Intelligent exhibition content recommendation method and system based on neural network

Abstract

The invention provides an exhibition content intelligent recommendation method and system based on a neural network, and relates to the technical field of data processing, and the method comprises the steps: collecting basic information, historical browsing records and real-time interaction behavior data of a user; performing feature extraction processing, and distinguishing static attribute features, historical behavior features and dynamic behavior features; an interest analysis result is dynamically generated through a neural network model, the static attribute features are used for judging basic interests of the user, the historical behavior features are used for describing long-term preferences of the user, the dynamic behavior features are used for capturing current attention key points of the user, and the interest analysis result comprises weight distribution of attention exhibition content of the user in different time periods; in combination with the data of the exhibition content library, performing screening and priority ranking on the exhibition content to obtain a personalized recommendation exhibition content set; outputting the personalized recommendation exhibition content set to a user interface; according to the invention, the accuracy of intelligent recommendation of exhibition content is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for intelligent recommendation of exhibition content based on neural networks. Background Technology

[0002] In existing technologies, intelligent recommendations for exhibition content mainly rely on content matching methods based on user profiles. In practice, systems typically collect users' basic information, interests, and historical browsing data, then use traditional collaborative filtering or content filtering algorithms to analyze and match exhibition information, thereby pushing potentially interesting exhibition content to users. Some systems introduce shallow neural network models to improve recommendation accuracy, but the overall approach still relies heavily on static features, with limited dynamic behavior mining.

[0003] During large-scale exhibitions, user interests and focuses constantly change with different activities and booths. Existing recommendation methods, lacking in-depth processing of real-time behavioral data, often fail to capture these changes in user needs in a timely manner. For example, at a technology exhibition, if a user temporarily focuses on a particular emerging technology area, the system may still recommend content based on past interests, resulting in information that does not match the user's current needs and negatively impacting the actual exhibition experience. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent recommendation of exhibition content based on neural networks, in order to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for intelligent recommendation of exhibition content based on neural networks, the method comprising: Collect users' basic information, browsing history, and real-time interaction data, and integrate them to form the user's original dataset; Feature extraction processing is performed on the user's original dataset to distinguish between static attribute features, historical behavior features, and dynamic behavior features, resulting in a multi-dimensional feature set. Based on a multi-dimensional feature set, interest analysis results are dynamically generated. Static attribute features are used to determine the user's basic interests, historical behavior features are used to characterize the user's long-term preferences, and dynamic behavior features are used to capture the user's current focus. The interest analysis results include the weight distribution of the exhibition content that the user focuses on in different time periods. Based on the results of the interest analysis and combined with the data from the exhibition content library, the exhibition content is filtered and prioritized to obtain a personalized set of recommended exhibition content. Output a personalized set of recommended exhibition content to the user interface, and update the user's original dataset in real time based on the user's actual feedback on the recommended exhibition content.

[0006] Preferably, feature extraction processing is performed on the original user dataset to distinguish between static attribute features, historical behavior features, and dynamic behavior features, resulting in a multi-dimensional feature set, including: The basic information in the user's original dataset is parsed to identify and extract static attributes related to the exhibition, and these attributes are standardized to form structured static attribute feature data. The historical browsing records in the user's original dataset are analyzed one by one to count the exhibition content visited by the user and the corresponding dwell time. The data is then normalized and grouped by feature to form structured historical behavior feature data. The real-time interactive behavior data in the user's original dataset is sorted by timestamp, and different types of real-time interactive behavior are divided into preset time windows. The number of times different types of user interactive behavior occur for various exhibition contents is counted to form dynamic behavioral feature data. Static attribute feature data, historical behavior feature data, and dynamic behavior feature data are merged to form a multi-dimensional feature set.

[0007] Preferably, the neural network model includes: The input layer receives static attribute features, historical behavior features, and dynamic behavior features extracted from the user's original dataset and encodes them into an input vector. The basic interest score hidden layer is used to match static attribute features with exhibition content tags in the exhibition content library at the semantic or keyword level, calculate the similarity score for each piece of exhibition content, and generate a basic interest score. The hidden layer for dynamic weight extraction is used to count the frequency of exhibition content involved in each type of interactive behavior in the dynamic behavior features. The number of times each piece of exhibition content appears in the user operation is recorded as the initial value of dynamic weight. The initial values ​​of dynamic weight of all exhibition content are normalized to form dynamic interest distribution data. The fusion weighted hidden layer is used to generate an interest score for each piece of exhibition content by weighting it according to its corresponding basic interest score and the weight value in the dynamic interest distribution data through a preset first ratio factor. The output layer is used to sort the various exhibition contents from high to low according to their interest scores, and generate the sorting results as the interest analysis results.

[0008] Preferably, based on the interest analysis results and combined with data from the exhibition content library, the exhibition content is filtered and prioritized to obtain a personalized recommended set of exhibition content, including: It iterates through all exhibition projects in the exhibition content library, compares the exhibition content tags of each project with the high-interest items in the interest analysis results, and selects the exhibition projects that match the user's current interests as a candidate exhibition content set. For each item in the candidate exhibition content set, the following scoring steps are performed sequentially: Extract the interest score corresponding to the content of the exhibition from the interest analysis results, and use it as the original interest score value; The number of interactions with the exhibition content among all users is counted and used as the original popularity score. Calculate the distance between the start time of the exhibition content and the current time, and obtain the original time score value based on the linear reciprocal relationship; The min-max normalization method is used to normalize the interest score, popularity score, and time score of all candidate exhibition content, so that each score is mapped to the interval between zero and one, resulting in normalized interest score, normalized popularity score, and normalized time score. According to the preset second proportional factor, the normalized interest score, normalized popularity score and normalized time score are multiplied by the corresponding proportional factor, and the resulting weighted values ​​are superimposed to obtain the priority score of the exhibition content. Based on priority scores, the exhibition content in the candidate exhibition content set is sorted from high to low, and a preset number of exhibition content with the highest priority scores are selected and organized into a personalized recommended exhibition content set, with a unique number attached.

[0009] Preferably, the user's original dataset is updated in real time based on the user's actual feedback on the recommended exhibition content, including: Real-time monitoring of user interaction behavior for each exhibition content in the personalized recommended exhibition content set; recording operation type, operation time and exhibition content number involved, forming an interaction behavior feedback dataset; Based on the interactive behavior feedback dataset, the cumulative statistics of user interactions with the same exhibition content at different time periods are compiled to form a feedback behavior statistics table. The data in the feedback behavior statistics table is merged and updated with the historical and dynamic behavior data in the user's original dataset.

