Artificial intelligence-based smart campus service push method and system

Abstract

The present invention provides a method and system for pushing smart campus services based on artificial intelligence. The method obtains service push training examples labeled with service acceptance description portraits, and performs feature extraction on the service usage trajectory data and campus service resource data in the service push training examples to generate multiple service knowledge vectors. The attention coefficient of each service knowledge vector is obtained using an attention network, and the service acceptance portrait is estimated in combination with an estimation network to generate service acceptance estimation portraits corresponding to multiple service push nodes. By calculating the network training error, the neuron weight information in the service acceptance portrait estimation network is optimized until the network converges. The method can accurately make smart campus service push decisions for the target push object, improve the accuracy of service push and user satisfaction, and realize efficient and personalized push of smart campus services.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based smart campus service push method and system. Background Art

[0002] With the rapid development of information technology and the widespread application of artificial intelligence, the construction of smart campuses has become a key trend in modern education. By integrating advanced information technology and intelligent management systems, smart campuses provide teachers and students with a more convenient and efficient service experience. However, delivering these services accurately and personalized to users remains a challenge in smart campus development.

[0003] Traditional service push methods are often based on simple rules or historical user behavior data. While these methods can achieve a certain degree of automated service push, they lack a deep understanding of users' current needs and service acceptance. As a result, the pushed services are often inaccurate and difficult to meet users' personalized needs. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for pushing smart campus services based on artificial intelligence, the method comprising:

[0005] Obtaining a service push training sample for a smart campus service, wherein the service push training sample is annotated with a service acceptance description portrait corresponding to a target service push node of the service push training sample;

[0006] Performing feature extraction on the service usage trajectory data and campus service resource data in the service push training example to generate a plurality of service knowledge vectors included in the service push training example;

[0007] Obtaining an attention coefficient of each service knowledge vector through an attention network, and performing a service acceptance profile estimation based on the multiple service knowledge vectors included in the service push training sample and the attention coefficient of each service knowledge vector through an estimation network, thereby generating a service acceptance estimation profile of each service push node among the multiple service push nodes corresponding to the service push training sample;

[0008] Calculate the network training error based on the service acceptance description portrait of the target service push node corresponding to the service push training sample and the service acceptance estimation portrait of each service push node corresponding to the service push training sample;

[0009] Based on the network training error, the neuron weight information used to estimate the service push training sample corresponding to the target service push node in the service acceptance portrait estimation network is optimized until the network convergence requirement is met, and then a smart campus service push decision is made for the target push object based on the service acceptance portrait estimation network, wherein the service acceptance portrait estimation network includes the attention network and the estimation network.

[0010] On the other hand, an embodiment of the present invention also provides an artificial intelligence-based smart campus service push system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present application obtains service push training samples marked with service acceptance description portraits, and performs feature extraction on the service usage trajectory data and campus service resource data in the service push training samples to generate multiple service knowledge vectors. The attention network is used to obtain the attention coefficient of each service knowledge vector, and the service acceptance portrait is estimated in combination with the estimation network to generate service acceptance estimation portraits corresponding to multiple service push nodes. By calculating the network training error, the neuron weight information in the service acceptance portrait estimation network is optimized until the network converges. This method can accurately make smart campus service push decisions for the target push object, improve the accuracy of service push and user satisfaction, and realize efficient and personalized push of smart campus services. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the artificial intelligence-based smart campus service push method provided by an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the hardware architecture of the artificial intelligence-based smart campus service push system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an artificial intelligence-based smart campus service push method provided by an embodiment of the present invention. The artificial intelligence-based smart campus service push method is introduced in detail below.

[0015] Step S110 , obtaining a service push training sample of a smart campus service, wherein the service push training sample is annotated with a service acceptance description portrait corresponding to a target service push node of the service push training sample.

[0016] In this embodiment, a smart campus environment includes a variety of services, such as online course learning services, campus event notification services, and library loan services. Assuming a smart campus service push system is being built, a large number of service push training examples are required to train the model. Taking the online course learning service as an example, service push training examples can be historical records of online course learning service pushes for different student groups over a period of time. These historical records contain a wide range of information, such as the basic information of the students being pushed (grade, major, etc.), the time of push, and the campus network environment at the time.

