AI intelligent recommendation method and system applied to online patent operation digital platform

Abstract

The invention provides an AI intelligent recommendation method and system applied to an online patent operation digital platform, and the method comprises the steps: obtaining user full-amount behavior data and patent full-period attribute data, carrying out the demand intention analysis and value dimension modeling, generating a user demand intention feature set and a patent value dimension feature set, and carrying out the calculation of the user demand intention feature set and the patent value dimension feature set; and on this basis, an AI intelligent recommendation model is used to carry out multi-dimensional association matching on the two, an initial recommendation candidate set is generated, dynamic adjustment is carried out based on real-time interaction information of the user, and finally an optimized patent recommendation sequence is generated and output to a user terminal interface, so that precision and dynamic patent recommendation are realized, and the user experience is improved. And the patent operation efficiency and the user experience are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of digital technology, and more specifically, to an AI intelligent recommendation method and system applied to an online patent operation digital platform. Background Art

[0002] In the current digital age, the field of patent operations is facing increasing challenges in data processing and accurate recommendation. Existing patent operation platforms often focus on simple patent information display and basic search functions, lacking in-depth mining and utilization of user behavior data, and comprehensive integration and analysis of patent attribute information throughout the entire life cycle. When using these patent operation platforms, users often need to spend a lot of time manually screening patent resources that meet their needs, and the recommendation results are often difficult to accurately match the user's potential needs with the actual value of the patent. In addition, existing technologies have also failed to effectively combine real-time user interaction information to dynamically optimize recommendation results, resulting in difficulty in achieving optimal recommendation results. Summary of the Invention

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an AI intelligent recommendation method applied to an online patent operation digital platform, the method comprising: Obtaining full user behavior data and patent lifecycle attribute data, including historical user interaction records, demand expression records, and scenario context records; and obtaining full patent lifecycle attribute data, including patent technology evolution records, legal status change records, and market value fluctuation records; Performing demand intention analysis on the full amount of user behavior data to generate a user demand intention feature set, wherein the user demand intention feature set includes explicit demand features and potential development demand features at the current stage; Performing value dimension modeling on the patent full-cycle attribute data to generate a patent value dimension feature set, wherein the patent value dimension feature set includes a technology continuity feature, a legal stability feature, and a market value-added feature; Calling the AI ​​intelligent recommendation model to perform multi-dimensional correlation matching processing on the user demand intention feature set and the patent value dimension feature set to generate an initial recommendation candidate set, wherein the initial recommendation candidate set includes patent resource identifiers associated with the user demand features; Based on the real-time user interaction information, the initial recommendation candidate set is adaptively adjusted to generate a dynamically optimized patent recommendation sequence, and the patent recommendation sequence is output to the user terminal interface to trigger patent operation interaction operations.

[0004] On the other hand, an embodiment of the present invention also provides an AI intelligent recommendation system applied to an online patent operation digital platform, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0005] Based on the above aspects, the embodiment of the present invention obtains the full amount of user behavior data and patent full-cycle attribute data, and conducts in-depth analysis and modeling to construct a user demand intention feature set and a patent value dimension feature set. On this basis, the AI ​​intelligent recommendation model is used to perform multi-dimensional correlation matching on the user demand intention feature set and the patent value dimension feature set, thereby achieving accurate positioning of the recommendation results, which not only meets the user's current explicit needs, but also effectively explores and meets the user's potential development needs. At the same time, by dynamically adjusting the initial recommendation candidate set in combination with the user's real-time interaction information, the real-time and adaptability of the recommendation sequence are ensured, so that the recommendation results can be dynamically optimized as user needs change, thereby significantly improving the intelligence level of patent operations and promoting the effective allocation and utilization of patent resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a schematic diagram of the execution flow of the AI ​​intelligent recommendation method applied to the online patent operation digital platform provided by an embodiment of the present invention.

[0007] Figure 2 This is a schematic diagram of exemplary hardware and software components of an AI intelligent recommendation system applied to an online patent operation digital platform provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0008] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an AI intelligent recommendation method applied to an online patent operation digital platform provided by an embodiment of the present invention. The following is a detailed introduction to the AI ​​intelligent recommendation method applied to the online patent operation digital platform.

[0009] Step S110: Obtain the full amount of user behavior data and patent full cycle attribute data, wherein the full amount of user behavior data includes user historical interaction records, demand expression records and scenario context records, and the full cycle attribute data of the patent includes patent technology evolution records, legal status change records and market value fluctuation records.

[0010] This embodiment aims to provide basic data for the intelligent recommendation system of an online patent management digital platform, requiring comprehensive collection of key information related to users and patents. First, on the user side, user behavior patterns, demand expressions, and operational scenarios are the primary basis for understanding their patent needs. By collecting comprehensive behavioral data, a precise profile of user needs can be constructed. Regarding the patents themselves, their technical development trajectory, legal status changes, and market value fluctuations are also key factors in measuring their value and potential. Therefore, based on the above objectives, the following sub-steps are specifically performed to complete data acquisition. It is important to note that during the data collection process, explicit user authorization must be obtained. Only after obtaining authorization and consent can the collection and subsequent data analysis operations be performed. To protect user privacy, anonymization and desensitization technologies are used during the data collection phase to process sensitive data. Information that directly identifies the user is encrypted into an irreversible hash value using a hash function. For sensitive fields such as mobile phone numbers and email addresses, key features are retained and some content is hidden. Furthermore, obtaining attribute data throughout the patent lifecycle requires legal authorization and permission.

[0011] Step S111: Extract the time stamp, stay duration and access object information of the user's visit to the patent details page from the platform interaction log storage module as the user's historical interaction record.

[0012] In this embodiment, the interaction log storage module continuously records the user's interaction behavior with the patent details page in the platform, and stores the key data of each user access operation in chronological order. During the extraction process, the data query statement can be used to filter and locate the stored data of the interaction log storage module. For example, by analyzing the index information and metadata of the log file, the records related to the user's access to the patent details page can be quickly determined. For the access time mark, it can be extracted based on the timestamp field in the log. The timestamp field can be accurate to the second level or even finer time granularity, so as to clarify the user's interaction with the specific patent at different times. The length of stay is determined by calculating the difference between the timestamps of the user entering and leaving the patent details page. This calculation process is implemented using code logic, for example, automatically triggering the calculation program at the end of the access. The access object information can be obtained from the page identification field in the log. The page identification field uniquely corresponds to each patent details page in the platform. Its information usually covers the basic identification of the patent, the field to which it belongs, etc., which facilitates the subsequent feature analysis and classification of the patents accessed by the user.

[0013] Step S112: extract the technical requirement description text, legal status consultation text and market value assessment text submitted by the user from the platform demand feedback module as a record of the user's demand expression.

[0014] The platform's demand feedback module is a key channel for users to express their patent needs and questions. This module stores and manages user feedback input in text form. Natural language processing technology is used to process the text data in the platform's demand feedback module during data extraction. First, a text segmentation algorithm is used to segment the large amount of feedback content by topic, distinguishing different types of text segments, such as technical requirements, legal status consultation, and market value assessment. Next, lexical and grammatical analysis is performed on each segment to remove irrelevant modifiers and grammatical errors, extracting the core content. For example, for technical requirement descriptions, keywords and phrases describing the patent's technical features, performance indicators, and application scenarios are identified. For legal status consultations, expressions related to patent statuses such as grant, maintenance, invalidation, and pledge are highlighted. For market value assessments, information on patent prices, market prospects, and investment value is extracted. Text mining algorithms are used to extract and organize this valuable information, ultimately forming structured data storage to facilitate subsequent semantic analysis and feature extraction.

[0015] Step S113: extract the terminal type information, access period information and user historical interaction records during the user operation from the platform device information module as the user scenario context record.

[0016] The platform device information module is the data source for managing user device operation information and interaction time. This module associates and stores user device information and operation records, establishing a correspondence between device identification codes and user account information. When extracting terminal type information, judgments are made based on information such as the device's operating system identification, hardware model, and device characteristics. For example, the operating system version number is used to distinguish different operating system types for mobile terminals, and hardware parameters are used to determine whether the terminal is a desktop terminal or a smart device terminal. For access time period information, by parsing the timestamp information in the log, the time when the user accesses the platform is divided into different time period identifiers, such as daytime, nighttime, and peak hours. Custom divisions can be made based on the activity level of platform services and user usage habits. User historical interaction records are retrieved from the module's associated storage area, filtered and integrated based on user account and device information, and the user's interaction records across different devices and time periods are aggregated to form complete user scenario context information.

[0017] Step S114: extracting the patent family technology development path, subsequent patent citation relationship and related technology iteration information from the patent technology document database as a patent technology evolution record.

[0018] The patent technology document database is a core repository for storing all types of information related to patent technologies. It contains a large number of original patent application documents, examination documents, and technical reports. To extract the technological development path of a patent family, a graph algorithm is used to construct a tree structure of the patent family, leveraging the patent family identifier and technical classification code. Starting from a core patent in the database, all patents in its family are searched. The position and hierarchical relationships within the tree structure are determined based on the patent's application date and technological relevance, thereby depicting the technological development trajectory of the patent family. Semantic matching technology and knowledge graph reasoning methods are used to extract subsequent patent citation relationships. The citations within the patent text are analyzed to identify the logical relationships between cited and citing patents. These relationships are then mapped into the patent knowledge graph to visually demonstrate the technological inheritance and innovation connections between patents. Information on related technological iterations is extracted by comparing the technical features and performance indicators of patents from different periods within the same technological field. For example, a feature vector matching algorithm is used to calculate the similarities and differences in technical features between different patents, identifying key points and trends in technological updates and improvements.

