A cloud computing-based online learning data processing method and cloud platform

Abstract

The present invention discloses an online learning data processing method and cloud platform based on cloud computing, belonging to the field of data processing technology. The present invention mines the initial user behavior attribute vector of the learning behavior data to be processed and the historical user behavior attribute vector associated therewith, performs iterative linkage behavior preference embedding, and finally obtains personalized resource push recommendation features based on the correlation characteristics between the linkage behavior preference embedding vector and the initial user behavior attribute. This effectively solves the problems existing in the traditional method, can provide users with more accurate and personalized learning resource push, improve user learning efficiency and satisfaction, and also contribute to resource optimization and quality and efficiency improvement of platform operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to an online learning data processing method and a cloud platform based on cloud computing. Background Art

[0002] In today's era of rapid digital development, online learning has become an important way for people to acquire knowledge. With the continuous emergence of online learning platforms and the rapid growth in the number of users, how to effectively provide users with personalized learning resources has become a technical problem that needs to be solved urgently.

[0003] Traditional online learning resource push methods have many limitations. On the one hand, many platforms push resources based on simple analysis of users' current learning behavior data. For example, similar course resources are pushed only based on the viewing time or correct answer rate of the course that the user is currently studying. This method ignores the user's historical learning behavior, and the user's learning preference is a long-term process. The historical learning behavior contains a lot of valuable information for judging the user's learning preference. This leads to the possibility that the pushed resources may deviate greatly from the user's actual needs and fail to meet the user's personalized learning needs.

[0004] On the other hand, even if some platforms take historical learning behaviors into account, they lack an effective integration and analysis method. The processing of historical and current learning behavior data is often isolated, and there is no mechanism to organically combine the two to deeply explore users' behavioral preferences. Moreover, when analyzing user behavioral preferences, most platforms use a relatively rough method and lack an accurate description of behavioral preferences. For example, users are simply divided into theoretical learners and practical learners, but the specific preferences of users in different learning concepts or links cannot be analyzed in detail.

[0005] In addition, due to the lack of accurate analysis of user behavior preferences, the push of learning resources is often generalized and cannot be customized according to the unique needs of each user. This not only reduces the user's learning efficiency, because the user may receive a large number of irrelevant learning resources and need to spend extra time to screen; it also affects the user's satisfaction and loyalty to the online learning platform. From the perspective of platform operation, generalized resource push also leads to a waste of platform resources. The platform needs to store and transmit a large number of learning resources that may not be used by users, increasing operating costs. Summary of the invention

[0006] The present invention provides an online learning data processing method and a cloud platform based on cloud computing, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0007] The present invention provides an online learning data processing method based on cloud computing, which is applied to an online learning cloud platform. The method comprises: obtaining learning behavior data to be processed, and mining the initial user behavior attributes of the learning behavior data to be processed to obtain the initial user behavior attribute vector of the learning behavior data to be processed; determining the historical user behavior attributes associated with the learning behavior data to be processed, and mining the historical user behavior attributes to obtain the historical user behavior attribute vector; the historical user behavior attributes are obtained by identifying the target learning behavior data corresponding to the learning behavior data to be processed and not including the push interest items; performing at least one round of iterative linkage behavior preference embedding based on the initial user behavior attribute vector and the historical user behavior attribute vector to obtain the linkage behavior of the at least one round of iteration. The linkage behavior preference embedding vector output by preference embedding; in each round of iterative linkage behavior preference embedding task, key behavior element sampling is performed on the target initial user behavior attribute vector and the target historical user behavior attribute vector input in the corresponding iteration cycle to obtain the historical user behavior attribute vector output in the corresponding iteration cycle, and key behavior element sampling is performed on the historical user behavior attribute vector output in the corresponding iteration cycle and the target initial user behavior attribute vector to obtain the initial user behavior attribute vector output in the corresponding iteration cycle; based on the linkage behavior preference embedding vector, a learning resource interest description for the learning behavior data to be processed is obtained; according to the correlation characteristics between the learning resource interest description and the initial user behavior attributes, a resource push recommendation feature for the learning behavior data to be processed is obtained.

[0008] The present invention provides an online learning cloud platform, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0009] The present invention provides a readable storage medium, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0010] The present invention mines the initial user behavior attribute vector of the learning behavior data to be processed and the historical user behavior attribute vector associated therewith, performs iterative linkage behavior preference embedding, and finally obtains personalized resource push recommendation features based on the correlation characteristics between the linkage behavior preference embedding vector and the initial user behavior attribute. This effectively solves the problems existing in the traditional method, can provide users with more accurate and personalized learning resource push, improve user learning efficiency and satisfaction, and also contribute to resource optimization and quality and efficiency improvement of platform operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of an online learning data processing method based on cloud computing provided by the present invention.

[0012] Figure 2 This is a structural schematic diagram of an online learning cloud platform provided by the present invention. DETAILED DESCRIPTION

[0013] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0015] Figure 1 A cloud computing-based online learning data processing method is shown, which is applied to an online learning cloud platform. The method includes the following steps 110 to 130.

[0016] Step 110: The online learning cloud platform obtains the learning behavior data to be processed, and mines the initial user behavior attributes of the learning behavior data to be processed to obtain the initial user behavior attribute vector of the learning behavior data to be processed; determines the historical user behavior attributes associated with the learning behavior data to be processed, and mines the historical user behavior attributes to obtain the historical user behavior attribute vector.

[0017] The historical user behavior attributes are obtained by identifying the target learning behavior data that does not include the pushed interest items and corresponds to the learning behavior data to be processed.

[0018] In the present invention, the core idea of ​​step 110 is to obtain data and mine behavior attribute vectors, which is explained in two aspects below.

[0019] 1. Obtaining the learning behavior data to be processed and mining the initial user behavior attribute vector

[0020] The online learning cloud platform must first obtain the learning behavior data to be processed. In actual online learning scenarios, this data can be diverse. For example, on a large online education platform, the learning behavior data to be processed may include information such as the user's viewing time in the course video, the number of pauses, whether the course notes were taken, whether the course discussion was participated in, etc. The online learning cloud platform mines the initial user behavior attributes of these learning behavior data to be processed, thereby obtaining the initial user behavior attribute vector.

[0021] For example, a user who is learning a programming course watched an introductory course video about Python programming. He watched the video for 30 minutes, paused it 5 times, and took a few notes. The online learning cloud platform will quantify and encode these behavior attributes. For example, the viewing time can be converted into a numerical value according to a certain ratio, the number of pauses can be directly used as a numerical value, and the number of notes can also be quantified as a numerical value, and then combined into an initial user behavior attribute vector, such as [0.5, 5, 0.2] (this is an illustrative example, the actual situation will be more complicated).

[0022] 2. Determine historical user behavior attributes and mine their vectors

[0023] The online learning cloud platform also needs to determine the historical user behavior attributes associated with the learning behavior data to be processed. The historical user behavior attributes in the present invention are obtained by identifying the target learning behavior data corresponding to the learning behavior data to be processed, excluding the push interest items. On the online learning platform, the historical user behavior attributes may be the user's behavior when studying other programming courses in the past, such as the study time and the completed homework status when studying the Java course before.

[0024] For a user who previously studied Java, he completed 80% of the course assignments, watched the course videos for an average of 40 minutes, and rarely participated in course discussions. The online learning cloud platform also mines these historical behavior attributes to obtain a historical user behavior attribute vector, such as [0.8, 0.67, 0.1] (note that 0.67 is the ratio of 40 minutes of viewing time to a standard time, which is also an example).

[0025] Step 120: The online learning cloud platform performs at least one round of iterative linkage behavior preference embedding based on the initial user behavior attribute vector and the historical user behavior attribute vector to obtain a linkage behavior preference embedding vector output by the at least one round of iterative linkage behavior preference embedding.

[0026] Among them, in each round of iterative linkage behavior preference embedding task, key behavior element sampling is performed on the target initial user behavior attribute vector and the target historical user behavior attribute vector input in the corresponding iteration period to obtain the historical user behavior attribute vector output in the corresponding iteration period, and key behavior element sampling is performed on the historical user behavior attribute vector output in the corresponding iteration period and the target initial user behavior attribute vector to obtain the initial user behavior attribute vector output in the corresponding iteration period.

[0027] In the present invention, step 120 focuses on embedding linkage behavior preferences based on vectors, which is introduced below through three exemplary sub-steps.

[0028] 1.1 Overview of the Iteration Process

[0029] The online learning cloud platform embeds the linkage behavior preference for at least one round of iteration based on the initial user behavior attribute vector and the historical user behavior attribute vector. In each round of iteration, there is a specific task process.

[0030] 2. Sampling of key behavioral elements in each iteration

[0031] 1. Sampling of initial user behavior attribute vector and historical user behavior attribute vector

[0032] In each round of iterative linkage behavior preference embedding task, key behavior element sampling is performed on the target initial user behavior attribute vector and the target historical user behavior attribute vector input in the corresponding iteration cycle.

[0033] Taking the example of programming course learning mentioned above, in the first round of iteration, for the initial user behavior attribute vector [0.5, 5, 0.2] of the Python course and the historical user behavior attribute vector [0.8, 0.67, 0.1] of the Java course, the online learning cloud platform may determine the key behavior elements according to a certain algorithm. For example, the online learning cloud platform considers the viewing time and homework completion status as key behavior elements, and then resamples or adjusts the weights of these elements. For example, after sampling, an adjusted historical user behavior attribute vector [0.7, 0.7, 0.1] is obtained.

[0034] 2. Sampling again to obtain the adjusted initial user behavior attribute vector

[0035] Next, the key behavior elements are sampled for the historical user behavior attribute vector and the target initial user behavior attribute vector output in the corresponding iteration cycle to obtain the initial user behavior attribute vector output in the corresponding iteration cycle. For example, based on the adjusted historical user behavior attribute vector [0.7, 0.7, 0.1] and the original Python course initial user behavior attribute vector [0.5, 5, 0.2], a new initial user behavior attribute vector [0.6, 4, 0.2] is obtained after sampling again.

[0036] 3. Multiple rounds of iteration

[0037] This process will be repeated for at least one round. As the number of iterations increases, the initial user behavior attribute vector and the historical user behavior attribute vector are constantly adjusted and optimized, gradually approaching a vector representation that can accurately reflect the user's learning behavior preferences. For example, after three rounds of iterations, the final linkage behavior preference embedding output linkage behavior preference embedding vector may be [0.65, 4.5, 0.15]. This vector comprehensively considers the user's initial behavior in the current learning (Python course) and the behavior in the historical learning (Java course), and more accurately reflects the user's behavior preference in programming learning.

[0038] Step 130: The online learning cloud platform obtains a learning resource interest description for the learning behavior data to be processed based on the linkage behavior preference embedding vector; and obtains a resource push recommendation feature for the learning behavior data to be processed based on the correlation characteristics between the learning resource interest description and the initial user behavior attributes.

[0039] In the present invention, step 130 is one of the core invention points, focusing on the specific implementation method of obtaining the learning resource interest description and resource push suggestion features based on the vector.

[0040] 1. Obtaining learning resource interest description based on linkage behavior preference embedding vector

[0041] The online learning cloud platform obtains the interest description of learning resources for the learning behavior data to be processed based on the linkage behavior preference embedding vector. Continuing with programming learning as an example, if the linkage behavior preference embedding vector is [0.65, 4.5, 0.15], the online learning cloud platform may interpret this vector as indicating that the user is more interested in the basic theory part of programming learning (corresponding to a certain behavior element related to the basic theory represented by the first value), and has a certain interest in the practical operation part (corresponding to the behavior element related to the practical operation represented by the middle value), but has less interest in the interactive communication part (corresponding to the behavior element related to the interactive communication represented by the last value).

[0042] Based on such a vector, the online learning cloud platform can generate a description of learning resource interests as "the user is interested in basic programming theory and practical operations, has low interest in interactive communication, and is suitable for pushing more learning resources related to theoretical explanations and practical cases."

[0043] 2. Obtaining resource push suggestion features based on relevance features

[0044] The online learning cloud platform obtains the resource push suggestion feature for the learning behavior data to be processed based on the correlation characteristics between the learning resource interest description and the initial user behavior attributes. In the previous example, the initial user behavior attribute vector [0.5, 5, 0.2] indicates that the viewing time is not very long and there are few notes. Combined with the learning resource interest description, the resource push suggestion feature may be "push a concise version of basic theoretical knowledge, and match it with simple practical cases, while reducing the push of interactive communication resources, such as reducing the push of forum resources."