[0017] The service acceptance description portrait is a description of the acceptance level of each service push training example at the target service push node (for example, a specific course type push node, such as the advanced mathematics course push node). For training examples of online course learning services, the service acceptance description portrait may include multi-dimensional descriptions such as whether the student clicked to view the course details, whether the student actually registered for the course, and the duration of study. If a student not only clicked to view the course details after receiving the advanced mathematics course push, but also registered and continued to study for a long time, then the service acceptance description portrait at this advanced mathematics course push node will show a high level of acceptance. By collecting many such service push training examples and their corresponding service acceptance description portraits of the target service push nodes, basic data is provided for subsequent model training.

[0018] Step S120 : performing feature extraction on the service usage trajectory data and campus service resource data in the service push training example to generate a plurality of service knowledge vectors included in the service push training example.

[0019] Continuing with the example of online course learning services, service usage trajectory data may include students' course-related operation trajectories on the smart campus platform. For example, the time when students log in to the platform, the length of time they stay on different course pages, whether they participate in course-related discussion forum interactions, etc. First, these service usage trajectory data are cleaned to remove possible error records (such as abnormal login time records caused by network fluctuations), and some meaningless operation records are denoised (such as extremely short page stays caused by frequent misclicks). Then, normalization is performed to convert data of different magnitudes into the same magnitude for subsequent analysis. Next, key behavioral features are extracted from these processed trajectory data. For example, if the average length of time students stay on a specific course page exceeds a certain threshold, this may indicate that students have a high degree of attention to the course, which is a key behavioral feature.

[0020] Meanwhile, regarding campus service resource data, for example, resources related to online course learning services, including course videos, course handouts, and course assignments, are semantically analyzed. For example, the lecture content in the course videos is analyzed to extract key knowledge points. Keyword extraction is performed, such as extracting keywords such as "limit" and "derivative" from advanced mathematics course videos. These keywords are then converted into vectorized representations to generate corresponding resource feature vectors. Finally, the key behavioral features extracted from the service usage trajectory data are combined with the resource feature vectors obtained from the campus service resource data to generate individual service knowledge vectors containing service information and resource semantics. For example, a service knowledge vector might include a student's level of interest in the advanced mathematics course (derived from the service usage trajectory data) and vector representations of key knowledge points in the course (derived from the campus service resource data). Such a service knowledge vector can comprehensively reflect the relevant information in the service push training examples.

[0021] Step S130: Obtain the attention coefficient of each service knowledge vector through the attention network, and estimate the service acceptance portrait based on the multiple service knowledge vectors included in the service push training sample and the attention coefficient of each service knowledge vector through the estimation network, and generate a service acceptance estimation portrait of each service push node in the multiple service push nodes corresponding to the service push training sample.

[0022] In the online course learning service scenario of a smart campus, assume there are multiple service knowledge vectors containing various information related to students and courses. The attention network begins to work to capture the relationships between these service knowledge vectors. Using the attention network's global attention mechanism, similarity is calculated for each service knowledge vector with all other service knowledge vectors. For example, a service knowledge vector representing a student's interest in an advanced mathematics course and the characteristics of the course resources is similar to a similar vector representing a student's interest in an English course. This calculation yields a global similarity matrix, which reflects the importance of each service knowledge vector in the global context. If a service knowledge vector has a high degree of similarity with multiple other vectors related to improving academic performance, then it is considered to be highly important in the global context.

[0023] This global similarity matrix is ​​then used to perform a weighted summation of the service knowledge vectors to obtain an enhanced service knowledge vector that incorporates global context information. Next, each enhanced service knowledge vector is segmented to generate multiple sub-vectors. For example, an enhanced service knowledge vector for advanced mathematics courses is segmented according to dimensions such as course content, learning time, and learning effect. The local attention mechanism based on the attention network applies a local attention function to each sub-vector and identifies the local focus of attention by calculating the correlation between the internal elements of the sub-vector. For example, in the sub-vector about learning time, the local focus of attention is identified as 7-9 pm every day, which is the concentrated time period for students to study advanced mathematics courses. The local focus of attention of each sub-vector is merged to generate a local focus of attention vector for the entire service knowledge vector.

