An intelligent recommendation method and system for educational books

Abstract

The present invention provides an intelligent recommendation method and system for educational books, and relates to the field of data processing technology. The system includes a style analysis module, a group clustering module, a target user analysis module, a rating list generation module, a recommendation parameter analysis module, and an educational book recommendation module. The present invention extracts the user's long-term preference for different content styles by analyzing the user's historical interaction data, generates a style rating list representing the user group's preference to effectively capture the user's long-term preference, and further evaluates the user's cognitive load and learning status at different learning stages by analyzing the interaction state parameters in the user's historical interaction data, generates a state influence parameter reflecting the user's current state and obtains a state reference list, and finally analyzes the matching degree between different books and users in combination with the user's long-term preference and learning status, thereby achieving accurate recommendation of educational books.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an intelligent recommendation method and system for educational books. Background Art

[0002] Existing educational book recommendation systems usually adopt a matching method based on historical behavior data for personalized recommendations, usually focusing on the popularity of books or a single content dimension (such as book ratings or user views). However, this recommendation method is prone to show insufficient recommendation effects when facing users with diverse needs. In educational scenarios, the content of books is often diverse, covering a variety of styles such as theoretical explanations, case analyses, and illustration demonstrations, and different users have large differences in their preferences for these content styles. Recommendation methods based solely on book popularity or local user preferences are difficult to fully reflect users' actual needs for various styles of content, and often lead to recommendation results that are inconsistent with users' comprehensive preferences, reducing the accuracy of recommendations and user satisfaction.

[0003] In addition, books with different content styles will have different effects on users' learning cognitive load. For example, users may need to spend more cognitive resources when learning the theoretical explanation part, but may feel relaxed when learning the diagram demonstration or case analysis part. Some traditional recommendation systems ignore the differences in users' cognitive loads for different content styles during the reading process, resulting in ineffective matching of users' needs in different learning states, which further affects the accuracy and applicability of recommendations.

[0004] The multi-dimensional content preferences and cognitive states of users have not been fully explored and utilized, which makes the recommendation system obviously insufficient in meeting personalized needs. Therefore, the current educational book recommendation system lacks comprehensive analysis of users' fine-grained preferences and cognitive states, making it difficult to provide highly accurate and personalized recommendation services. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes an intelligent recommendation method and system for educational books, which provides users with highly accurate and personalized educational book recommendation services by comprehensively analyzing users' fine-grained preferences and cognitive states.

[0006] To achieve the above object, the present invention provides a first aspect of an intelligent recommendation method for educational books, comprising:

[0007] Collecting first historical interaction data of multiple platform users on educational books, extracting first content style preference data of each platform user from the first historical interaction data, and extracting a first style preference reference vector of each platform user based on the first content style preference data;

[0008] Cluster multiple platform users based on style preference feature vectors to obtain multiple user clusters, extract the group preference feature vector of each user cluster based on the cluster center of each user cluster, and determine the style score list of each user cluster based on the group preference feature vector;

[0009] Acquire second historical interaction data of the target user, extract second content style preference data of the target user from the second historical interaction data, the content style preference data including score data of multiple style tags, and extract a second style preference reference vector of the target user based on the second content style preference data;

[0010] Determine whether there is a target cluster corresponding to the target user among the multiple user clusters, and if so, determine a rating reference list for the target user according to the user cluster to which the target user belongs; otherwise, determine a rating reference list for the target user based on the second style preference reference vector of the target user;

[0011] Acquire the educational book query needs of the target user, determine multiple target recommendation objects according to the educational book query needs of the target user, and calculate first recommendation parameters of the multiple target recommendation objects based on the style reference list of the target user;

[0012] Extracting the target user's book interaction state parameters from the second historical interaction data, extracting the state influence parameters of each style tag from the book interaction state parameters and constructing a state reference list, and calculating the second recommendation parameters of multiple target recommendation objects based on the state reference list;

[0013] Books are recommended to a target user using a plurality of target recommendation objects according to the first recommendation parameter and the second recommendation parameter.

[0014] Preferably, extracting the first content style preference data of each platform user from the first historical interaction data, and extracting the first style preference reference vector of each platform user based on the first content style preference data includes:

[0015] Determine multiple educational book interactive events in the first historical interactive data, extract the timestamp and interactive book preference data of each educational book interactive event, the interactive book preference data includes the interactive scores of multiple style tags in the educational book interactive event, and use the following formula to integrate the interactive book preference data of multiple educational book interactive events:

[0016] Get the first content style preference data corresponding to each set of first historical interaction data: In the formula, Indicates the total scores of platform users corresponding to multiple style tags. Indicates that the platform user is Educational book interactive event about The interactive score of style tags, Indicates that the platform user is Educational book interactive event about The decay factor of the interaction score of style labels, where Indicates the current time. Indicates The timestamp of each educational book interaction event, is the attenuation base, which is a constant less than 1. represents the total number of educational book interaction events, Indicates the total number of style tags;

[0017] The above formula is used to calculate the total score of each style label, and the first content style preference data corresponding to each group of first historical interaction data is generated. The first content style preference data includes the total scores corresponding to multiple style labels. Based on the first content style preference data, the first style preference reference vector of each platform user is extracted.