[0045] In another online learning scenario, such as a foreign language learning platform. The user's pending learning behavior data shows his learning behavior in the word memorization course. The initial user behavior attribute vector may include the speed of memorizing words, the frequency of review, etc. The historical user behavior attributes may be his behavior in the previous grammar course. After the previous steps, after obtaining the linkage behavior preference embedding vector, the learning resource interest description may be "the user has a demand for fast memorization and efficient review of words, and has a general interest in in-depth understanding of grammar rules", and the resource push suggestion feature may be "push more resources on word memorization skills, such as associative memory courses, and simple review plan making tools, and reduce the push of resources for in-depth grammar explanations."

[0046] It is worth mentioning that step 130 plays a vital role in the design of the entire online learning cloud platform. It is based on the mining of user behavior attribute vectors in the previous step 110 and the embedding of linkage behavior preferences in step 120. It aims to generate a learning resource interest description for the learning behavior data to be processed by in-depth analysis of the obtained linkage behavior preference embedding vector, and further obtain accurate resource push recommendation features based on the correlation characteristics between the learning resource interest description and the initial user behavior attributes. The implementation of this step will directly affect whether the online learning platform can provide users with personalized and accurate learning resource push based on their learning behavior and preferences, thereby improving their learning experience and learning effect.

[0047] The online learning cloud platform obtains the learning resource interest description for the learning behavior data to be processed based on the linkage behavior preference embedding vector. This process first involves the interpretation of the linkage behavior preference embedding vector. Taking programming learning as an example, the linkage behavior preference embedding vector is [0.65, 4.5, 0.15]. Each value in this vector does not exist in isolation, but is closely related to the different behavioral elements of the user in the programming learning process.

[0048] For the first value 0.65 in the vector, the online learning cloud platform will correspond it to the behavioral elements related to the basic theory part of programming learning. This means that this value reflects a certain degree of inclination or preference of the user in basic theory learning. For example, in programming learning, the basic theory part may include knowledge content such as data types, variable definitions, and grammatical rules. If this value is high, it may mean that when the user is learning the introductory Python programming course or the previous Java course, he or she has shown a high degree of attention to the learning of these basic theoretical knowledge, such as watching the part involving basic theory explanations more carefully when watching the video, or being more proficient in the application of basic theoretical knowledge in the process of completing homework.

[0049] The middle value of 4.5 is associated with behavioral factors related to the practical operation part. In programming learning, practical operations include writing code, debugging programs, solving actual programming problems, etc. This value reflects the user's preference for these practical operations. The size of the value may imply the user's level of involvement, interest level, or proficiency in practical operations during the learning process. For example, if this value is relatively high, it may mean that the user actively participated in code writing exercises in previous learning, and took the initiative to try different solutions when encountering problems, showing a strong interest in the practical operation part.

[0050] The last value, 0.15, is related to the behavioral elements related to the interactive communication part. In the online programming learning environment, interactive communication includes participating in course forum discussions, exchanging codes with classmates or teachers, and participating in programming learning groups. The lower value of 0.15 indicates that the user has a lower interest in these interactive communications, and may have rarely participated in course discussions or interacted with others about programming in the previous learning process.

[0051] As for generating learning resource interest descriptions, the online learning cloud platform can generate learning resource interest descriptions by analyzing the correspondence between each value and different behavioral elements in the linkage behavior preference embedding vector. In the above example of programming learning, according to the vector [0.65, 4.5, 0.15], the learning resource interest description generated by the online learning cloud platform is "The user is interested in the basic theory and practical operation of programming, has low interest in interactive communication, and is suitable for pushing more theoretical explanations and practical case-related learning resources." This learning resource interest description is a general expression of the user's learning preference, which provides an important basis for the generation of subsequent resource push suggestion features.

[0052] Furthermore, obtaining resource push suggestion features based on correlation features can be expanded through the following three aspects.

[0053] 1. Definition and connotation of correlation characteristics

[0054] The correlation feature refers to the intrinsic connection and degree of association between the learning resource interest description and the initial user behavior attributes. It is a concept that comprehensively considers multiple factors and aims to find the logical relationship between the user's learning preference (reflected by the learning resource interest description) and the user's actual initial behavior in the learning process (reflected by the initial user behavior attributes). This logical relationship is not a simple one-to-one mapping, but a complex relationship network of multiple factors interacting with each other.

[0055] The initial user behavior attributes include multiple dimensions, such as viewing time, number of pauses, and note-taking in programming learning examples. These different dimensions of behavior attributes may affect each other and have different degrees of correlation with different parts of the learning resource interest description. For example, shorter viewing time may be related to the speed of understanding theoretical knowledge, while fewer notes may imply a need for simplicity in learning resources, which are all related to the theoretical explanation part of the learning resource interest description.

[0056] The description of learning resource interest is a comprehensive summary of the user's interest in multiple behavioral elements. This comprehensiveness makes its correlation with the initial user behavior attributes complex. For example, when the description of learning resource interest indicates that the user is interested in practical operations, the number of pauses in the initial user behavior attributes may be related to the user's pauses in thinking when encountering problems in practical operations, while the completion of homework directly reflects the user's actual ability and investment in practical operations, which all reflect the correlation characteristics between the two.

[0057] 2.2 The process of generating resource push suggestion features

[0058] In the programming learning example mentioned above, the initial user behavior attribute vector is [0.5, 5, 0.2], where 0.5 indicates a relatively short viewing time, 5 indicates a high number of pauses, and 0.2 indicates a low number of notes. Combined with the previously generated learning resource interest description "This user is interested in programming basic theory and practical operations, but has low interest in interactive communication, and is suitable for pushing more theoretical explanations and practical case-related learning resources", let's analyze the generation process of resource push suggestion features in detail.

[0059] Since the viewing time is relatively short, this may mean that users want to acquire key theoretical knowledge in a shorter time. Therefore, combined with the interest in programming basic theory in the learning resource interest description, the theoretical knowledge part of the resource push suggestion feature should be pushed in a concise version. For example, push a brief and refined summary of programming theory knowledge points, avoid lengthy and complex theoretical explanation videos or documents, to meet the user's needs to acquire theoretical knowledge within a limited viewing time.

[0060] Although the number of pauses in the initial user behavior attribute vector is relatively high, considering that users are interested in practical operations, this may mean that users have encountered many problems that need to be thought about during the practical operation process. Therefore, in the resource push suggestion feature, for the practical operation part, simple practical cases should be pushed. These practical cases can be representative small programming projects, each of which comes with detailed operation steps and explanations to help users better understand and master practical operation skills, while not frustrating users because the cases are too complicated.

[0061] The small number of notes in the initial user behavior attribute vector and the low interest in interactive communication in the description of learning resource interests indicate that users may not rely too much on interactive communication to assist learning, and do not pay much attention to detailed records of knowledge. Therefore, in the resource push suggestion features, it is necessary to reduce the push of interactive resources, such as reducing the push of programming course forum links or online learning group event invitations, etc., to avoid pushing resources that users are not interested in, thereby improving the accuracy of resource push.

[0062] 3. Diversity and flexibility of resource push suggestion features

[0063] The generation of resource push suggestion features is not limited to programming learning, but is also applicable to other subject areas and is diverse and flexible. For example, in the field of language learning, if the learning resource interest description shows that the user is interested in vocabulary accumulation and oral practice, and the initial user behavior attributes show that the user is less focused in listening practice and rarely takes language learning notes. Then the resource push suggestion feature may be to push concise vocabulary memory cards and practical oral practice scene videos, while reducing the push of listening practice materials, and not pushing grammar explanation materials that require a lot of notes.

[0064] Within the same subject, different learning stages will also affect the resource push suggestion features. Taking mathematics learning as an example, for students at the elementary level, if the learning resource interest description shows an interest in basic operations and simple geometric figures, the initial user behavior attributes indicate that the problem-solving speed in classroom exercises is slow and reference materials are often checked. Then the resource push suggestion can be to push the fast calculation skills of basic operations and the animation demonstration video of simple geometric figures, and provide detailed reference materials for problem-solving steps to help students improve their problem-solving speed. For students at the advanced level, if the learning resource interest description shows an interest in complex mathematical models and mathematical applications, the initial user behavior attributes show that they have a high degree of concentration when reading advanced mathematical literature but rarely participate in mathematical discussion groups. Then the resource push suggestion feature may be to push the latest mathematical research results and case analysis materials of complex mathematical models, while reducing the push of mathematical discussion group activities.

[0065] In addition, the resource push suggestion feature also needs to consider the personalized needs of users. Even under the same learning resource interest description and initial user behavior attributes, different users may have different needs due to personal learning habits, learning goals, and other factors. For example, some users may prefer to learn on mobile devices, so the resource push suggestion feature needs to consider pushing learning resources suitable for viewing and operation on mobile devices, such as learning tools in the form of mobile applications or learning materials in the form of short videos.

[0066] It can be seen that step 130 obtains the learning resource interest description based on the linkage behavior preference embedding vector, and obtains the resource push suggestion feature based on the correlation feature between the learning resource interest description and the initial user behavior attribute, which provides a comprehensive and accurate method for pushing learning resources to users for the online learning cloud platform. This method fully considers the various behavioral factors of users in the learning process, and can provide learning resource push suggestions that meet user needs based on the user's personalized learning preferences and actual learning behavior. Whether in different subject areas, different learning stages or meeting user personalized needs, it has demonstrated its high adaptability and effectiveness, thus providing strong support for improving the quality and efficiency of online learning.

[0067] In summary, by mining the initial user behavior attribute vector of the learning behavior data to be processed and the historical user behavior attribute vector associated with it, the online learning cloud platform can fully and deeply understand the user's learning behavior pattern. This comprehensiveness is reflected in not only focusing on the current learning behavior, but also combining historical learning behavior. For example, on an online language learning platform, the user is currently studying an advanced English course (learning behavior data to be processed), and its initial behavior attribute vector is mined, while considering the previous study of the intermediate English course (historical user behavior attributes). This helps the online learning cloud platform to more accurately grasp the behavioral characteristics of users throughout the learning process, overcoming the one-sided problem caused by only analyzing current behavior.

[0068] Secondly, it is very meaningful to perform at least one round of iterative linkage behavior preference embedding operations based on the initial user behavior attribute vector and the historical user behavior attribute vector. Sampling key behavior elements in each round of iteration can continuously optimize and refine the understanding of user behavior preferences. Taking online programming course learning as an example, each round of iteration may enable the online learning cloud platform to more accurately determine the user's preference for different programming concepts (such as algorithm complexity, code readability and other key behavior elements). As the iteration proceeds, the obtained linkage behavior preference embedding vector can more accurately reflect the user's real learning preference. Compared with the traditional one-time analysis method, this iterative method greatly improves the accuracy of user behavior preference analysis.

[0069] Furthermore, the learning resource interest description is obtained based on the linkage behavior preference embedding vector, and the resource push recommendation feature is obtained based on the correlation feature between the learning resource interest description and the initial user behavior attribute. This process realizes highly personalized learning resource push. In the field of online education, the learning needs and preferences of different users vary greatly. In the above way, the online learning cloud platform can tailor suitable learning resource push recommendations for each user. For example, for a programming learner who understands theoretical knowledge quickly but has less practical operations, the online learning cloud platform will push more practice-related learning resources, while for users who are relatively balanced in theory and practice, comprehensive learning resources will be pushed. This personalized push can significantly improve the user's learning efficiency, because the user receives resources that match their learning preferences and actual behaviors, reducing the time and energy waste of users in screening a large number of irrelevant resources, and at the same time can also improve the user's interest and satisfaction in learning.

[0070] Finally, from the perspective of the operation of online learning platforms, this precise resource push helps improve the utilization rate of platform resources. Instead of blindly pushing a large number of general learning resources to users, the platform can push them in a targeted manner according to the actual needs of users. This not only saves the platform's resource storage and transmission costs, but also improves user loyalty to the platform, because users can get a better learning experience on the platform, thereby attracting more users to use the platform and promoting the sustainable development of the online learning platform.

[0071] In some preferred embodiments, step 120 involves sampling key behavior elements for the target initial user behavior attribute vector and the target historical user behavior attribute vector input for the corresponding iteration cycle to obtain the historical user behavior attribute vector output for the corresponding iteration cycle, including: obtaining the first user operation interaction vector and the first platform response interaction vector based on the target initial user behavior attribute vector input for the corresponding iteration cycle, and obtaining the first learning resource relationship vector based on the target historical user behavior attribute vector input for the corresponding iteration cycle; sampling user group behavior elements for the first user operation interaction vector, the first platform response interaction vector, and the first learning resource relationship vector to obtain the historical user behavior attribute vector output for the corresponding iteration cycle.