[0024] Then, a multi-level attention fusion strategy based on the attention network is employed to define multiple levels of attention heads, such as the semantic level (focusing on the semantic information of the course content), the behavioral level (focusing on students' learning behavior), and the resource level (focusing on the utilization of course resources). These attention heads are simultaneously applied to each local focus vector, resulting in a multi-level attention representation. These multi-level attention representations are then fused through a weighted summation to generate a final service knowledge vector that incorporates the multi-level attention information. Predefined dynamic adjustment rules are then applied to the attention coefficient of each final service knowledge vector. For example, a time-decay-based adjustment rule might reduce the attention coefficient of a service knowledge vector representing a course learned a long time ago. A user preference-based adjustment rule might adjust the attention coefficient of service knowledge vectors related to mathematics courses if a student indicates a lack of interest in mathematics in their profile. A service popularity-based adjustment rule might increase the attention coefficient of a service knowledge vector related to a course that has recently become popular on campus. The sum or maximum value of all dynamically adjusted attention coefficients is calculated, and each attention coefficient is divided by this sum or maximum value to obtain the normalized attention coefficient.

[0025] The service acceptance portrait is estimated through the estimation network. For each service knowledge vector, the service knowledge vector and the attention coefficient corresponding to the service knowledge vector are weighted and calculated to generate a target weighted knowledge vector corresponding to each service knowledge vector. Assume that the first network function unit in the estimation network is a multidimensional fusion output unit and the second network function unit is a sequence response unit. The service acceptance portrait is estimated based on the multiple target weighted knowledge vectors corresponding to the service push training sample by the multidimensional fusion output unit, and the first service acceptance estimation portrait of the service push training sample corresponding to each service push node (such as different course push nodes) is generated. At the same time, the service acceptance portrait is estimated based on the multiple target weighted knowledge vectors corresponding to the service push training sample by the sequence response unit, and the second service acceptance estimation portrait of the service push training sample corresponding to each service push node is generated. In this way, the service acceptance estimation portrait of the service push training sample corresponding to each service push node in the multiple service push nodes is obtained.

[0026] Step S140 , calculating the network training error according to the service acceptance description portrait of the target service push node corresponding to the service push training sample and the service acceptance estimation portrait of each service push node corresponding to the service push training sample.

[0027] Taking the online course learning service in a smart campus as an example, let's assume the target service push node is the advanced mathematics course push node. The service acceptance profile for this target service push node in the service push training example includes accurate information on actual student acceptance, such as the percentage of students who have registered for the advanced mathematics course and their progress within the course. The estimated service acceptance profile for each service push node (including the advanced mathematics course push node and other course push nodes) in the service push training example is an estimate obtained through the estimation network described above.

[0028] For example, on the advanced mathematics course push node, the service acceptance profile shows that 30% of students have actually registered for the course and the average learning progress has reached 50%. The first service acceptance estimation profile obtained through the estimation network shows that 25% of students are expected to register for the course and the average learning progress is 45%. The second service acceptance estimation profile shows that 28% of students are expected to register for the course and the average learning progress is 48%. Based on these actual values ​​and estimated values, a specific error calculation method (such as mean square error) is used to calculate the network training error. The error is calculated for each dimension (registration ratio, learning progress, etc.), and these errors are combined to obtain the network training error on the advanced mathematics course push node. Similarly, this calculation is performed on other course push nodes to obtain the network training error of the entire service push training sample on each service push node.

[0029] Step S150: Optimize the neuron weight information in the service acceptance portrait estimation network used to estimate the service push training sample corresponding to the target service push node based on the network training error until the network convergence requirement is met, and make a smart campus service push decision for the target push object based on the service acceptance portrait estimation network, wherein the service acceptance portrait estimation network includes the attention network and the estimation network.

[0030] In the context of a smart campus online course learning service, the neuron weights in the service acceptance profile estimation network are optimized based on the network training error calculated on the advanced mathematics course push node. Assume that the attention network in the service acceptance profile estimation network contains multiple neurons, each of which plays a different role in processing service knowledge vectors, such as calculating the global similarity matrix and local attention focus. The first network functional unit (multidimensional fusion output unit) and the second network functional unit (sequential response unit) in the estimation network also contain numerous neurons.

[0031] First, the target neuron weight information in the service acceptance portrait estimation network is shielded and annotated. The target neuron weight information here refers to the other neuron weight information except the neuron weight information used to estimate the push node corresponding to the service push training sample of the advanced mathematics course in the fully connected mapping unit of the estimation network. For example, in the attention network, the neuron weight information related to other courses (non-advanced mathematics courses) is shielded and annotated. Then, based on the network training error, the unshielded and annotated neuron weight information in the service acceptance portrait estimation network is optimized. For example, if it is found that the error in calculating the global similarity matrix for the service knowledge vector related to the advanced mathematics course is large, the weights of the relevant neurons in the attention network are adjusted to improve the calculation accuracy in this regard.