[0018] Preferably, the rating reference list for the target user also includes:

[0019] Based on the second style preference reference vector and the plurality of group preference feature vectors, the distances between the target user and the plurality of user clusters are respectively calculated. If there is at least one user cluster among the plurality of user clusters whose distance to the target user is less than a preset similarity threshold, it is considered that there is a target cluster corresponding to the target user among the plurality of user clusters, and the user cluster with the shortest distance is selected from the user clusters whose distance to the target user is less than the preset similarity threshold and recorded as the target cluster, and the style score list of the target cluster is recorded as the score reference list of the target user;

[0020] Otherwise, a reference score for each style tag is determined according to the second style preference reference vector of the target user, and a score reference list for the target user is constructed according to the reference scores of multiple style tags.

[0021] Preferably, the first recommended parameter and the second recommended parameter further include:

[0022] Determine the interaction scores corresponding to the multiple style tags of each target recommendation object, perform weighted analysis on the multiple interaction scores according to the style tag score corresponding to each style tag in the style reference list of the target user, and calculate the first recommendation parameter of each target recommendation object;

[0023] Determine the interaction status data corresponding to each educational interaction event in the second historical interaction data, extract the book interaction status parameters of each book interaction feature from the interaction status data, process the book interaction status parameters according to the interactive book preference data of each educational book interaction event, and obtain the interaction status analysis result of each educational book interaction event, fuse multiple interaction status analysis results to obtain the state influence parameter of each style label, construct a state reference list containing the state influence parameters corresponding to multiple style labels, perform weighted analysis on the interaction scores corresponding to multiple style labels of each target recommendation object according to the state influence parameter corresponding to each style label in the state reference list of the target user, and calculate the second recommendation parameter of each target recommendation object.

[0024] Preferably, recommending books to a target user from a plurality of target recommendation objects according to the first recommendation parameter and the second recommendation parameter includes:

[0025] The recommendation score of each target recommendation object is calculated according to the weight parameters corresponding to the first recommendation parameter and the second recommendation parameter respectively, the recommendation priority of multiple target recommendation objects is sorted according to the recommendation score, and books are recommended to the target user based on the recommendation priority order of the multiple target recommendation objects.

[0026] Preferably, the style score list further includes:

[0027] Cluster multiple platform users based on the DBSCAN clustering algorithm to obtain multiple user clusters, and calculate the cluster center of each user cluster;

[0028] For any user cluster, calculate the distance between each object in the user cluster and the cluster center, arrange all objects in ascending order based on the distance between the object and the cluster, draw a curve showing the cumulative number of individuals changing with distance, determine the target distance point in the curve based on the elbow method, record the objects within the target distance point as the core group of the user cluster, calculate the mean of the first style preference reference vector of the core group and record it as the group preference feature vector of the user cluster, determine the style label score of each style label based on the elements corresponding to each style label in the group preference feature vector, and construct a style score list containing the style label scores corresponding to multiple style labels.

[0029] A second aspect of the present invention provides an intelligent recommendation system for educational books, which is used to implement the above-mentioned intelligent recommendation method for educational books, comprising:

[0030] A style analysis module, used to collect first historical interaction data of multiple platform users on educational books, extract first content style preference data of each platform user from the first historical interaction data, and extract a first style preference reference vector of each platform user based on the first content style preference data;

[0031] A group clustering module is used to cluster multiple platform users based on style preference feature vectors to obtain multiple user clusters, extract the group preference feature vector of each user cluster based on the cluster center of each user cluster, and determine the style score list of each user cluster based on the group preference feature vector;

[0032] A target user analysis module obtains second historical interaction data of the target user, extracts second content style preference data of the target user from the second historical interaction data, wherein the content style preference data includes rating data of multiple style tags, and extracts a second style preference reference vector of the target user based on the second content style preference data;

[0033] A rating list generation module is used to determine whether there is a target cluster corresponding to the target user among the multiple user clusters. If so, a rating reference list for the target user is determined according to the user cluster to which the target user belongs; otherwise, a rating reference list for the target user is determined based on the second style preference reference vector of the target user.

[0034] A recommendation parameter analysis module is used to obtain the educational book query needs of the target user, determine multiple target recommendation objects according to the educational book query needs of the target user, calculate the first recommendation parameters of the multiple target recommendation objects based on the style reference list of the target user, extract the book interaction state parameters of the target user from the second historical interaction data, extract the state influence parameters of each style label from the book interaction state parameters and construct a state reference list, and calculate the second recommendation parameters of the multiple target recommendation objects based on the state reference list;

[0035] The educational book recommendation module is used to recommend books to target users based on a first recommendation parameter and a second recommendation parameter for multiple target recommendation objects.

[0036] Preferably, for the educational book recommendation module, recommending books to target users from multiple target recommendation objects according to the first recommendation parameter and the second recommendation parameter includes:

[0037] The recommendation score of each target recommendation object is calculated according to the weight parameters corresponding to the first recommendation parameter and the second recommendation parameter respectively, the recommendation priority of multiple target recommendation objects is sorted according to the recommendation score, and books are recommended to the target user based on the recommendation priority order of the multiple target recommendation objects.

[0038] The present invention has the following beneficial effects:

[0039] The present invention extracts users' long-term preferences for different content styles by analyzing users' historical interaction data, and generates a style score list representing the preferences of user groups based on clustering of these preferences, which can effectively capture users' long-term preferences. It also further evaluates users' cognitive load and learning status at different learning stages by analyzing interaction state parameters in users' historical interaction data, generates state influence parameters reflecting users' current status and obtains a state reference list. Finally, it analyzes the matching degree between different books and users in combination with users' long-term preferences and learning status, thereby achieving accurate recommendation of educational books and ensuring that the recommended content not only meets users' long-term preferences, but also adapts to users' learning status and learning needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a flow chart of an intelligent recommendation method for educational books in an embodiment of the present invention.

