A privacy data desensitization method for a scientific and technological achievement transformation platform
By using BERT word segmentation and semantic clustering to construct a generalized hierarchical tree in a technology transfer platform, and dynamically adjusting the K value in combination with sensitivity and rarity indicators, the problem of over-generalization or under-generalization in the traditional K-anonymity algorithm during generalization desensitization is solved, thus achieving efficient privacy data protection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INFOSOFT CO LTD
- Filing Date
- 2025-12-09
- Publication Date
- 2026-07-07
AI Technical Summary
Traditional K-anonymization algorithms struggle to identify semantic information between different keywords in technology transfer platforms, leading to overgeneralization or undergeneralization during desensitization, thus failing to effectively protect privacy data.
The BERT word segmentation model is used for text segmentation and keyword extraction. A generalized hierarchical tree is constructed through semantic clustering. Data privacy indicators are constructed by combining the sensitivity and rarity of keywords. The K value of the K-anonymity algorithm is dynamically adjusted to perform adaptive generalization desensitization.
This approach protects privacy while preventing data over-generalization and information loss, ensuring the validity and usability of the data.
Smart Images
Figure CN121659360B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data security technology, specifically to a method for de-identifying privacy data for technology transfer platforms. Background Technology
[0002] The digital economy has developed rapidly in recent years, and digital information data, as a core element, is becoming increasingly important in areas such as enterprise operations and the protection of scientific and technological achievements. However, due to platform management issues, the problem of privacy data leakage during the sharing and transfer of information data within systems is becoming increasingly serious. This is especially true in technology transfer platforms, which involve confidential and private data from numerous research institutions, enterprises, investors, and other parties. Leakage of such data could lead to competitive risks and legal liabilities. Therefore, protecting the security of this private information is crucial. While multi-level data encryption is effective in improving data security, simple encryption inevitably impacts the efficiency of data retrieval and usage.
[0003] Data anonymization involves masking and perturbing sensitive fields in data, preserving its business value while making it difficult to reverse engineer, thus providing a powerful method for data security. Among data anonymization methods, the K-anonymization algorithm, a commonly used technique, is widely applied for data anonymization and privacy protection during the data release phase. However, in the diverse and complex information data of technology transfer platforms, traditional K-anonymization algorithms, which group data based on numerical values and categories, struggle to identify data with similar semantics. This leads to a lack of effective equivalence levels in the grouped data, making it difficult to set a suitable k value for generalized anonymization. Furthermore, the K-anonymization algorithm's reliance on a fixed k value to control generalization strength can result in over-generalization or under-generalization of data with varying sensitivities within complex information, leading to significant information loss or ineffective privacy protection after anonymization. Summary of the Invention
[0004] To address the aforementioned technical problems, this application provides a privacy data anonymization method for technology transfer platforms, thereby resolving the existing issues.
[0005] This application proposes a privacy data desensitization method for technology transfer platforms, which adopts the following technical solution:
[0006] One embodiment of this application provides a method for de-identifying privacy data for a technology transfer platform, the method comprising the following steps:
[0007] Step 1: Collect scientific and technological achievement data from the technology transfer platform, and perform word segmentation and keyword extraction on the text content;
[0008] Step 2: Perform semantic clustering on keywords based on their feature vectors to construct a generalized hierarchical tree;
[0009] Step 3: Determine the sensitivity of keywords based on the number of nodes traversed by the shortest reachable path between corresponding nodes in the generalization hierarchy tree and the co-occurrence frequency of keywords; determine the rarity of keywords based on the ratio of the inverse document frequency of a keyword to the maximum inverse document frequency in its cluster; and obtain data privacy by weighted fusion of keyword sensitivity and rarity.
[0010] Step 4: Divide the data into equal-length intervals according to the data privacy index, and assign an adaptive K value to each group when generalizing using the K-anonymity algorithm, so as to perform generalized desensitization processing on the keywords in the scientific and technological achievement data.
[0011] Preferably, the word segmentation uses the BERT word segmentation model to process each sentence of the text content, and outputs all sub-words in the text and their semantic feature vectors.
[0012] Preferably, the keyword extraction uses the TF-IDF method to process all sub-words of the scientific and technological achievement data, and extracts the first preset number of keywords from each scientific and technological achievement data.