[0032] This optimization process is continued until the network convergence requirement is met, that is, the network training error is reduced to an acceptable range. Once the network convergence requirement is met, the service acceptance profile estimation network can be used to make smart campus service push decisions for the target push object (such as a specific student group). For example, for a new group of students, their service usage trajectory vectors (such as their learning trajectory in other courses) and the campus service resource semantic vectors (such as the semantic vector representation of the course content) of multiple campus service resource data to be pushed (such as different online course resources) are obtained. Based on these vectors, the service acceptance profile estimation network generates a service acceptance estimation portrait of each service push node (such as a different course push node) corresponding to the campus service resource data to be pushed. Based on these service acceptance estimation portraits, the target campus service resource data (such as the online course that is most suitable for a certain student group) and the service push node corresponding to the target campus service resource data (such as the push timing and method of the course) are determined from the multiple campus service resource data to be pushed. Finally, based on the target campus service resource data and the service push node corresponding to the target campus service resource data, the target push object is pushed to the smart campus service, for example, the most suitable online course is pushed to a specific student group at the right time and in the right way.

[0033] Based on the above steps, the embodiment of the present application obtains service push training samples marked with service acceptance description portraits, and performs feature extraction on the service usage trajectory data and campus service resource data in the service push training samples to generate multiple service knowledge vectors. The attention network is used to obtain the attention coefficient of each service knowledge vector, and the service acceptance portrait is estimated in combination with the estimation network to generate service acceptance estimation portraits corresponding to multiple service push nodes. By calculating the network training error, the neuron weight information in the service acceptance portrait estimation network is optimized until the network converges. This method can accurately make smart campus service push decisions for the target push object, improve the accuracy of service push and user satisfaction, and realize efficient and personalized push of smart campus services.

[0034] In a possible implementation, step S150 includes:

[0035] Step S151, shielding and marking the target neuron weight information in the service acceptance portrait estimation network, wherein the target neuron weight information is other neuron weight information except the neuron weight information used to estimate the target service push node corresponding to the service push training sample in the fully connected mapping unit of the estimation network.

[0036] Step S152: Optimizing the weight information of neurons that are not shielded and marked in the service acceptance profile estimation network based on the network training error.

[0037] In this embodiment, in the online course learning service scenario of a smart campus, after obtaining the network training error through the previous steps, the service acceptance portrait estimation network begins to optimize the neuron weight information of the target service push node (such as the advanced mathematics course push node) used to estimate the service push training sample.

[0038] The first step is to shield and annotate the target neuron weight information in the service acceptance portrait estimation network. The service acceptance portrait estimation network is a complex structure that includes an attention network and an estimation network. In the estimation network, there is a fully connected mapping unit. This unit plays a key role in estimating that the training examples for service push correspond to the advanced mathematics course push node, and specific neuron weight information. The target neuron weight information to be shielded and annotated here is the neuron weight information other than these specific neuron weight information. For example, in the attention network part, there are some neurons that are used to process the calculation of service knowledge vectors related to other courses (such as English courses, physics courses, etc.). These neuron weight information is not directly related to the estimation of the advanced mathematics course push node, so it is necessary to shield and annotate it.

[0039] Next, based on the network training error, the unmasked and annotated neuron weights in the service acceptance profile estimation network are optimized. Since the network training error was previously calculated, this error reflects the difference between the estimated value at the advanced mathematics course push node and the actual service acceptance profile. If, during the previous service acceptance profile estimation process, it was found that the results obtained by the attention network for the service knowledge vector related to the advanced mathematics course deviated significantly from the actual situation in certain steps (such as calculating the global similarity matrix under the global attention mechanism or determining the local focus of attention under the local attention mechanism), this indicates that the relevant neuron weights may need to be adjusted. For example, if it is found that the global similarity matrix calculation of the service knowledge vector related to the advanced mathematics course causes the course difficulty factor to be over- or under-emphasized due to the setting of certain neuron weights, thereby affecting the overall service acceptance estimate, the unmasked and annotated neuron weights can be optimized and adjusted based on the feedback from the network training error. In the first network functional unit (such as the multi-dimensional fusion output unit) and the second network functional unit (such as the sequence response unit) in the estimation network, there is also neuron weight information related to the estimation of the advanced mathematics course push node. If a certain neuron weight in these units is found to cause a large error when fusing information of different dimensions or processing information in sequence, it can also be optimized based on the network training error. By optimizing the neuron weight information that is not shielded and marked in this way, the accuracy of the service acceptance portrait estimation network in estimating the advanced mathematics course push node can be gradually improved, thereby improving the performance of the entire smart campus service push system.