[0010] Preferably, the static attribute features are semantically or keyword-level matched with the exhibition content tags in the exhibition content library, and a similarity score is calculated for each piece of exhibition content to generate a basic interest score, including: Map each attribute in the user's static attribute feature vector to a set of static keywords; For each piece of exhibition content in the exhibition content library, extract its content tags and break them down into keywords to form a set of keywords to be matched; The static keyword set is matched with the keyword set to be matched to obtain the matching degree; Based on the matching degree, the number of identical keywords is counted, and a first fixed score is assigned to each identical keyword. The scores of all identical keywords are then summed. For keywords that are not completely identical but have a similarity higher than a preset similarity threshold, a second fixed score is assigned. The sum of the first fixed score and the second fixed score is used as the original similarity score between the exhibition content and the user's static attribute features; Divide the original similarity scores of all exhibition content by the number of corresponding exhibition content tag keywords to obtain the normalized basic interest score.

[0011] Preferably, frequency statistics are performed on the exhibition content involved in each type of interactive behavior in the dynamic behavior characteristics, and the number of times each piece of exhibition content appears in the user's operation is recorded as the initial value of the dynamic weight, including: Based on dynamic behavioral characteristics, all user interaction records within each time window are extracted and categorized according to exhibition content and interaction type; For each type of interactive behavior in each exhibition content, the number of times the behavior occurs within each time window is counted, and different weight values ​​are assigned to different types of interactive behavior. The weighted frequency of each interaction type is summed under the same exhibition content to obtain the total weighted frequency of a single exhibition content within the time window; The initial dynamic weight of each exhibition content is obtained by summing the weighted frequency of each behavior across all time windows.

[0012] Preferably, the initial dynamic weights of all exhibition content are standardized and normalized to ensure that all dynamic weight values ​​are within a uniform range, forming dynamic interest distribution data, including: Based on the initial dynamic weights of all exhibition content, determine the maximum and minimum values ​​among them; For each exhibition content, the initial dynamic weight value is combined with the maximum and minimum values ​​in the initial dynamic weight value, and then the standard normalization is performed to map all dynamic weight values ​​to the interval between zero and one, so as to obtain the normalized dynamic weight value. The normalized dynamic weight values ​​of all exhibition content are merged to form dynamic interest distribution data.

[0013] Preferably, for each piece of exhibition content, an interest score is generated by weighting and summing it according to its corresponding basic interest score and the weight value in the dynamic interest distribution data using a preset first proportional factor. This first proportional factor is used to adjust the influence intensity of static and dynamic features, including: Extract the basic interest score and weight value from the dynamic interest distribution data for each piece of exhibition content; Based on the first proportional factor, the basic interest score and the dynamic interest weight value are multiplied by the corresponding proportional factor to obtain two weighted values. The two weighted values ​​are added together to obtain the interest score for the exhibition content.

[0014] Secondly, a neural network-based intelligent recommendation system for exhibition content, the system comprising: The data acquisition module is used to collect users' basic information, historical browsing records, and real-time interaction behavior data, and integrate them to form the user's original dataset; The feature extraction module is used to extract features from the user's original dataset, distinguishing between static attribute features, historical behavior features, and dynamic behavior features to obtain a multi-dimensional feature set. The interest analysis module is used to dynamically generate interest analysis results based on a multi-dimensional feature set. Static attribute features are used to determine the user's basic interests, historical behavior features are used to characterize the user's long-term preferences, and dynamic behavior features are used to capture the user's current focus. The interest analysis results include the weight distribution of the exhibition content that the user focuses on in different time periods. The content filtering and sorting module is used to filter and prioritize exhibition content based on the results of interest analysis and data from the exhibition content library, so as to obtain a personalized set of recommended exhibition content. The recommended output module is used to output a personalized set of recommended exhibition content to the user interface; The feedback update module is used to update the user's original dataset in real time based on the user's actual feedback on the recommended exhibition content.

[0015] The above-described solution of the present invention has at least the following beneficial effects: This invention collects and integrates users' basic information, historical browsing records, and real-time interaction behavior data to construct a raw user dataset covering static attributes, long-term interests, and immediate focus, significantly improving the comprehensiveness and dynamism of user profiles. Based on this, a multi-dimensional feature extraction method is employed to organically combine static attribute features, historical behavior features, and dynamic behavior features, inputting them into a neural network model for deep modeling. The neural network model can automatically identify and fuse complex relationships between different features, dynamically generating interest analysis results, and achieving accurate modeling and time-series tracking of user interest states.

[0016] Compared to traditional recommendation algorithms that rely solely on static user profiles and historical data and lag in responding to real-time behavioral changes, this invention can capture changes in user interests in real time and automatically adjust the priority of recommended content. The system can not only determine basic interests based on static attribute features and characterize long-term user preferences using historical behavioral features, but also reflect the user's immediate focus at the exhibition through dynamic behavioral features. By jointly filtering and prioritizing the interest analysis results with the exhibition content library, the system can output personalized recommendations that highly match the user's current needs, and continuously iterate and optimize based on actual user feedback, achieving a closed-loop recommendation system and intelligent self-adaptation.

[0017] For example, in response to users' temporary interests arising from emerging technological trends at large exhibitions, the system can use neural network models to promptly identify and dynamically increase the recommendation weight of relevant content. This ensures that the information users receive always aligns with their changing interests, significantly improving the problems of lagging content and weak relevance in existing recommendation systems. In summary, this invention not only improves the real-time performance, accuracy, and personalization of exhibition content recommendations but also greatly enhances users' exhibition experience and information acquisition efficiency, demonstrating significant practical value and promotional significance. Attached Figure Description

[0018] Figure 1 This is a flowchart of an intelligent recommendation method for exhibition content based on neural networks provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] like Figure 1 As shown, embodiments of the present invention propose an intelligent recommendation method for exhibition content based on neural networks, the method comprising: Collect users' basic information, browsing history, and real-time interaction data, and integrate them to form the user's original dataset; Feature extraction processing is performed on the user's original dataset to distinguish between static attribute features, historical behavior features, and dynamic behavior features, resulting in a multi-dimensional feature set. Based on a multi-dimensional feature set, interest analysis results are dynamically generated. Static attribute features are used to determine the user's basic interests, historical behavior features are used to characterize the user's long-term preferences, and dynamic behavior features are used to capture the user's current focus. The interest analysis results include the weight distribution of the exhibition content that the user focuses on in different time periods. Based on the results of the interest analysis and combined with the data from the exhibition content library, the exhibition content is filtered and prioritized to obtain a personalized set of recommended exhibition content. Output a personalized set of recommended exhibition content to the user interface, and update the user's original dataset in real time based on the user's actual feedback on the recommended exhibition content.