[0041] Figure 2 The figure is a schematic diagram of the structure of an intelligent recommendation system for educational books in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make those skilled in the art better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0043] Figure 1 FIG. 1 is a flow chart of an intelligent recommendation method for educational books in an embodiment of the present invention. Figure 1 , the method comprises the following steps:

[0044] S1. Collect first historical interaction data of multiple platform users on educational books, extract first content style preference data of each platform user from the first historical interaction data, and extract a first style preference reference vector of each platform user based on the first content style preference data.

[0045] In this embodiment, the historical interaction data includes various interactive behaviors between users and educational books, such as borrowing, browsing, rating, commenting and other data. Among them, educational books can refer to books involving teaching, learning and practice content of multiple disciplines, such as books covering multiple subject areas and can be used for course teaching, providing students or researchers with specific information for in-depth understanding of a certain field, such as books including theory, application and actual case analysis, etc. According to the content of the books, these teaching books can be summarized and summarized into theoretical explanations, exercise analysis, case analysis, illustration demonstration and other contents. Each book may contain one or more of these contents, which can be recorded as the style tags of the books.

[0046] It is worth noting that there are differences in the content involved in different teaching books, and the relative proportion or relative importance of each part of the content in the book can be characterized based on the review data of the teaching book. For example, after borrowing one of the books, the user thinks that the theoretical explanation content is relatively excellent, so the book can be marked with a label about the theoretical explanation, and some students think that the legend demonstration content of this book is more exciting, so the legend demonstration label of this book can be marked. Finally, the total number of marks of each style label in all the review data can be counted to determine the importance of multiple platform users to each part of the content in this book or the importance of perception, so as to quantify the relative importance of different contents in each book, for example, the theoretical explanation is finally determined to be 4.7 points, the problem analysis is 2.8 points, and the case analysis is 2.5 points. In this way, the importance of different users to each style label of the book can be determined. Those skilled in the art can also analyze the comments or rating data of each book based on natural language processing and other technologies to determine the relationship between the importance of each part of the content in the book by the user, which is not specifically limited in this embodiment.

[0047] In this embodiment, the platform can periodically analyze the review data of each book to determine the relative importance of each part of the content in the book, and regard it as the relative proportion of the content corresponding to multiple style tags in the book. Therefore, for the historical interaction data, the relative proportion data of the content corresponding to multiple style tags in each book can be extracted, and the data of multiple books borrowed by users can be summarized to obtain the first content style preference data of each platform user. The content style preference data includes the score data of multiple style tags, such as the relative proportion of the content corresponding to each style tag. Specifically, the sum of the relative proportions of the content corresponding to multiple style tags can be normalized to obtain the quantitative interaction score of each style tag to characterize the user's preference for different styles in past reading behaviors. Therefore, based on the quantified interaction score component of each style tag, the first style preference reference vector is obtained. The style preference reference vector can represent the user's comprehensive preference for different styles to reflect the user's preference distribution for different style tags during the reading process. Each dimension in the vector corresponds to a style tag, and its value represents the user's preference for the style tag.

[0048] S2. Cluster multiple platform users based on style preference feature vectors to obtain multiple user clusters, extract the group preference feature vector of each user cluster based on the cluster center of each user cluster, and determine the style score list of each user cluster based on the group preference feature vector.

[0049] In this embodiment, the first style preference reference vector of each platform user is used as the clustering basis to perform cluster analysis on multiple platform users, with the purpose of classifying users with similar reading preferences into the same cluster. Through the multiple user clusters obtained by clustering, users can be divided into multiple groups, and users in each group show similar preferences for certain content styles. The group preference feature vector of each user cluster is determined by the cluster center of each cluster, which represents the overall preference of users in the cluster for different style labels. Each dimension in the group preference feature vector corresponds to a style label, and its value represents the overall preference of users in this group for the label. A style score list for each user cluster is generated based on the group preference feature vector. The style score list includes the style label score corresponding to each style label. The style label score can be specifically determined based on the value corresponding to each element in the group preference feature vector. For example, the value corresponding to each element is directly recorded as the style label score. The style score list can reflect the correlation between the preference degrees of groups within the cluster for different style labels.

[0050] S3. Acquire second historical interaction data of the target user, extract second content style preference data of the target user from the second historical interaction data, and extract a second style preference reference vector of the target user based on the second content style preference data.

[0051] In this embodiment, the second historical interaction data of the target user includes but is not limited to the target user's interaction behavior with different educational books on the platform or the most recent reading record, which contains the target user's learning characteristics at the current stage. Based on these interaction data, the system can extract the target user's second content style preference data, which is the result of analyzing the preference distribution of the style tags of the educational books borrowed by the target user. After the style preference related data contained in the second historical interaction data of the target user is aggregated to obtain the second historical interaction data, the target user's interaction score for each style tag can be obtained, thereby constructing the target user's second style preference reference vector, which is used to reflect the user's needs or content preferences at the current learning stage.

[0052] S4. Determine whether there is a target cluster corresponding to the target user among the multiple user clusters. If so, determine a rating reference list for the target user based on the user cluster to which the target user belongs. Otherwise, determine a rating reference list for the target user based on the second style preference reference vector of the target user.

[0053] In this embodiment, the system determines whether the target user already belongs to an existing user cluster. This determination process is based on the second style preference reference vector of the target user. Specifically, the distances between the target user and multiple user clusters are calculated based on the second style preference reference vector and multiple group preference feature vectors, respectively, to characterize the similarity between the target user and multiple user clusters.