[0040] In a possible implementation, the estimation network includes a first network function unit and a second network function unit, and step S120 includes:

[0041] Step S121: Clean, denoise, and normalize the service usage trajectory data to extract key behavior features. At the same time, perform semantic analysis, keyword extraction, and vectorization on the campus service resource data to generate corresponding resource feature vectors. The key behavior features and the resource feature vectors are fused to generate various service knowledge vectors containing service information and resource semantics.

[0042] Step S130 includes:

[0043] In step S131, the global attention mechanism of the attention network is used to capture the global dependency between each service knowledge vector. Specifically, the similarity between each service knowledge vector and all other service knowledge vectors is calculated to obtain a global similarity matrix, which reflects the importance of each service knowledge vector in the global context. The global similarity matrix is ​​then used to perform weighted summation on the service knowledge vectors to obtain an enhanced service knowledge vector that integrates global context information.

[0044] Step S132, each enhanced service knowledge vector is segmented to generate multiple sub-vectors, and a local attention function is applied to each sub-vector based on the local attention mechanism of the attention network. The local attention function identifies the local attention focus by calculating the correlation between the internal elements of the sub-vector, and merges the local attention focus of each sub-vector to generate a local attention focus vector for the entire service knowledge vector.

[0045] In step S133, the multi-level attention fusion strategy of the attention network is adopted to define multiple levels of attention heads, each of which is used to capture information at a corresponding level.

[0046] In step S134, the attention head is applied to each local focus vector simultaneously to obtain multiple levels of attention representation, and the multiple levels of attention representation are fused by weighted summation to generate a final service knowledge vector that integrates multiple levels of attention information, wherein the levels include semantic level, behavioral level and resource level.

[0047] Step S135, applying predefined dynamic adjustment rules to the attention coefficient of each final service knowledge vector to obtain a dynamically adjusted attention coefficient, wherein the dynamic adjustment rules include adjustment rules based on time attenuation, adjustment rules based on user preferences, and adjustment rules based on service popularity.

[0048] Step S136, calculate the sum or maximum value of all dynamically adjusted attention coefficients, and divide each attention coefficient by the sum or maximum value to obtain a normalized attention coefficient.

[0049] Step S137: For each service knowledge vector, perform weighted calculation on the service knowledge vector and the attention coefficient corresponding to the service knowledge vector to generate a target weighted knowledge vector corresponding to each service knowledge vector.

[0050] Step S138: The first network function unit estimates a service acceptance profile based on multiple target weighted knowledge vectors corresponding to the service push training examples, and generates a first service acceptance estimation portrait corresponding to each service push node of the service push training examples.

[0051] Step S139: The second network function unit estimates a service acceptance profile based on multiple target weighted knowledge vectors corresponding to the service push training sample, and generates a second service acceptance estimation portrait corresponding to each service push node of the service push training sample.

[0052] Step S140 may include: calculating the network training error based on the service acceptance description portrait of the service push node corresponding to the service push training sample, the first service acceptance estimation portrait of each service push node corresponding to the service push training sample, and the second service acceptance estimation portrait of each service push node corresponding to the service push training sample.

[0053] Step S150 may include: optimizing the neuron weight information of the attention network based on the network training error, optimizing the neuron weight information in the first network function unit used to estimate the service push training sample corresponding to the target service push node, and optimizing the neuron weight information of the second network function unit.

[0054] In this embodiment, on the smart campus platform, the service usage trajectory data for online course learning services contains a lot of information. For example, the time records of students logging into the smart campus online course platform, the browsing time on different course pages, and the timestamps of course-related operations (such as clicking on course materials, submitting homework, participating in course discussions, etc.). First, these service usage trajectory data are cleaned to remove those obviously erroneous records, such as abnormal login times caused by system failures or unreasonable operation records caused by misoperation. Then, denoising is performed to eliminate small fluctuations that may interfere with the analysis. For example, occasional fast page switching may be a misoperation rather than a real learning behavior, and it needs to be removed. Then, normalization is performed to convert data of different magnitudes into the same magnitude. For example, the login time may be in seconds, and the homework submission time interval may be in days. Normalization enables it to be analyzed at the same scale. Key behavioral features are extracted from the processed service usage trajectory data. For example, if a student spends more than a certain amount of time (e.g., 5 hours) per week browsing course materials and completing homework in a specific online course (e.g., an advanced mathematics course), this is a key behavioral feature, indicating that the student has a high level of engagement in the course.