[0021] In this embodiment of the invention, basic user information, historical browsing records, and real-time interaction behavior data are collected and integrated to form a user raw dataset. This allows the recommendation system to comprehensively acquire the user's static attributes, long-term interests, and immediate concerns. By extracting multi-dimensional features from this data, static attribute features, historical behavior features, and dynamic behavior features can be simultaneously distinguished and modeled, significantly improving the detail and dynamism of the user profile. Inputting this multi-dimensional feature set into a neural network model leverages the neural network's ability to model complex nonlinear relationships to automatically generate interest analysis results. This not only accurately reflects the user's basic interests but also dynamically captures changes in the user's interests and concerns in the short term. The interest analysis results are further combined with data from the exhibition content library to achieve precise filtering and priority ranking of exhibition content, ultimately outputting a personalized set of recommended exhibition content. The recommendation results can iterate in real time based on user feedback, and the system has adaptive optimization capabilities. Through the above methods, the accuracy and real-time performance of personalized recommendations are significantly improved, ensuring that users continuously receive exhibition content recommendations that highly match their interests and are dynamically updated, greatly enhancing the user's exhibition viewing experience and information acquisition efficiency.

[0022] This involves collecting users' basic information, browsing history, and real-time interaction data, and integrating them to form a raw user dataset, specifically including: During the system initialization phase, basic information for each user is first collected, including but not limited to static attributes such as age range, gender, occupation type, industry, and city of residence. This process can be achieved through user registration, synchronization of third-party authentication information, or user self-supplementation.

[0023] Subsequently, the system continuously records the user's browsing history at the exhibition, collecting detailed information on the exhibition content the user visited, including the start and end times of each visit, the duration of stay, and the unique identifiers of the exhibition content involved.

[0024] Meanwhile, the system collects various interactive behavior data in real time when users actually use it, including but not limited to browsing, clicking, collecting, commenting, sharing and other behaviors. Each behavior is recorded synchronously with information such as the timestamp of the operation, the object of the operation (the exhibition content number), and the type of behavior.

[0025] All the aforementioned static attribute data, historical browsing history data, and real-time interaction behavior data, after being organized using a unified data structure and sorted chronologically, ultimately form the original user dataset for each user. This dataset provides a complete and structured data foundation for subsequent feature extraction and intelligent recommendation analysis.

[0026] In a preferred embodiment of the present invention, feature extraction processing is performed on the original user dataset to distinguish between static attribute features, historical behavior features, and dynamic behavior features, resulting in a multi-dimensional feature set, including: The basic information in the user's original dataset is parsed to identify and extract static attributes related to the exhibition, and these attributes are standardized to form structured static attribute feature data. The historical browsing records in the user's original dataset are analyzed one by one to count the exhibition content visited by the user and the corresponding dwell time. The data is then normalized and grouped by feature to form structured historical behavior feature data. The real-time interactive behavior data in the user's original dataset is sorted by timestamp, and different types of real-time interactive behavior are divided into preset time windows. The number of times different types of user interactive behavior occur for various exhibition contents is counted to form dynamic behavioral feature data. Static attribute feature data, historical behavior feature data, and dynamic behavior feature data are merged to form a multi-dimensional feature set.

[0027] In this embodiment of the invention, feature extraction processing is performed on the original user dataset to distinguish between static attribute features, historical behavior features, and dynamic behavior features, laying a solid data foundation for subsequent neural network interest modeling and personalized recommendations. By parsing and standardizing user basic information, the system can accurately and systematically extract static attribute information related to the exhibition, such as age range, occupation type, industry, and city of residence, thereby effectively avoiding problems such as inconsistent formats or ambiguous descriptions that may exist in the information collection process. Analyzing and normalizing historical browsing records one by one not only allows for a detailed reconstruction of the user's past attention distribution and interest preferences on various exhibition contents, but also further characterizes the user's sustained attention to different content by statistically analyzing indicators such as dwell time, forming high-quality historical behavior feature data. By processing real-time interactive behavior data by time window and statistically analyzing behavior types, the system can dynamically capture the user's real-time interest changes and behavioral hotspots during the exhibition viewing process, providing basic data support for subsequent in-depth modeling of dynamic behavior features. Ultimately, merging static attribute feature data, historical behavioral feature data, and dynamic behavioral feature data can fully integrate users' long-term, short-term, and immediate interests, providing comprehensive and structured data input for multi-dimensional feature modeling and subsequent vectorization processing of neural network input layers. This greatly improves the accuracy of personalized interest identification and the adaptability of the exhibition content recommendation system.

[0028] This involves parsing the basic information in the user's original dataset, identifying and extracting static attributes related to the exhibition, and standardizing them to form structured static attribute feature data, specifically including: First, the system reads and parses the basic information fields in the user's original dataset, identifying basic attributes such as name, age, gender, occupation, industry, and city of residence for common data formats (e.g., JSON, tables, database fields). For each attribute, the system converts the original description into a standard category according to preset standard encoding rules or an industry-standard dictionary. For example, "software engineer" can be categorized as "IT professional," and "Beijing" as "North China region."

[0029] During the parsing process, the system handles anomalies such as missing fields and non-standard expressions, ensuring data consistency through methods such as default value completion and synonym normalization. All valid static attributes are ultimately encoded in a fixed order and uniform format (such as numbering, stringing, or labeling), and output as structured static attribute feature data for subsequent feature fusion and model input.