[0019] Step S115: Synchronize the time series data of patent authorization status change, patent maintenance status change, patent invalidation status change, and patent pledge status change from the preset data interface as a record of patent legal status change.

[0020] The preset data interface is a bridge for data connection with external patent management agencies or related data sources. This interface regularly synchronizes information on changes in the legal status of patents from external data sources according to pre-defined protocols and formats. During the synchronization process, a strategy combining incremental updates and full updates is adopted. During the initial synchronization, the historical legal status records of all relevant patents are obtained through a full update; during each subsequent synchronization, the data update status is determined based on the timestamp, and only the data that has changed since the last synchronization is obtained, and the data is supplemented using an incremental update method. The synchronized time series data is formatted and standardized, and the date representation, status coding, and field naming are unified to ensure data consistency and compatibility. At the same time, a data verification mechanism is established to check the integrity and accuracy of the synchronized data. This is achieved by comparing data between different data sources and calculating the logical rationality of status changes.

[0021] Step S116: Collecting information on changes in historical patent transaction prices, regional expansion events of patent families, and migration information of technology application scenarios from the patent transaction market database as records of patent market value fluctuations.

[0022] The patent trading market database records patent market transactions and related events. To collect historical patent price change information, data crawling technology is used to comprehensively scan and capture transaction records in the database. Price data is sorted and organized according to patent identifiers and transaction times to construct price time series data. Furthermore, outlier processing is performed on the price data. Statistical analysis methods such as boxplots and standard deviation analysis are used to identify, correct, or eliminate abnormal price data caused by data entry errors or unusual market fluctuations. For geographical expansion events within a patent family, the database is queried for the geographical distribution and application dates of patents to identify the initial filings and subsequent expansions of patents in different regions. Geographic Information System (GIS) technology is used to visualize geographical expansion events, facilitating analysis of patent market penetration in different regions. Information on technology application scenario migration is obtained through in-depth analysis of text such as patent specifications and market research reports. Topic models and text classification algorithms are used to identify changes in the primary application scenarios of patents over different time periods, such as expansion from a traditional industry to an emerging one.

[0023] Step S117: Integrate the user historical interaction records, user demand expression records and user scenario context records into the user's full behavior data, and integrate the patent technology evolution records, patent legal status change records and patent market value fluctuation records into the patent full cycle attribute data.

[0024] When integrating all user behavior data, data fusion technology is used to merge historical user interaction records, demand expression records, and scenario context records. First, this data is associated based on the user's unique identifier to ensure that each user's data information is accurately merged. Then, different types of data are normalized to unify the data format and data range, eliminating dimensional differences between data. For example, for time stamps and duration data, they are converted to the same time unit and timestamp format; for text information, the encoding and character set are unified. Feature engineering methods are used to extract and represent features from the integrated data, converting it into a vector form suitable for processing by machine learning models.

[0025] Data fusion and normalization techniques are also used when integrating attribute data throughout the patent lifecycle. Records of patent technology evolution, legal status changes, and market value fluctuations are linked based on the patent's unique identifier. Different types of data are processed uniformly according to their physical meanings and business rules. For example, for time series data, the time granularity and data sampling frequency are unified; for graphic and text data, they are converted into feature vectors. Association analysis methods are used to explore potential relationships between different types of data and construct a comprehensive feature vector for the entire patent lifecycle, enabling subsequent comprehensive value assessment and recommendation analysis of the patent.

[0026] Step S120: performing demand intention analysis on the full amount of user behavior data to generate a user demand intention feature set, wherein the user demand intention feature set includes explicit demand features and potential development demand features at the current stage.

[0027] To accurately grasp users' actual patent needs and future demand trends, we analyze demand intent using the full set of user behavior data we've collected. Specifically, we conduct in-depth mining of this data from various perspectives, extracting explicit demand features that reflect current operational needs and potential demand features that reflect the direction demand is headed. These features are then combined to form a comprehensive set of user demand intent features. This process is accomplished through the following sub-steps.

[0028] Step S121: inputting the user historical interaction records into a time series feature extraction module, capturing the time dependency of user behavior through a time series analysis method, and generating user behavior time series features.

[0029] The time series feature extraction module is a tool specifically designed for processing time series data. It applies theories and algorithms related to time series analysis to analyze historical user interaction records. First, historical user interaction records are arranged in chronological order to form time-indexed series data. Models such as the Autoregressive Integrated Moving Average (ARIMA) model and the Long Short-Term Memory (LSTM) network are used to analyze this time series. Taking the ARIMA model as an example, the data is differencing-processed based on the stationarity test results of the time series to ensure that it meets stationarity requirements. Model parameters such as the autoregressive order p, differencing order d, and moving average order q are determined. Information criteria such as the Akaike Information Criterion (AIC) are used to select the optimal parameter combination. The trained model is then used to predict and analyze the dependencies between user behaviors at different time points, extracting features representing temporal patterns of behavior, such as the high frequency of user access to certain patents within a specific time period and periodic variations in access intervals. These features are quantified to form a user behavior time series feature vector. The dimensionality of this user behavior time series feature vector is determined by the number of different temporal pattern features extracted.

[0030] Step S122: Input the user demand expression record into the text semantic analysis module, extract the technical field keywords, legal status keywords and market value keywords that the user is concerned about through the semantic analysis method, and generate user demand semantic features.

[0031] The text semantic parsing module uses relevant technologies in natural language processing to process user demand expression records. First, the user demand expression records are pre-processed, and then further operations are performed according to specific sub-steps.

[0032] Step S1221: performing text segmentation processing on the user demand expression record to obtain a word set including technical terms, legal terms and market terms.

[0033] Use Chinese word segmentation algorithms, such as word segmentation tools from related technologies, to segment the text in the demand expression records. For cleaned text, these word segmentation tools can segment continuous text into independent tokens based on predefined lexicons and rules. For example, text describing new energy vehicle battery technology would be segmented into tokens such as "new energy vehicle" and "battery technology." During the word segmentation process, terminology from different fields is marked and identified. By maintaining specialized dictionaries of technical terminology, legal terminology, and market terminology, when the word segmentation results match words in these dictionaries, they are marked as terms in the corresponding fields, ultimately forming a token collection containing terms from different fields.

[0034] Step S1222: Input the word-unit set into a pre-trained language understanding model to generate a contextual semantic embedding vector for each word-unit.

[0035] Pre-trained language understanding models such as BERT and GPT can learn the semantic information of words in context. Token sets are formatted as input according to the model's requirements. For example, the BERT model adds special markers [CLS] and [SEP] before and after the token set. The model applies a multi-layer encoder to each token, capturing the semantic connections between tokens through an attention mechanism. During this process, the model maps tokens to a high-dimensional vector space and generates a contextual semantic embedding vector for each token based on the language representation rules learned by the model. The dimensionality of the vector is typically determined by the model architecture; for example, the dimensionality of BERT-base's word vectors is 768. These vectors not only capture the semantic information of the token itself but also take into account semantic variations within the context of the text.

[0036] Step S1223: Determine the importance of each word-unit embedding vector in the expression of user needs through the attention weight calculation method, and generate the word-unit attention weight distribution.

[0037] Attention weight calculation methods can focus on important tokens in a text. Common attention mechanisms, such as dot-product attention (Dot-ProductAttention), calculate the similarity between the query vector (typically the overall representation of the text) and the embedding vector of each token. The dot product operation is performed on the token embedding vector and the query vector, adjusted by a scaling factor, and then converted into a probability distribution using the softmax function to obtain the attention weight for each token. This attention weight reflects the importance of the token in the overall demand expression. Through multiple iterations and training, the parameters of the attention calculation are optimized to make the attention weight more accurately reflect the importance of the token, ultimately generating a token attention weight distribution.

[0038] Step S1224: Filter out word elements whose attention weight distribution meets the preset importance conditions as the technical field keywords, legal status keywords and market value keywords that the user is concerned about.

[0039] The preset importance condition can be to set an attention weight threshold. The attention weight of each word is compared with the attention weight threshold. When the attention weight of the word is greater than or equal to the attention weight threshold, the word is considered important. Based on the field labeling of the word before, the keywords belonging to the technical field, legal status and market value field are distinguished. For example, for the technical field, filter out words related to new materials, processes, etc.; for the legal field, filter out words related to patent authorization, transfer, etc.; for the market field, filter out words related to price trends, market demand, etc. In this way, keywords in different fields that users are concerned about are obtained.

[0040] Step S1225: performing weighted summation processing on the contextual semantic embedding vectors of the keywords to generate user demand semantic features containing the user demand semantic focus.

[0041] For the selected keywords, obtain their contextual semantic embedding vectors generated in the previous step. Weight the corresponding vectors according to the attention weight of each keyword, for example, assigning a larger weight coefficient to the vectors of keywords with higher attention weights. Summing the weighted vectors, adding the values ​​on each dimension, yields a comprehensive vector. This vector is the user demand semantic feature, representing the semantic focus of the user's demand expression. It contains information from various aspects, including technology, law, and market, and can be used for subsequent matching and analysis with patent features.

[0042] Step S123: extracting the association between terminal type and demand intensity, and the association between access time period and demand urgency according to the user scenario context record, and generating user scenario context features.

[0043] For example, step S1231: classify the terminal type information in the user scenario context record to generate terminal type identifiers for mobile terminals, desktop terminals, and smart device terminals. By identifying and classifying the operating system, hardware features, and other information of the terminal, a specific identifier is assigned to each terminal type. For example, a terminal using an Android or iOS system is identified as a mobile terminal and is given a corresponding "mobile terminal identifier"; a terminal using a Windows or macOS system and being a traditional desktop or notebook device is identified as a desktop terminal and is given a "desktop terminal identifier"; specific devices with intelligent interactive functions, such as smart watches, are classified as smart device terminals and are given a "smart device terminal identifier."