[0013] Preferably, the semantic clustering adopts a hierarchical clustering algorithm; when the number of keywords in both clusters is 1, the distance between the clusters is determined by the cosine similarity of the feature vectors between the two keywords in the two clusters; otherwise, the distance between the clusters is calculated using the Average-Link method.
[0014] Preferably, the feature vector of the keyword is determined by the mean of the semantic feature vectors of all corresponding sub-words after the text content is segmented.
[0015] Preferably, the method for constructing the generalization hierarchical tree is as follows: each cluster in the clustering results is connected to an external knowledge base to obtain the hypernym of each keyword within the cluster, so as to construct a generalization hierarchical tree for each cluster.
[0016] Preferably, the data privacy is a weighted fusion of sensitivity and rarity, with the sum of the weighting coefficients being 1.
[0017] Preferably, the sensitivity is calculated based on the sensitivity distance and co-occurrence frequency between keywords, and the sensitivity distance is measured by the number of nodes traversed by the shortest reachable path of the corresponding node in the generalization hierarchy tree.
[0018] Preferably, the adaptive K value is dynamically set according to the group to which the keyword belongs. The higher the data privacy division interval of the group, the larger the corresponding K value, so as to enhance the generalization strength.
[0019] Preferably, during the generalized desensitization process, special identifier data is deleted or masked, and keyword data is generalized based on a generalized hierarchical tree and an adaptive K value.
[0020] This application has at least the following beneficial effects:
[0021] This application addresses the problems of overgeneralization or undergeneralization when using traditional K-anonymization algorithms to generalize and desensitize data from technology transfer platforms. These problems stem from the algorithm's difficulty in identifying semantic information between different keywords and its reliance on a fixed k-value. This application analyzes the data characteristics of each keyword in the technology transfer data and performs semantic clustering to construct a generalization hierarchy tree. This results in a higher equivalent hierarchy for the grouped data, providing a basis for subsequent generalization. Then, a data privacy index is constructed based on the sensitivity and rarity characteristics of key information data to characterize the privacy level of each keyword. Finally, the adaptive k-value is dynamically adjusted based on the data privacy index when using the K-anonymization algorithm for generalization, achieving better generalization and desensitization. This ensures the privacy of high-risk sensitive data in the technology transfer data while avoiding overgeneralization that could lead to data loss. Attached Figure Description
[0022] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 A flowchart illustrating a privacy data desensitization method for a technology transfer platform provided in this application. Detailed Implementation
[0024] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a privacy data desensitization method for a technology transfer platform proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0026] The following description, in conjunction with the accompanying drawings, details a specific scheme for a privacy data desensitization method for technology transfer platforms provided in this application.
[0027] One embodiment of this application provides a method for de-identifying privacy data for a technology transfer platform.
[0028] Specifically, the following method for de-identifying privacy data for technology transfer platforms is provided. Please refer to [link / reference]. Figure 1 The method includes the following steps:
[0029] Step 1: Collect scientific and technological achievement data from the technology transfer platform, and perform word segmentation and keyword extraction on the text content.
[0030] This application acquires scientific and technological achievement data through a technology transfer platform. To facilitate subsequent processing and analysis of the acquired data, this application employs a word segmentation model to segment the text content of the data. The choice of word segmentation model can be determined by the implementer based on the implementation scenario, without special restrictions. In this embodiment, the BERT word segmentation model is used to process the data, splitting the text content of the scientific and technological achievement data into sub-words, each sub-word having a corresponding semantic feature vector. The input to the BERT word segmentation model is a single text sentence, and the output is the segmented text sentence, including all sub-words in the sentence and their semantic feature vector representations. The BERT word segmentation model is existing technology, and its specific segmentation process will not be elaborated further.
[0031] Regarding the acquired scientific and technological data, let's assume its quantity is... , of which Each scientific and technological achievement data is represented as For scientific and technological achievement data after word segmentation, the term frequency-inverse document frequency (TF-IDF) method is used to process all sub-words of the scientific and technological achievement data to obtain the keywords of the scientific and technological achievement data. The TF-IDF method is an existing technology, and its specific process will not be described in detail.
[0032] The keywords corresponding to each scientific and technological achievement data are represented as follows: ,in Indicates the first The first of the scientific and technological achievements One keyword, The number of keywords extracted from the scientific and technological achievement data can be set by the implementer according to the implementation scenario, without any special restrictions. In this embodiment... The size is 20.