[0030] This involves analyzing each historical browsing record in the user's original dataset, statistically analyzing the exhibition content visited by the user and the corresponding dwell time, and then normalizing and grouping these records to form structured historical behavioral feature data, specifically including: The system iterates through the user's browsing history, extracting key information such as the exhibition content identifier and start and end times for each access record. For multiple access records of the same exhibition content, the system can merge and statistically analyze them, accumulating the total access duration and number of visits for that content. To eliminate absolute differences in user historical behavior across different content, the system normalizes the number of visits and cumulative dwell time for all exhibition content (e.g., by proportional mapping based on total duration or maximum duration) to ensure all feature values ​​are within a comparable range.

[0031] Subsequently, the system groups and organizes the statistical results according to characteristics such as exhibition content category and exhibition theme, forming a historical behavior feature table. This table records users' historical interest preferences in various categories or exhibition content, and the final output is structured historical behavior feature data.

[0032] Specifically, the real-time interactive behavior data in the user's original dataset is sorted by timestamp, and different types of real-time interactive behavior are divided into preset time windows. The number of times different types of user interactive behavior occur for various exhibition contents is counted to form dynamic behavioral feature data, which specifically includes: The system first sorts all real-time user interaction data (such as clicks, favorites, comments, shares, etc.) in ascending order by the timestamp of the event. Based on preset time window parameters (such as every 5 minutes, 30 minutes, 1 hour), the system divides the data into several windows based on time slices. Within each time window, the system counts the number of occurrences of each piece of exhibition content according to different interaction types (such as clicks, favorites, etc.).

[0033] The statistical results are indexed by "Exhibition Content Number - Time Window - Behavior Type" and output as a dynamic behavior feature matrix. Each cell in the matrix represents the frequency of behavior for a specific content within a specific time period and under a specific interaction type, facilitating subsequent behavior weighting, time-series modeling, or interest trend analysis. All statistical results are output in a unified structure as dynamic behavior feature data for subsequent processing by the neural network model.

[0034] In a preferred embodiment of the present invention, the neural network model includes: The input layer receives static attribute features, historical behavior features, and dynamic behavior features extracted from the user's original dataset and encodes them into an input vector. The basic interest score hidden layer is used to match static attribute features with exhibition content tags in the exhibition content library at the semantic or keyword level, calculate the similarity score for each piece of exhibition content, and generate a basic interest score. The hidden layer for dynamic weight extraction is used to count the frequency of exhibition content involved in each type of interactive behavior in the dynamic behavior features. The number of times each piece of exhibition content appears in the user operation is recorded as the initial value of dynamic weight. The initial values ​​of dynamic weight of all exhibition content are normalized to form dynamic interest distribution data. The fusion weighted hidden layer is used to generate an interest score for each piece of exhibition content by weighting it according to its corresponding basic interest score and the weight value in the dynamic interest distribution data through a preset first ratio factor. The output layer is used to sort the various exhibition contents from high to low according to their interest scores, and generate the sorting results as the interest analysis results.

[0035] In this embodiment of the invention, by setting an input layer, a basic interest score hidden layer, a dynamic weight extraction hidden layer, a fusion weighted hidden layer, and an output layer in the neural network model, hierarchical perception and fusion modeling of multi-dimensional user features are achieved. The input layer can effectively receive and encode static attribute features, historical behavior features, and dynamic behavior features, ensuring that all user features are fully incorporated into the model analysis. The basic interest score hidden layer can perform semantic or keyword-level matching between static attribute features and exhibition content tags, realizing the automatic quantification of users' basic interests. The dynamic weight extraction hidden layer can perform frequency statistics, weighting, and normalization of various interactive behaviors of users in different time windows, effectively capturing the user's current real-time focus. The fusion weighted hidden layer weights and fuses the basic interest score and dynamic interest distribution data, realizing the automatic synthesis of multi-dimensional interest signals. The model can flexibly adjust the influence intensity of each feature according to a set first proportion factor. The output layer automatically sorts according to the comprehensive interest scores of various exhibition contents, ensuring that the recommended content can accurately and dynamically respond to changes in user needs. This hierarchical structure fully leverages the advantages of neural networks in modeling feature combinations and complex relationships, significantly improving the intelligent decision-making capabilities and recommendation relevance of recommendation systems in multi-source data environments.

[0036] In a preferred embodiment of the present invention, based on the results of interest analysis and combined with data from the exhibition content library, the exhibition content is filtered and prioritized to obtain a personalized recommended set of exhibition content, including: It iterates through all exhibition projects in the exhibition content library, compares the exhibition content tags of each project with the high-interest items in the interest analysis results, and selects the exhibition projects that match the user's current interests as a candidate exhibition content set. For each item in the candidate exhibition content set, the following scoring steps are performed sequentially: Extract the interest score corresponding to the content of the exhibition from the interest analysis results, and use it as the original interest score value; The number of interactions with the exhibition content among all users is counted and used as the original popularity score. Calculate the distance between the start time of the exhibition content and the current time, and obtain the original time score value based on the linear reciprocal relationship; The min-max normalization method is used to normalize the interest score, popularity score, and time score of all candidate exhibition content, so that each score is mapped to the interval between zero and one, resulting in normalized interest score, normalized popularity score, and normalized time score. According to the preset second proportional factor, the normalized interest score, normalized popularity score and normalized time score are multiplied by the corresponding proportional factor, and the resulting weighted values ​​are superimposed to obtain the priority score of the exhibition content. Based on priority scores, the exhibition content in the candidate exhibition content set is sorted from high to low, and a preset number of exhibition content with the highest priority scores are selected and organized into a personalized recommended exhibition content set, with a unique number attached.

[0037] In this embodiment of the invention, by combining the results of interest analysis with data from the exhibition content library, efficient filtering and prioritization of exhibition content are achieved. The system can traverse all exhibition projects and, by comparing exhibition content tags with high-interest items in the interest analysis results, filter out a set of candidate exhibition content that highly matches the user's current interests. For each candidate exhibition content, the system calculates an interest score, a popularity score, and a time score, and uses a maximum-minimum normalization method to ensure that each score value is within a uniform numerical range. Furthermore, it weights these scores using a set proportional factor to derive a priority score for each item. This multi-dimensional, normalized, and weighted scoring mechanism allows the system to not only reflect the user's personalized interests but also take into account the real-time popularity and time urgency of exhibition content, achieving scientific and reasonable ranking of recommended content. For example, a technical lecture that is about to start but has high popularity and clear user interest will be given higher priority, effectively improving the relevance and timeliness of the recommendations and bringing users a higher-value exhibition experience.