[0054] Specifically, if there is at least one user cluster among multiple user clusters whose distance to the target user is less than a preset similarity threshold, it is considered that there is a target cluster corresponding to the target user among the multiple user clusters, and the user cluster with the shortest distance to the target user is screened out from the user clusters whose distance to the target user is less than the preset similarity threshold and recorded as the target cluster, and the style score list of the target cluster is recorded as the score reference list for the target user.

[0055] Among them, by calculating the distance between the style preference reference vector of the target user and the group preference feature vector of each user cluster, for example, calculating the Euclidean distance between the two vectors, the smaller the distance, the closer the style preference of the target user is to the preference of the user group in the cluster. And by presetting the similarity threshold to define the similarity standard, only the user cluster whose distance between the two is less than the preset similarity threshold can be recorded as the target cluster. Compared with directly relying on the shortest distance to determine the target cluster, this method can avoid the situation where the distance between the cluster and the target user is the smallest but its value itself may still be relatively large, resulting in the cluster not really being very similar to the target user, thereby ensuring that there is sufficient similarity between the target user and the selected target cluster. After determining the target cluster, the style score list of the target cluster is used as the score reference list of the target user.

[0056] It is worth mentioning that the style score list of the target cluster is based on the common preferences of multiple users and integrates the historical interaction data of multiple users. If the historical data of a target user is sparse or insufficient, directly using his or her historical data may result in an incomplete score reference list and fail to accurately reflect the user's preferences. In this case, using a style score list with a high similarity to the target user as the score reference list for the target user can make up for the problem of insufficient data and provide a broader and more stable recommendation basis.

[0057] If the distance between the target user and all user clusters is greater than the preset similarity threshold, it means that the target user's preference is significantly different from the group preference in the existing clusters, and a sufficiently similar cluster cannot be found. In this case, the reference score of each style tag is determined based on the target user's second style preference reference vector, which reflects the target user's personal preference strength for each style tag. The target user's score reference list is constructed based on the reference scores of multiple style tags.

[0058] S5. Obtain the educational book query needs of the target user, determine multiple target recommendation objects according to the educational book query needs of the target user, and calculate first recommendation parameters of the multiple target recommendation objects based on the style reference list of the target user.

[0059] In this embodiment, a group of target recommendation objects, i.e., potential recommended books matching the query requirements, are determined according to the query requirements of the target user. These recommendation objects can be determined based on the fields of interest mentioned in the educational book query requirements of the target user. For example, if a customer wants to query books related to calculus, the platform can determine multiple target recommendation objects that meet the user's needs based on existing query and preliminary screening technologies. In order to analyze the degree of match between multiple target recommendation objects and the target user, a first recommendation parameter is calculated for each target recommendation object based on the target user's style reference list to measure the degree of match between the style of a book and the user's preference. For example, if the user prefers books with more theoretical explanations, and the theoretical explanation part of a book is relatively rich, the first recommendation parameter of the book will be higher.

[0060] Specifically, for the calculation of the first recommendation parameter, the interaction scores corresponding to the multiple style tags of each target recommendation object are first determined. The interaction scores are determined based on the aforementioned analysis of the review data of each book, and the recommendation value of the book on each style tag is characterized by the relative importance that a large number of users attach to the content corresponding to each style tag in the book. Finally, a weighted analysis is performed on the multiple interaction scores according to the style tag score corresponding to each style tag in the target user's style reference list, and the first recommendation parameter of each target recommendation object is calculated. Specifically, a weighted interaction score is calculated for each style tag, and the weight is determined by the target user's preference score for the style in the style reference list. The larger the weight, the stronger the user's interest in the content corresponding to the style tag, and the greater the influence of the interaction score. The more importance a large number of users attach to the content of the style tag in the book, the higher the calculated first recommendation parameter of the book on the style tag.

[0061] S6. Extract the target user's book interaction status parameters from the second historical interaction data, extract the status influence parameters of each style tag from the book interaction status parameters and construct a status reference list, calculate the second recommendation parameters of multiple target recommendation objects based on the status reference list, and recommend books to the target user based on the first recommendation parameters and the second recommendation parameters.

[0062] In this embodiment, the recommendation process is further refined, and the target user's book interaction state parameters are extracted from the second historical interaction data. The book interaction state parameters reflect the user's state performance data when reading books, such as reading time, page jump frequency, click speed, sliding speed, etc. These data can reflect the user's cognitive load or interest level when reading different content. And according to the interaction score of each book on multiple style tags, the state performance data of different books in the user's historical interaction data are summarized to obtain the state influence parameter of each style tag.

[0063] For example, a longer reading time may reflect that the user has spent a lot of cognitive resources in the book, especially when reading certain difficult or information-rich content, the user needs more time to process and understand. On the other hand, if the user stays on the book for too short a time, it may mean that they feel that the content is too difficult or boring and choose to skip it. The book interaction state parameter can reflect whether the user feels overburdened in the learning process. Since the user's preference for different style tags will affect the user's concentration when learning different content, the user's state performance data when reading the book can reflect the user's interaction preference for different styles of content to a certain extent. By summarizing the user's interaction state performance data on multiple books, the state influence parameter that is equivalent to describing the user's interaction preference for different styles of content is obtained, thereby constructing a state reference list containing the state influence parameter of each style tag. The state reference list reflects the degree of interaction matching between the target user's interaction state and the content corresponding to different style tags. Based on the state reference list, the second recommendation parameter is calculated for each recommendation object. The second recommendation parameter reflects the matching degree between the user's interaction state and the book style, thereby further adjusting the recommendation priority. Finally, the system will combine the first recommendation parameter and the second recommendation parameter to provide personalized book recommendations for the target user.