[0044] Step S1232: The access period information in the user scenario context record is divided into time periods to generate access period identifiers for daytime, nighttime, and peak time. A day is divided into different time periods based on the activity patterns of platform services and user usage habits. Referring to general work and rest schedules and platform traffic data, normal daytime work and study time is set as the daytime period and assigned a "daytime period identifier"; nighttime rest time is set as the nighttime period and assigned a "nighttime period identifier"; and peak time periods are determined based on platform traffic peak statistics and marked as "peak period identifiers."

[0045] Step S1233: Count the dwell time distribution in the historical interaction records of users under different terminal type identifiers, and calculate the demand intensity coefficient corresponding to each terminal type identifier. For each terminal type identifier, summarize the dwell time of users on key pages such as the patent details page under this type of terminal, and analyze its distribution pattern. Probability theory and statistical methods can be used, such as calculating indicators such as the average dwell time and the median dwell time. Based on the assumed relationship between dwell time and demand intensity, that is, the longer the dwell time, the higher the demand intensity, the dwell time indicator is mapped to the demand intensity coefficient to reflect the intensity of the user's demand for patents expressed through this terminal type.

[0046] Step S1234: Count the frequency of occurrence of urgent demand keywords in user demand expression records under different access time period identifiers, and calculate the demand urgency coefficient corresponding to each access time period. Pre-determine a series of urgent demand keywords, such as "urgently needed" and "immediately," through manual annotation and natural language processing technology. Keyword matching and frequency statistics are performed on user-submitted demand expression records under different access time period identifiers to obtain the frequency of occurrence of urgent demand keywords within different time periods. Based on the relationship between frequency and demand urgency, the frequency of occurrence is converted into a demand urgency coefficient to measure the urgency of user needs in each time period.

[0047] Step S1235: The terminal type identifier is mapped to the demand intensity coefficient, and the access period identifier is mapped to the demand urgency coefficient, generating scenario feature association information comprising a terminal-intensity association table and a period-urgency association table. The terminal type identifiers are mapped to the corresponding demand intensity coefficients in a tabular format to construct the terminal-intensity association table. Similarly, the access period identifiers are mapped to the corresponding demand urgency coefficients to form the period-urgency association table, thereby integrating the relationships between terminal type, access period, and demand features.

[0048] Step S1236: Generate user scenario context features based on the scenario feature association information. The information in the terminal-intensity association table and the time period-urgency association table is integrated and converted to form a unified feature vector. The vector elements correspond to the demand intensity and urgency information for different terminal types and access time periods, respectively. The vector is presented in a format that can be directly input into subsequent processing modules to ultimately generate user scenario context features.

[0049] Step S124: Input the user behavior temporal features, user demand semantic features and user scenario context features into the dynamic fusion module, dynamically assign the corresponding fusion weights to each of them through the feature weight adjustment mechanism, and generate explicit demand features that reflect the needs of the current operation stage.

[0050] The dynamic fusion module receives as input temporal features of user behavior, semantic features of user needs, and contextual features of user scenarios. Within the module, these features are fused using a hybrid model framework from machine learning. Different features are weighted by setting variable weight parameters, which determine the contribution of each feature to the fusion process.

[0051] The feature weight adjustment mechanism dynamically adjusts weights based on the current system environment and user feedback. For example, when the latest user interaction behavior shows a significant impact on demand due to changes in temporal patterns, the system increases the weight of the temporal features of the user's behavior; if the user's expression reflects more technical and legal concerns, the weight of the semantic features of the user's needs is correspondingly increased. A gradient descent-based optimization algorithm is used to continuously update the weight parameters according to the preset objective function and loss function. The objective function can be to maximize the match between the fused features and the user's actual needs. The loss function is used to measure the difference between the current fused features and the objective. Through iterative optimization, the value of the loss function is gradually reduced, and ultimately a suitable weight allocation scheme is obtained. The weighted features are then weightedly fused, and the corresponding elements of each feature vector are weighted and summed according to the weights to generate explicit demand features that reflect the needs of the current operation stage.

[0052] Step S125: performing trend analysis on the historical output features of the dynamic fusion module, extracting the evolution direction information of user needs over time, and generating potential development demand features reflecting the development direction of the needs.

[0053] In this embodiment, the explicit demand feature vectors output by the dynamic fusion module at multiple historical time points are collected to form a feature time series, which is processed using trend analysis methods in time series analysis, such as linear trend analysis and nonlinear trend analysis.

[0054] Taking linear trend analysis as an example, the values ​​along each dimension of the feature vector are treated as a time series. A trend line is fitted using the least squares method, and the slope of the trend line is calculated. This slope represents the changing trend of the feature in that dimension over time. A positive slope indicates an increasing trend, while a negative slope indicates a decreasing trend. By analyzing trends across all dimensions, a set of trend value vectors is generated, reflecting the evolution of user needs across each dimension over time.

[0055] To more accurately capture the evolving trends of user demand, in addition to linear trend analysis, nonlinear trend analysis methods such as polynomial fitting or wavelet analysis can also be employed. Polynomial fitting can create more complex trend curves and better handle feature sequences with nonlinear variations. In polynomial fitting, the order of the polynomial is determined based on the characteristics of the sequence, and the coefficients are solved using the least squares method to obtain the fitted trend curve equation. Analysis of this equation reveals the changing trends of the features. Wavelet analysis can decompose feature sequences at different time scales, breaking them down into subsequences with different frequency components. By analyzing high- and low-frequency subsequences, it is possible to distinguish short-term fluctuations from long-term trends in demand.

[0056] The resulting trend value vector is further processed and transformed so that it can be reflected in a meaningful feature space. The trend value vector can be mapped to a new feature dimension, or the original feature vector can be adjusted based on the trend value to generate a new feature vector. In this new feature vector, the value of each element represents the changing trend of user demand along the corresponding dimension. By formatting and standardizing this feature vector to meet the requirements of subsequent system processing, the new feature vector ultimately becomes a potential development demand feature that reflects the direction of demand development.

[0057] Step S126: Dimensionally combine the explicit demand features and the potential development demand features to generate a user demand intention feature set.

[0058] Explicit demand characteristics and potential development demand characteristics each reflect user needs from different perspectives. Combining them dimensionally can integrate the information of both and form a more comprehensive expression of user needs.

[0059] First, analyze the dimensional structure of explicit demand features and potential development demand features. Since the two may have different numbers of dimensions and feature meanings, dimensional matching and adjustment are necessary. If some dimensions of the two have similar semantic meanings, these dimensions can be merged and integrated; for those with different dimensions, these can be retained and the dimensionality of the feature vector expanded. For example, explicit demand features may include a dimension of user attention to current technical features, while potential development demand features include a dimension of the future development direction of the technical features. These two dimensions can be combined to form a new dimension that reflects user demand for technical features throughout their lifecycle.

[0060] The processed explicit demand feature vector and potential development demand feature vector are combined using vector concatenation. The elements of the potential development demand feature vector are sequentially added to the explicit demand feature vector to form a new, longer feature vector, in which each element represents information about user needs in different aspects and time scales.

[0061] Post-processing, such as normalization, is performed on the combined vectors to ensure that they are of the same scale and comparable across all dimensions. Based on the value range of each element in the vector, a normalization algorithm, such as min-max normalization or Z-score normalization, is used to normalize the vectors so that the values ​​of the vector elements are mapped to a set interval. The resulting normalized vector is the user demand intent feature set, which fully reflects the user's current demand status and the direction of future demand development.

[0062] Step S130: Perform value dimension modeling on the patent full-cycle attribute data to generate a patent value dimension feature set, wherein the patent value dimension feature set includes a technology continuity feature, a legal stability feature, and a market value-added feature.

[0063] To comprehensively assess the value of patents, it is necessary to model and process the patent's full-cycle attribute data from multiple value dimensions. The following describes the process of modeling and generating features for each dimension: Step S131: Based on the patent's technological evolution record, extract the patent's position in the technological evolution chain and subsequent technological derivative features through a technology citation network analysis method to generate technological continuity features. For example, this step can be further refined into the following sub-steps: Step S1311: extracting the forward citing patent set and the backward cited patent set of the patent from the patent technology evolution record.

[0064] Patent technology evolution records contain a large amount of citation relationship information between patents. Through data mining and screening, we identify each patent's forward citing patents (i.e., other patents cited by the patent) and organize them into a forward citing patent collection. We also identify backward cited patents (i.e., other patents that cite the patent) to form a backward cited patent collection. During the extraction process, we must ensure data accuracy and verify each citation relationship to avoid errors or omissions.

[0065] Step S1312: Based on the forward citing patent set and the backward cited patent set, a technology citation network including the patent and its forward and backward citing patents is constructed, wherein the nodes in the technology citation network represent patents and the edges represent citation relationships.

[0066] With this patent at the core, all patents in the forward-citing patent set and the backward-cited patent set are treated as nodes, and corresponding edges are established based on the citation relationships between them. Using graph theory and graph database technology, this technical citation network is constructed, enabling the network structure to clearly demonstrate the technical connections between patents. During the construction process, attention must be paid to the integrity and consistency of the network, ensuring that each patent node and citation edge accurately reflects the actual technical relationship.

[0067] Step S1313: Calculate the centrality index of the patent in the technology citation network, where the centrality index includes degree centrality, betweenness centrality, and closeness centrality.