[0038] This involves iterating through all exhibition projects in the exhibition content library, comparing the exhibition content tags of each project with the high-interest items in the interest analysis results, and selecting exhibition projects that match the user's current interests as a candidate exhibition content set, specifically including: The system reads each exhibition item sequentially from the exhibition content library. For each piece of exhibition content, the system first extracts its corresponding content tags, which may include attribute information such as theme category, related fields, keywords, and technical direction.

[0039] Subsequently, the system identifies the user's high-interest items, which are typically content categories or keywords that rank highly in the interest analysis results and have a high weight distribution. The system then compares the exhibition content tags with the user's high-interest items. The comparison can be achieved using various algorithms, such as keyword exact matching, semantic similarity judgment, and tag intersection statistics. If a certain exhibition content tag contains words or categories that are identical to or highly semantically related to the high-interest items, the exhibition content is determined to match the user's current interests.

[0040] All exhibition content items deemed to match user interests are collected and organized by the system into a candidate exhibition content set. This set serves as the basis for subsequent priority scoring and the final recommendation list, ensuring that the recommendation range highly aligns with users' real-time interest changes.

[0041] The preset second scaling factor specifically includes: During the personalized recommendation and ranking phase, the system calculates a normalized interest score, a normalized popularity score, and a normalized time score for each item in the candidate exhibition content set. To flexibly adjust the influence weight of different scores in the final priority score, the system is designed with a second proportional factor for each of the three scores.

[0042] The optimal configuration of the three scoring weights can be determined under different business scenarios through methods such as historical data backtesting and A / B testing. For example, if recommended content should prioritize responding to user interests, a higher proportion of the interest score can be allocated; if it is necessary to balance the popularity of exhibition content with time urgency, the proportion of the popularity score or time score can be increased accordingly.

[0043] In the actual recommendation process, the system multiplies the three normalized scores of each exhibition item by the corresponding scaling factor, and then sums the three weighted results to obtain the priority score of that exhibition item. By reasonably setting the second scaling factor, the system can dynamically adapt to different user needs and exhibition business objectives, thereby improving the scientific nature and flexibility of the recommendation ranking.

[0044] In a preferred embodiment of the present invention, the user's original dataset is updated in real time based on the user's actual feedback on the recommended exhibition content, including: Real-time monitoring of user interaction behavior for each exhibition content in the personalized recommended exhibition content set; recording operation type, operation time and exhibition content number involved, forming an interaction behavior feedback dataset; Based on the interactive behavior feedback dataset, the cumulative statistics of user interactions with the same exhibition content at different time periods are compiled to form a feedback behavior statistics table. The data in the feedback behavior statistics table is merged and updated with the historical and dynamic behavior data in the user's original dataset.

[0045] In this embodiment of the invention, by real-time monitoring of user interaction behavior for each exhibition content in the personalized recommendation content set, the system can continuously and accurately record the actual operation type, operation time, and exhibition content number involved, thereby dynamically reflecting the user's current interest changes and behavioral feedback. This real-time feedback data is automatically aggregated and organized by the system to construct an interactive behavior feedback dataset, providing the latest and most authentic user behavior basis for subsequent interest modeling. The system further accumulates and statistically analyzes multiple interactions with the same exhibition content over different time periods, which not only quantifies the user's sustained attention to specific exhibition content but also subdivides the contribution of different types of operations to overall interest preferences, achieving a comprehensive capture of the user's interest evolution process. By merging and updating the feedback behavior statistics table with the original historical and dynamic behavior data, the system ensures that the user profile closely reflects the user's actual interests and behavioral trajectory at all times, significantly improving the recommendation system's response speed and adaptability to changes in user interests. For example, when a user frequently clicks or favorites a new exhibition content, the system can immediately incorporate this behavior into the user data structure, prompting timely adjustments to subsequent recommendation results to better match the user's immediate needs and personalized preferences. The overall solution enables timely and dynamic updates of user interest data, effectively overcoming the limitations of existing technologies that can only rely on static or historical interest data and cannot reflect users' true preferences in a timely manner. It provides the exhibition content recommendation system with powerful self-learning and intelligent optimization capabilities.

[0046] Specifically, based on the interactive behavior feedback dataset, the system accumulates and statistically analyzes the instances where users interact with the same exhibition content multiple times within different time periods, forming a feedback behavior statistics table, which includes: After receiving the user interaction behavior feedback dataset, the system first parses each interaction record in the dataset, extracting information such as the exhibition content number, operation type (e.g., click, favorite, comment, share), and the specific time the operation occurred. Based on preset time period division rules (e.g., every 5 minutes, 30 minutes, 1 hour, etc.), all operation times are categorized into the corresponding time intervals.

[0047] For the same exhibition content, the system accumulates and statistically analyzes all interactive behaviors of the same user on that content, based on the operation type and time period. For example, if a user clicks and favorites multiple times on a certain exhibition content at different time periods, the system counts the number of each type of operation in each time period and then adds them up.

[0048] The statistical process supports parallel processing of multiple behavior types, and a separate cumulative counter can be set for each type of operation. Finally, the system organizes the statistical results into a feedback behavior statistics table, indexed by "Exhibition Content Number - Operation Type - Time Period". This statistics table can be saved in a structured data format (such as a database table or data frame), with fields including exhibition content number, operation type, time period identifier, and cumulative number of operations.

[0049] Through the above processing, the feedback behavior statistics table can not only reflect users' continuous interest and behavioral intensity in the same exhibition content at different time periods, but also provide high-quality time-series data input for subsequent user profile updates, interest weight adjustments, and dynamic training of recommendation models.

[0050] In a preferred embodiment of the present invention, static attribute features are matched semantically or keyword-wise with exhibition content tags in the exhibition content library, and a similarity score is calculated for each piece of exhibition content to generate a basic interest score, including: Map each attribute in the user's static attribute feature vector to a set of static keywords; For each piece of exhibition content in the exhibition content library, extract its content tags and break them down into keywords to form a set of keywords to be matched; The static keyword set is matched with the keyword set to be matched to obtain the matching degree; Based on the matching degree, the number of identical keywords is counted, and a first fixed score is assigned to each identical keyword. The scores of all identical keywords are then summed. For keywords that are not completely identical but have a similarity higher than a preset similarity threshold, a second fixed score is assigned. The sum of the first fixed score and the second fixed score is used as the original similarity score between the exhibition content and the user's static attribute features; Divide the original similarity scores of all exhibition content by the number of corresponding exhibition content tag keywords to obtain the normalized basic interest score.