[0072] In detail, the user group behavior element sampling is performed on the first user operation interaction vector, the first platform response interaction vector and the first learning resource relationship vector to obtain the historical user behavior attribute vector output by the corresponding iteration cycle, including: based on the first user operation interaction vector, the first platform response interaction vector and the first learning resource relationship vector, the user group behavior element sampling is performed based on the feature focus strategy to obtain the first user group behavior element vector; the first user group behavior element vector and the first learning resource relationship vector are feature integrated to obtain a target integration vector; the target integration vector is soft-threshold activated at least once to obtain a behavior state activation vector; the behavior state activation vector and the target integration vector are feature integrated to obtain the historical user behavior attribute vector output by the corresponding iteration cycle.

[0073] In the online learning cloud platform, the key part of step 120 involves specific sampling of the vector input in the iteration cycle to obtain the output historical user behavior attribute vector. The following is a detailed description of this process.

[0074] (I) Obtaining related vectors from input vectors

[0075] 1. Obtain the relevant vector from the target initial user behavior attribute vector

[0076] The first user operation interaction vector and the first platform response interaction vector are obtained according to the target initial user behavior attribute vector input in the corresponding iteration cycle. In actual operation, the target initial user behavior attribute vector is a quantitative representation of the various attributes of the user in the initial learning behavior. For example, the target initial user behavior attribute vector is [0.3, 0.5, 0.2] (only an example in the present invention), where 0.3 may represent the user's operation frequency in a certain learning link, 0.5 represents a certain complexity of the operation, and 0.2 represents the degree of deviation between the operation and the expected result, etc. Based on such a target initial user behavior attribute vector, the first user operation interaction vector and the first platform response interaction vector can be obtained by an adaptive algorithm.

[0077] In actual application, the conversion can be performed according to the proportional relationship of each value in the vector and the pre-set rules. For example, for the value 0.3 in the target initial user behavior attribute vector, if under the corresponding rules, its quantitative relationship with the user operation interaction is multiplied by 2, then the corresponding value in the first user operation interaction vector may be 0.6. At the same time, according to a certain association rule between the value 0.5 and the platform response interaction (for example, multiplied by 1.5), the corresponding value in the first platform response interaction vector may be 0.75. Of course, this is just an exemplary explanation for the purpose of illustrating the conversion process, and the actual algorithm will be more complicated.

[0078] 2. Obtain the first learning resource relationship vector from the target historical user behavior attribute vector

[0079] The first learning resource relationship vector is obtained according to the target historical user behavior attribute vector input in the corresponding iteration cycle. The target historical user behavior attribute vector is also a quantitative representation of the user's past learning behavior. For example, the target historical user behavior attribute vector is [0.4, 0.6, 0.1], where 0.4 may represent the frequency of contact with a certain type of learning resource in historical learning, 0.6 represents the degree of mastery of these resources, and 0.1 represents the error rate when using these resources. Through an adaptive algorithm, based on these values ​​and their intrinsic connection with the relationship between learning resources, the first learning resource relationship vector can be obtained. For example, according to the association rule between the frequency of contact and the resource type preference in the resource relationship (such as multiplying by 3), if 0.4 corresponds to the frequency of contact with a certain type of learning resource, then in the first learning resource relationship vector, the value related to the resource type preference may be 1.2.

[0080] 2. Sampling user group behavior elements based on multiple vectors

[0081] 1. Based on the feature focus strategy, the user group behavior elements are sampled to obtain the first user group behavior element vector

[0082] The user group behavior elements are sampled for the first user operation interaction vector, the first platform response interaction vector, and the first learning resource relationship vector, and the operation is first performed based on the feature focus strategy. The feature focus strategy is a method that aims to highlight the elements that have a key impact on the user group behavior. For example, among the numerous user operation interactions, platform response interactions, and learning resource relationship elements, there may be some elements that play a decisive role in the behavior pattern of the entire user group.

[0083] For example, in the scenario of an online mathematics learning platform, the value in the first user operation interaction vector is related to the user's operation in the process of solving mathematics problems (such as the number of answering steps, the frequency of using tools, etc.), the value in the first platform response interaction vector is related to the platform's feedback on the user's answering (such as the number of prompts, the timing of displaying the correct answer, etc.), and the value in the first learning resource relationship vector is related to the user's use of mathematics learning materials (such as textbooks, exercises, video tutorials, etc.). The feature focus strategy may determine the number of answering steps, the number of platform prompts, and the frequency of textbook use as key user group behavior elements. Through adaptive algorithms (such as screening and combining according to the weight of each element in the overall behavior pattern), these key elements are combined into the first user group behavior element vector. For example, after calculation, the first user group behavior element vector is [0.5, 0.3, 0.2], where 0.5 may be the weight value of the number of answering steps after algorithm conversion, 0.3 is the weight value after the number of platform prompts, and 0.2 is the weight value after the frequency of textbook use.

[0084] 2. Perform feature integration on the first user group behavior element vector and the first learning resource relationship vector to obtain the target integration vector

[0085] It is a key step to integrate the features of the first user group behavior element vector and the first learning resource relationship vector. Feature integration aims to fuse vector information from different sources to obtain more comprehensive and integrated information. For example, the first user group behavior element vector is [0.5, 0.3, 0.2], and the first learning resource relationship vector is [1.2, 0.8, 0.1]. Through a weighted summation algorithm (this is just an example algorithm, and it may be more complicated in practice), for example, multiply each value in the first user group behavior element vector by 2, multiply each value in the first learning resource relationship vector by 1.5, and then add the values ​​in the corresponding positions to obtain the target integration vector. That is: [(0.5*2+1.2*1.5), (0.3*2+0.8*1.5), (0.2*2+0.1*1.5)]=[2.8, 1.8, 0.55]. This target integration vector integrates the information of user group behavior elements and learning resource relationships.

[0086] 3. Perform at least one soft threshold activation on the target integrated vector to obtain the behavior state activation vector

[0087] The purpose of performing at least one soft threshold activation on the target integrated vector is to further mine the behavioral state information hidden in the vector. Soft threshold activation is a nonlinear transformation operation that can adjust the values ​​in the vector according to a preset threshold. For example, the soft threshold is set to 0.5. For 2.8 in the target integrated vector [2.8, 1.8, 0.55], if it is greater than 0.5, a certain calculation (such as multiplying by 0.8) is performed to obtain 2.24; for 1.8, if it is greater than 0.5, multiply by 0.7 to obtain 1.26; for 0.55, if it is greater than 0.5, multiply by 0.6 to obtain 0.33. After such a soft threshold activation operation, the behavioral state activation vector [2.24, 1.26, 0.33] is obtained. This process can be performed multiple times, each time using a different threshold or the same threshold but a different calculation rule to more finely mine the behavioral state information.

[0088] 4. Perform feature integration on the behavior state activation vector and the target integration vector to obtain the historical user behavior attribute vector outputted in the corresponding iteration cycle

[0089] Finally, the behavior state activation vector and the target integration vector are feature integrated to obtain the historical user behavior attribute vector output in the corresponding iteration cycle. This step again fuses the behavior state information after soft threshold activation with the original target integration vector information. For example, the behavior state activation vector is [2.24, 1.26, 0.33], and the target integration vector is [2.8, 1.8, 0.55]. Through a weighted summation algorithm similar to the previous one (such as multiplying each value in the behavior state activation vector by 1.2, multiplying each value in the target integration vector by 0.8, and then adding the values ​​at the corresponding positions), the historical user behavior attribute vector output in the corresponding iteration cycle is obtained. For example, after calculation, [(2.24*1.2+2.8*0.8), (1.26*1.2+1.8*0.8), (0.33*1.2+0.55*0.8)]=[4.256, 2.712, 0.796] is obtained. This output historical user behavior attribute vector will serve as an important basis for the next round of iteration or subsequent processing.

[0090] It can be seen that this embodiment mines the historical user behavior attribute vector outputted by the corresponding iteration cycle from the target initial user behavior attribute vector and the target historical user behavior attribute vector. Its beneficial effects are reflected in many aspects. First, by obtaining the first user operation interaction vector and the first platform response interaction vector from the target initial user behavior attribute vector, and obtaining the first learning resource relationship vector from the target historical user behavior attribute vector, it is possible to effectively integrate information such as the user's operation behavior, the platform's response feedback, and the relationship between learning resources, providing a comprehensive data basis for subsequent analysis. Secondly, sampling user group behavior elements based on the feature focus strategy helps to focus on the elements that have a key impact on the user group behavior, avoid blind analysis in a large amount of data, and improve the pertinence and accuracy of the analysis. Furthermore, operations such as feature integration and soft threshold activation can further mine the potential information in the data, so that the historical user behavior attribute vector finally obtained can more accurately reflect the user's behavior state, which helps to more accurately perform operations such as resource push and user behavior analysis in the entire online learning cloud platform, thereby improving the user's learning experience and the operating efficiency of the online learning cloud platform.

[0091] In some further optional embodiments, step 120 involves sampling key behavior elements for the historical user behavior attribute vector outputted by the corresponding iteration cycle and the target initial user behavior attribute vector to obtain the initial user behavior attribute vector outputted by the corresponding iteration cycle, including: obtaining a second user operation interaction vector and a second platform response interaction vector based on the historical user behavior attribute vector outputted by the corresponding iteration cycle, and obtaining a second learning resource relationship vector based on the target initial user behavior attribute vector inputted by the corresponding iteration cycle; sampling user group behavior elements for the second user operation interaction vector, the second platform response interaction vector and the second learning resource relationship vector to obtain the initial user behavior attribute vector outputted by the corresponding iteration cycle.

[0092] In detail, the user group behavior element sampling is performed on the second user operation interaction vector, the second platform response interaction vector and the second learning resource relationship vector to obtain the initial user behavior attribute vector outputted by the corresponding iteration cycle, including: based on the second user operation interaction vector, the second platform response interaction vector and the second learning resource relationship vector, user group behavior element sampling is performed based on the feature focus strategy to obtain the second user group behavior element vector; the second user group behavior element vector is feature integrated with the second learning resource relationship vector to obtain the initial user behavior attribute vector outputted by the corresponding iteration cycle.

[0093] In this embodiment, key behavior element sampling is performed on the historical user behavior attribute vector and the target initial user behavior attribute vector output from the corresponding iteration cycle to obtain the initial user behavior attribute vector output from the corresponding iteration cycle. An exemplary implementation method is given below, which will be explained in detail below.

[0094] (I) Obtaining the corresponding vector from the related vector

[0095] 1. Obtain the relevant vector from the historical user behavior attribute vector output by the corresponding iteration cycle

[0096] The second user operation interaction vector and the second platform response interaction vector are obtained according to the historical user behavior attribute vector outputted in the corresponding iteration cycle. For example, the historical user behavior attribute vector outputted in the corresponding iteration cycle is [0.6, 0.4, 0.3] (the values ​​in this embodiment are only examples). The values ​​in this vector reflect the quantitative results of different attributes of the user in the historical learning behavior. Among them, 0.6 may represent the frequency of a certain behavior in the historical learning process, 0.4 may represent the quantification of the degree of mastery of a certain knowledge, and 0.3 may represent a quantitative indicator of interaction with learning resources, etc.

[0097] The second user operation interaction vector and the second platform response interaction vector are obtained from the historical user behavior attribute vector through an adaptive algorithm. For example, there is an algorithm rule that converts each numerical value in the historical user behavior attribute vector according to its predefined relationship with the user operation interaction and the platform response interaction. If according to the rule, the numerical value 0.6 has a certain quantitative relationship with the user operation interaction as multiplied by 1.5, then the corresponding numerical value in the second user operation interaction vector may be 0.9. At the same time, for the numerical value 0.4, if the association rule with the platform response interaction is multiplied by 2, then the corresponding numerical value in the second platform response interaction vector may be 0.8. This is just an illustrative algorithm example. The actual algorithm is much more complicated and based on a deeper behavioral analysis logic.

[0098] 2. Obtain the second learning resource relationship vector from the target initial user behavior attribute vector input in the corresponding iteration cycle

[0099] The second learning resource relationship vector is obtained according to the target initial user behavior attribute vector input in the corresponding iteration cycle. For example, the target initial user behavior attribute vector is [0.5, 0.3, 0.2], where 0.5 may represent the operation frequency in the initial learning behavior, 0.3 may represent the complexity of the operation, and 0.2 may represent the degree of deviation between the operation result and the expectation, etc. The second learning resource relationship vector is also obtained through an adaptive algorithm. For example, there is an algorithm based on the association between learning resources and user behavior. For the value 0.5, if it has a certain predefined relationship with the resource acquisition frequency in the learning resource relationship (for example, multiplied by 3), then the value related to the resource acquisition frequency in the second learning resource relationship vector may be 1.5. In a similar manner, according to 0.3 and 0.2 and their association rules with other learning resource relationship elements, a complete second learning resource relationship vector is obtained.