[0068] Degree centrality measures the number of edges associated with a patent node—that is, the number of connections it has with other patents—and reflects the patent's local influence within the technology network. Betweenness centrality is calculated by counting the number of shortest paths passing through the patent node, reflecting the patent's role as an intermediary in the dissemination and exchange of technological information. Closeness centrality measures the average shortest path length from the patent node to all other nodes in the network, measuring the proximity of the patent to other patents within the network. When calculating these centrality metrics, a breadth-first search algorithm can be used to determine the shortest path to ensure accuracy.

[0069] Step S1314: Determine the core position information or edge position information of the patent in the technology chain based on the centrality index.

[0070] The calculated degree centrality, betweenness centrality, and closeness centrality indices are comprehensively analyzed. High values ​​for these centrality indices indicate that the patent occupies a core position in the technology citation network, significantly driving the development and dissemination of technology. Conversely, low values ​​for these centrality indices indicate that the patent is peripheral to the technology chain, with relatively little influence on overall technological development. By setting appropriate thresholds, the position of patents can be clearly delineated, providing accurate positioning information for subsequent analysis.

[0071] Step S1315: Count the proportion of the backward cited patents of the patent that belong to the same technical field as the derivative capability indicator within the technical field.

[0072] The backward-cited patent collection is classified and annotated according to the technical field to which the patent belongs. The number of backward-cited patents belonging to the same technical field is counted, and their proportion in the entire backward-cited patent collection is calculated. This proportion reflects the influence and derivative ability of the patent in its technical field. The higher the proportion, the more subsequent research and technological innovation the patent can trigger in that technical field.

[0073] Step S1316: Count the proportion of cross-technical field patents in the backward cited patents of the patent as an indicator of the derivation capability between technical fields.

[0074] We also categorize the backward-cited patent set, this time focusing on cross-technical areas. We count the number of backward-cited patents belonging to different technical fields and calculate their proportion in the entire backward-cited patent set. This proportion reflects the patent's ability to transcend its own technical field and influence and inspire other technical fields. A higher proportion indicates a patent's greater cross-technical derivative capabilities and technological spillover effects.

[0075] Step S1317: Integrate the core position information or edge position information, the derivative capability indicators within the technical field, and the derivative capability indicators between technical fields into a technology continuity feature.

[0076] The core or edge position information, as well as the derivative capability indicators within and between technical fields, obtained through the above steps, are processed and encoded in a unified format to form a multi-dimensional feature vector. During the integration process, the dimensions and scales of each indicator are ensured to be consistent, allowing this feature vector to comprehensively and accurately describe the patent's sustainability and influence during technological evolution. The resulting technological continuity features can be used in subsequent patent value assessment and recommendation systems, providing users with more targeted patent information.

[0077] Step S132: Input the patent legal status change record into the legal value assessment module, calculate the probability information of the patent maintaining a stable state within a preset time period in the future through the state transition probability analysis method, and generate a legal stability feature.

[0078] Patent legal status change records contain information about the legal status of a patent at different points in time, such as granted, maintained, invalidated, and pledged. This information is fundamental data for assessing the legal stability of a patent. The Legal Value Assessment module conducts in-depth analysis of these records.

[0079] In the state transition probability analysis method, a finite state machine model is first constructed to describe the transition relationship between patent legal statuses. This finite state machine contains all possible legal status nodes, such as "Authorization Valid," "Authorization Invalid," and "Pledged," and there are edges between nodes representing state transitions. Based on the historical records of patent legal status changes, the number of transitions from one state to another is counted. For example, the number of transitions from "Authorization Valid" to "Authorization Invalid" within a set time period is counted. Using these statistics, the frequency of state transitions is calculated.

[0080] The state transition frequencies are normalized so that they satisfy the property of probability, that is, the sum of all transition probabilities from a state is 1. These normalized state transition frequencies form the state transition probability matrix. Each element in the matrix represents the probability of transitioning from one specific state to another.

[0081] To calculate the probability of a patent maintaining a stable state within a predetermined future time period, the prediction principle of a Markov chain is employed. Based on the assumption that the future state depends solely on the current state, a Markov chain predicts the probability distribution of states at different future time points by continuously multiplying the state transition probability matrix from the current legal status of the patent. For example, if a patent is currently in the "granted valid" state, by multiplying the vector representing that state by the state transition probability matrix, the probability of each state at the next time point is calculated. Similarly, the probability of each state within a predetermined time period can be predicted.

[0082] Define stability states. For example, combinations of states such as "valid authorization and no pledge" can be defined as stability states. From the predicted state probability distribution, extract the probability values ​​of the patent being in these stability states. These probability values ​​are organized and formatted to form legal stability characteristics.

[0083] Step S133: Input the patent market value fluctuation record into the market value prediction module, extract the fluctuation pattern of patent value as market demand changes through time series analysis method, and generate market value-added characteristics. Step S1331: performing time series decomposition processing on the historical transaction price change information in the patent market value fluctuation record to separate the trend change part, the seasonal fluctuation part and the random fluctuation part.

[0084] Time series decomposition is a common method for processing data with a temporal order. It can help us break down complex price change sequences into their components, thereby better understanding the patterns of price fluctuations. When decomposing historical transaction price change information in patent market value fluctuation records, either an additive model or a multiplicative model can be used. In the additive model, the original time series is represented as the sum of trend changes, seasonal fluctuations, and random fluctuations; in the multiplicative model, the original time series is the product of these three components. Generally, when the amplitude of price fluctuations is relatively stable, the additive model is more appropriate; when the amplitude of price fluctuations increases over time, the multiplicative model can more accurately reflect the characteristics of the sequence.

[0085] To isolate the trend, you can use a moving average or a fitted trend equation. The moving average smoothes the price series by calculating the average price over a set time window, eliminating short-term random fluctuations and thus determining the general price trend. For example, a moving average can be calculated for three, five, or even longer periods. The time window chosen needs to be adjusted based on the characteristics of the price series. A fitted trend equation involves mathematically modeling the price series, such as using linear regression or polynomial regression, to find an equation that describes the long-term price trend. In polynomial regression, the order of the polynomial is determined based on the trend shape of the series, and the coefficients of the equation are solved using the least squares method to obtain a fitted trend curve.

[0086] Seasonal fluctuations reflect seasonal or cyclical changes in prices. Seasonal fluctuations can be isolated using the seasonal index method. First, the average price for each seasonal cycle (such as annually, quarterly, etc.) is calculated. These averages are then compared with the average price for all cycles to obtain a seasonal index for each season. The seasonal index indicates the degree to which prices in that season deviate from the average price. The seasonal fluctuation component can be isolated by dividing the original price series by the trend component and then multiplying it by the inverse of the seasonal index.

[0087] Random fluctuations are price fluctuations that cannot be explained by trend changes and seasonal fluctuations and are generally attributed to chance factors or unexpected events. After separating the trend and seasonal fluctuations, the remainder of the original time series is the random fluctuation, subtracting the trend from the original time series and dividing it by the seasonal fluctuation (in a multiplicative model) or subtracting the trend and seasonal fluctuations (in an additive model). Statistical analysis of the random fluctuations can be used to assess the uncertainty in price fluctuations.

[0088] Step S1332: extrapolate and predict the trend change part through the time series prediction method to generate a basic value prediction sequence of the patent within a preset future time period.

[0089] After determining the price trend, it is necessary to extrapolate and forecast it to obtain information on the underlying value of the patent within a predetermined future time period. Common time series forecasting methods include exponential smoothing and the Autoregressive Integrated Moving Average (ARIMA) model. Exponential smoothing assigns different weights to historical data, performing a weighted average, with recent data receiving a higher weight and further data receiving a lower weight. Its advantages are simple calculations and the ability to quickly track data trends. In exponential smoothing, an appropriate smoothing coefficient is selected based on the characteristics of the data. A larger smoothing coefficient increases the weight of recent data, and the model becomes more responsive to data changes.

[0090] The ARIMA model is a more complex but generally more accurate forecasting method that consists of three parts: autoregression (AR), differencing (I), and moving average (MA). The autoregression part represents the linear relationship between the current value and the historical value, the differencing part is used to process non-stationary time series to make them stationary, and the moving average part takes into account the correlation between error terms. When using the ARIMA model for forecasting, it is necessary to first test the stationarity of the trend change part. If the series is not stationary, it is differentiated to make it stationary. Then, the appropriate AR, I, and MA orders are selected through information criteria (such as AIC or BIC), and the model parameters are estimated and fitted. Finally, the fitted ARIMA model is used to predict the fundamental value within a preset time period in the future to generate a fundamental value forecast series.

[0091] Step S1333: extracting the correlation between the regional expansion events of the same patent family and the transaction price fluctuations in the patent market value fluctuation records, and generating a regional expansion impact parameter of the regional expansion on the value.

[0092] Detailed analysis and mining of patent market value fluctuations reveals relevant information about regional expansion events within the same patent family. Regional expansion events include the first application and authorization of patents in different countries or regions. These regional expansion events are arranged chronologically and correlated with corresponding transaction price fluctuations.

[0093] An event study can be used to analyze the relationship between geographic expansion events and trading price fluctuations. Observe price fluctuations over a period of time before and after an event. For example, consider the time when a patent is first granted in a new region as the event point. Analyze the fluctuations and trends in trading prices over a period of one, three, or even longer periods before and after the event.