[0051] In this embodiment of the invention, a basic interest scoring hidden layer of a neural network model achieves deep feature fusion and semantic matching between static attribute features and exhibition content tags. This hidden layer not only accepts the static attribute features encoded by the input layer, but also automatically embeds multidimensional tags such as user age, occupation, and industry within the network, and efficiently aligns them with the keyword set formed by the exhibition content tags. Through parameter training, the network assigns higher weights to perfectly matching keywords and adaptively scores semantically similar but not identical tags using a built-in similarity calculation submodule. The basic interest scoring hidden layer can employ a fully connected layer or a self-attention mechanism structure to enhance sensitivity and generalization ability regarding keyword relevance. Finally, this hidden layer automatically outputs a normalized basic interest score, reflecting both stable user interests and adapting to changes in different user tag structures, thus providing an accurate interest expression basis for subsequent interest fusion layers and final recommendation ranking. For example, if a user's occupation tag is "software engineer," even if the exhibition content tag is "IT development" or "programming," the network can still make intelligent recommendations through deep semantic matching, rather than being limited to literal consistency, thus greatly improving the semantic understanding depth and adaptability of personalized recommendations.

[0052] In a preferred embodiment of the present invention, frequency statistics are performed on the exhibition content involved in each type of interactive behavior in the dynamic behavior features, and the number of times each piece of exhibition content appears in the user operation is recorded as the initial value of the dynamic weight, including: Based on dynamic behavioral characteristics, all user interaction records within each time window are extracted and categorized according to exhibition content and interaction type; For each type of interactive behavior in each exhibition content, the number of times the behavior occurs within each time window is counted, and different weight values ​​are assigned to different types of interactive behavior. The weighted frequency of each interaction type is summed under the same exhibition content to obtain the total weighted frequency of a single exhibition content within the time window; The initial dynamic weight of each exhibition content is obtained by summing the weighted frequency of each behavior across all time windows.

[0053] In this embodiment of the invention, the dynamic weight extraction hidden layer, as the core behavior modeling unit in the neural network model, can effectively integrate and analyze the real-time interactive behavior characteristics of users. This hidden layer can combine a fully connected layer and a temporal feature extraction layer structure to uniformly vectorize and encode various user interactions such as browsing, clicking, and saving in different time windows, and automatically learn the influence of different interaction types and behavior frequencies on exhibition content recommendations using network parameters. The dynamic weight extraction hidden layer can structurally embed a normalization processing module to achieve automatic proportional adjustment of behavior weights, ensuring that the final output dynamic interest distribution data remains within a uniform numerical range regardless of the number of user behaviors. Through the temporal stacking of multiple time windows and the interaction of behavioral features, the network can efficiently capture the changing trends of users' short-term interests and increase the weight of recently high-frequency or high-weight interactive content in the recommendation ranking. For example, if a user clicks on a new exhibition theme multiple times in a short period, even with limited historical data, the dynamic weight extraction hidden layer can promptly increase the recommendation priority of that content, achieving immediate response and content adjustment to interest hotspots, greatly enhancing the sensitivity of the recommendation system to real-time changes in user interests and the timeliness of the recommendation results.

[0054] Specifically, for each type of interactive behavior within each exhibition content, the number of occurrences of each behavior is counted within each time window, and different weight values ​​are assigned to different interactive behavior types, including: Interactive behavior organization and time window segmentation: First, all real-time user interaction data (such as clicks, favorites, comments, shares, etc.) is sorted according to the timestamps of the interaction events. Then, based on preset time window lengths (such as every 5 minutes, 30 minutes, 1 hour, etc.), all interaction events are divided into different time window intervals. For each exhibition content, the system records all interaction events and their types that occur within each time window.

[0055] Behavior type statistics: Within each time window, for each piece of exhibition content, the system counts the number of occurrences of various interactive behaviors. For example, for "Exhibition A", the system can count the number of clicks, favorites, comments, etc. within a certain time window, and save the statistical results as a behavior frequency vector.

[0056] Neural network input encoding: The frequency vector of behaviors obtained by the above window statistics is used as the dynamic behavior feature input and encoded into the dynamic behavior feature input layer of the neural network.

[0057] During the network training phase, the neural network model can automatically learn the weights of different behavior types on content recommendation based on historical user data. For example, in some scenarios, the act of saving may reflect interest more than the act of clicking, and the neural network can dynamically adjust the weight parameters of each behavior type through training.

[0058] Differentiated weight assignment: For each type of interaction behavior, the system has trainable weight parameters in the neural network structure. The model automatically optimizes the weight parameters of each behavior type through backpropagation algorithm based on the actual recommendation results, so that the final recommendation interest score can fully reflect the real business contribution of each type of behavior.

[0059] For example, the model can automatically increase the weight of "commenting" behavior if historical data shows that commenting behavior is strongly correlated with content preferences.

[0060] Output weighted behavioral features: The neural network weights the frequency of behaviors with automatically learned weight parameters to generate a dynamic interest contribution value for each exhibition content within each time window, providing a feature basis for subsequent recommendation scoring or comprehensive interest modeling.

[0061] In a preferred embodiment of the present invention, the initial dynamic weight values ​​of all exhibition content are standardized and normalized to ensure that all dynamic weight values ​​are within a uniform numerical range, forming dynamic interest distribution data, including: Based on the initial dynamic weights of all exhibition content, determine the maximum and minimum values ​​among them; For each exhibition content, the initial dynamic weight value is combined with the maximum and minimum values ​​in the initial dynamic weight value, and then the standard normalization is performed to map all dynamic weight values ​​to the interval between zero and one, so as to obtain the normalized dynamic weight value. The normalized dynamic weight values ​​of all exhibition content are merged to form dynamic interest distribution data.