[0100] 2. Sampling user group behavior elements based on multiple vectors

[0101] 1. Based on the feature focus strategy, the user group behavior elements are sampled to obtain the second user group behavior element vector

[0102] The user group behavior elements are sampled for the second user operation interaction vector, the second platform response interaction vector, and the second learning resource relationship vector, and the operation is performed based on the feature focus strategy. The feature focus strategy aims to identify and select elements that have a key influence on the user group behavior. For example, in an online learning environment, the value in the second user operation interaction vector is related to the user's operation behavior in course learning (such as video viewing operations, homework submission operations, etc.), the value in the second platform response interaction vector is related to the platform's feedback on user operations (such as scoring, prompt information, etc.), and the value in the second learning resource relationship vector is related to the user's use and association of learning resources (such as downloading learning materials, using learning tools, etc.).

[0103] The feature focus strategy determines which elements are key user group behavior elements based on pre-set standards. For example, the number of pauses in video viewing, platform ratings, and the frequency of downloading learning materials may be determined as key elements. These key elements are then combined into a second user group behavior element vector through an adaptive algorithm. For example, after calculation, the second user group behavior element vector is [0.4, 0.3, 0.2], where 0.4 may be the weight value of the number of pauses after algorithm conversion, 0.3 is the weight value after platform rating conversion, and 0.2 is the weight value after learning material download frequency conversion.

[0104] 2. Perform feature integration on the second user group behavior element vector and the second learning resource relationship vector to obtain the initial user behavior attribute vector outputted in the corresponding iteration cycle

[0105] The purpose of integrating the second user group behavior element vector and the second learning resource relationship vector is to synthesize the information of the two to obtain the initial user behavior attribute vector outputted in the corresponding iteration cycle. For example, the second user group behavior element vector is [0.4, 0.3, 0.2], and the second learning resource relationship vector is [1.5, 0.8, 0.3]. Through a feature integration algorithm (such as a weighted summation algorithm), for example, multiply each value in the second user group behavior element vector by 2, multiply each value in the second learning resource relationship vector by 1.5, and then add the values ​​in the corresponding positions. That is: [(0.4*2+1.5*1.5), (0.3*2+0.8*1.5), (0.2*2+0.3*1.5)]=[3.05, 2.1, 0.85]. This vector is the initial user behavior attribute vector outputted in the corresponding iteration cycle, which will serve as the basis for the next round of iteration or subsequent processing (such as resource push suggestions, etc.).

[0106] In summary, firstly, vectors related to user operation interaction, platform response interaction and learning resource relationship are obtained from the historical user behavior attribute vector and the target initial user behavior attribute vector outputted in the corresponding iteration cycle, which helps to fully integrate various types of information related to user learning behavior and make fuller use of data. Secondly, sampling of user group behavior elements based on feature focusing strategy can highlight key elements, reduce the interference of unnecessary information, and improve the efficiency and accuracy of analysis. Finally, the second user group behavior element vector and the second learning resource relationship vector are feature integrated to obtain the initial user behavior attribute vector outputted in the corresponding iteration cycle. This method can integrate key information from different sources, so that the obtained initial user behavior attribute vector can more accurately reflect the user's behavior characteristics, thereby providing a more reliable data basis for the online learning cloud platform in subsequent resource push, user behavior analysis and other operations.

[0107] In some examples, the at least one round of iteration includes several iteration cycles; the linkage behavior preference embedding of at least one round of iteration is performed based on the initial user behavior attribute vector and the historical user behavior attribute vector to obtain the linkage behavior preference embedding vector output by the linkage behavior preference embedding of the at least one round of iteration, including: under the linkage behavior preference embedding task of the first iteration cycle among the several iteration cycles, key behavior element sampling is performed on the initial user behavior attribute vector and the historical user behavior attribute vector input in the first iteration cycle to obtain the historical user behavior attribute vector output in the first iteration cycle, and key behavior element sampling is performed on the historical user behavior attribute vector and the initial user behavior attribute vector output in the first iteration cycle to obtain the initial user behavior attribute vector output in the first iteration cycle; in each round of iteration starting from the second iteration cycle among the several iteration cycles, the linkage behavior preference embedding of the corresponding iteration cycle is performed based on the historical user behavior attribute vector and the initial user behavior attribute vector output in the previous iteration cycle to obtain the historical user behavior attribute vector and the initial user behavior attribute vector output in the corresponding iteration cycle, until the linkage behavior preference embedding of the last iteration cycle among the several iteration cycles outputs the linkage behavior preference embedding vector of the several iteration cycles.

[0108] In the online learning cloud platform, the specific implementation method of performing at least one round of iterative linkage behavior preference embedding based on the initial user behavior attribute vector and the historical user behavior attribute vector to obtain the linkage behavior preference embedding vector is as follows.

[0109] 1.1 Operations of the First Iteration Cycle

[0110] 1. Sampling key behavior elements for the initial user behavior attribute vector and the historical user behavior attribute vector to obtain the historical user behavior attribute vector output in the first iteration cycle

[0111] In the linkage behavior preference embedding task of the first iteration cycle among several iteration cycles, the key behavior elements must first be sampled for the initial user behavior attribute vector and the historical user behavior attribute vector input in the first iteration cycle. For example, the initial user behavior attribute vector is [0.3, 0.5, 0.2], and the historical user behavior attribute vector is [0.4, 0.6, 0.1]. Here, 0.3 in the initial user behavior attribute vector may represent the frequency of a certain behavior of the user in the current learning, 0.5 represents the complexity of the behavior, and 0.2 represents the degree of deviation from the expected result; 0.4 in the historical user behavior attribute vector may represent the frequency of resource acquisition in past learning, 0.6 represents the degree of mastery of past learning content, and 0.1 represents the error rate in past learning.

[0112] Key behavior elements are sampled through an adaptive algorithm. This algorithm may be based on predefined rules for the importance of behavior elements. For example, behavior frequency, resource acquisition frequency, and the degree of mastery of learning content are determined as key behavior elements. For 0.3 (behavior frequency) in the initial user behavior attribute vector, it may be multiplied by 1.5 to obtain 0.45 according to the rule; for 0.4 (resource acquisition frequency) in the historical user behavior attribute vector, it is multiplied by 2 to obtain 0.8, and 0.6 (degree of mastery) is multiplied by 1.5 to obtain 0.9. These processed values ​​are then combined (this embodiment is only an illustrative combination method, and the actual algorithm is more complex) to obtain the historical user behavior attribute vector output by the first iteration cycle, for example, [0.8, 0.9, 0.1].

[0113] 2. Sampling key behavior elements for the historical user behavior attribute vector and the initial user behavior attribute vector output in the first iteration cycle to obtain the initial user behavior attribute vector output in the first iteration cycle

[0114] Next, key behavior element sampling is performed for the historical user behavior attribute vector [0.8, 0.9, 0.1] and the initial user behavior attribute vector [0.3, 0.5, 0.2] output in the first iteration cycle. Also based on the adaptive algorithm, for example, this algorithm redefines the key behavior elements as the adjusted behavior complexity, mastery level, and degree of deviation from the expected results. For the 0.9 (degree of mastery) in the historical user behavior attribute vector output in the first iteration cycle, multiply it by 1.2 according to the rule to obtain 1.08, for the 0.5 (behavior complexity) in the initial user behavior attribute vector, multiply it by 1.3 to obtain 0.65, and multiply 0.2 (degree of deviation) by 1.1 to obtain 0.22. Then combine these values ​​(also a schematic combination) to obtain the initial user behavior attribute vector output in the first iteration cycle, for example, [0.65, 1.08, 0.22].

[0115] (II) Operations starting from the second iteration cycle

[0116] In each round of iteration starting from the second iteration cycle among several iteration cycles, the linkage behavior preference embedding of the corresponding iteration cycle is performed based on the historical user behavior attribute vector and the initial user behavior attribute vector output by the previous iteration cycle to obtain the historical user behavior attribute vector and the initial user behavior attribute vector output by the corresponding iteration cycle.

[0117] For example, for the second iteration cycle, the historical user behavior attribute vector [0.8, 0.9, 0.1] output in the previous iteration cycle (i.e., the first iteration cycle) and the initial user behavior attribute vector [0.65, 1.08, 0.22] are used as the basis. Operations are performed through a specific linkage behavior preference embedding algorithm. This algorithm may involve re-evaluation and adjustment of the values ​​in the vector to reflect the dynamic changes in user behavior preferences.

[0118] For example, this algorithm adjusts the values ​​according to the long-term trend and short-term fluctuations of user behavior. For 0.8 in the historical user behavior attribute vector, according to the analysis of long-term trends and short-term fluctuations (through complex calculation logic), it may be multiplied by 1.1 to get 0.88; for 0.9, it is multiplied by 1.05 to get 0.945; 0.1 is multiplied by 1.2 to get 0.12. For 0.65 in the initial user behavior attribute vector, according to the rule, it is multiplied by 1.08 to get 0.702; 1.08 is multiplied by 1.03 to get 1.1124; 0.22 is multiplied by 1.15 to get 0.253. These adjusted values ​​are then combined into the historical user behavior attribute vector [0.88, 0.945, 0.12] and the initial user behavior attribute vector [0.702, 1.1124, 0.253] output by the second iteration cycle.

[0119] This process continues in subsequent iterations, and each iteration performs similar operations based on the vector output in the previous iteration, continuously adjusting and optimizing the values ​​in the vector until the last iteration in several iterations. In the last iteration, after a series of adjustments and optimizations, the final output vector is the linkage behavior preference embedding vector of several iterations. This linkage behavior preference embedding vector combines the initial user behavior attribute vector and the historical user behavior attribute vector from the initial input, and after multiple rounds of iterative adjustments, it can more accurately reflect the user's behavior preferences.

[0120] It can be seen that through multiple rounds of iterations, the final linkage behavior preference embedding vector is gradually adjusted from the initial initial user behavior attribute vector and the historical user behavior attribute vector. This gradual optimization process can dig deeper into the user's behavior preferences and avoid the one-sidedness that may be caused by a one-time analysis. In each round of iteration, sampling and adjustment are performed for different key behavior elements, making the adjustment of the vector more targeted and able to accurately capture the changes in user preferences at different learning stages and in different behaviors. The complete iteration process from the first iteration cycle to the last iteration cycle ensures that the analysis of user behavior preferences is a dynamic and continuous process, so that the final linkage behavior preference embedding vector can comprehensively and accurately reflect the user's behavior preferences throughout the learning process, providing a more reliable data basis for subsequent operations such as pushing learning resources.

[0121] Under an alternative design idea, step 130 described in which the learning resource interest description for the learning behavior data to be processed is obtained based on the linkage behavior preference embedding vector, including: sampling user group behavior elements for the initial user behavior attribute vector and the historical user behavior attribute vector included in the linkage behavior preference embedding vector to obtain the learning resource interest description for the learning behavior data to be processed.

[0122] Furthermore, step 130 as described above obtains resource push suggestion features for the learning behavior data to be processed based on the correlation features between the learning resource interest description and the initial user behavior attributes, including: determining the correlation features between the learning resource interest description and the initial user behavior attributes; identifying potential interests based on the correlation features to obtain potential interest weight features; and determining resource push suggestion features for the learning behavior data to be processed based on the potential interest weight features.

[0123] In this embodiment, in order to obtain the learning resource interest description for the learning behavior data to be processed based on the linkage behavior preference embedding vector, a method of sampling user group behavior elements based on the initial user behavior attribute vector and the historical user behavior attribute vector included in the linkage behavior preference embedding vector is adopted.

[0124] For example, the initial user behavior attribute vector in the linkage behavior preference embedding vector is [0.5, 0.3, 0.2], and the historical user behavior attribute vector is [0.4, 0.6, 0.1]. The numerical values ​​here have specific meanings. For example, in the initial user behavior attribute vector, 0.5 may indicate the user's attention to a certain knowledge module in the current learning behavior, 0.3 may indicate the user's operational proficiency in the current learning, and 0.2 may indicate the user's interaction frequency with the current learning resources. In the historical user behavior attribute vector, 0.4 may indicate the user's contact frequency with similar knowledge modules in past learning, 0.6 may indicate the degree of mastery in past learning, and 0.1 may indicate the depth of interaction with related learning resources in past learning.