[0094] To measure the impact of regional expansion on value, statistical indicators of price changes before and after the regional expansion event are calculated. Indicators such as the price increase or decrease, as well as the standard deviation of price changes, can be calculated. These statistical indicators are then processed comprehensively, for example, by linearly combining the increase, decrease, and standard deviation according to pre-set weights to generate a comprehensive regional expansion impact parameter. This parameter reflects the extent and manner in which the regional expansion event impacted patent transaction prices, such as whether it was a positive price-increasing effect or a negative price-suppressing effect.

[0095] Step S1334: extract the correlation between the technology application scenario migration event and the transaction price fluctuation in the patent market value fluctuation record, and generate the scenario migration impact parameter of the technology application scenario migration on the value.

[0096] Similar to the process of analyzing regional expansion events, we conduct in-depth research on patent market value fluctuations to identify technology application scenario migration events. Technology application scenario migration events refer to the transfer of a patent's application scenario from one field to another, which can have a significant impact on the patent's market value.

[0097] Use text mining and semantic analysis techniques to identify changes in technology application scenarios within documents such as patent specifications and market research reports. For example, by analyzing keywords and phrases describing application scenarios within patent texts, we can identify shifts in patent application scenarios. We then match and correlate these technology application scenario migration events with transaction price fluctuations.

[0098] Regression analysis is used to quantify the relationship between technology application scenario migration events and transaction price fluctuations. A regression model is constructed using transaction price fluctuations as the dependent variable and the characteristics of technology application scenario migration events as independent variables. Independent variables can include characteristics such as the target domain of migration and the timing of migration. By estimating the parameters of the regression model, the coefficient of influence of each independent variable on the transaction price is obtained.

[0099] These impact coefficients are combined and processed to form scenario migration impact parameters that can comprehensively reflect the impact of technology application scenario migration on patent value. This parameter takes into account multiple aspects of technology application scenario migration and has a stronger ability to explain fluctuations in patent market value.

[0100] Step S1335: Perform weighted fusion processing on the basic value prediction sequence, the regional expansion impact parameter, and the scenario migration impact parameter to generate description information on the fluctuation pattern of patent value as market demand changes.

[0101] Before performing weighted fusion, the base value forecast sequence, regional expansion impact parameters, and scenario migration impact parameters must be preprocessed to ensure consistency in dimension and scale. The base value forecast sequence is a time series, typically containing forecast values ​​at multiple future time points. The regional expansion impact parameters and scenario migration impact parameters are comprehensive impact indicators and must be converted into a form that matches the base value forecast sequence in terms of time.

[0102] By assigning time weights to the regional expansion impact parameters and scenario migration impact parameters, they can have corresponding impact values ​​at each point in time. Based on the actual situation and empirical judgment of the patent market, the weights of the basic value prediction sequence, regional expansion impact parameters, and scenario migration impact parameters in the integration process are determined. For example, the near-term value of a patent may be more influenced by the trend of the basic value, while the long-term value may be more influenced by regional expansion and technology application scenario migration. Therefore, the weights of the three can be adjusted according to the time span.

[0103] Each component is weighted and summed (if weighted addition is reasonable) or weighted concatenated (if weighted concatenation is more reasonable) according to its assigned weights. When weighted addition is performed, the value of each time point in the basic value forecast sequence is added to the values ​​of the regional expansion impact parameter and scenario migration impact parameter at the corresponding time point according to the weights to obtain a comprehensive value sequence. This comprehensive value sequence describes the fluctuation pattern of patent value as market demand changes, and comprehensively considers the impact of factors such as basic value trends, regional expansion, and technology application scenario migration on patent value.

[0104] Step S1336: Calculate the value growth possibility information of the patent in the future time window based on the fluctuation pattern description information as a market value-added feature.

[0105] Analyze the information describing the fluctuation pattern and determine its statistical characteristics. For example, calculate statistical indicators such as the mean, standard deviation, and rate of change of the comprehensive value series. The mean can reflect the average value level of the patent in the future time window, the standard deviation can measure the degree of value fluctuation, and the rate of change can reflect the trend of value change over time.

[0106] Based on these statistical indicators, a rule can be developed to assess the likelihood of value growth. For example, a threshold can be set. If the rate of change in the comprehensive value series exceeds the threshold over a set period of time, and the mean shows an upward trend with a relatively small standard deviation, it indicates that the patent value has a high probability of growth. Probabilistic statistical methods can also be used to establish a probabilistic model for the likelihood of value growth by studying and simulating historical data.

[0107] In probabilistic models, machine learning algorithms, such as logistic regression or decision trees, can be used. These algorithms use the statistical indicators in the fluctuation pattern description as input features and the output label of whether or not the patent value is increasing. By training on large amounts of historical data, the relationship between features and value growth is learned, thereby predicting the probability of patent value growth within a future time window.

[0108] The calculated value growth possibility information is sorted and formatted to form market value-added characteristics.

[0109] Step S134: Input the technical continuity features, legal stability features and market value-added features into the value association analysis module, extract the synergistic enhancement relationship between the features of each dimension through the feature co-occurrence analysis method, and generate a patent value dimension feature set with multi-dimensional association information.

[0110] The Value Relationship Analysis module primarily explores the potential connections between technical continuity, legal stability, and market value-added characteristics. For example, feature co-occurrence analysis focuses on identifying elements that co-occur across different dimensions, thereby revealing their inherent connections and synergistic mechanisms.

[0111] To accurately identify co-occurrence information for these three features, preprocessing is required. This preprocessing process includes operations such as data cleaning, normalization, and dimension matching. Data cleaning aims to remove noise and erroneous information from the feature data to improve data quality. Normalization unifies feature values ​​of different ranges and scales into a relatively consistent range to eliminate the impact of dimensional differences on subsequent analysis. Dimension matching ensures that the dimensions of each feature vector are consistent, allowing for effective comparison and analysis.

[0112] After preprocessing, the analysis is gradually expanded to examine the co-occurrence of three features, starting with the co-occurrence of two features. When analyzing the co-occurrence of two features, it is possible that patents at the core of a technology and with strong technological derivatives tend to have a more stable legal status. This may be because high-quality technology patents typically follow more rigorous legal procedures during their development, reducing the risk of legal disputes. Regarding the technical continuity and market value-added characteristics, patents with high derivatives within a technology field and significant cross-disciplinary impact generally have higher market value-added potential. This is because such patents can expand application scenarios and market space, attracting more commercial investment and partnership opportunities. Similarly, there may be a correlation between legal stability and market value-added characteristics. Patents with a stable legal status can instill greater confidence in investors and market participants, thereby driving up the patent's market value.

[0113] Co-occurrence analysis of the three characteristics reveals that when a patent excels in all three areas—technical continuity, legal stability, and market value—it may indicate a patent with extremely high overall value. For example, patents in cutting-edge technology fields that occupy a core position in technological evolution, possess extensive technological derivative capabilities, enjoy solid legal protection, and exhibit significant market value growth often become popular investment targets.

[0114] To quantify these co-occurrence relationships, association rule mining algorithms are used. These algorithms automatically discover association patterns and rules between features from large amounts of data and calculate the strength and confidence of the associations. Association strength reflects the frequency of co-occurrence between features, while confidence indicates the probability that one feature will also appear when another feature appears. By screening and optimizing these association rules, significant co-occurrence relationships are identified. For example, by setting thresholds for association strength and confidence, only rules with high association strength and confidence are retained, ensuring that the mined co-occurrence relationships are authentic, reliable, and of practical value.

[0115] The mined co-occurrence relationship information is encoded and integrated into each feature vector. Each co-occurrence relationship can be assigned a new feature dimension, with the strength and confidence of the co-occurrence relationship serving as the value of that dimension. This process forms a patent value dimension feature set with multi-dimensional correlation information. This feature set is no longer a simple combination of individual features, but rather contains structured information about the synergistic relationships between features in each dimension, enabling a more comprehensive and accurate description of the patent's overall value.

[0116] Step S140: Calling the AI ​​intelligent recommendation model to perform multi-dimensional correlation matching processing on the user demand intention feature set and the patent value dimension feature set to generate an initial recommendation candidate set, which includes patent resource identifiers associated with the user demand features.

[0117] The AI ​​intelligent recommendation model is a core tool for accurately matching user needs with patent resources. Before matching, feature sets must be preprocessed and aligned to ensure consistency across time and space, thereby improving matching accuracy and effectiveness.

[0118] Step S141: Input the user demand intention feature set and the patent value dimension feature set into the feature alignment module of the AI ​​intelligent recommendation model, make the time resolution of the two consistent through time dimension alignment processing, and match the feature dimensions of the user demand intention feature set and the patent value dimension feature set through space dimension alignment processing.

[0119] The feature alignment module in the AI ​​intelligent recommendation model needs to address the differences in time and space between the user demand intention feature set and the patent value dimension feature set. Specifically, in terms of time dimension alignment, the time range and granularity involved in the user demand intention feature set and the patent value dimension feature set may be different. User demand may focus more on the current and future situation, while the patent value dimension feature may contain historical data and forecast information over a longer time span. In order to make the two comparable in time, the time resolution needs to be unified. Specifically, according to the specific circumstances of the time scales of the two, appropriate time intervals are selected for sampling and alignment. For example, if the time resolution of user demand is calculated on a monthly basis, while the time resolution of the patent value dimension feature is calculated on a quarterly basis, the data of the patent value dimension feature can be resampled to monthly data, or the user demand data can be aggregated on a quarterly basis to keep the two consistent in time dimension.