[0064] Exemplarily, a recommendation score for each target recommendation object is calculated according to weight parameters corresponding to the first recommendation parameter and the second recommendation parameter, and multiple target recommendation objects are ranked according to the recommendation scores, and book recommendations are made to the target user based on the recommendation priority order of the multiple target recommendation objects.

[0065] In this embodiment, the first recommendation parameter is obtained by weighting the preference score corresponding to each style tag in the style reference list of the target user and the interaction score of each style tag of the target recommended object, reflecting the matching degree between the target user and the recommended object in terms of content style. In other words, the first recommendation parameter mainly measures whether the content of the target recommended object meets the user's personal preferences and preference tags. The second recommendation parameter takes into account the state parameters of the target user in the process of interacting with the book, such as the user's current learning state, learning progress, cognitive load, etc. These state parameters can reflect the differences in user needs at different learning stages. Therefore, the second recommendation parameter mainly focuses on whether the target recommended object is suitable for the user's current learning state, ensuring that the recommended content can match the user's current cognitive level and needs. For example, when the user feels that the cognitive load is heavy, the system may tend to recommend books with relatively easy or easy-to-understand content, while when the user has more time and energy, it recommends more challenging books. The weight parameters corresponding to the first recommendation parameter and the second recommendation parameter are used to characterize the relative importance of the first recommendation parameter and the second recommendation parameter in the recommendation score. The weight parameters can be adjusted according to the actual scenario. For example, in some cases, the user's style preference may be the main basis for recommendation, so the weight of the first recommendation parameter is larger, while in other cases, the user's current learning status is more critical, and the weight of the second recommendation parameter will be larger.

[0066] According to the weight parameters corresponding to the first recommendation parameter and the second recommendation parameter, the first recommendation parameter and the second recommendation parameter are weighted and summed, and the result is used as the recommendation score of the target recommendation object, which not only considers the user's historical interest preferences, but also combines the user's current actual needs and status, so that the recommendation results are more in line with the user's immediate learning scenario. Finally, all target recommendation objects are sorted according to the recommendation score of each target recommendation object to determine their recommendation priority. The higher the score, the more it will be considered to match the user's needs and preferences, and thus have a higher priority. This sorting process ensures that the system first recommends books that best meet the user's preferences and current needs, and recommends books to the target user in order of the sorted recommendation priority. Users will give priority to books that best match their style preferences and learning status. This design of recommendation order can effectively improve the user's book selection efficiency and ensure the relevance and practicality of the recommended content.

[0067] The recommendation method provided by the embodiment of the present invention not only makes recommendations based on the user's long-term preferences (such as preference for content style), but also combines the user's immediate status (such as current learning progress or cognitive load) to achieve dynamic personalized recommendations. This recommendation strategy based on comprehensive scoring ensures that the recommended content is in line with the user's long-term interests and can meet their learning needs at different stages, thereby further improving the accuracy of the recommendation and user experience.

[0068] As one implementation process, in step S1, first content style preference data of each platform user is extracted from the first historical interaction data, and a first style preference reference vector of each platform user is extracted based on the first content style preference data, specifically including:

[0069] Determine multiple educational book interactive events in the first historical interactive data, where educational book interactive events include but are not limited to browsing, reading, and scoring of books, and each interactive behavior is associated with specific time information, thereby extracting the timestamp and interactive book preference data of each educational book interactive event. Exemplarily, one borrowing behavior of a user, including browsing, reading, and commenting, constitutes an educational book interactive event.

[0070] The interactive book preference data includes the interactive scores of multiple style tags in the educational book interactive events. Specifically, the interactive book preference data of multiple educational book interactive events can be integrated using the following formula to obtain the first content style preference data corresponding to each group of first historical interactive data, where: In the formula, Indicates the total scores of platform users corresponding to multiple style tags. Indicates that the platform user is Educational book interactive event about The interactive score of style tags, Indicates that the platform user is Educational book interactive event about The decay factor of the interaction score of each style tag reflects the trend of the score decay of different style tags of educational books over time. Specifically, the user's earlier interaction events have less impact on the current preference, while the newer interaction data has a greater impact. This time decay mechanism ensures that the recommendation system can capture the user's recent interest changes. Indicates the current time. Indicates The timestamp of each educational book interaction event, is the attenuation base, which is a constant less than 1. represents the total number of educational book interaction events, Indicates the total number of style tags;

[0071] The above formula is used to calculate the total score of each style tag, and the first content style preference data corresponding to each set of first historical interaction data is generated. The first content style preference data contains the total scores corresponding to multiple style tags, which represents the user's preference for different style tags in all historical interaction events. Through the above formula, the total preference score of platform users for each style tag can be calculated under the weight of time decay. In other words, the most recent interaction event will contribute more to the score, while the earlier interaction event will have less impact. Finally, the first style preference reference vector of each platform user is extracted based on the first content style preference data.