[0125] The operation is performed through a specific user group behavior element sampling algorithm. This algorithm is based on the analysis logic of user group behavior and aims to find out the behavior elements that have an important impact on the description of learning resource interest. For example, the algorithm may determine that the degree of attention, degree of mastery and frequency of interaction are key user group behavior elements. For 0.5 (degree of attention) in the initial user behavior attribute vector, according to the algorithm rules, it may be multiplied by 1.5 to obtain 0.75; for 0.6 (degree of mastery) in the historical user behavior attribute vector, it may be multiplied by 1.2 to obtain 0.72; 0.2 (interaction frequency) is multiplied by 2 to obtain 0.4. Then these processed values ​​are combined (this embodiment is only a schematic combination method, and the actual algorithm is more complicated), for example, combined into [0.75, 0.72, 0.4]. According to this combination vector and the predefined mapping rules, the description of learning resource interest for the learning behavior data to be processed can be obtained. For example, if the value is above 0.5, it means that the interest in the element is high, then the description of the learning resource interest may be "the user has a high interest in the relevant knowledge module, has a good foundation for knowledge mastery, and has a high tendency to interact with the learning resource."

[0126] After obtaining the learning resource interest description, the resource push suggestion feature for the learning behavior data to be processed should be obtained based on the correlation feature between the learning resource interest description and the initial user behavior attribute. First, the correlation feature between the two should be determined.

[0127] Continuing with the above example, the initial user behavior attribute vector is [0.5, 0.3, 0.2], and the learning resource interest description is "the user has a high interest in the relevant knowledge module, has a good foundation for knowledge mastery, and has a high tendency to interact with the learning resource". The process of determining the correlation feature involves analyzing the logical connection between the two. For example, 0.5 (the degree of attention to a certain knowledge module) in the initial user behavior attribute vector has a direct positive correlation with "having a high interest in the relevant knowledge module" in the learning resource interest description; 0.3 (operation proficiency) and "having a good foundation for knowledge mastery" may have an indirect correlation because operation proficiency may be a reflection of the degree of knowledge mastery; 0.2 (interaction frequency with the current learning resource) and "having a high tendency to interact with the learning resource" have a direct correlation. Through this correlation analysis of each value and each part of the learning resource interest description, the entire correlation feature is determined.

[0128] After determining the correlation features, potential interest identification is performed for these correlation features. Potential interest identification is to further explore the hidden and incompletely reflected interest tendencies of users in terms of learning resource interests. For example, according to the correlation features determined earlier, it is found that although 0.3 (operation proficiency) in the initial user behavior attribute vector has an indirect correlation with the degree of knowledge mastery, through a deeper analysis (which may involve a comprehensive analysis of multiple data such as user historical operation records and learning progress), it is found that if the operation proficiency can be improved, the user may be interested in more advanced knowledge modules. After this potential interest tendency is identified, it will be assigned a potential interest weight feature. For example, through a specific algorithm (this algorithm is based on the analysis of a large amount of user data and the training of a machine learning model), the potential interest weight feature assigned to this potential interest tendency is 0.2. This weight feature indicates the importance of this potential interest in the overall resource push recommendation.

[0129] Finally, the resource push suggestion feature for the learning behavior data to be processed is determined based on the potential interest weight feature. For example, based on the potential interest weight feature 0.2 obtained earlier, and the weights corresponding to other directly related relationships (for example, the weight of the degree of attention corresponding to 0.5 is 0.3, and the weight of the interaction frequency corresponding to 0.2 is 0.5), the resource push suggestion feature can be determined. If the sum of the weights is 1, then in the resource push, the push ratio for learning resources related to the degree of attention (such as in-depth explanation materials of related knowledge modules) may be 0.3, the push ratio for learning resources related to the interaction frequency (such as interactive learning tools) may be 0.5, and the push ratio for learning resources related to potential interests (such as guiding materials for improving operational proficiency and then learning advanced knowledge modules) may be 0.2. In this way, the resource push suggestion feature for the learning behavior data to be processed is determined.

[0130] This embodiment first obtains the learning resource interest description by sampling the user group behavior elements of the initial user behavior attribute vector and the historical user behavior attribute vector in the linkage behavior preference embedding vector, which can comprehensively integrate the user's current and historical learning behaviors, so that the learning resource interest description more accurately reflects the user's interest tendencies. Secondly, the process of determining the correlation characteristics between the learning resource interest description and the initial user behavior attributes and identifying potential interests to obtain potential interest weight characteristics deeply explores the user's learning needs, not only considering the surface correlation, but also discovering potential interest tendencies, thereby improving the accuracy of resource push. Finally, the resource push suggestion feature is determined based on the potential interest weight feature, so that the resource push can be reasonably allocated according to the importance of different user interest tendencies, so as to better meet the user's learning needs and improve the user's learning experience.

[0131] In some further exemplary embodiments, the method further comprises: performing at least one round of iterative behavior preference path identification on the correlation features to obtain a learning behavior preference path.

[0132] Based on this, determining the resource push recommendation features for the learning behavior data to be processed based on the potential interest weight features includes: combining the potential interest weight features with the learning behavior preference path to obtain a personalized learning behavior portrait; determining the resource push recommendation features for the learning behavior data to be processed based on the personalized learning behavior portrait.

[0133] In this embodiment, at least one round of iterative behavior preference path identification is first performed on the correlation features to obtain the learning behavior preference path. This process is based on the correlation features between the previously determined learning resource interest description and the initial user behavior attributes. The correlation features reflect the association between different learning behavior attributes and resource interests of the user, and the behavior preference path identification aims to mine the evolution path of the user's behavior preference in the learning process from these associations.

[0134] For example, the determined correlation features include the following relationship: between the initial user behavior attribute vector [0.5, 0.3, 0.2] (in this embodiment, 0.5 represents the degree of attention to a certain knowledge module, 0.3 represents the operation proficiency, and 0.2 represents the frequency of interaction with learning resources) and the description of learning resource interest "the user has a high interest in the relevant knowledge module, has a good foundation for knowledge mastery, and has a high tendency to interact with learning resources", the degree of attention is positively correlated with the interest in the knowledge module, the operation proficiency is indirectly correlated with the knowledge mastery, and the interaction frequency is positively correlated with the interaction tendency. These correlation features form the basis for identifying the behavior preference path.

[0135] In each iteration, a specific behavior preference path identification algorithm is used. This algorithm conducts in-depth analysis of each relationship in the correlation feature and constructs a behavior preference path based on historical trends in user behavior data, learning progress, and the mutual influence between different behavior attributes. For example, in the first iteration, for the positive correlation between attention level and interest in knowledge modules, the algorithm may construct a path from attention level to in-depth learning of knowledge modules based on the data of users in the past multiple learning processes (e.g., as attention level increases, the frequency of in-depth learning of related knowledge modules in subsequent learning also increases).

[0136] At the same time, for the indirect correlation between operational proficiency and knowledge mastery, the algorithm may analyze the impact path of changes in operational proficiency on knowledge mastery at different learning stages. For example, if the knowledge mastery shows a nonlinear growth during the process of improving operational proficiency from low to high (for example, it grows slowly at first, and then grows rapidly after a certain operational proficiency threshold), the algorithm will identify this special path relationship.

[0137] For the positive correlation between interaction frequency and interaction tendency, the algorithm may consider how changes in interaction frequency affect the user's tendency to choose different types of learning resources, thereby constructing a path from interaction frequency to different resource types. After multiple rounds of iterations (for example, three rounds of iterations), the path analysis results of these different relationships are combined to obtain a complete learning behavior preference path. For example, this learning behavior preference path may be: starting with a focus on knowledge modules, gradually deepening the knowledge as operational proficiency improves, and in this process, the increase in interaction frequency leads to an increased tendency to choose specific types of learning resources (such as practical case resources).

[0138] After obtaining the learning behavior preference path, the potential interest weight feature and the learning behavior preference path should be combined to obtain a personalized learning behavior profile. The potential interest weight feature is obtained by identifying potential interests based on the correlation feature, which indicates the importance of the user's potential interest in the overall resource push recommendation.

[0139] For example, the potential interest weight feature obtained previously is 0.2, and the corresponding potential interest is to improve operational proficiency and then learn advanced knowledge modules. This potential interest weight feature is fused with the learning behavior preference path. In the learning behavior preference path "starting from paying attention to knowledge modules, gradually mastering knowledge in depth as operational proficiency improves, and in this process, the increase in interaction frequency promotes the increase in the tendency to choose specific types of learning resources (such as practical case resources)", operational proficiency is related to potential interest. The fusion process may be achieved through a weighted fusion algorithm.

[0140] For example, each node in the learning behavior preference path (such as attention level, operation proficiency improvement, knowledge mastery, interaction frequency increase, resource selection tendency) has an initial weight (these weights may be obtained based on statistical analysis of a large amount of user learning behavior data), such as the initial weight of attention level is 0.3, the initial weight of operation proficiency improvement is 0.25, the initial weight of knowledge mastery is 0.2, the initial weight of interaction frequency increase is 0.15, and the initial weight of resource selection tendency is 0.1. When integrating the potential interest weight feature, since the potential interest is related to the operation proficiency improvement, the potential interest weight feature 0.2 may be allocated to the weight of the operation proficiency improvement according to a certain ratio (such as 0.5), so that the weight of the operation proficiency improvement becomes 0.25+0.2*0.5=0.35.

[0141] Through this fusion mechanism, a personalized learning behavior portrait is constructed. This portrait is not only a simple description of the user's learning behavior preference path, but also incorporates the potential interest weight characteristics, which can more comprehensively and accurately reflect the user's personalized learning behavior characteristics. For example, a personalized learning behavior portrait may be described as follows: the user pays attention to the knowledge module with a weight of 0.3, the importance weight of the improvement of operation proficiency in the overall learning behavior is 0.35, and this increase in weight is due to the existence of potential interest (improving operation proficiency and then learning advanced knowledge modules), the weight of knowledge mastery is 0.2, the weight of increased interaction frequency is 0.15, and when the interaction frequency increases, there is a tendency to choose a specific type of learning resource (such as practical case resources) with a weight of 0.1.

[0142] Determine the recommended features for resource push for the learning behavior data to be processed based on the personalized learning behavior portrait. First, determine the resource type based on the weight distribution in the portrait. In the above personalized learning behavior portrait, the weight of improving operation proficiency is relatively high, which is 0.35. This means that in resource push, we should focus on pushing learning resources that help improve operation proficiency, such as operation practice tutorials for a certain knowledge module, skill improvement guides, etc.

[0143] At the same time, the weight of knowledge mastery is 0.2, and a certain proportion of learning resources that help consolidate knowledge need to be pushed, such as knowledge summary manuals, review materials, etc. The weight of interaction frequency increase is 0.15, and some learning resources that encourage users to increase interaction can be pushed, such as online learning community communication activity recommendations, interactive learning games, etc. The weight of resource selection tendency is 0.1, and specific types of learning resources (such as practical case resources) also need to be pushed according to this weight ratio.

[0144] In addition to determining the resource type, the resource push ratio must also be determined based on the weight ratio. For example, if a total of 10 learning resources are to be pushed, calculate according to the above weight ratio. The number of resource pushes related to improved operational proficiency may be 10*0.35=3.5 (this embodiment can be rounded to 3 or 4 according to actual conditions), the number of resource pushes related to knowledge mastery is 10*0.2=2, the number of resource pushes related to increased interaction frequency is 10*0.15=1.5 (rounded to 1 or 2), and the number of pushes for specific types of learning resources (such as practical case resources) is 10*0.1=1. In this way, the resource push recommendation features for the learning behavior data to be processed are determined, so that resource push can be accurately pushed according to the user's personalized learning behavior portrait.

[0145] With this design, we first obtain the learning behavior preference path by iteratively identifying the behavior preference path based on the correlation features, which can deeply explore the evolution of the user's behavior preference in the learning process, rather than just focusing on the surface correlation, so as to have a deeper understanding of the user's learning behavior. Secondly, we combine the potential interest weight features with the learning behavior preference path to build a personalized learning behavior portrait. This fusion method integrates the potential interest into the overall learning behavior analysis, so that the portrait reflects the user's personalized characteristics more comprehensively and accurately. Finally, based on the personalized learning behavior portrait, we determine the resource push suggestion features, which can accurately push resources according to the user's personalized learning needs, improve the efficiency of users in obtaining suitable learning resources, and thus enhance the user's learning experience.

[0146] In some optional embodiments, the mining of the initial user behavior attributes of the learning behavior data to be processed to obtain the initial user behavior attribute vector of the learning behavior data to be processed includes: extracting the initial user behavior attributes of the learning behavior data to be processed, and performing at least one round of iterative one-hot vector extraction on the initial user behavior attributes to obtain the initial user behavior attribute vector of the learning behavior data to be processed;

[0147] Furthermore, the determining of historical user behavior attributes associated with the learning behavior data to be processed, and mining of the historical user behavior attributes to obtain a historical user behavior attribute vector, includes: determining the historical user behavior attributes associated with the learning behavior data to be processed from a user behavior attribute set, the user behavior attribute set including at least one shared historical user behavior attribute identified based on target learning behavior data; performing at least one round of iterative one-hot vector extraction on the historical user behavior attributes to obtain a historical user behavior attribute vector.