[0033] Specifically, special identifier data in scientific and technological achievement data will not be processed using the TF-IDF method. Examples include patent numbers, patent names, inventors, inventor's affiliated institutions, and addresses in patent data; and company names, addresses, return on investment, and revenue figures in corporate annual report data. A total of five types of special identifier data will be extracted. The quantity of this data can be set by the implementer based on the data types in the implementation scenario, without any special restrictions. The special identifier data refers to identifiers in the K-anonymity algorithm, while the keyword data is a set of quasi-identifiers in the K-anonymity algorithm.
[0034] Step 2: Perform semantic clustering on keywords based on their feature vectors to construct a generalized hierarchical tree.
[0035] K-anonymity, a fundamental technology in the field of privacy protection, aims to conceal individual data within published data, preventing the leakage of sensitive personal information after data disclosure. However, in scientific and technological achievements data, many data points involve similar fields but differ in their textual representations. This leads traditional K-anonymity algorithms to group and generalize anonymization based solely on numerical differences. When processing massive amounts of scientific and technological achievements data, specialized data in certain fields may exhibit extensive generalization levels. Furthermore, traditional K-anonymity algorithms use a fixed k value globally for generalization and anonymization. This creates a conflict between the generalization levels of specialized scientific and technological data and highly unique data. Using a high k value can lead to overgeneralization of highly unique data, resulting in significant data distortion and information loss; conversely, a low k value can lead to insufficient generalization of some data, failing to effectively anonymize it.
[0036] To address the aforementioned issues, this application performs semantic clustering of key information data in all scientific and technological achievement data by analyzing the data features of each sub-word and its feature vector in the data. This results in the grouped data having a higher equivalence level, facilitating subsequent generalization processing. Then, a data privacy index is constructed based on the sensitivity and rarity characteristics of the key information data to characterize the privacy level of each keyword data. Finally, the k-value is dynamically adjusted based on the data privacy index to achieve better generalization and desensitization. This approach ensures the privacy of high-risk sensitive data in the scientific and technological achievement data while avoiding the loss of data validity due to excessive generalization.
[0037] Technology transfer platforms often involve information on scientific and technological achievements from many disciplines and fields, enabling the integration of information between multiple technology transfer companies and investment institutions. The cross-combination of keywords across many fields results in multiple keywords within the same technical field. Traditional K-anonymity algorithms group keywords based solely on string similarity, failing to identify semantic similarity between different technical fields. This can easily lead to the classification of these keywords into different equivalence levels, resulting in overly detailed keyword category distinctions in some generalized groups. Consequently, it becomes difficult to determine a reasonable k value for generalization desensitization based on the generalization level of each group.
[0038] Specifically, semantic clustering is performed based on the contextual semantic feature vectors of keywords extracted from various scientific and technological achievements data, thereby enabling reasonable generalization and grouping of keywords in a massive amount of technical fields according to their semantic information.
[0039] The feature vector of each keyword is obtained by analyzing the semantic feature vector of its corresponding sub-word in the original segmented text data of each scientific and technological achievement. The first scientific and technological achievement data The feature vector of each keyword is denoted as... , The specific representation is as follows:
[0040] for In the A total of [number] scientific and technological achievements data appeared. Secondly, it appeared in the data of scientific and technological achievements. The semantic feature vector of the sub-word corresponding to the next time is ,So Repeat this process to obtain the feature vectors of all keywords in all scientific and technological achievements.
[0041] Furthermore, each keyword in each scientific and technological achievement data is treated as an independent cluster, resulting in a total of There are several clusters; then, the semantic distance between each cluster is calculated using a distance metric function, which is: ,in , They represent the first The first of the scientific and technological achievements data The first keyword, the first Feature vectors of each keyword , Here is the formula for calculating cosine similarity; each time, the two closest clusters are found and merged into a new cluster. When a cluster contains multiple keywords, the Average-Link method is used to calculate the semantic distance between the keywords in the two clusters; the target number of clusters is... Its size can be set by the implementer according to the number of professional fields in the implementation scenario. In this embodiment, The value is 500, when the number of clusters in the clustering result reaches... Clustering ends at this point. Hierarchical clustering is an existing technique, and its specific process will not be elaborated further. Final output Each cluster contains several semantically similar keywords.
