accuracy of identifying the potential for patent conversion, but also helps improve the success rate of patent conversion in practical applications.

[0166] Through quantified effect data, the present invention demonstrates its significant advantages in identifying patent transformation potential, which can effectively improve identification accuracy, prediction accuracy and resource utilization efficiency, and provide scientific and reliable support for patent management and transformation.

[0167] This Example 2 selects 500 patent data from the public patent database, involving the technical fields of artificial intelligence, material science and new energy. Incomplete patent entries (such as patents with missing citation information) are removed, and 470 patents are finally retained. Each patent contains the number of citations and a list of other patents cited, which is used to construct the basic data of the citation network.

[0168] This example uses the Louvain algorithm to cluster the citation network and divides the 470 patents into five theme clusters. Each theme cluster represents a technical direction, such as "artificial intelligence model optimization", "new material development", "green energy application", etc. In each theme cluster, the TF-IDF algorithm is used to extract high-frequency keywords to represent the core technical content of each theme.

[0169] A multi-dimensional scoring is conducted based on indicators in the three dimensions of technology, economy and law, including 9 indicators such as the number of citations.

[0170] The standardized index value and weighted scoring formula are used to calculate the conversion potential score of each patent. Taking "Patent A" as an example, its technical index (weight 0.35) scores 0.9, economic index (weight 0.25) scores 0.8, legal index (weight 0.15) scores 0.85, citation span score (weight 0.15) scores 0.7, and the second score (weight 0.1) scores 0.8.

[0171] The optimized comprehensive score S ′ for:

[0172] S ′ =(0.35×0.9)+(0.25×0.8)+(0.15×0.85)+(0.15×0.7)+(0.1×0.8)

[0173] =0.8275

[0174] The third embodiment also provides a personalized recommendation system for university patents, including:

[0175] The data collection and preprocessing module is used to obtain and preprocess university patent data to obtain preprocessed university patent data;

[0176] The citation network construction module is used to construct a citation network based on the preprocessed university patent data, taking each patent as a node and the citation relationship between patents as an edge;

[0177] Patent theme identification module, which is used to cluster the citation network and divide all patents into multiple theme clusters, each of which represents a technical direction;

[0178] The text similarity calculation module is used to input the text of each patent and the text of each enterprise demand into the BERT model, output their corresponding vector representations, and calculate the text similarity between each patent and each enterprise demand;

[0179] A time difference matrix building module, used to build a time difference matrix based on the time difference between each patent and other patents when it was last cited;

[0180] The citation span score calculation module is used to sum the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent;

[0181] The scoring module is used to score all patents according to the scoring indicators of different dimensions and obtain the score of each patent in each dimension;

[0182] The multi-dimensional comprehensive evaluation module is used to determine the weights of the citation span score and each dimension based on the citation span score of each patent and the score in each dimension, and to obtain the comprehensive score of each patent through weighted calculation;

[0183] The recommendation module is used to generate a patent recommendation list that meets the needs of each enterprise based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent.

[0184] The data collection and preprocessing module is used to obtain and preprocess university patent data to obtain preprocessed university patent data; the university patent data includes: citation information of different patents, the number of citations, the time of citation, text, patent owner information, and the number of IPC classification numbers;

[0185] The citation network construction module is used to construct a citation network based on the preprocessed university patent data, taking each patent as a node and the citation relationship between patents as an edge;

[0186] Patent theme identification module, which is used to cluster the citation network and divide all patents into multiple theme clusters, each of which represents a technical direction;

[0187] An enterprise text similarity calculation module is used to calculate the text similarity between each patent and each enterprise requirement based on the text of each patent and the text of each enterprise requirement;

[0188] A time difference matrix building module, used to build a time difference matrix based on the time difference between each patent and other patents when it was last cited;

[0189] The citation span score calculation module is used to sum the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent;

[0190] Patent text similarity calculation module, used to calculate the text similarity between each patent and other patents after text vectorization;

[0191] A first score calculation module, used to calculate the first score of each patent based on the number of citations, texts and IPC classification numbers of each patent;

[0192] A second score calculation module, used for calculating a second score of each patent based on the maximum value of text similarities between each patent and other patents and the first score of each patent;

[0193] A comprehensive scoring module is used to determine the weights of the citation span score and the second score of each patent according to the citation span score and the second score, and to obtain the comprehensive score of each patent through weighted calculation;

[0194] The recommendation module is used to generate a patent recommendation list that meets the needs of each enterprise based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent.