[0148] Among them, determining the historical user behavior attributes associated with the learning behavior data to be processed from the user behavior attribute set includes: determining the user behavior attribute set; performing behavior attribute association on the initial user behavior attributes and each shared historical user behavior attribute in the user behavior attribute set to obtain a behavior attribute association result; and determining the shared historical user behavior attribute corresponding to the behavior attribute association result indicating that the behavior attribute has learning resource relevance as the historical user behavior attribute associated with the learning behavior data to be processed.

[0149] When mining the initial user behavior attributes of the learning behavior data to be processed to obtain the initial user behavior attribute vector, the initial user behavior attributes of the learning behavior data to be processed must first be extracted. For example, for a course learning data (learning behavior data to be processed) on an online learning platform, it may contain multiple types of user behavior information. This information may include the user's learning time in the course (in minutes), the correct rate of answering questions, the number of times participating in course discussions, etc. For example, in a specific learning process, the user's learning time is 30 minutes, the correct rate of answering questions is 70%, and the number of times participating in course discussions is 2. These are the initial user behavior attributes extracted from the learning behavior data to be processed.

[0150] After the initial user behavior attributes are extracted, at least one round of iterative one-hot vector extraction is performed on these initial user behavior attributes to obtain the initial user behavior attribute vector. One-hot vector extraction is a method of converting discrete attribute values ​​into vector form to facilitate subsequent data analysis and processing.

[0151] For example, for the attribute of 30 minutes of study time, several time intervals are set, such as 0-15 minutes, 15-45 minutes, 45-60 minutes, etc. If 30 minutes falls within the interval of 15-45 minutes, then in the first round of iterative one-hot vector extraction, a vector will be generated for this attribute, such as [0, 1, 0] (indicating that it belongs to the second interval). For the correct answer rate of 70%, for example, several correct rate intervals are set, such as 0-30%, 30-70%, 70-100%, then the corresponding vector may be [0, 1, 0]. For the number of times of participating in course discussions 2 times, for example, intervals such as 0-1 times, 1-3 times, 3-5 times, etc. are set, and the corresponding vector is [0, 1, 0].

[0152] Then, in subsequent iterations (for example, two rounds of iterations), these vectors may be further adjusted according to other rules. For example, in the second round of iterations, the vector of learning time may be adjusted according to the average learning time of the entire course. If the average learning time is 40 minutes, then 30 minutes is relatively short, and the values ​​in the vector may be fine-tuned to obtain [0, 0.8, 0.2] (indicating that it is closer to the lower limit in the range of 15-45 minutes). Similarly, similar adjustments are made to the vectors of the correct answer rate and the number of times participated in course discussions. Finally, these iteratively adjusted vectors are combined to obtain the initial user behavior attribute vector of the learning behavior data to be processed, such as [0, 0.8, 0.2, 0, 1, 0, 0, 1, 0].

[0153] Exemplarily, in terms of mining historical user behavior attribute vectors, the user behavior attribute set must first be determined. This user behavior attribute set contains at least one shared historical user behavior attribute identified based on the target learning behavior data. For example, in an online learning cloud platform, the user behavior attribute set may be various behavior attribute information collected from all users in past course learning (target learning behavior data). This information may include study time, answering of questions, interaction with teachers or classmates, etc. when studying different courses in the past. For example, the user behavior attribute set contains relevant behavior attributes of multiple users when studying mathematics, Chinese, English and other courses.

[0154] Then the initial user behavior attributes are associated with each shared historical user behavior attribute in the user behavior attribute set to obtain the behavior attribute association result. Continuing with the above example, the initial user behavior attributes of 30 minutes of learning time, 70% correct answer rate, and 2 times of participating in course discussions will be associated with various historical learning behavior attributes in the user behavior attribute set. For example, for a historical mathematics course learning record, the learning time is 40 minutes, the correct answer rate is 80%, and the number of times participating in course discussions is 3 times. The association results between them are determined by a certain algorithm (such as calculating the similarity of numerical values ​​or logical relationships). The degree of association may be comprehensively judged based on factors such as the difference in learning time, the closeness of the correct answer rate, and the difference in the number of discussions.

[0155] Then, the shared historical user behavior attribute corresponding to the behavior attribute association result indicating that the behavior attribute has learning resource relevance is determined as the historical user behavior attribute associated with the learning behavior data to be processed. For example, if in the behavior attribute association result, it is found that a certain historical English course learning record has certain similarities with the current learning behavior data to be processed in terms of learning time, correct answer rate, and number of discussions, and this historical learning record has common learning resource requirements with the current course being studied (for example, both require reading a large amount of text materials), then the relevant behavior attribute in this historical English course learning record is determined as the historical user behavior attribute associated with the learning behavior data to be processed.

[0156] After determining the historical user behavior attributes associated with the learning behavior data to be processed, perform at least one round of iterative one-hot vector extraction on these historical user behavior attributes to obtain the historical user behavior attribute vector. For example, the historical user behavior attributes associated with the learning behavior data to be processed are that the learning time for the past English course was 40 minutes, the correct answer rate was 80%, and the number of times the course discussion was participated in was 3 times.

[0157] For a study time of 40 minutes, according to the previously set time interval range (0-15 minutes, 15-45 minutes, 45-60 minutes), in the first round of iterative one-hot vector extraction, the corresponding vector is [0, 1, 0]. For an answer accuracy of 80%, according to the set accuracy interval (0-30%, 30-70%, 70-100%), the corresponding vector is [0, 0, 1]. For a course discussion participation of 3 times, according to the set interval (0-1 times, 1-3 times, 3-5 times), the corresponding vector is [0, 1, 0].

[0158] In subsequent iterations (for example, two rounds of iterations), these vectors are also adjusted according to other rules. For example, in the second round of iterations, adjustments may be made based on the degree of correlation with the current learning behavior data to be processed. If it is found that the history English course learning is highly correlated with the current learning, the values ​​in the vector may be strengthened. For example, the vector of learning time is adjusted to [0, 1.2, -0.2] (indicating that it is closer to the upper limit in the range of 15-45 minutes and has been strengthened to a certain extent), the vector of the correct answer rate is adjusted to [0, 0, 1.2], and the vector of the number of times participated in course discussions is adjusted to [0, 1.2, -0.2]. Finally, these iteratively adjusted vectors are combined to obtain the historical user behavior attribute vector, such as [0, 1.2, -0.2, 0, 0, 1.2, 0, 1.2, -0.2].

[0159] The beneficial effects of this embodiment are mainly reflected in the following aspects: First, by mining the initial user behavior attribute vector and the historical user behavior attribute vector through the unique hot vector extraction method, different types of user behavior attributes can be represented in a unified vector form, which is convenient for subsequent calculation and analysis in the entire data processing online learning cloud platform. Secondly, when determining the historical user behavior attributes associated with the learning behavior data to be processed, the historical learning behavior information related to the current learning behavior can be accurately found through the behavior attribute association method, avoiding the blind use of all historical data and improving the pertinence of data utilization. Finally, performing at least one round of iterative operations can continuously optimize the vector representation, so that the vector can more accurately reflect the user's behavior characteristics. Whether it is the initial user behavior attribute vector or the historical user behavior attribute vector, it can provide a more reliable data basis for subsequent user behavior analysis, resource push and other operations.

[0160] In some independent embodiments, after obtaining the resource push recommendation characteristics for the learning behavior data to be processed based on the correlation characteristics between the learning resource interest description and the initial user behavior attributes, the method also includes: using the resource push recommendation characteristics to generate a learning resource push strategy for the user learning terminal corresponding to the learning behavior data to be processed; and pushing online learning resources to the user learning terminal based on the learning resource push strategy.

[0161] In some other independent embodiments, the resource push suggestion feature is used to generate a learning resource push strategy for the user learning terminal corresponding to the learning behavior data to be processed, including: obtaining the push mode representation information corresponding to the resource push suggestion feature and the page output state description corresponding to the user learning terminal, performing quantitative mapping on the page output state description, and obtaining the page output state quantitative knowledge corresponding to the page output state description; the function used in the quantitative mapping is a soft threshold activation function; using a push annotation feature to push-annotate the page output state quantitative knowledge and the push mode representation information respectively, and associating the results obtained from the push annotation with policy nodes to obtain policy node association results; generating a report on the policy node association results to obtain a push execution report; the function used in the report generation is a soft threshold activation function; the push execution report indicates the push rules for the user learning terminal.

[0162] In detail, the push mode representation information corresponding to the resource push suggestion feature and the page output state description corresponding to the user learning terminal are obtained, including: obtaining the resource push suggestion feature and the user learning terminal, and converting the user learning terminal into a page output state description; the page output state description is used to represent the page push feedback feature of the user learning terminal; if the resource push suggestion feature is a first type of suggestion feature, a first push task keyword is generated according to the resource push suggestion feature, and the first push task keyword is determined as the push mode representation information; if the resource push suggestion feature is a second type of suggestion feature, the resource push suggestion feature is feature pooled to obtain a push suggestion pooling vector, the first push task keyword is generated according to the resource push suggestion feature, and the first push task keyword is associated with the push suggestion pooling vector by a strategy node to obtain the push mode representation information; the push suggestion pooling vector is used to represent the derived learning resource demand of the resource push suggestion feature; the push mode representation information is used to represent the page push feedback feature of the resource push suggestion feature.

[0163] In detail, the first push task keyword and the push suggestion pooling vector are associated with a strategy node to obtain the push mode characterization information, including: de-dimensionalizing the push suggestion pooling vector to obtain a target pooling feature; the target pooling feature is used to represent the importance of the resource push suggestion feature in each online learning link; and the target pooling feature is optimized using the first push task keyword to obtain the push mode characterization information.

[0164] In order to facilitate those skilled in the art to clearly and completely implement the above four progressive embodiments, an exemplary implementation is provided below for explanation.

[0165] 1. Generating learning resource push strategies based on resource push suggestion features

[0166] 1. Obtain push mode representation information and page output status description and quantify page output status description

[0167] When using the resource push suggestion feature to generate a learning resource push strategy for the user learning terminal corresponding to the learning behavior data to be processed, it is necessary to first obtain the push mode representation information corresponding to the resource push suggestion feature and the page output state description corresponding to the user learning terminal. For example, the resource push suggestion feature may be a vector containing multiple numerical values, such as [0.3, 0.5, 0.2] (each numerical value in this embodiment represents the weight or importance of different learning resource types in the push, etc.). The page output state description of the user learning terminal may include information such as the resolution of the terminal screen, the type of content currently displayed (such as the ratio of text, video, and images), and the interactive elements of the page (such as the number of buttons, menu structure, etc.).

[0168] After obtaining the resource push suggestion feature and the user learning terminal, the user learning terminal is converted into a page output state description, which is used to represent the page push feedback feature of the user learning terminal. If the resource push suggestion feature is a first-class suggestion feature (for example, the resource push suggestion feature mainly focuses on the push of a single type of learning resource, such as push suggestions only for video learning resources), then the first push task keyword is generated according to the resource push suggestion feature. For example, for the push suggestion feature of video learning resources, the first push task keyword may be "video resource optimization push", and this first push task keyword is determined as the push mode representation information.

[0169] If the resource push suggestion feature is a second type of suggestion feature (for example, the resource push suggestion feature involves comprehensive push of multiple types of learning resources), the resource push suggestion feature is feature pooled to obtain a push suggestion pooling vector. For example, for the resource push suggestion feature [0.3, 0.5, 0.2], a push suggestion pooling vector is obtained by a feature pooling algorithm (such as a weighted summation algorithm), such as [1] (this embodiment is only an exemplary description). At the same time, a first push task keyword is generated based on the resource push suggestion feature, such as "comprehensive learning resource push". Then, the first push task keyword is associated with the push suggestion pooling vector by a policy node to obtain push mode characterization information.

[0170] The page output state description is quantitatively mapped to obtain the page output state quantitative knowledge corresponding to the page output state description, where the function used for the quantitative mapping is the soft threshold activation function. For example, for the screen resolution in the page output state description, if it is 1920x1080, it can be converted into a quantitative value, such as 0.8, by the soft threshold activation function (the value of this embodiment is obtained according to the specific definition and rules of the soft threshold activation function, and different resolutions will correspond to different quantitative values). For the type of content displayed on the page, if the text accounts for 60%, the video accounts for 30%, and the image accounts for 10%, they are also converted into quantitative values, such as 0.6, 0.3, and 0.1, respectively, by the soft threshold activation function, and the quantitative knowledge of the page output state is obtained by combining [0.8, 0.6, 0.3, 0.1].