[0120] In terms of spatial dimension alignment, the main issue to be addressed is the matching of the number and meaning of feature dimensions. The user demand intention feature set and the patent value dimension feature set may have different feature vector dimensions, which may be due to different emphases in the feature extraction process. For example, the user demand intention feature may focus more on dimensions such as the user's subjective preferences and the urgency of the need, while the patent value dimension feature may focus more on objective dimensions such as technology, law, and market. In order to match the two in spatial dimensions, it is first necessary to analyze and compare the feature dimensions and find dimensions with similar meanings to merge. For those dimensions that are unique to the user demand intention feature set, you can try to find related proxy dimensions or expand them in the patent value dimension feature set; for dimensions that are unique to the patent value dimension feature set, if they are important for user demand matching, you can introduce them into the user demand intention feature set. Through the above-mentioned dimension adjustment and matching, the consistency of the two feature sets in spatial dimensions is ensured.

[0121] Step S142: In the matching calculation module of the AI ​​intelligent recommendation model, the correlation matching degree between the user demand intention feature set and the feature set of each patent value dimension is calculated. The correlation matching degree is used to represent the degree of correlation between user demand and patent value.

[0122] The matching calculation module is the core component for calculating the correlation matching degree. In order to more accurately calculate the correlation between user needs and patent value, the user demand intention feature set and the patent value dimension feature set are further subdivided.

[0123] First, the user demand intention feature set is decomposed into a set of explicit demand sub-features at the current time point and a set of potential demand sub-features in future time windows. The explicit demand sub-feature set reflects the needs explicitly expressed by users at the current operational stage, such as their immediate requirements for specific technical fields, legal status, and market value. The potential demand sub-feature set reflects the trend and direction of user demand development over time and provides a prediction of possible future user needs.

[0124] Similarly, the feature set of the patent value dimension is decomposed into the basic value sub-feature set at the current time point and the value-added potential sub-feature set in the future time window. The basic value sub-feature set represents the basic value of the patent in terms of technology, law, and market at the current moment, such as the current technological level, the stability of the legal status, and the market transaction price. The value-added potential sub-feature set reflects the possibility and trend of the patent's value growth in the future, such as the development prospects of technology, the potential for market expansion, and the sustainability of legal protection.

[0125] Calculate the degree of match between different sub-feature sets. First, calculate the degree of match between the explicit demand sub-feature set and the basic value sub-feature set. For example, this can be achieved by calculating the similarity between two vectors. Common similarity calculation methods include cosine similarity and Euclidean distance. Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them. A value closer to 1 indicates greater similarity. Euclidean distance calculates the distance between two vectors in space. A smaller distance indicates greater similarity. The calculated similarity value is used as the matching degree for the current stage.

[0126] Next, the degree of match between the potential demand sub-feature set and the value-added potential sub-feature set is calculated. Because potential demand and value-added potential involve future uncertainty, the influence of time needs to be considered. A time impact factor can be introduced to adjust the weight of the future-stage match. The time impact factor decreases as the future time window lengthens. This is because the further out in time, the greater the uncertainty in the prediction, so its impact on the future-stage match should be appropriately reduced. A time-weighted approach is used to calculate the match between the potential demand sub-feature set and the value-added potential sub-feature set to obtain the future-stage match.

[0127] Finally, the current-stage matching degree and the adjusted future-stage matching degree are comprehensively calculated. Based on their importance in the overall matching, appropriate weights are assigned to each. The two matching degrees are weighted and summed to obtain the final correlation matching degree. This correlation matching degree comprehensively and accurately reflects the correlation between user needs and patent value.

[0128] Step S143: Filter the patent resource identifiers corresponding to the patent value dimension feature set whose correlation matching degree meets the preset matching conditions as the initial candidate subset.

[0129] The preset matching condition is set based on the system's design goals and business needs. It is a threshold for measuring the degree of correlation matching. After the correlation matching degree is calculated, the correlation matching degree corresponding to each patent value dimension feature set is compared with the preset matching condition one by one. If the correlation matching degree is greater than or equal to the preset threshold, the patent is considered to have a high degree of match with the user's needs, and its corresponding patent resource identifier is screened out. These screened patent resource identifiers constitute the initial candidate subset, which represents a portion of all patents that have a high initial match with the user's needs.

[0130] Step S144: performing legal stability filtering on the initial candidate subset, excluding patent resource identifiers whose legal stability characteristics do not meet preset requirements, and generating a legally valid candidate subset.

[0131] Although the initial candidate subset is selected based on correlation matching, there may be some patents with unstable legal status. To ensure that the patents recommended to users have reliable legal protection, the initial candidate subset needs to be filtered for legal stability.

[0132] The legal stability characteristics corresponding to each patent in the initial candidate subset are obtained from the patent full-cycle attribute data. The preset requirement is a legal stability threshold set based on legal risk assessment and business needs. For example, it can be set that the probability of a patent maintaining a stable state within a preset period in the future must reach a certain level. The legal stability characteristics of each patent are compared with the preset requirements. If the legal stability characteristics of a patent do not meet the preset requirements, it means that the patent has a high legal risk, and its corresponding patent resource identifier is excluded from the initial candidate subset. After the above filtering process, the remaining patent resource identifiers constitute the legally valid candidate subset, which ensures that the included patents have high stability and reliability at the legal level.

[0133] Step S145: performing a technology continuity sorting process on the patent resource identifiers in the legally valid candidate subset, retaining the patent resource identifiers whose technology continuity characteristics meet the preset standards, and generating an initial recommended candidate set.

[0134] Technological continuity is a key indicator for measuring the value and development potential of patent technology. Patent resource identifiers within the legally valid candidate subset are ranked based on each patent's corresponding technological continuity characteristics. These characteristics encompass information such as a patent's position in the technological evolution chain and its technological derivative capabilities.

[0135] Before sorting, it is necessary to first clarify the specific conditions that meet the preset standards. The preset standards can be determined based on technology development trends, market demand and business strategies. For example, it can be set that the patent must be at the core of technology evolution and have strong derivative capabilities within the technology field and cross-domain derivative capabilities. On the basis of meeting the preset standards, the patent resource identifiers are sorted in descending order according to the characteristics of technology continuity, that is, the patents with stronger technology continuity are ranked higher. After the sorting is completed, the patent resource identifiers that meet the preset standards are selected from the sorted results. These patent resource identifiers constitute the final initial recommended candidate set. The patents in this initial recommended candidate set are not only legally stable, but also have strong continuity and development potential in technology, and can better meet the user's demand for high-quality patents.

[0136] Step S150: Adaptively adjust the initial recommendation candidate set based on the user's real-time interaction information to generate a dynamically optimized patent recommendation sequence, and output the patent recommendation sequence to the user terminal interface to trigger patent operation interaction operations.

[0137] When users see the patent resource identifiers in the initial set of recommended candidates on the terminal interface, their real-time interactive behavior will reflect their feedback on these recommendations and their latest needs. Leveraging this real-time interactive information, we can adjust and optimize the initial set of recommended candidates, making the recommendations more tailored to the user's actual needs. This dynamic adaptation process is described in detail below.

[0138] Step S151: After displaying the patent resource identifiers in the initial recommendation candidate set on the user terminal interface, the user's viewing operation data, collection operation data and consultation operation data for each recommended patent are collected as real-time user interaction information.

[0139] The user terminal interface is an important interface for users to interact with the recommendation system. After displaying the patent resource identifiers in the initial recommendation candidate set on this interface, the system will collect real-time data on the user's interaction behavior with these patents. Viewing operation data records which recommended patents the user has viewed, as well as information such as the viewing duration and number of times. This information can reflect the user's initial interest in a specific patent. Collection operation data indicates that the user is particularly interested in certain patents and has collected them for further understanding later. Consultation operation data reflects the user's intention to have a deeper understanding of the patent or to cooperate, such as consulting on the patent's technical details, legal status, market price, and other aspects. These viewing operation data, collection operation data, and consultation operation data are integrated to form real-time user interaction information.

[0140] Step S152: Input the real-time user interaction information into the feedback learning module of the AI ​​intelligent recommendation model, update the weights of the potential development demand features of the user demand intention feature set through the model parameter adjustment mechanism, and enhance the feature expression of the user's actual concern needs.

[0141] The AI ​​intelligent recommendation model's feedback learning module leverages real-time user interaction information to dynamically adjust model parameters to better reflect changes in user needs. This model parameter adjustment mechanism is based on optimization algorithms used in machine learning. Its core goal is to continuously adjust model parameters to make the model's recommendations more accurate and personalized.

[0142] The real-time user interaction information contains the actual feedback of users on different patents. These feedbacks can reveal the real interests and potential needs of users hidden behind the original needs. In order to better capture this information, the feedback learning module will focus on the potential development demand features in the user demand intention feature set. By analyzing the real-time interaction information, it is found that when users show special interest in certain types of patents, it means that the demand dimensions corresponding to these patents may be an important direction for the development of users' future needs. At this time, the weights of these related demand dimensions in the potential development demand features are increased through the model parameter adjustment mechanism. For example, if a user frequently views and collects patents in a specific technical field, the weight of the potential demand features related to this technical field will increase. In this way, the feature expression of the needs that users actually pay attention to is enhanced, so that the user demand intention feature set can more accurately reflect the real needs of users at this stage.

[0143] Step S153: recalculating the correlation matching degree of each patent resource identifier in the initial recommendation candidate set through the matching calculation module according to the updated user demand intention feature set, and generating updated matching degree ranking information.

[0144] After updating the user demand intent feature set, the correlation matching degree of each patent resource identifier in the initial recommendation candidate set needs to be recalculated. The updated user demand intent feature set is re-entered into the matching calculation module and matched with the value dimension feature set of each patent in the initial recommendation candidate set. The matching calculation process is similar to step S142 described above, namely, first decomposing the feature set, calculating the current stage matching degree and the future stage matching degree respectively, and then comprehensively calculating the correlation matching degree.