[0195] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0196] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0197] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0198] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0199] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A personalized recommendation method for university patents, characterized in that: include: Obtain university patent data and preprocess it to obtain preprocessed university patent data; wherein the university patent data includes: citation information of different patents, the number of citations, the time of citation, text, patent owner information, and the number of IPC classification numbers; Based on the preprocessed university patent data, each patent is taken as a node and the citation relationship between patents as an edge to construct a citation network; Cluster the citation network and divide all patents into multiple subject clusters, each of which represents a technical direction; Based on the text of each patent and the text of each enterprise demand, calculate the text similarity between each patent and each enterprise demand; Construct a time difference matrix based on the time difference between each patent and other patents when it was last cited; Sum the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent; After text vectorization of the texts of each patent and other patents, the text similarity between each patent and other patents is calculated; Calculate the first score for each patent based on the number of citations, texts and IPC classification codes of each patent; Calculate the second score of each patent based on the maximum value of the text similarity between each patent and other patents and the first score of each patent; According to the citation span score and the second score of each patent, the weights of the citation span score and the second score are determined, and the comprehensive score of each patent is obtained by weighted calculation; Based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent, a patent recommendation list that meets the needs of each enterprise is generated.

2. The personalized recommendation method for university patents according to claim 1 is characterized in that: The calculation formula for each element in the time difference matrix is: Among them, t ij is the element in the i-th row and j-th column of the time difference matrix, d ij For the i-th patent P i and patent P j The time difference of the last citation, d ij =t j -t i , t j is the time when the jth patent was last cited, t i is the time when the i-th patent was last cited, e is a natural constant, α is the time decay factor, q ij is the quality weighting factor of the i-th patent and the j-th patent, t current is the current time.

3. The personalized recommendation method for university patents according to claim 2 is characterized in that: The row elements corresponding to each patent in the time difference matrix are summed to obtain the citation span score of each patent. The calculation formula is: Among them, Y i is the citation span score of the i-th patent, n is the total number of patents, t ij is the element in the i-th row and j-th column of the time difference matrix.

4. The personalized recommendation method for university patents according to claim 1 is characterized in that: The first score of each patent is calculated based on the number of citations, texts and IPC classification numbers of each patent. The calculation formula is: S1=ω1·C norm +ω2·N(K)+ω3·S IPC +ω4·Sim, Among them, S1 is the first score of the current patent, ω1, ω2, ω3, and ω4 are the weights of the number of citations, the weight of keyword scarcity, the weight of technology branch concentration, and the weight of feature deviation, respectively, and C norm is the normalized value of the number of citations of the current patent, N(K) is the keyword scarcity score of the current patent, μ is the frequency of the keywords of the current patent appearing in other patents, S IPC is the normalized value of the number of IPC classification numbers to which the current patent belongs, Sim = 1-CosineSimilarity(F,G), Sim is the feature deviation between the current patent and the patent that cites it, CosineSimilarity(.) is the cosine similarity, F is the vector representation of the current patent text, and G is the vector representation of the patent text that cites it.

5. The personalized recommendation method for university patents according to claim 4 is characterized in that: Based on the maximum value of the text similarity between each patent and other patents and the first score of each patent, the second score of each patent is calculated using the following formula: Among them, S2 is the second score of the current patent, max(Similarity) is the maximum value of the text similarity between the current patent and other patents, Similarity is the similarity between the vector representation of the current patent text and the vector representation of other patent texts, is the first scoring weight, and S1 is the first scoring of the current patent.