[0042] Furthermore, a knowledge base in the field of Natural Language Inference (NLI) is integrated into the platform, including but not limited to open-source knowledge bases such as ConceptNet and Wikipedia. Then, based on the prior knowledge in the knowledge base, the hypernyms of each keyword are found, ultimately obtaining the generalization hierarchy tree of each cluster, which serves as the generalization basis for subsequent keyword desensitization.
[0043] The generalized grouping obtained in this way considers the contextual semantic features of keywords in various scientific and technological achievement data for clustering, avoiding the problem of over-generalization or under-generalization when generalizing with some highly unique data by relying solely on string similarity, and improving the stability and usability of subsequent data desensitization.
[0044] Step 3: Determine the keyword sensitivity based on the number of nodes traversed by the shortest reachable path between corresponding nodes in the generalization hierarchy tree and the co-occurrence frequency between keywords; determine the keyword rarity based on the ratio of the keyword's inverse document frequency to the maximum inverse document frequency in its cluster; and obtain the data privacy by weighted fusion of keyword sensitivity and rarity.
[0045] Traditional K-anonymity algorithms use a globally uniform fixed k value for generalization. This can effectively protect user privacy in simple scenarios such as medical data. However, in technology transfer platforms, where the data volume is large and the categories are complex, a globally uniform k value can easily lead to overgeneralization of some highly unique data, resulting in significant data loss. After generalization, the data may not be usable or may be insufficiently generalized. User information can still be located based on the generalized terms, making it impossible to effectively desensitize the data for privacy.
[0046] Furthermore, this application uses the characteristics of keywords in scientific and technological achievement data to analyze their sensitivity and rarity in order to construct data privacy indicators.
[0047] In scientific and technological achievements, attempting to intersect technologies or themes from different fields can yield new insights into solving technical challenges. In scientific and technological achievement data, this cross-combination can be reflected in keyword similarity and co-occurrence frequency. When two semantically very different keywords co-occur frequently, it usually indicates that the combination of these two technical fields constitutes the main content of the scientific and technological achievement. When these two keywords are combined, because they represent emerging scientific and technological achievements and are easily retrieved for specific sources, their sensitivity is high. Therefore, stronger generalization and desensitization are needed for such keywords to prevent the leakage of privacy information. Thus, keyword sensitivity can serve as one indicator for adjusting the degree of generalization and desensitization, providing stronger privacy data desensitization for emerging and sensitive scientific and technological achievement keywords to protect their privacy and security.
[0048] In summary, the first The first of the scientific and technological achievements data The sensitivity of each keyword is expressed as follows:
[0049]
[0050]
[0051]
[0052] in Indicates the first The first of the scientific and technological achievements data The and the first Sensitive distance between keywords Indicates the first The first of the scientific and technological achievements data The and the first The number of nodes traversed by the shortest reachable path of each keyword to its corresponding node in the generalization hierarchy tree. The function is a logarithmic function, used to scale the distance of keywords in the generalization hierarchy tree, so as to avoid the value being too large and having too much impact on sensitivity. Indicates the first The first of the scientific and technological achievements data The and the first Co-occurrence sensitivity index among keywords Indicates the first The first of the scientific and technological achievements data The and the first The co-occurrence frequency of each keyword in the original text is calculated by counting whether they appear simultaneously in each paragraph. To conduct statistics, Indicates the first The first of the scientific and technological achievements data The total frequency of each keyword. Indicates the first The first of the scientific and technological achievements data The total frequency of each keyword. This is a function that takes the minimum value. Indicates the first The first of the scientific and technological achievements data Sensitivity to each keyword.
[0053] When the The and the first The greater the distance between two keywords in the generalization hierarchy tree, the more likely they are keywords from two different technical fields. The larger it is. Then it means the first and the The co-occurrence frequency of two keywords represents the percentage of the frequency of any one of them that appears least frequently, indicating the sensitivity of the relationship between the two keywords. The larger the value, the more it indicates the number of... and the These keywords are highly relevant in the data on scientific and technological achievements.