[0062] In this embodiment of the invention, the dynamic weight normalization hidden layer in the neural network model can automatically normalize the initial dynamic weight values ​​of each exhibition content output by the dynamic weight extraction hidden layer. This hidden layer can employ batch normalization or layer normalization structures, ensuring that the dynamic weights of all exhibition content are adjusted to a uniform numerical range within the network, typically between zero and one. This neural network normalization structure effectively eliminates the impact of differences in the magnitude of behavior among different exhibition content, ensuring the comparability and stability of the dynamic interest distribution of each content during subsequent interest fusion and ranking. The introduction of the normalization hidden layer can also optimize the model's training efficiency and convergence speed, improving the overall generalization ability and robustness of the neural network. For example, for a piece of content with an abnormally high dynamic weight value due to a recent surge in user behavior, the normalization hidden layer can automatically pull its data distribution back, preventing a single piece of content from completely dominating the recommendation results due to a short-term surge in behavior, achieving a reasonable balance in the distribution of interest weights, and enhancing the fairness and diversity of the system.

[0063] In a preferred embodiment of the present invention, for each piece of exhibition content, an interest score is generated by weighting and adding it according to its corresponding basic interest score and the weight value in the dynamic interest distribution data through a preset first proportional factor, including: Extract the basic interest score and weight value from the dynamic interest distribution data for each piece of exhibition content; Based on the first proportional factor, the basic interest score and the dynamic interest weight value are multiplied by the corresponding proportional factor to obtain two weighted values. The two weighted values ​​are added together to obtain the interest score for the exhibition content.

[0064] In this embodiment of the invention, the fusion weighted hidden layer in the neural network model can automatically receive and integrate basic interest scores and normalized dynamic interest distribution data, and achieve multi-feature weighted fusion using a preset first ratio factor. This hidden layer can adopt a multi-input fully connected structure or introduce a weighted sum layer to achieve flexible adjustment of the ratio between static and dynamic features. The parameters of the fusion weighted hidden layer can be automatically learned through backpropagation during the model training phase to obtain the optimal weighting factor, or flexibly set according to actual business needs, achieving balanced control of static interests and dynamic behaviors in different scenarios. The output is a comprehensive interest score for each exhibition content, fully reflecting the dynamic fusion of long-term user preferences and real-time interests. This structure allows the system to dynamically adjust the ranking of recommended content based on the user's actual interest state, improving the personalization and timeliness of recommendations. For example, if a user's static interests are relatively clear but their recent behavior has changed significantly, this hidden layer can adjust the comprehensive interest score in a timely manner according to the fusion weights, ensuring that the recommended content not only matches the user's deep preferences but also flexibly responds to real-time attention, greatly improving the intelligent adaptability and user satisfaction of the recommendation system.

[0065] Embodiments of the present invention also provide an intelligent recommendation system for exhibition content based on neural networks, the system comprising: The data acquisition module is used to collect users' basic information, historical browsing records, and real-time interaction behavior data, and integrate them to form the user's original dataset; The feature extraction module is used to extract features from the user's original dataset, distinguishing between static attribute features, historical behavior features, and dynamic behavior features to obtain a multi-dimensional feature set. The interest analysis module is used to dynamically generate interest analysis results based on a multi-dimensional feature set. Static attribute features are used to determine the user's basic interests, historical behavior features are used to characterize the user's long-term preferences, and dynamic behavior features are used to capture the user's current focus. The interest analysis results include the weight distribution of the exhibition content that the user focuses on in different time periods. The content filtering and sorting module is used to filter and prioritize exhibition content based on the results of interest analysis and data from the exhibition content library, so as to obtain a personalized set of recommended exhibition content. The recommended output module is used to output a personalized set of recommended exhibition content to the user interface; The feedback update module is used to update the user's original dataset in real time based on the user's actual feedback on the recommended exhibition content.

[0066] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0067] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0068] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0069] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation of exhibition content based on neural networks, characterized in that, The method includes: Collect users' basic information, browsing history, and real-time interaction data, and integrate them to form the user's original dataset; Feature extraction processing is performed on the user's original dataset to distinguish between static attribute features, historical behavior features, and dynamic behavior features, resulting in a multi-dimensional feature set. A multi-dimensional feature set is input into a neural network model to dynamically generate interest analysis results. Static attribute features are used to determine the user's basic interests, historical behavior features are used to characterize the user's long-term preferences, and dynamic behavior features are used to capture the user's current focus. The interest analysis results include the weight distribution of the exhibition content that the user focuses on in different time periods. Based on the results of the interest analysis and combined with the data from the exhibition content library, the exhibition content is filtered and prioritized to obtain a personalized set of recommended exhibition content. Output a personalized set of recommended exhibition content to the user interface, and update the user's original dataset in real time based on the user's actual feedback on the recommended exhibition content.

2. The intelligent recommendation method for exhibition content based on neural networks according to claim 1, characterized in that, Feature extraction is performed on the original user dataset, distinguishing between static attribute features, historical behavior features, and dynamic behavior features, resulting in a multi-dimensional feature set, including: The basic information in the user's original dataset is parsed to identify and extract static attributes related to the exhibition, and these attributes are standardized to form structured static attribute feature data. The historical browsing records in the user's original dataset are analyzed one by one to count the exhibition content visited by the user and the corresponding dwell time. The data is then normalized and grouped by feature to form structured historical behavior feature data. The real-time interactive behavior data in the user's original dataset is sorted by timestamp, and different types of real-time interactive behavior are divided into preset time windows. The number of times different types of user interactive behavior occur for various exhibition contents is counted to form dynamic behavioral feature data. Static attribute feature data, historical behavior feature data, and dynamic behavior feature data are merged to form a multi-dimensional feature set.

3. The intelligent recommendation method for exhibition content based on neural networks according to claim 1, characterized in that, The neural network model includes: The input layer receives static attribute features, historical behavior features, and dynamic behavior features extracted from the user's original dataset and encodes them into an input vector. The basic interest score hidden layer is used to match static attribute features with exhibition content tags in the exhibition content library at the semantic or keyword level, calculate the similarity score for each piece of exhibition content, and generate a basic interest score. The hidden layer for dynamic weight extraction is used to count the frequency of exhibition content involved in each type of interactive behavior in the dynamic behavior features. The number of times each piece of exhibition content appears in the user operation is recorded as the initial value of dynamic weight. The initial values ​​of dynamic weight of all exhibition content are normalized to form dynamic interest distribution data. The fusion weighted hidden layer is used to generate an interest score for each piece of exhibition content by weighting it according to its corresponding basic interest score and the weight value in the dynamic interest distribution data through a preset first ratio factor. The output layer is used to sort the various exhibition contents from high to low according to their interest scores, and generate the sorting results as the interest analysis results.