[0145] Due to changes in the user's demand intent feature set, especially the weight adjustment of potential development demand features, the calculated correlation matching degree will also change accordingly. The patent resource identifiers in the initial recommendation candidate set are re-ranked according to the new correlation matching degree, generating updated matching ranking information, which reflects the latest degree of match between each patent and user needs after considering real-time interactive feedback from users.

[0146] Step S154: Extract the patent feature pattern corresponding to the user's attention operation in the real-time user interaction information, and use the feature pattern matching method to filter patent resource identifiers that are not included in the initial recommended candidate set but meet the patent feature pattern from the platform patent database to generate a supplementary candidate subset. For example, this step can be further refined into the following sub-steps: Step S1541: Statistically analyze the viewing operation data, collection operation data, and consultation operation data in the user's real-time interactive information to determine the set of patent resource identifiers corresponding to the user's attention operation. By analyzing the viewing operation data, count the patents that the user has viewed more times; analyze the collection operation data to identify the patents that have been collected by the user; and organize the consultation operation data to find out the patents that the user has consulted. These patent resource identifiers related to the user's attention operation are integrated to form a set. In this process, corresponding weights can also be assigned based on the importance of different operations. For example, the collection operation may better reflect the user's attention than a simple viewing operation, and the priority of the patent corresponding to the collection operation in the set can be appropriately increased.

[0147] Step S1542: Extract the patent value dimension feature set corresponding to the patent resource identifier set of the user's attention operation, and calculate the mean of the patent value dimension feature set as the user preference feature pattern. Based on the patent resource identifier set, obtain the patent value dimension feature set corresponding to each patent from the platform database. Perform statistical calculations on each dimension in the patent value dimension feature set, add up the values ​​of each dimension and take the average value. Finally, obtain the user preference feature pattern, which represents the user's overall preference for patent features reflected in attention operations such as viewing, collecting, and consulting.

[0148] Step S1543: Traverse all patent resource identifiers in the platform patent database and calculate the degree of match between their patent value dimension feature set and the user preference feature pattern. For each patent resource identifier in the platform patent database, obtain its corresponding patent value dimension feature set. Use a similarity calculation method, such as cosine similarity or Euclidean distance, to compare the feature set with the user preference feature pattern to calculate the degree of match between them. During the calculation process, it is necessary to ensure that the dimensions of the feature set are consistent in order to perform accurate matching calculations.

[0149] Step S1544: Filter out patent resource identifiers whose matching degree meets the preset conditions and is not included in the initial recommended candidate set as a supplementary candidate subset. Compare the calculated matching degree with the preset conditions, which can be reasonably set according to business needs and user behavior data. Only select patent resource identifiers whose matching degree meets or exceeds the conditions. At the same time, check whether these identifiers exist in the initial recommended candidate set, and exclude identifiers that are already in it. Aggregate the remaining patent resource identifiers that meet the requirements to form a final supplementary candidate subset, which can provide users with more patent options that may meet their interests.

[0150] Step S155: Perform market value-added verification on the supplementary candidate subset, exclude patent resource identifiers whose market value-added characteristics do not meet the preset requirements, and generate a final candidate set. For example, this step can be refined into the following sub-steps: Step S1551: Extract the market value-added features corresponding to each patent resource identifier in the supplementary candidate subset from the patent lifecycle attribute data. The patent lifecycle attribute data contains a wealth of information about various patents. Patent resource identifiers are used to accurately locate and filter the data and extract the market value-added features corresponding to each patent in the supplementary candidate subset. This process ensures the accuracy and completeness of the data extraction, strictly matching according to the patent resource identifier to avoid incorrect association or omission of feature information. All market value-added features, such as information describing the fluctuation pattern of patent value as it changes with market demand and information on the possibility of value growth, are extracted.

[0151] Step S1552: Obtain the preset market value-added verification rules, which include a value growth possibility threshold and a fluctuation pattern stability requirement. The preset market value-added verification rules are formulated based on the platform's operating strategy, market analysis, and past business experience. The value growth possibility threshold is a pre-set boundary used to measure the possibility of future value growth of a patent. Patents below this threshold may not meet the requirements in terms of market value-added potential. The fluctuation pattern stability requirement involves the smoothness and predictability of the patent market value fluctuation pattern. For example, it can be defined in terms of the amplitude and frequency of fluctuations, reflecting the degree of stable development of the patent market value in the future. Obtain these rules from the system's configuration parameters to ensure that the rules are obtained accurately and to ensure the uniformity of subsequent verification standards.

[0152] Step S1553: Perform threshold judgment processing on the value growth possibility information in the market value-added characteristics, and screen out patent resource identifiers whose value growth possibility meets the value growth possibility threshold. Compare the value growth possibility information in the extracted market value-added characteristics with the set value growth possibility threshold one by one. For the value growth possibility information corresponding to each patent resource identifier, determine whether it reaches or exceeds the threshold. If the conditions are met, the patent resource identifier is retained; if the threshold is not reached, it is marked as not meeting the requirements. In this way, a set of patent resource identifiers that meet the requirements in terms of value growth possibility is preliminarily screened out, and patents with poor value growth potential are removed.

[0153] Step S1554: Perform stability analysis on the fluctuation pattern description information in the market value-added characteristics to determine whether it meets the fluctuation pattern stability requirements. For the fluctuation pattern description information corresponding to the patent resource identifiers retained after preliminary screening, stability analysis is performed from multiple dimensions. For example, the price fluctuation amplitude in the fluctuation pattern is statistically analyzed to determine whether the frequency of the fluctuation shows regularity. It is also possible to observe whether there are abnormal and drastic changes in the fluctuation. According to the preset fluctuation pattern stability requirements, evaluate whether the fluctuation pattern description information meets the standards. If the fluctuation pattern is too drastic and there are too many unstable factors, it is considered that the patent does not meet the stability requirements; if the fluctuation is relatively stable and highly predictable, it is determined to meet the rules.

[0154] Step S1555: retaining the patent resource identifier that satisfies both the value growth possibility threshold requirement and the fluctuation pattern stability requirement as a verified patent resource identifier.

[0155] After determining the value growth threshold and analyzing the stability of the fluctuation pattern, only those patent resource identifiers that meet both the value growth threshold and the fluctuation pattern stability requirements are retained. These verified identifiers are further identified and separated from the previous screening process to form verified patent resource identifiers, which contain patents that have both growth potential and stability in terms of market value-added.

[0156] Step S1556: Integrate the verified patent resource identifiers into a final candidate set. The patent resource identifiers retained after rigorous screening and verification are aggregated and integrated to form a complete set, the final candidate set. This set of patents represents patents that demonstrate good market value-added performance, possessing both significant potential for value growth and relatively stable market performance, and can provide users with patent recommendations with greater investment potential and market value. The final candidate set is organized and structured to facilitate efficient prioritization and display output.

[0157] Step S156: Prioritize the final candidate set according to the updated matching ranking information, generate a dynamically optimized patent recommendation sequence, and output the patent recommendation sequence to the user terminal interface to trigger patent operation interaction.

[0158] After obtaining the final candidate set, it is necessary to prioritize these patents in the recommendation list so that the patents that best meet the user's needs can be displayed first. This process is based on the updated matching ranking information. The updated matching ranking information is recalculated after considering real-time user feedback and reflects the latest matching degree between each patent and the user's needs.

[0159] Each patent resource identifier in the final candidate set is associated with the updated match ranking information, and the patents are sorted in descending order of match. Patents with higher match scores are ranked higher in the sequence, indicating that these patents are more in line with the user's current needs and preferences. Through this priority sorting process, a dynamically optimized patent recommendation sequence is generated.

[0160] The generated patent recommendation sequence contains a series of prioritized patent resource identifiers, each carrying the corresponding patent information. This patent recommendation sequence is then transmitted to the user terminal interface for display. The user terminal interface can present these patents in intuitive formats such as lists and cards, making them easy for users to browse and view. When users see these recommended patents on the terminal interface, further patent operation interactions are triggered, such as clicking to view patent details, inquiring about patent-related information, and communicating patent transaction intentions. This completes the business chain from user demand acquisition to patent recommendation and then to actual business interaction.

[0161] Figure 2 A schematic diagram illustrates exemplary hardware and software components of an AI intelligent recommendation system 100 for an online patent operation digital platform, which can implement the concepts of the present application, as provided in some embodiments of the present application. For example, the processor 120 can be used in the AI ​​intelligent recommendation system 100 for an online patent operation digital platform and perform the functions described in the present application.

[0162] The AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform can be a general-purpose server or a special-purpose server, both of which can be used to implement the AI ​​intelligent recommendation method applied to the online patent operation digital platform of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0163] For example, the AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform may include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROMs, or RAMs, or any combination thereof. Exemplarily, the AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform may also include program instructions stored in ROM, RAM, or other types of non-temporary storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform also includes an I / O interface 150 between the computer and other input and output devices.

[0164] For ease of explanation, only one processor is described in the AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform. However, it should be noted that the AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform in this application may also include multiple processors, so the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the AI ​​intelligent recommendation system 100 applied to the online patent operation digital platform executes step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0165] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the AI ​​intelligent recommendation method applied to the online patent operation digital platform as described above is implemented.