[0055] Meanwhile, for campus service resource data, taking online course learning resources as an example, including course videos, course handouts, and course assignments, semantic analysis is performed on course videos. This involves recognizing the audio content in the videos and converting it into text. The semantic information in this text is then analyzed, such as identifying key concepts and theories in the video explanations. For course handouts, keyword extraction is performed. For example, keywords closely related to the core course content, such as "function," "limit," and "derivative," are extracted from the advanced mathematics course handouts. These extracted keywords and related information from the course assignments are then vectorized to generate corresponding resource feature vectors. For example, the keyword "function" is converted into a vector whose dimensions may be related to the entire smart campus knowledge system. Each element in the vector represents the characteristic value of this keyword at different knowledge dimensions. Finally, key behavioral features extracted from service usage trajectory data are combined with resource feature vectors obtained from campus service resource data to generate individual service knowledge vectors containing service information and resource semantics. For example, a service knowledge vector might include a student's participation in an advanced mathematics course (key behavioral characteristics) and a vector representation of the core knowledge points in the course (such as functions and limits) (resource feature vector). Such a service knowledge vector can comprehensively reflect the relevant information in the service push training examples, providing a rich data foundation for subsequent analysis.

[0056] In the context of online course learning services in smart campuses, multiple service knowledge vectors exist, each containing different aspects of information. The global attention mechanism of the attention network is used to capture the global dependencies between these service knowledge vectors. For example, one service knowledge vector represents a student's learning behavior and course resource characteristics for an advanced mathematics course, while another service knowledge vector represents similar information for an English course. A global similarity matrix is ​​constructed by calculating the similarity between each service knowledge vector and all other service knowledge vectors. Specifically, for the service knowledge vector representing the advanced mathematics course, its similarity is calculated with the service knowledge vectors representing all other courses, such as English, physics, and computer science. This similarity can be calculated based on a distance metric between vectors, such as Euclidean distance or cosine similarity. The resulting global similarity matrix reflects the importance of each service knowledge vector in the global context. If a service knowledge vector has high similarity with multiple vectors related to improved academic performance—for example, a service knowledge vector representing active student participation in course discussions and high course resource utilization has high similarity with other similar service knowledge vectors related to active learning—then it is considered highly important in the global context. This global similarity matrix is ​​then used to perform a weighted summation of the service knowledge vectors to obtain an enhanced service knowledge vector that incorporates global context information. For example, for the service knowledge vector of an advanced mathematics course, its original vector representation is weighted and adjusted based on its importance in the global similarity matrix, resulting in an enhanced service knowledge vector that incorporates global context information (such as its relationship with learning behaviors and resource utilization in other courses).

[0057] Each enhanced service knowledge vector is segmented. For example, an enhanced service knowledge vector for a higher mathematics course is segmented according to dimensions such as course content, study time, and learning outcomes, generating multiple sub-vectors. The attention network-based local attention mechanism applies a local attention function to each sub-vector. For example, the study time sub-vector contains internal elements such as the distribution of students' time spent studying higher mathematics courses in different time periods. The local attention function identifies local attentional focuses by calculating the correlation between the internal elements of the sub-vectors. For example, analysis reveals that students spend the most time studying higher mathematics courses between 7:00 and 9:00 p.m. each day, and their learning efficiency is relatively high during this time period (as reflected by indicators such as homework completion quality). Therefore, this time period is the local attentional focus of the study time sub-vector. This operation is repeated for each sub-vector, and the local attentional focuses of each sub-vector are then merged to generate a local attentional focus vector for the entire service knowledge vector. For example, the local attentional focuses of sub-vectors such as course content, study time, and learning outcomes are merged to form a local attentional focus vector that comprehensively reflects the key information of the higher mathematics course service knowledge vector at the local level.