[0072] As one implementation process, the style score list in step S2 further includes:

[0073] Based on the DBSCAN clustering algorithm, multiple platform users are clustered to obtain multiple user clusters. DBSCAN is a density-based clustering algorithm that can automatically discover high-density user groups and identify multiple clusters. Through this clustering method, the system can group users with similar preferences in the same cluster, which is convenient for subsequent preference feature extraction and analysis. After obtaining multiple user clusters, the cluster center of each user cluster is calculated. The cluster center is determined by calculating the mean of the style preference reference vectors of all users in the cluster. Calculate the cluster center of each user cluster;

[0074] For any user cluster, the distance between each object in the user cluster and the cluster center is calculated, where the distance can be calculated using Euclidean distance or other appropriate metrics. The smaller the distance, the closer the user's preference is to the cluster center, and the more representative the overall preference of the cluster is. After calculating the distance between each platform user and the cluster center, all objects are arranged in ascending order based on the distance between the object and the cluster, and a curve showing the cumulative number of individuals changing with distance is drawn. This curve can reflect the distribution of users in the cluster and show how the cumulative number of users closest to the cluster center changes with distance. Then, the target distance point in the curve is determined based on the elbow method. The elbow rule is a commonly used method for determining the inflection point in the curve, that is, when the growth rate of the cumulative number of users slows down significantly, the corresponding distance point is the target distance point. Users after this point are often far away from the cluster center and are users who deviate from the cluster core, while users within the target distance point are closer to the cluster center and can better represent the core preference characteristics of the cluster.

[0075] Users within the target distance point determined by the elbow method are recorded as the core group of the user cluster. These users are at a small distance from the cluster center and represent the group with the most consistent preferences in the cluster and the most representative of the overall preferences of the cluster. The style preference reference vectors of these users will be used to further calculate the overall preference characteristics of the cluster.

[0076] Specifically, the objects within the target distance point are recorded as the core group of the user cluster, the mean of the first style preference reference vector of the core group is calculated and recorded as the group preference feature vector of the user cluster, each element in the vector corresponds to a specific style label, and the value of the element reflects the overall preference of the core group for the style label. Finally, the style label score of each style label is determined according to the element corresponding to each style label in the group preference feature vector, and a style score list containing style label scores corresponding to multiple style labels is constructed.

[0077] As one implementation process, for the second recommended parameter in step S6, the calculation process includes:

[0078] Determine the interaction status data corresponding to each educational interaction event in the second historical interaction data, extract the book interaction status parameters of each book interaction feature from the interaction status data, process the book interaction status parameters according to the interactive book preference data of each educational book interaction event, and obtain the interaction status analysis results of each educational book interaction event.

[0079] In this embodiment, the interaction status data includes data related to multiple book interaction characteristics such as the user's reading time, page jump frequency, click speed, and sliding speed. For any educational book interaction event, the book interaction status parameter of the book interaction characteristic is specifically the actual parameter value of the book interaction characteristic.

[0080] After normalizing multiple book interaction features in the educational book interaction event, the weighted sum of multiple book interaction features is performed by pre-assigning weights to each book interaction feature, so that the cognitive load score corresponding to each educational book interaction event can be obtained, and then based on the cognitive load score, the interaction status score of each style tag in the interactive book preference data is determined. Exemplarily, the interaction status score is obtained by multiplying the interaction score of the style tag by the cognitive load score corresponding to the educational book interaction event. Thus, the interaction status analysis result containing the interaction status scores of multiple style tags for each educational book interaction event is obtained.

[0081] The results of multiple interaction status analyses are integrated to obtain the status influence parameters of each style label, and a status reference list containing the status influence parameters corresponding to multiple style labels is constructed. According to the status influence parameters corresponding to each style label in the target user's status reference list, a weighted analysis is performed on the interaction scores corresponding to multiple style labels of each target recommendation object, and the second recommendation parameter of each target recommendation object is calculated.

[0082] In this embodiment, the interaction status scores corresponding to each style tag in the multiple interaction status analysis results are accumulated, and the accumulated parameters are normalized to obtain the state influence parameter of each style tag, and the state reference list is constructed based on the multiple state influence parameters. The state influence parameter indicates the matching degree of the content under the style tag to the user's current learning state or interaction state. The larger the state influence parameter, the higher the matching degree of the content under the style tag to the user's current learning state or interaction state. Exemplarily, if the user's interaction state under a certain style tag indicates that his cognitive load is low (for example, the user stays on the content of the tag for a long time, the jump frequency is low, and the reading speed is moderate), the state influence parameter will be high. This indicates that the user feels relatively relaxed and comfortable when learning in this style, and can digest the content well. Based on the state influence parameter of each style tag, the style tag of the target recommendation object is weighted to generate a second recommendation parameter. This step combines the user's state parameter with the style preference data to ensure that the recommended content is not only suitable for the user's long-term interests, but also can adapt to the user's current state and needs in real time.

[0083] An intelligent recommendation method for educational books provided by an embodiment of the present invention extracts users' long-term preferences for different content styles by analyzing users' historical interaction data, and generates a style score list representing the preferences of user groups based on clustering of these preferences. This method can effectively capture users' long-term preferences, and further evaluate users' cognitive load and learning status at different learning stages by analyzing interaction state parameters in users' historical interaction data, generate state influence parameters reflecting users' current status and obtain a state reference list. Finally, the matching degree between different books and users is analyzed in combination with users' long-term preferences and learning status, thereby achieving accurate recommendation of educational books and ensuring that the recommended content not only meets users' long-term preferences, but also adapts to users' learning status and learning needs.