[0166] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. An AI intelligent recommendation method applied to an online patent operation digital platform, characterized in that: The method comprises: Obtaining full user behavior data and patent lifecycle attribute data, including historical user interaction records, demand expression records, and scenario context records; and obtaining full patent lifecycle attribute data, including patent technology evolution records, legal status change records, and market value fluctuation records; Performing demand intention analysis on the full amount of user behavior data to generate a user demand intention feature set, wherein the user demand intention feature set includes explicit demand features and potential development demand features at the current stage; Performing value dimension modeling on the patent full-cycle attribute data to generate a patent value dimension feature set, wherein the patent value dimension feature set includes a technology continuity feature, a legal stability feature, and a market value-added feature; Calling the AI ​​intelligent recommendation model to perform multi-dimensional correlation matching processing on the user demand intention feature set and the patent value dimension feature set to generate an initial recommendation candidate set, wherein the initial recommendation candidate set includes patent resource identifiers associated with the user demand features; Based on the real-time user interaction information, the initial recommendation candidate set is adaptively adjusted to generate a dynamically optimized patent recommendation sequence, and the patent recommendation sequence is output to the user terminal interface to trigger patent operation interaction operations.

2. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 1 is characterized in that: The acquisition of full user behavior data and full-cycle patent attribute data includes: Extract the time stamp, duration of stay, and access object information of the user's visit to the patent details page from the platform interaction log storage module as the user's historical interaction record; Extract the technical requirement description text, legal status consultation text, and market value assessment text submitted by users from the platform's demand feedback module as a record of user demand expression; Extract the terminal type information, access time period information and user interaction history records during user operations from the platform device information module as user scenario context records; Extract the patent family technology development path, subsequent patent citation relationship and related technology iteration information from the patent technology document database as a record of patent technology evolution; Synchronize the time series data of patent authorization status changes, patent maintenance status changes, patent invalidation status changes, and patent pledge status changes from the preset data interface as a record of patent legal status changes; Collect information on historical patent transaction price changes, regional expansion events of patent families, and migration of technology application scenarios from the patent trading market database as a record of patent market value fluctuations; The user historical interaction records, user demand expression records and user scenario context records are integrated into the user's full behavior data, and the patent technology evolution records, patent legal status change records and patent market value fluctuation records are integrated into the patent full cycle attribute data.

3. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 1 is characterized in that: The performing demand intention analysis on the full amount of user behavior data to generate a user demand intention feature set includes: Inputting the user historical interaction records into a time series feature extraction module, capturing the time dependency of user behavior through a time series analysis method, and generating user behavior time series features; Inputting the user demand expression record into a text semantic parsing module, extracting technical field keywords, legal status keywords, and market value keywords of interest to the user through a semantic analysis method, and generating user demand semantic features; Extracting the association between terminal type and demand intensity, and the association between access time period and demand urgency according to the user scenario context record to generate user scenario context features; Input the user behavior temporal features, user demand semantic features, and user scenario context features into a dynamic fusion module, dynamically assign corresponding fusion weights to each through a feature weight adjustment mechanism, and generate explicit demand features that reflect the needs of the current operation stage; Performing trend analysis on the historical output features of the dynamic fusion module to extract information on the evolution direction of user needs over time and generate potential development demand features that reflect the development direction of demand; The explicit demand features and the potential development demand features are dimensionally combined to generate a user demand intention feature set.

4. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 3 is characterized in that: The user demand expression record is input into the text semantic parsing module, and the technical field keywords, legal status keywords and market value keywords of the user's concern are extracted through the semantic analysis method to generate the user demand semantic features, including: Performing text segmentation processing on the user demand expression record to obtain a word element set including technical terms, legal terms and market terms; Inputting the word-unit set into a pre-trained language understanding model to generate a contextual semantic embedding vector for each word-unit; The importance of each word-unit embedding vector in expressing user needs is determined by the attention weight calculation method, and the word-unit attention weight distribution is generated; Filter out the words whose attention weight distribution meets the preset importance conditions as the technical field keywords, legal status keywords, and market value keywords that the user is concerned about; The contextual semantic embedding vectors of the keywords are weighted and summed to generate user demand semantic features containing the semantic focus of the user demand.

5. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 1 is characterized in that: The value dimension modeling process of the patent full-cycle attribute data to generate a patent value dimension feature set includes: Based on the patent technology evolution record, the position of the patent in the technology evolution chain and the subsequent technology derivative characteristics are extracted through the technology citation network analysis method to generate technology continuity characteristics; Input the patent legal status change record into the legal value assessment module, calculate the probability information of the patent maintaining a stable state in the future preset period of time through the state transition probability analysis method, and generate legal stability characteristics; Inputting the patent market value fluctuation record into the market value prediction module, extracting the fluctuation pattern of patent value as market demand changes through time series analysis method, and generating market value-added characteristics; The technical continuity characteristics, legal stability characteristics and market value-added characteristics are input into the value association analysis module, and the synergistic enhancement relationship between the characteristics of each dimension is extracted through the feature co-occurrence analysis method to generate a patent value dimension feature set with multi-dimensional association information.

6. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 5 is characterized in that: The patent market value fluctuation record is input into the market value prediction module, and the fluctuation pattern of the patent value as the market demand changes is extracted through the time series analysis method to generate market value-added characteristics, including: Performing time series decomposition processing on the historical transaction price change information in the patent market value fluctuation record to separate the trend change portion, the seasonal fluctuation portion, and the random fluctuation portion; Extrapolate and predict the trend change part by using the time series prediction method to generate a basic value prediction sequence of the patent within a preset future period; Extracting the correlation between the regional expansion events and transaction price fluctuations of the same patent family in the patent market value fluctuation records, and generating a regional expansion impact parameter on value; Extracting the correlation between technology application scenario migration events and transaction price fluctuations in the patent market value fluctuation records, and generating scenario migration impact parameters of technology application scenario migration on value; Perform weighted fusion processing on the basic value prediction sequence, the regional expansion impact parameter, and the scenario migration impact parameter to generate information describing the fluctuation pattern of patent value as market demand changes; Based on the fluctuation pattern description information, the value growth possibility information of the patent in the future time window is calculated as the market value-added feature.

7. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 1 is characterized in that: The calling of the AI ​​intelligent recommendation model to perform multi-dimensional correlation matching processing on the user demand intention feature set and the patent value dimension feature set to generate an initial recommendation candidate set includes: Input the user demand intention feature set and the patent value dimension feature set into the feature alignment module of the AI ​​intelligent recommendation model, align the time dimension to make the time resolution of the two consistent through time dimension alignment, and align the feature dimensions of the user demand intention feature set and the patent value dimension feature set through space dimension alignment; In the matching calculation module of the AI ​​intelligent recommendation model, the correlation matching degree between the user demand intention feature set and the feature set of each patent value dimension is calculated. The correlation matching degree is used to represent the degree of correlation between user demand and patent value; Filter the patent resource identifiers corresponding to the patent value dimension feature set whose correlation matching degree meets the preset matching conditions as the initial candidate subset; Performing legal stability filtering on the initial candidate subset, excluding patent resource identifiers whose legal stability characteristics do not meet preset requirements, and generating a legally valid candidate subset; The patent resource identifiers in the legally valid candidate subset are sorted by technology continuity, and the patent resource identifiers whose technology continuity characteristics meet the preset standards are retained to generate an initial recommended candidate set.

8. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 7 is characterized in that: In the matching calculation module of the AI ​​intelligent recommendation model, the correlation matching degree between the user demand intention feature set and the feature set of each patent value dimension is calculated. The correlation matching degree is used to represent the degree of correlation between user demand and patent value, including: In the matching calculation module of the AI ​​intelligent recommendation model, the user demand intention feature set is decomposed into an explicit demand sub-feature set at the current time point and a potential demand sub-feature set in the future time window; Decompose the feature set of patent value dimension into the basic value sub-feature set at the current time point and the value-added potential sub-feature set in the future time window; Calculate the matching degree between the explicit demand sub-feature set and the basic value sub-feature set as the matching degree of the current stage; Calculate the matching degree between the potential demand sub-feature set and the value-added potential sub-feature set as the matching degree in the future stage; Adjusting the weight of the future stage matching degree by a time impact factor, wherein the time impact factor decreases as the future time window is extended; The current stage matching degree and the adjusted future stage matching degree are comprehensively calculated to generate the associated matching degree.

9. The AI ​​intelligent recommendation method applied to the online patent operation digital platform according to claim 7 is characterized in that: The adaptive adjustment processing of the initial recommendation candidate set based on the real-time user interaction information to generate a dynamically optimized patent recommendation sequence includes: After displaying the patent resource identifiers in the initial recommendation candidate set on the user terminal interface, collecting the user's viewing operation data, collection operation data, and consultation operation data for each recommended patent as real-time user interaction information; Input the user's real-time interaction information into the feedback learning module of the AI ​​intelligent recommendation model, and update the weights of the potential development demand features of the user's demand intention feature set through the model parameter adjustment mechanism to enhance the feature expression of the user's actual concern needs; Recalculating the correlation matching degree of each patent resource identifier in the initial recommendation candidate set through the matching calculation module according to the updated user demand intention feature set, and generating updated matching ranking information; Extract patent feature patterns corresponding to user attention operations in real-time user interaction information, and use feature pattern matching methods to filter patent resource identifiers from the platform's patent database that are not included in the initial recommended candidate set but meet the patent feature pattern, to generate a supplementary candidate subset; Performing market value-added verification on the supplementary candidate subset, excluding patent resource identifiers whose market value-added characteristics do not meet preset requirements, and generating a final candidate set; The final candidate set is prioritized according to the updated matching ranking information to generate a dynamically optimized patent recommendation sequence.

10. An AI intelligent recommendation system applied to an online patent operation digital platform, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI ​​intelligent recommendation method applied to the online patent operation digital platform as described in any one of claims 1 to 9 above.

Images