4. The intelligent recommendation method for exhibition content based on neural networks according to claim 1, characterized in that, Based on the interest analysis results and combined with data from the exhibition content library, the exhibition content is filtered and prioritized to obtain a personalized set of recommended exhibition content, including: It iterates through all exhibition projects in the exhibition content library, compares the exhibition content tags of each project with the high-interest items in the interest analysis results, and selects the exhibition projects that match the user's current interests as a candidate exhibition content set. For each item in the candidate exhibition content set, the following scoring steps are performed sequentially: Extract the interest score corresponding to the content of the exhibition from the interest analysis results, and use it as the original interest score value; The number of interactions with the exhibition content among all users is counted and used as the original popularity score. Calculate the distance between the start time of the exhibition content and the current time, and obtain the original time score value based on the linear reciprocal relationship; The min-max normalization method is used to normalize the interest score, popularity score, and time score of all candidate exhibition content, so that each score is mapped to the interval between zero and one, resulting in normalized interest score, normalized popularity score, and normalized time score. According to the preset second proportional factor, the normalized interest score, normalized popularity score and normalized time score are multiplied by the corresponding proportional factor, and the resulting weighted values ​​are superimposed to obtain the priority score of the exhibition content. Based on priority scores, the exhibition content in the candidate exhibition content set is sorted from high to low, and a preset number of exhibition content with the highest priority scores are selected and organized into a personalized recommended exhibition content set, with a unique number attached.

5. The intelligent recommendation method for exhibition content based on neural networks according to claim 1, characterized in that, Based on actual user feedback on recommended exhibition content, the original user dataset is updated in real time, including: Real-time monitoring of user interaction behavior for each exhibition content in the personalized recommended exhibition content set; recording operation type, operation time and exhibition content number involved, forming an interaction behavior feedback dataset; Based on the interactive behavior feedback dataset, the cumulative statistics of user interactions with the same exhibition content at different time periods are compiled to form a feedback behavior statistics table. The data in the feedback behavior statistics table is merged and updated with the historical and dynamic behavior data in the user's original dataset.

6. The intelligent recommendation method for exhibition content based on neural networks according to claim 3, characterized in that, Static attribute features are matched semantically or keyword-wise with exhibition content tags in the exhibition content library. A similarity score is calculated for each piece of exhibition content to generate a basic interest score, including: Map each attribute in the user's static attribute feature vector to a set of static keywords; For each piece of exhibition content in the exhibition content library, extract its content tags and break them down into keywords to form a set of keywords to be matched; The static keyword set is matched with the keyword set to be matched to obtain the matching degree; Based on the matching degree, the number of identical keywords is counted, and a first fixed score is assigned to each identical keyword. The scores of all identical keywords are then summed. For keywords that are not completely identical but have a similarity higher than a preset similarity threshold, a second fixed score is assigned. The sum of the first fixed score and the second fixed score is used as the original similarity score between the exhibition content and the user's static attribute features; Divide the original similarity scores of all exhibition content by the number of corresponding exhibition content tag keywords to obtain the normalized basic interest score.

7. The intelligent recommendation method for exhibition content based on neural networks according to claim 3, characterized in that, Frequency statistics are performed on the exhibition content involved in each type of interactive behavior in the dynamic behavior characteristics. The number of times each piece of exhibition content appears in user operations is recorded as the initial value of the dynamic weight, including: Based on dynamic behavioral characteristics, all user interaction records within each time window are extracted and categorized according to exhibition content and interaction type; For each type of interactive behavior in each exhibition content, the number of times the behavior occurs within each time window is counted, and different weight values ​​are assigned to different types of interactive behavior. The weighted frequency of each interaction type is summed under the same exhibition content to obtain the total weighted frequency of a single exhibition content within the time window. The initial dynamic weight of each exhibition content is obtained by summing the weighted frequency of each behavior across all time windows.

8. The intelligent recommendation method for exhibition content based on neural networks according to claim 3, characterized in that, The initial dynamic weights of all exhibition content are standardized and normalized to form dynamic interest distribution data, including: Based on the initial dynamic weights of all exhibition content, determine the maximum and minimum values ​​among them; For each exhibition content, the initial dynamic weight value is combined with the maximum and minimum values ​​in the initial dynamic weight value, and then the standard normalization is performed to map all dynamic weight values ​​to the interval between zero and one, so as to obtain the normalized dynamic weight value. The normalized dynamic weight values ​​of all exhibition content are merged to form dynamic interest distribution data.

9. The intelligent recommendation method for exhibition content based on neural networks according to claim 3, characterized in that, For each piece of exhibition content, an interest score is generated by weighting and summing it based on its corresponding basic interest score and the weight value in the dynamic interest distribution data using a preset first proportional factor. This score includes: Extract the basic interest score and weight value from the dynamic interest distribution data for each piece of exhibition content; Based on the first proportional factor, the basic interest score and the dynamic interest weight value are multiplied by the corresponding proportional factor to obtain two weighted values. The two weighted values ​​are added together to obtain the interest score for the exhibition content.

10. A neural network-based intelligent recommendation system for exhibition content, characterized in that: The system, used in any one of claims 1 to 9, comprises: The data acquisition module is used to collect users' basic information, historical browsing records, and real-time interaction behavior data, and integrate them to form the user's original dataset; The feature extraction module is used to extract features from the user's original dataset, distinguishing between static attribute features, historical behavior features, and dynamic behavior features to obtain a multi-dimensional feature set. The interest analysis module is used to dynamically generate interest analysis results based on a multi-dimensional feature set. Static attribute features are used to determine the user's basic interests, historical behavior features are used to characterize the user's long-term preferences, and dynamic behavior features are used to capture the user's current focus. The interest analysis results include the weight distribution of the exhibition content that the user focuses on in different time periods. The content filtering and sorting module is used to filter and prioritize exhibition content based on the results of interest analysis and data from the exhibition content library, so as to obtain a personalized set of recommended exhibition content. The recommended output module is used to output a personalized set of recommended exhibition content to the user interface; The feedback update module is used to update the user's original dataset in real time based on the user's actual feedback on the recommended exhibition content.

Images