[0084] Figure 2 FIG. 1 is a schematic diagram of a structure of an intelligent recommendation system for educational books in an embodiment of the present invention. Figure 2 The system comprises:

[0085] A style analysis module, used to collect first historical interaction data of multiple platform users on educational books, extract first content style preference data of each platform user from the first historical interaction data, and extract a first style preference reference vector of each platform user based on the first content style preference data;

[0086] Specifically, multiple educational book interaction events in the first historical interaction data are determined, and the timestamp and interactive book preference data of each educational book interaction event are extracted. The interactive book preference data includes the interactive scores of multiple style tags in the educational book interaction event. The interactive book preference data of multiple educational book interaction events are integrated using the following formula:

[0087] Get the first content style preference data corresponding to each set of first historical interaction data: In the formula, Indicates the total scores of platform users corresponding to multiple style tags. Indicates that the platform user is Educational book interactive event about The interactive score of style tags, Indicates that the platform user is Educational book interactive event about The decay factor of the interaction score of style labels, where Indicates the current time. Indicates The timestamp of each educational book interaction event, is the attenuation base, which is a constant less than 1. represents the total number of educational book interaction events, Indicates the total number of style tags;

[0088] The above formula is used to calculate the total score of each style label, and the first content style preference data corresponding to each group of first historical interaction data is generated. The first content style preference data includes the total scores corresponding to multiple style labels. Based on the first content style preference data, the first style preference reference vector of each platform user is extracted.

[0089] A group clustering module is used to cluster multiple platform users based on style preference feature vectors to obtain multiple user clusters, extract the group preference feature vector of each user cluster based on the cluster center of each user cluster, and determine the style score list of each user cluster based on the group preference feature vector;

[0090] A target user analysis module obtains second historical interaction data of the target user, extracts second content style preference data of the target user from the second historical interaction data, wherein the content style preference data includes rating data of multiple style tags, and extracts a second style preference reference vector of the target user based on the second content style preference data;

[0091] A rating list generation module is used to determine whether there is a target cluster corresponding to the target user among the multiple user clusters. If so, a rating reference list for the target user is determined according to the user cluster to which the target user belongs; otherwise, a rating reference list for the target user is determined based on the second style preference reference vector of the target user.

[0092] Based on the second style preference reference vector and the plurality of group preference feature vectors, the distances between the target user and the plurality of user clusters are respectively calculated. If there is at least one user cluster among the plurality of user clusters whose distance to the target user is less than a preset similarity threshold, it is considered that there is a target cluster corresponding to the target user among the plurality of user clusters, and the user cluster with the shortest distance is selected from the user clusters whose distance to the target user is less than the preset similarity threshold and recorded as the target cluster, and the style score list of the target cluster is recorded as the score reference list of the target user;

[0093] Otherwise, a reference score for each style tag is determined according to the second style preference reference vector of the target user, and a score reference list for the target user is constructed according to the reference scores of multiple style tags.

[0094] A recommendation parameter analysis module is used to obtain the educational book query needs of the target user, determine multiple target recommendation objects according to the educational book query needs of the target user, calculate the first recommendation parameters of the multiple target recommendation objects based on the style reference list of the target user, extract the book interaction state parameters of the target user from the second historical interaction data, extract the state influence parameters of each style label from the book interaction state parameters and construct a state reference list, and calculate the second recommendation parameters of the multiple target recommendation objects based on the state reference list;

[0095] The educational book recommendation module is used to recommend books to target users based on a first recommendation parameter and a second recommendation parameter.

[0096] Specifically, the recommendation score of each target recommendation object is calculated according to the weight parameters corresponding to the first recommendation parameter and the second recommendation parameter respectively, the recommendation priorities of multiple target recommendation objects are sorted according to the recommendation scores, and books are recommended to the target user based on the recommendation priority order of the multiple target recommendation objects.

[0097] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An intelligent recommendation method for educational books, characterized in that: include: Collecting first historical interaction data of multiple platform users on educational books, extracting first content style preference data of each platform user from the first historical interaction data, and extracting a first style preference reference vector of each platform user based on the first content style preference data; Cluster multiple platform users based on style preference feature vectors to obtain multiple user clusters, extract the group preference feature vector of each user cluster based on the cluster center of each user cluster, and determine the style score list of each user cluster based on the group preference feature vector; Acquire second historical interaction data of the target user, extract second content style preference data of the target user from the second historical interaction data, the content style preference data including score data of multiple style tags, and extract a second style preference reference vector of the target user based on the second content style preference data; Determine whether there is a target cluster corresponding to the target user among the multiple user clusters, and if so, determine a rating reference list for the target user according to the user cluster to which the target user belongs; otherwise, determine a rating reference list for the target user based on the second style preference reference vector of the target user; Acquire the educational book query needs of the target user, determine multiple target recommendation objects according to the educational book query needs of the target user, and calculate first recommendation parameters of the multiple target recommendation objects based on the style reference list of the target user; Extracting the target user's book interaction state parameters from the second historical interaction data, extracting the state influence parameters of each style tag from the book interaction state parameters and constructing a state reference list, and calculating the second recommendation parameters of multiple target recommendation objects based on the state reference list; Books are recommended to a target user using a plurality of target recommendation objects according to the first recommendation parameter and the second recommendation parameter.

2. The intelligent recommendation method for educational books according to claim 1, characterized in that: Extracting first content style preference data of each platform user from the first historical interaction data, and extracting a first style preference reference vector of each platform user based on the first content style preference data, including: Determine multiple educational book interactive events in the first historical interactive data, extract the timestamp and interactive book preference data of each educational book interactive event, the interactive book preference data includes the interactive scores of multiple style tags in the educational book interactive event, and use the following formula to integrate the interactive book preference data of multiple educational book interactive events: Get the first content style preference data corresponding to each set of first historical interaction data: In the formula, Indicates the total scores of platform users corresponding to multiple style tags. Indicates that the platform user is Educational book interactive event about The interactive score of style tags, Indicates that the platform user is Educational book interactive event about The decay factor of the interaction score of style labels, where Indicates the current time. Indicates The timestamp of each educational book interaction event, is the attenuation base, which is a constant less than 1. represents the total number of educational book interaction events, Indicates the total number of style tags; The above formula is used to calculate the total score of each style label, and the first content style preference data corresponding to each group of first historical interaction data is generated. The first content style preference data includes the total scores corresponding to multiple style labels. Based on the first content style preference data, the first style preference reference vector of each platform user is extracted.