[0054] Final statistics The first of the scientific and technological achievements data Keyword and others The combination of sensitive distance and co-occurrence sensitivity index among the keywords is used to represent the first... The first of the scientific and technological achievements data Sensitivity to each keyword. When When the number of occurrences is large, the co-occurrence frequency of these two keywords in the text should theoretically be low, but it is affected by... The high frequency of co-occurrence of these two keywords suggests a possible fusion of technological innovations from different technical fields within the scientific and technological achievement data. In the data of scientific and technological achievement platforms, such data with novel keyword combinations is more easily retrieved. and Under the combined effect, The larger it is, the more it indicates the first... The higher the sensitivity of a keyword, the stronger the subsequent generalization and desensitization required. Conversely, when... When the frequency is large, the co-occurrence frequency of these two keywords is very low, that is... It is also relatively small, even approaching 0, and ultimately The smaller the value, the lower the sensitivity of the keyword. When the frequency is low, it indicates that the two keywords belong to the same field and are semantically similar. Even if they co-occur frequently, this is a normal phenomenon in scientific and technological achievement data, and ultimately its sensitivity will be affected. It won't be too high either.
[0055] Furthermore, the rarity of the keyword itself is also one of the indicators of the keyword's privacy. In the above hierarchical clustering results, a cluster represents some keywords with similar semantics, which can be understood as these keywords belonging to the same technical field. If a certain keyword appears very rarely in all scientific and technological achievement data in this field, it indicates that the keyword has a high rarity in this technical field, is easy to expose its source, and requires further generalization and desensitization.
[0056] Based on the above analysis, the first... The first of the scientific and technological achievements data The rarity of a keyword can be represented as:
[0057]
[0058] in, Indicates the first The first of the scientific and technological achievements data Inverse document frequency of each keyword, , This indicates the total number of scientific and technological achievements data. Indicates the presence of keywords The number of documents. Assuming... Belongs to the hierarchical clustering of the first In each cluster, There are a total of Based on the clustering results, the keywords in this cluster are related to... If they belong to the same technical field, then Keywords The first The rarity is calculated by finding the maximum inverse document frequency of all keywords within a cluster. The smaller the value, the higher the identifiability; if a keyword only appears in a few specific scientific and technological achievements, it is easily identifiable and its source can be located through the word, but the privacy risk is relatively high. Conversely, a larger value indicates a higher identifiability. The larger the value, the more frequently the keyword appears in a large amount of scientific and technological data, indicating that it is a common term and poses a lower risk to privacy.
[0059] Furthermore, According to the All of the scientific and technological achievements data The sensitivity of each keyword is normalized by taking its maximum and minimum values to obtain the normalized sensitivity. The maximum and minimum values are normalized to the existing technology, and the specific process will not be elaborated here.
[0060] No. The first of the scientific and technological achievements data The data privacy of a keyword can be represented as:
[0061]
[0062] in , The first The first of the scientific and technological achievements data The normalized sensitivity and rarity preset weight parameters of each keyword satisfy... Its size can be set by the implementer according to the implementation scenario, without special restrictions. In this embodiment, , The values are set to 0.6 and 0.4 respectively. This constructed data privacy index can be comprehensively represented based on the sensitivity and rarity of a keyword. Keywords with higher sensitivity and rarity have higher privacy, requiring more generalized desensitization to prevent privacy leaks; conversely, keywords with lower sensitivity and rarity have lower privacy. The smaller the value, the lower the possibility of data privacy leaks. Therefore, it's important to avoid overgeneralizing these keywords, which could lead to data loss and loss of usability.
[0063] Repeat the above steps to obtain the data privacy index for all keywords.
[0064] Step 4: Divide the data into equal-length intervals according to the data privacy index, and assign an adaptive K value to each group when generalizing using the K-anonymity algorithm, so as to perform generalized desensitization processing on the keywords in the scientific and technological achievement data.
[0065] Grouped according to the data privacy index of all keywords, a total of Different keyword groups , The size can be set by the implementer according to the implementation scenario, without special restrictions. In this embodiment... The value is 5. The data privacy index described in this application ranges from (0,1), so when grouping, equal-length intervals are divided according to the data privacy index value, as follows: , , , , .
[0066] The adaptive k-value in the K-anonymity algorithm for generalizing to different keyword groups is expressed as:
[0067]
[0068] in This is the initial value for anonymity in the K-anonymity algorithm. Its size can be set by the implementer according to the implementation scenario, without special restrictions. In this embodiment, The value is 3.
[0069] Finally, special identifier data in the acquired scientific and technological achievement data is directly hidden by deletion or masking operations in the K-anonymity algorithm. Keyword data in the scientific and technological achievement data is desensitized by the generalized hierarchical tree and adaptive k value constructed above. The data desensitization process of the K-anonymity algorithm is an existing technology, and its specific process will not be described in detail.