[0171] 2. Push annotation, policy node association and report generation

[0172] The push annotation feature is used to push and annotate the page output state quantitative knowledge and push mode representation information respectively. For example, the push annotation feature can be some pre-defined rules or labels. For the page output state quantitative knowledge [0.8, 0.6, 0.3, 0.1], if the push annotation feature stipulates that the screen resolution quantization value greater than 0.5 is a high-resolution mark, then this value is marked as "high resolution"; for the quantitative value of the content type, it is marked as "high text ratio" according to different rules. For the push mode representation information, if it is "video resource optimization push", it is marked as "video resource push mode" according to the push annotation feature.

[0173] The results obtained by push annotation are associated with policy nodes to obtain policy node association results. For example, the annotation results of quantitative knowledge of page output states such as "high resolution" and "high text ratio" are associated with the annotation results of push mode representation information such as "video resource push mode" according to predefined policy node association rules. These rules may be obtained based on the testing and analysis of a large number of user learning terminals. For example, it is stipulated that in the page output state of high resolution and high text ratio, the association result of the video resource push mode is to give priority to pushing high-definition video tutorials with text descriptions.

[0174] The report generation of the policy node association result is performed to obtain a push execution report. The function used for report generation is a soft threshold activation function. For example, the policy node association result is a complex result containing various association relationships and operation suggestions. The soft threshold activation function converts it into a more concise and clear push execution report. This push execution report indicates the push rules for the user's learning terminal. For example, the report may stipulate that a specific type of learning resources should be pushed in a certain order and proportion, as well as requirements for the display format and interaction method of the resources.

[0175] 2. Online learning resource push based on learning resource push strategy

[0176] After receiving the push execution report, online learning resources are pushed to the user's learning terminal based on this learning resource push strategy. For example, the push execution report stipulates that a 10-minute high-definition video tutorial must be pushed first, and relevant text descriptions are displayed below the video, and an interactive button is provided on the side of the page for users to ask questions. Then the online learning resource push online learning cloud platform will follow this rule, select the appropriate high-definition video tutorial from the resource library, combine it with the corresponding text description, and push it on the page of the user's learning terminal according to the specified layout and interactive method.

[0177] (III) Detailed process of associating the strategy nodes with the push suggestion pooling vector for the first push task keyword to obtain the push mode representation information

[0178] 1. De-dimensionalization to obtain target pooling features

[0179] The push suggestion pooling vector is de-dimensionalized to obtain the target pooling feature, which is used to represent the importance of the resource push suggestion feature in each online learning link. For example, the push suggestion pooling vector is [1]. For example, this vector is obtained by weighted summing the resource push suggestion features [0.3, 0.5, 0.2], and the dimension of this value is related to a certain standard dimension in the online learning link (for example, the dimension of the difficulty coefficient of the learning resource). Through de-dimensionalization (for example, dividing by a standard value or scaling according to a certain ratio), the target pooling feature is obtained, for example, [0.5] (the numerical value of this embodiment is schematic, and is actually obtained according to a specific de-dimensionalization algorithm), and this numerical value represents the relative importance in a certain online learning link.

[0180] 2. Optimize the push mode representation information

[0181] The target pooling feature is optimized using the first push task keyword to obtain push mode representation information. For example, the first push task keyword is "comprehensive learning resource push", and the target pooling feature [0.5] is optimized. The optimization process may be to combine the target pooling feature with the task represented by the first push task keyword according to the semantics of the first push task keyword and the predefined optimization rules. For example, the optimization rule is to multiply the target pooling feature by a coefficient related to "comprehensive learning resource push" (this coefficient is obtained based on the analysis and experience of comprehensive learning resource push, such as 1.2), and obtain 0.5*1.2=0.6. This 0.6 is associated with "comprehensive learning resource push", and the push mode representation information is obtained by combining them, for example, it can be expressed as "comprehensive learning resource push-0.6", where 0.6 represents a certain optimized importance or weight under this push mode.

[0182] With this design, a learning resource push strategy is generated by obtaining the push mode representation information corresponding to the resource push suggestion features and the page output status description of the user's learning terminal, and performing a series of operations such as quantitative mapping, push annotation, strategy node association and report generation. This can fully consider the resource push suggestion features and the characteristics of the user's learning terminal, so that the push strategy not only meets the user's demand for learning resources, but also adapts to the display and interaction characteristics of the terminal.

[0183] For different types of resource push suggestion features (the first category and the second category), different methods are used to determine the push mode representation information, which increases the flexibility and applicability of the method and is able to handle various complex resource push situations.

[0184] In the process of associating strategy nodes, especially the operation of the first push task keyword and the push suggestion pooling vector, we can deeply explore the importance of resource push suggestion features in each online learning link, and optimize them according to the task keywords, so that the push mode representation information can more accurately reflect the needs and characteristics of resource push.

[0185] Pushing online learning resources to user learning terminals based on such a carefully constructed learning resource push strategy can improve the accuracy and effectiveness of resource push, thereby enhancing the user's learning experience, and also help to improve the utilization rate of online learning resources.

[0186] To sum up the above content, the following is an introduction through a specific algorithm example.

[0187] 1. Overview of the overall architecture

[0188] 1. Input layer

[0189] 1. Initial user behavior attribute vector input

[0190] For the initial user behavior attribute vector obtained by mining the learning behavior data to be processed (for example, a vector obtained by one-hot vector extraction, such as [0.8, 0.67, 0.1], etc.), it is input into the node as part of the neural network input layer. Each element of the vector corresponds to an input node, and the number of nodes in this part of the input layer is equal to the dimension of the initial user behavior attribute vector.

[0191] 2. Historical user behavior attribute vector input

[0192] Similarly, the mined historical user behavior attribute vector (such as [0.7, 0.7, 0.1], etc.) is also used as another part of the input nodes of the input layer. The number of this part of the input nodes is equal to the dimension of the historical user behavior attribute vector.

[0193] (II) Hidden Layer

[0194] 1. Key behavioral elements sampling related layer

[0195] The key behavior element sampling process involved in step 120 can be implemented by the hidden layer of the neural network.

[0196] For the sampling operation in each round of iteration (such as the sampling of the initial user behavior attribute vector and the historical user behavior attribute vector), the neurons in the hidden layer can adjust the weights and recombine the input vectors according to predefined rules (such as based on feature focusing strategies, etc.).

[0197] Taking the fully connected layer as an example, each neuron is connected to all nodes in the input layer, and the activation function of the neuron can be the ReLU (RectifiedLinearUnit) function, which can convert negative values ​​to 0, thereby increasing the nonlinear expression ability of the network. When sampling key behavioral elements, the weights of the neurons can be initialized based on the determined key behavioral elements (such as viewing time and homework completion as key behavioral elements) and continuously adjusted during training.

[0198] 2. Iterative Optimization Layer

[0199] Since there is at least one round of iterative linkage behavior preference embedding, multiple sub-layers are set in the hidden layer to implement each round of iterative operation.

[0200] In each round of iteration, the historical user behavior attribute vector and the initial user behavior attribute vector output from the previous round are used as the input of the next round, and are further adjusted and optimized through the neurons in the hidden layer. This iterative structure is similar to the time step iteration in the recurrent neural network (RNN), but here it is an iterative optimization for the behavior attribute vector.

[0201] To prevent overfitting, a Dropout layer can be added to the hidden layer to randomly set the output of some neurons to 0, so that the network does not rely too much on certain neurons during training, thereby improving the generalization ability of the network.

[0202] (III) Output layer

[0203] 1. Linked behavior preference embedding vector output

[0204] After multiple rounds of iterations, the final output of the hidden layer is the linkage behavior preference embedding vector (e.g. [0.65, 4.5, 0.15], etc.), which is used as the output of the output layer. The number of neurons in the output layer is equal to the dimension of the linkage behavior preference embedding vector.

[0205] 2. Output related to description of learning resource interests and resource push suggestion features

[0206] Furthermore, based on the output vector, a learning resource interest description and resource push suggestion features are generated through subsequent neural network layers or a separate processing module (which can be regarded as a part of the neural network).

[0207] For example, for the description of learning resource interests, a classifier-type neural network structure can be constructed, which takes the linkage behavior preference embedding vector as input and outputs the probabilities of different interest categories (such as the level of interest in basic theory, practical operation, interactive communication, etc.), and then generates a description of learning resource interests based on these probabilities.

[0208] For resource push suggestion features, the fully connected layer in the neural network can be used to calculate the correlation features between the learning resource interest description and the initial user behavior attributes. For example, the learning resource interest description and the initial user behavior attribute vector are concatenated as input, and the output is the weights or decision results of different resource push suggestions (such as pushing a concise version of basic theoretical knowledge, with simple practical cases, reducing the push of interactive communication resources, etc.).

[0209] 2. Algorithm Training Process

[0210] 1. Data preparation

[0211] Collect a large amount of online learning behavior data, covering the learning behavior data of different users in various courses, such as viewing time, number of pauses, homework completion status, etc., and use it as training data.

[0212] These data are processed according to the method of step 110 above to obtain the initial user behavior attribute vector and the historical user behavior attribute vector, as well as the corresponding known learning resource interest description and resource push suggestion features (which can be determined by manual labeling or based on existing platform strategies).

[0213] 2. Loss function definition

[0214] For the output of the linkage behavior preference embedding vector, the mean square error (MSE) can be defined as the loss function. Assume that the output linkage behavior preference embedding vector is (hat{y}_i), the known correct linkage behavior preference embedding vector (in the training data) is (y_i), and the number of samples is (n). The calculation formula of the mean square error is: (MSE=frac{1}{n}sum_{i=1}^{n}(y_{i}-hat{y}_{i})^2).

[0215] For the output of learning resource interest description and resource push suggestion features, appropriate loss functions can be defined according to the specific output form. For example, if it is a classification problem (such as different interest categories in learning resource interest description), the cross entropy loss function can be used; if it is a regression problem (such as the resource push ratio in the resource push suggestion feature), the mean square error or mean absolute error (MAE) can be used.

[0216] 3. Optimization algorithm selection

[0217] You can choose Stochastic Gradient Descent (SGD) and its variants, such as Adagrad, Adadelta, Adam and other optimization algorithms to minimize the loss function. These optimization algorithms can adaptively adjust the learning rate according to different gradient conditions and accelerate the convergence of the network.

[0218] During the training process, the error of the loss function is back-propagated from the output layer to the input layer through the back-propagation algorithm, and the weights of the neurons in the hidden layer and the input layer are adjusted to gradually reduce the loss function until the predetermined convergence condition is reached (such as the loss function value is less than a certain threshold or no longer decreases significantly within a certain number of iterations).

[0219] 3. Specific Algorithm Operation Examples

[0220] 1. Processing of initial user behavior attribute vector and historical user behavior attribute vector (step 110)

[0221] The initial user behavior attribute vector is (boldsymbol{X}=[x_1, x_2, x_3]), and the historical user behavior attribute vector is (boldsymbol{Y}=[y_1, y_2, y_3]). In the input layer, the elements of (boldsymbol{X}) and (boldsymbol{Y}) are mapped to the input nodes respectively.

[0222] In the first hidden layer (key behavior element sampling related layer), let the weight matrix of the neuron be (boldsymbol{W}_1) (used to process (boldsymbol{X}) and (boldsymbol{Y})), and the bias vector be (boldsymbol{b}_1).

[0223] After this layer of calculation, the output is:

[0224] (boldsymbol{Z}_1=f(boldsymbol{W}_1[boldsymbol{X}, boldsymbol{Y}]+boldsymbol{b}_1));

[0225] Where (f) is the activation function (such as ReLU).

[0226] 2. Iterative Linked Behavior Preference Embedding (Step 120)

[0227] In the iterative process, when the (t)th iteration is performed, the historical user behavior attribute vector output in the previous round is (boldsymbol{Y}_{t-1}), and the initial user behavior attribute vector is (boldsymbol{X}_{t-1}).

[0228] In the iterative optimization layer of the hidden layer, let the weight matrix be (boldsymbol{W}_t) and the bias vector be (boldsymbol{b}_t). Then the output of this round is

[0229] (boldsymbol{Z}_t=f(boldsymbol{W}_t[boldsymbol{X}_{t-1}, boldsymbol{Y}_{t-1}]+boldsymbol{b}_t)). This process is repeated multiple times (at least one iteration) until the final linkage behavior preference embedding vector is obtained.