6. The personalized recommendation method for university patents according to claim 4 is characterized in that: Obtain university patent data and preprocess it to obtain preprocessed university patent data, including: After word segmentation, stop word removal and word form restoration, the university patent data is formatted to obtain the preprocessed university patent data.

7. The personalized recommendation method for university patents according to claim 1 is characterized in that: The text similarity between each patent and each enterprise requirement is calculated based on the text of each patent and the text of each enterprise requirement, including: Decompose each module of each patent to obtain decomposition information of each module in each patent; the modules include: invention title, abstract, claims, and specification; Structuring the decomposition information of each module in each patent to generate structured input data for each module; The structured input data of all modules of each patent are passed through the set prompt word template to generate the prompt words of each patent; the prompt words of each patent are input into the large language model respectively, and the characteristic summary of each patent is output; The feature summary of each patent and the text of each enterprise requirement are converted into their corresponding vector representations, and the text similarity between each patent and each enterprise requirement is calculated.

8. The personalized recommendation method for university patents according to claim 1 is characterized in that: Also includes: Obtain the scores of each dimension, determine the weights of the citation span score, the second score and each dimension according to the citation span score, the second score and the scores in each dimension of each patent, and perform weighted calculation to obtain the optimized comprehensive score of each patent; Among them, the scoring indicators of different dimensions include: technical indicators, economic indicators and legal indicators; technical indicators include: the number of citations, patent life cycle and the number of patents in the technical field to which the patent belongs; economic indicators include: R&D funds, patent owner type and the market size of the technical field to which the patent belongs; legal indicators include: patent authorization status, patent family size and patent protection history.

9. The personalized recommendation method for university patents according to claim 1 is characterized in that: According to the citation span score and the second score of each patent, the weights of the citation span score and the second score are determined, and the comprehensive score of each patent is obtained by weighted calculation. The calculation formula is: S i =w Y ·Y i +w I ·I i , Among them, S i is the comprehensive score of the i-th patent, w Y With w I are the citation span scoring weight and the second scoring weight, respectively, i Score the citation span of the i-th patent, w I is the second score of the i-th patent.

10. A personalized recommendation system for university patents, characterized in that: include: The data collection and preprocessing module is used to obtain and preprocess university patent data to obtain preprocessed university patent data; the university patent data includes: citation information of different patents, the number of citations, the time of citation, text, patent owner information, and the number of IPC classification numbers; The citation network construction module is used to construct a citation network based on the preprocessed university patent data, taking each patent as a node and the citation relationship between patents as an edge; Patent theme identification module, which is used to cluster the citation network and divide all patents into multiple theme clusters, each of which represents a technical direction; An enterprise text similarity calculation module is used to calculate the text similarity between each patent and each enterprise requirement based on the text of each patent and the text of each enterprise requirement; A time difference matrix building module, used to build a time difference matrix based on the time difference between each patent and other patents when it was last cited; The citation span score calculation module is used to sum the elements of the corresponding row of each patent in the time difference matrix to obtain the citation span score of each patent; Patent text similarity calculation module, used to calculate the text similarity between each patent and other patents after text vectorization; A first score calculation module, used to calculate the first score of each patent based on the number of citations, texts and IPC classification numbers of each patent; A second score calculation module, used for calculating a second score of each patent based on the maximum value of text similarities between each patent and other patents and the first score of each patent; A comprehensive scoring module is used to determine the weights of the citation span score and the second score of each patent according to the citation span score and the second score, and to obtain the comprehensive score of each patent through weighted calculation; The recommendation module is used to generate a patent recommendation list that meets the needs of each enterprise based on the subject cluster corresponding to each patent, the text similarity between each patent and each enterprise's needs, and the comprehensive score of each patent.

Images