[0058] A multi-level attention fusion strategy based on an attention network is employed to define multiple levels of attention heads: semantic, behavioral, and resource. Taking online course learning services as an example, the semantic-level attention heads primarily focus on the semantic information of the course content. For example, in advanced mathematics courses, the attention heads focus on the semantic relationships between mathematical concepts and theorems. The behavioral-level attention heads focus on students' learning behaviors, such as their study schedules and the frequency of their participation in course discussions. The resource-level attention heads focus on the utilization of course resources, such as the number of course video views and the number of course handout downloads. These attention heads are applied simultaneously to each local attention focus vector, resulting in a multi-level attention representation. For example, for a local attention focus vector in an advanced mathematics course, the semantic-level attention heads generate attention representations regarding the semantics of the course content, the behavioral-level attention heads generate attention representations regarding learning behavior, and the resource-level attention heads generate attention representations regarding course resource utilization. These multi-level attention representations are then fused through a weighted summation to generate a final service knowledge vector that incorporates information from multiple levels of attention. This final service knowledge vector integrates information from multiple aspects, such as course content semantics, student learning behavior, and course resource utilization, and can more comprehensively and accurately reflect the relevant situation of online course learning services.

[0059] Predefined dynamic adjustment rules are applied to the attention coefficient of each final service knowledge vector. In the online course learning service of the smart campus, dynamic adjustment rules include time decay-based adjustment rules, user preference-based adjustment rules, and service popularity-based adjustment rules. Regarding the time decay-based adjustment rule, if the course learning represented by a service knowledge vector occurred a long time ago, for example, a student had a learning record for an online course a year ago, the importance of this record may decrease over time, and its attention coefficient will be adjusted accordingly based on time decay. Regarding the user preference-based adjustment rule, if a student indicates a lack of interest in mathematics in their profile, the attention coefficient of the service knowledge vector related to mathematics courses will be adjusted accordingly. For example, if a service knowledge vector is related to an advanced mathematics course, and the student's preference indicates a lack of interest in mathematics, the attention coefficient of this service knowledge vector will be reduced to reflect the impact of this user preference on service acceptance. Regarding the service popularity-based adjustment rule, if a course has recently become popular on campus, for example, an emerging computer technology course has attracted a lot of attention and study on campus, the attention coefficient of the service knowledge vector related to this course may be increased. The sum or maximum value of all dynamically adjusted attention coefficients is calculated, and each attention coefficient is divided by this sum or maximum value to obtain the normalized attention coefficient. This normalized attention coefficient can reflect the relative importance of each service knowledge vector on a unified scale, providing accurate weight information for subsequent service acceptance profile estimation.

[0060] For each service knowledge vector in the Smart Campus Online Course Learning Service, a weighted calculation is performed on the service knowledge vector and its corresponding attention coefficient to generate a target weighted knowledge vector corresponding to each service knowledge vector. For example, for a service knowledge vector related to an advanced mathematics course, which contains information such as course participation and course resource characteristics, this service knowledge vector is multiplied by its attention coefficient obtained in the previous step to obtain a target weighted knowledge vector. This target weighted knowledge vector, to a certain extent, highlights the importance of information related to the advanced mathematics course within the entire service knowledge system. It is the result of a comprehensive consideration of the content of the service knowledge vector itself and its importance in both the global and local contexts.

[0061] Assume that the first network function unit is a multi-dimensional fusion output unit, and the service acceptance portrait is estimated based on the multiple target weighted knowledge vectors corresponding to the service push training sample through this multi-dimensional fusion output unit, and the first service acceptance estimation portrait corresponding to each service push node (such as different course push nodes) of the service push training sample is generated. In the scenario of smart campus online course learning service, the multi-dimensional fusion output unit will comprehensively consider the multi-dimensional information in each target weighted knowledge vector, such as students' learning behavior in different courses, course resource characteristics, and the importance weight of this information. For the advanced mathematics course push node, the multi-dimensional fusion output unit will fuse the information in these vectors based on multiple target weighted knowledge vectors related to the advanced mathematics course, and output an estimated portrait of the service acceptance at this node. This portrait may contain multi-dimensional estimated information such as the student's acceptance probability of the advanced mathematics course push and possible learning participation.

[0062] The service acceptance portrait is estimated based on multiple target weighted knowledge vectors corresponding to the service push training samples through the second network function unit (assuming it is a sequence response unit), and a second service acceptance estimation portrait corresponding to each service push node of the service push training sample is generated. In the online course learning service, the sequence response unit will process the target weighted knowledge vectors in a certain order. For example, according to the chronological order of course learning or the increasing order of course difficulty, etc. For the advanced mathematics course push node, the sequence response unit will process the target weighted knowledge vectors related to the advanced mathematics course according to this order, taking into account the response of students to the course push at different learning stages or different course difficulties, thereby generating a second service acceptance estimation portrait for this node. This portrait may be different from the first service acceptance estimation portrait in some aspects. For example, it may focus more on the impact of sequential factors in the learning process on service acceptance.