[0230] 3. Generate learning resource interest description and resource push suggestion features (step 130)

[0231] For the linkage behavior preference embedding vector (boldsymbol{Z}) (output after multiple rounds of iterations), it is input into the neural network part used to generate learning resource interest description.

[0232] Let the weight matrix of this part be (boldsymbol{W}_{desc}) and the bias vector be (boldsymbol{b}_{desc}).

[0233] Based on this, the probability vector output as the interest description can be: (boldsymbol{P}=g(boldsymbol{W}_{desc}boldsymbol{Z}+boldsymbol{b}_{desc})), where (g) can be a Softmax function (if it is a classification problem), and the learning resource interest description is generated based on (boldsymbol{P}).

[0234] Then, the learning resource interest description and the initial user behavior attribute vector (which can be processed again, such as re-encoding, etc.) are concatenated into a vector (boldsymbol{Q}) and input into the neural network part used to generate resource push suggestion features. Assuming the weight matrix is ​​(boldsymbol{W}_{push}) and the bias vector is (boldsymbol{b}_{push}), the output is a resource push suggestion feature vector, which can be:

[0235] (bold symbol {R}

[0236] =h(boldsymbol{W}_{push}boldsymbol{Q}+boldsymbol{b}_{push})), where (h) is determined according to the specific output form (such as linear function or other activation function).

[0237] 4. Advantages and adaptability of the algorithm

[0238] 1. Advantages

[0239] The nonlinear characteristics of neural networks enable them to handle the relationship between complex user behavior data very well. For example, there may be nonlinear correlations between user behavior attributes such as viewing time and number of pauses, and neural networks can capture this relationship through hidden layer neurons and activation functions.

[0240] The iterative optimization structure can gradually dig deeper into the user's behavioral preferences. Each round of iteration is equivalent to a more refined adjustment of the user's behavioral preferences, similar to how humans gradually clarify their preferences in the process of continuous learning, so that the final output linkage behavior preference embedding vector is more accurate.

[0241] For the generation of interest descriptions and resource push suggestion features for different types of learning resources, the structure and parameters of the neural network can be flexibly adjusted according to specific task requirements, and it has strong adaptability.

[0242] 2. Adaptability

[0243] When faced with online learning data from different subject areas (such as programming learning, foreign language learning, mathematics learning, etc.), since neural networks are trained based on large amounts of data, as long as there is enough data in each subject area for training, they can adapt to user behavior patterns and learning resource needs in different disciplines.

[0244] For users at different learning stages (such as beginner and advanced stages), the neural network can automatically adjust the output learning resource interest descriptions and resource push suggestion features by learning the differences in behavioral data of users at different stages to meet the personalized needs of different stages.

[0245] With the development of online learning platforms and the continuous updating of user behavior data, neural networks can adapt to new data patterns and changes in user needs through retraining or online learning (for example, using incremental learning techniques).

[0246] Further, Figure 2 Schematic diagram of the structure of an online learning cloud platform 200 provided by the present invention. Figure 2 The online learning cloud platform 200 shown includes a processor 210, which can call and run a computer program from a memory to implement the method of the present invention.

[0247] Alternatively, if Figure 2 As shown, the online learning cloud platform 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method of the present invention.

[0248] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0249] Alternatively, if Figure 2 As shown, the online learning cloud platform 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, may send information or data to other devices, or receive information or data sent by other devices.

[0250] Optionally, the online learning cloud platform 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the present invention, which will not be described here for the sake of brevity.

[0251] It should be understood that the processor of the present invention may be an integrated circuit chip having signal processing capabilities.

[0252] It is understood that the memory in the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the system (cloud platform) and method described herein is intended to include but is not limited to suitable types of memory.

[0253] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0254] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0255] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0256] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the present invention and the scope of protection of the present invention, all of which are within the protection of the present invention.

Claims

1. A cloud computing-based online learning data processing method, characterized in that: The method is implemented by an online learning cloud platform, and the method includes: Acquire the learning behavior data to be processed, and mine the initial user behavior attributes of the learning behavior data to be processed to obtain the initial user behavior attribute vector of the learning behavior data to be processed; determine the historical user behavior attributes associated with the learning behavior data to be processed, and mine the historical user behavior attributes to obtain the historical user behavior attribute vector; the historical user behavior attributes are obtained by identifying the target learning behavior data corresponding to the learning behavior data to be processed that does not include the pushed interest items; Perform at least one round of iterative linkage behavior preference embedding according to the initial user behavior attribute vector and the historical user behavior attribute vector to obtain a linkage behavior preference embedding vector output by the at least one round of iterative linkage behavior preference embedding; in each round of iterative linkage behavior preference embedding task, perform key behavior element sampling on the target initial user behavior attribute vector and the target historical user behavior attribute vector input in the corresponding iteration cycle to obtain the historical user behavior attribute vector output in the corresponding iteration cycle, and perform key behavior element sampling on the historical user behavior attribute vector output in the corresponding iteration cycle and the target initial user behavior attribute vector to obtain the initial user behavior attribute vector output in the corresponding iteration cycle; Based on the linkage behavior preference embedding vector, a learning resource interest description for the learning behavior data to be processed is obtained; based on the correlation characteristics between the learning resource interest description and the initial user behavior attribute, a resource push suggestion feature for the learning behavior data to be processed is obtained; After obtaining a resource push suggestion feature for the learning behavior data to be processed based on the correlation feature between the learning resource interest description and the initial user behavior attribute, the method further includes: using the resource push suggestion feature to generate a learning resource push strategy for the user learning terminal corresponding to the learning behavior data to be processed; and pushing online learning resources to the user learning terminal based on the learning resource push strategy; The method of utilizing the resource push suggestion feature to generate a learning resource push strategy for the user learning terminal corresponding to the learning behavior data to be processed includes: obtaining the push mode representation information corresponding to the resource push suggestion feature and the page output state description corresponding to the user learning terminal, performing quantitative mapping on the page output state description, and obtaining the page output state quantitative knowledge corresponding to the page output state description; the function adopted for the quantitative mapping is a soft threshold activation function; using a push annotation feature to perform push annotation on the page output state quantitative knowledge and the push mode representation information respectively, and performing policy node association on the results obtained by the push annotation to obtain a policy node association result; generating a report on the policy node association result to obtain a push execution report; the function adopted for the report generation is a soft threshold activation function; the push execution report indicates the push rules for the user learning terminal.

2. The method according to claim 1, characterized in that The key behavior element sampling is performed on the target initial user behavior attribute vector and the target historical user behavior attribute vector input in the corresponding iteration cycle to obtain the historical user behavior attribute vector output in the corresponding iteration cycle, including: Obtaining a first user operation interaction vector and a first platform response interaction vector according to the target initial user behavior attribute vector input in the corresponding iteration cycle, and obtaining a first learning resource relationship vector according to the target historical user behavior attribute vector input in the corresponding iteration cycle; User group behavior element sampling is performed on the first user operation interaction vector, the first platform response interaction vector, and the first learning resource relationship vector to obtain a historical user behavior attribute vector outputted in the corresponding iteration cycle.

3. The method according to claim 2, characterized in that The sampling of user group behavior elements for the first user operation interaction vector, the first platform response interaction vector, and the first learning resource relationship vector to obtain a historical user behavior attribute vector outputted in a corresponding iteration period includes: Based on the first user operation interaction vector, the first platform response interaction vector, and the first learning resource relationship vector, sampling user group behavior elements based on a feature focus strategy to obtain a first user group behavior element vector; Performing feature integration on the first user group behavior element vector and the first learning resource relationship vector to obtain a target integration vector; Performing at least one soft threshold activation on the target integrated vector to obtain a behavior state activation vector; The behavior state activation vector and the target integration vector are feature integrated to obtain a historical user behavior attribute vector outputted in the corresponding iteration cycle.

4. The method according to claim 1, characterized in that The key behavior element sampling is performed on the historical user behavior attribute vector outputted in the corresponding iteration cycle and the target initial user behavior attribute vector to obtain the initial user behavior attribute vector outputted in the corresponding iteration cycle, including: Obtain a second user operation interaction vector and a second platform response interaction vector according to the historical user behavior attribute vector outputted in the corresponding iteration cycle, and obtain a second learning resource relationship vector according to the target initial user behavior attribute vector inputted in the corresponding iteration cycle; User group behavior element sampling is performed on the second user operation interaction vector, the second platform response interaction vector, and the second learning resource relationship vector to obtain an initial user behavior attribute vector outputted in the corresponding iteration cycle.

5. The method according to claim 4, characterized in that The sampling of user group behavior elements for the second user operation interaction vector, the second platform response interaction vector, and the second learning resource relationship vector to obtain an initial user behavior attribute vector output in the corresponding iteration cycle includes: Based on the second user operation interaction vector, the second platform response interaction vector, and the second learning resource relationship vector, sampling user group behavior elements based on a feature focus strategy to obtain a second user group behavior element vector; The second user group behavior element vector and the second learning resource relationship vector are feature integrated to obtain an initial user behavior attribute vector output in the corresponding iteration cycle.

6. The method according to claim 1, characterized in that The at least one round of iteration includes a plurality of iteration cycles; The step of performing at least one round of iterative linkage behavior preference embedding according to the initial user behavior attribute vector and the historical user behavior attribute vector to obtain a linkage behavior preference embedding vector output by the at least one round of iterative linkage behavior preference embedding includes: In the linkage behavior preference embedding task of the first iteration cycle among the several iteration cycles, key behavior element sampling is performed on the initial user behavior attribute vector and the historical user behavior attribute vector inputted by the first iteration cycle to obtain the historical user behavior attribute vector outputted by the first iteration cycle, and key behavior element sampling is performed on the historical user behavior attribute vector outputted by the first iteration cycle and the initial user behavior attribute vector to obtain the initial user behavior attribute vector outputted by the first iteration cycle; In each round of iteration starting from the second iteration cycle among the several iteration cycles, the linkage behavior preference embedding of the corresponding iteration cycle is performed based on the historical user behavior attribute vector and the initial user behavior attribute vector outputted in the previous iteration cycle, and the historical user behavior attribute vector and the initial user behavior attribute vector outputted in the corresponding iteration cycle are obtained, until the linkage behavior preference embedding of the last iteration cycle among the several iteration cycles outputs the linkage behavior preference embedding vector of the several iteration cycles.

7. The method according to claim 1, characterized in that The method of obtaining a learning resource interest description for the learning behavior data to be processed based on the linkage behavior preference embedding vector includes: sampling user group behavior elements for the initial user behavior attribute vector and the historical user behavior attribute vector included in the linkage behavior preference embedding vector to obtain a learning resource interest description for the learning behavior data to be processed.

8. The method according to claim 1, characterized in that The step of obtaining resource push suggestion features for the learning behavior data to be processed based on the correlation features between the learning resource interest description and the initial user behavior attributes includes: Determining correlation characteristics between the learning resource interest description and the initial user behavior attributes; Performing potential interest identification on the correlation features to obtain potential interest weight features; Determining resource push suggestion features for the learning behavior data to be processed according to the potential interest weight features; The method further comprises: Perform at least one round of iterative behavior preference path identification on the correlation features to obtain a learning behavior preference path; The determining of resource push suggestion features for the learning behavior data to be processed based on the potential interest weight features includes: Combining the potential interest weight feature with the learning behavior preference path to obtain a personalized learning behavior portrait; Determine resource push suggestion features for the learning behavior data to be processed based on the personalized learning behavior portrait.

9. The method according to claim 1, characterized in that The mining of the initial user behavior attributes of the learning behavior data to be processed to obtain the initial user behavior attribute vector of the learning behavior data to be processed includes: extracting the initial user behavior attributes of the learning behavior data to be processed, and performing at least one round of iterative one-hot vector extraction on the initial user behavior attributes to obtain the initial user behavior attribute vector of the learning behavior data to be processed; The determining of the historical user behavior attributes associated with the learning behavior data to be processed, and mining the historical user behavior attributes to obtain the historical user behavior attribute vector, includes: determining the historical user behavior attributes associated with the learning behavior data to be processed from a user behavior attribute set, wherein the user behavior attribute set includes at least one shared historical user behavior attribute obtained based on target learning behavior data identification; performing at least one round of iterative one-hot vector extraction on the historical user behavior attributes to obtain the historical user behavior attribute vector; The determining, from the user behavior attribute set, the historical user behavior attribute associated with the learning behavior data to be processed comprises: Determine a user behavior attribute set; associate the initial user behavior attribute with each shared historical user behavior attribute in the user behavior attribute set to obtain a behavior attribute association result; The shared historical user behavior attribute corresponding to the behavior attribute association result indicating that the behavior attribute has learning resource relevance is determined as the historical user behavior attribute associated with the learning behavior data to be processed.

10. An online learning cloud platform, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images