3. The intelligent recommendation method for educational books according to claim 2, characterized in that: The rating reference list for target users also includes: Based on the second style preference reference vector and the plurality of group preference feature vectors, the distances between the target user and the plurality of user clusters are respectively calculated. If there is at least one user cluster among the plurality of user clusters whose distance to the target user is less than a preset similarity threshold, it is considered that there is a target cluster corresponding to the target user among the plurality of user clusters, and the user cluster with the shortest distance is selected from the user clusters whose distance to the target user is less than the preset similarity threshold and recorded as the target cluster, and the style score list of the target cluster is recorded as the score reference list of the target user; Otherwise, a reference score for each style tag is determined according to the second style preference reference vector of the target user, and a score reference list for the target user is constructed according to the reference scores of multiple style tags.

4. The intelligent recommendation method for educational books according to claim 3, characterized in that: For the first recommended parameter and the second recommended parameter, it also includes: Determine the interaction scores corresponding to the multiple style tags of each target recommendation object, perform weighted analysis on the multiple interaction scores according to the style tag score corresponding to each style tag in the style reference list of the target user, and calculate the first recommendation parameter of each target recommendation object; Determine the interaction status data corresponding to each educational interaction event in the second historical interaction data, extract the book interaction status parameters of each book interaction feature from the interaction status data, process the book interaction status parameters according to the interactive book preference data of each educational book interaction event, and obtain the interaction status analysis result of each educational book interaction event, fuse multiple interaction status analysis results to obtain the state influence parameter of each style label, construct a state reference list containing the state influence parameters corresponding to multiple style labels, perform weighted analysis on the interaction scores corresponding to multiple style labels of each target recommendation object according to the state influence parameter corresponding to each style label in the state reference list of the target user, and calculate the second recommendation parameter of each target recommendation object.

5. The intelligent recommendation method for educational books according to claim 4, characterized in that: Recommending books to a target user using a plurality of target recommendation objects according to the first recommendation parameter and the second recommendation parameter includes: The recommendation score of each target recommendation object is calculated according to the weight parameters corresponding to the first recommendation parameter and the second recommendation parameter respectively, the recommendation priority of multiple target recommendation objects is sorted according to the recommendation score, and books are recommended to the target user based on the recommendation priority order of the multiple target recommendation objects.

6. The intelligent recommendation method for educational books according to claim 5, characterized in that: For style score lists, also include: Cluster multiple platform users based on the DBSCAN clustering algorithm to obtain multiple user clusters, and calculate the cluster center of each user cluster; For any user cluster, calculate the distance between each object in the user cluster and the cluster center, arrange all objects in ascending order based on the distance between the object and the cluster, draw a curve showing the cumulative number of individuals changing with distance, determine the target distance point in the curve based on the elbow method, record the objects within the target distance point as the core group of the user cluster, calculate the mean of the first style preference reference vector of the core group and record it as the group preference feature vector of the user cluster, determine the style label score of each style label based on the elements corresponding to each style label in the group preference feature vector, and construct a style score list containing the style label scores corresponding to multiple style labels.

7. An intelligent recommendation system for educational books, characterized in that: The system is used to implement the intelligent recommendation method for educational books as described in any one of claims 1 to 6, comprising: A style analysis module, used to collect first historical interaction data of multiple platform users on educational books, extract first content style preference data of each platform user from the first historical interaction data, and extract a first style preference reference vector of each platform user based on the first content style preference data; A group clustering module is used to cluster multiple platform users based on style preference feature vectors to obtain multiple user clusters, extract the group preference feature vector of each user cluster based on the cluster center of each user cluster, and determine the style score list of each user cluster based on the group preference feature vector; A target user analysis module obtains second historical interaction data of the target user, extracts second content style preference data of the target user from the second historical interaction data, wherein the content style preference data includes rating data of multiple style tags, and extracts a second style preference reference vector of the target user based on the second content style preference data; A rating list generation module is used to determine whether there is a target cluster corresponding to the target user among the multiple user clusters. If so, a rating reference list for the target user is determined according to the user cluster to which the target user belongs; otherwise, a rating reference list for the target user is determined based on the second style preference reference vector of the target user. A recommendation parameter analysis module is used to obtain the educational book query needs of the target user, determine multiple target recommendation objects according to the educational book query needs of the target user, calculate the first recommendation parameters of the multiple target recommendation objects based on the style reference list of the target user, extract the book interaction state parameters of the target user from the second historical interaction data, extract the state influence parameters of each style label from the book interaction state parameters and construct a state reference list, and calculate the second recommendation parameters of the multiple target recommendation objects based on the state reference list; The educational book recommendation module is used to recommend books to target users based on a first recommendation parameter and a second recommendation parameter.

8. An intelligent recommendation system for educational books according to claim 7, characterized in that: For the educational book recommendation module, multiple target recommendation objects are recommended to target users according to the first recommendation parameter and the second recommendation parameter, including: The recommendation score of each target recommendation object is calculated according to the weight parameters corresponding to the first recommendation parameter and the second recommendation parameter respectively, the recommendation priority of multiple target recommendation objects is sorted according to the recommendation score, and books are recommended to the target user based on the recommendation priority order of the multiple target recommendation objects.

Images