[0070] This completes the desensitization of privacy data on the technology transfer platform.
[0071] The above technical features constitute the preferred embodiment of this application, which has strong adaptability and the best implementation effect. Unnecessary technical features can be added or removed according to actual needs to meet the needs of different situations.
Claims
1. A method for desensitizing privacy data for technology transfer platforms, characterized in that, The method includes the following steps: Step 1: Collect scientific and technological achievement data from the technology transfer platform, and perform word segmentation and keyword extraction on the text content; Step 2: Perform semantic clustering on keywords based on their feature vectors to construct a generalized hierarchical tree; Step 3: Determine the sensitivity of keywords based on the number of nodes traversed by the shortest reachable path between corresponding nodes in the generalization hierarchy tree and the co-occurrence frequency of keywords; determine the rarity of keywords based on the ratio of the inverse document frequency of a keyword to the maximum inverse document frequency in its cluster; and obtain data privacy by weighted fusion of keyword sensitivity and rarity. Step 4: Divide the data into equal-length intervals according to the data privacy index, and assign an adaptive K value to each group when using the K-anonymity algorithm for generalization, so as to perform generalized desensitization processing on the keywords in the scientific and technological achievement data. Among them, the The first of the scientific and technological achievements data The sensitivity of each keyword is expressed as follows: in Indicates the first The first of the scientific and technological achievements data The and the first Sensitive distance between keywords Indicates the first The first of the scientific and technological achievements data The and the first The number of nodes traversed by the shortest reachable path of each keyword to its corresponding node in the generalization hierarchy tree. It is a logarithmic function; Indicates the first The first of the scientific and technological achievements data The and the first Co-occurrence sensitivity index among keywords Indicates the first The first of the scientific and technological achievements data The and the first The co-occurrence frequency of each keyword in the original text, Indicates the first The first of the scientific and technological achievements data The total frequency of each keyword. Indicates the first The first of the scientific and technological achievements data The total frequency of each keyword. The function is for finding the minimum value; Indicates the first The first of the scientific and technological achievements data Sensitivity to each keyword; The number of keywords extracted from scientific and technological achievement data.
2. The privacy data anonymization method for technology transfer platforms as described in claim 1, characterized in that, The word segmentation uses the BERT word segmentation model to process each sentence of the text content and outputs all sub-words in the text and their semantic feature vectors.
3. The privacy data anonymization method for a technology transfer platform as described in claim 2, characterized in that, The keyword extraction method uses the TF-IDF method to process all sub-words of the scientific and technological achievement data and extract the first preset number of keywords from each scientific and technological achievement data.
4. The privacy data anonymization method for technology transfer platforms as described in claim 1, characterized in that, The semantic clustering adopts a hierarchical clustering algorithm; when the number of keywords in two clusters is 1, the distance between the clusters is determined by the cosine similarity of the feature vectors between the two keywords in the two clusters; otherwise, the distance between the clusters is calculated using the Average-Link method.
5. A method for desensitizing privacy data for a technology transfer platform as described in claim 4, characterized in that, The feature vector of the keyword is determined by the mean of the semantic feature vectors of all corresponding sub-words after the text content is segmented.
6. The privacy data anonymization method for a technology transfer platform as described in claim 1, characterized in that, The method for constructing the generalization hierarchical tree is as follows: each cluster in the clustering results is connected to an external knowledge base, and the hypernyms of each keyword in the cluster are obtained to construct the generalization hierarchical tree for each cluster.
7. The privacy data anonymization method for a technology transfer platform as described in claim 1, characterized in that, The data privacy is a weighted fusion of sensitivity and rarity, with the sum of the weighting coefficients being 1.
8. A method for desensitizing privacy data for a technology transfer platform as described in claim 7, characterized in that, The sensitivity is calculated based on the sensitivity distance and co-occurrence frequency between keywords. The sensitivity distance is measured by the number of nodes traversed by the shortest reachable path of the corresponding node in the generalization hierarchy tree.
9. A method for desensitizing privacy data for a technology transfer platform as described in claim 1, characterized in that, The adaptive K value is dynamically set according to the group to which the keyword belongs. The higher the data privacy division interval of the group, the larger the corresponding K value, so as to enhance the generalization strength.