[0063] The network training error is calculated based on the service acceptance profile of the service push node (e.g., the advanced mathematics course push node) corresponding to the service push training example, the first service acceptance estimation profile corresponding to each service push node, and the second service acceptance estimation profile corresponding to each service push node. In the smart campus online course learning service scenario, the service acceptance profile includes accurate information on actual student acceptance, such as the percentage of students who have actually registered for the advanced mathematics course and their actual learning progress within the course. The first service acceptance estimation profile and the second service acceptance estimation profile are estimates obtained through the estimation network. For example, on the advanced mathematics course push node, the service acceptance profile shows that 30% of students have actually registered for the course and their average learning progress has reached 50%, while the first service acceptance estimation profile shows that 25% of students are expected to register for the course and their average learning progress is 45%, and the second service acceptance estimation profile shows that 28% of students are expected to register for the course and their average learning progress is 48%. Based on these actual and estimated values, a specific error calculation method (such as mean square error) is used to calculate the network training error. The error is calculated for each dimension (enrollment ratio, learning progress, etc.), and these errors are combined to obtain the network training error on the advanced mathematics course push node. Similarly, the calculation is performed on the other course push nodes to obtain the network training error on each service push node for the entire service push training example.

[0064] Based on the network training error, the neuron weight information in the service acceptance portrait estimation network used to estimate the service push training samples corresponding to the target service push node (such as the advanced mathematics course push node) is optimized. Specifically, the neuron weight information of the attention network is optimized because the attention network plays a key role in processing service knowledge vectors, calculating attention coefficients, etc. If the previous analysis found that the attention network had errors in calculating the global similarity matrix or the local focus vector, resulting in inaccurate final service acceptance estimation, the neuron weights in the attention network are adjusted according to the network training error. For example, if it is found that the weight setting of a certain neuron is unreasonable when calculating the global similarity between the advanced mathematics course and other courses, resulting in a deviation in the overall similarity calculation, thereby affecting the service acceptance estimation, the weight of this neuron is optimized.

[0065] At the same time, the neuron weight information used to estimate the service push training examples corresponding to the advanced mathematics course push nodes in the first network functional unit (multi-dimensional fusion output unit) is optimized. If the multi-dimensional fusion output unit makes errors when fusing multi-dimensional information to generate the first service acceptance estimation profile, for example, if the neuron weights for certain key dimensions (such as the dimension of the impact of course difficulty on service acceptance) are improperly set, resulting in a significant deviation between the estimated value and the actual service acceptance description profile, these neuron weights are adjusted based on the network training error.

[0066] In addition, the neuron weight information of the second network functional unit (sequence response unit) is optimized. If there is an error in the sequence response unit when processing the target weighted knowledge vector in sequence to generate the second service acceptance estimation portrait, for example, when considering the impact of the course learning order on service acceptance, some neuron weights are not set accurately, resulting in the estimation result not being consistent with the actual situation, these neuron weights are adjusted according to the network training error. By comprehensively optimizing the neuron weight information in the service acceptance portrait estimation network in this way, the accuracy of the network in estimating the service acceptance of each service push node can be gradually improved, thereby improving the overall performance of the smart campus service push system.

[0067] In a possible implementation, the first network function unit is a multi-dimensional fusion output unit, and the second network function unit is a sequence response unit or a feature interaction relationship unit.

[0068] In a possible implementation, the estimation network further includes a third network function unit, and step S130 may further include:

[0069] The third network function unit estimates the service acceptance portrait based on multiple target weighted knowledge vectors corresponding to the service push training sample, and generates a third service acceptance estimation portrait corresponding to each service push node of the service push training sample, wherein the second network function unit is a sequence response unit and the third network function unit is a feature interaction relationship unit.

[0070] The method of calculating the network training error based on the service acceptance description portrait of the service push node corresponding to the service push training sample, the first service acceptance estimation portrait of each service push node corresponding to the service push training sample, and the second service acceptance estimation portrait of each service push node corresponding to the service push training sample includes: calculating the network training error based on the service acceptance description portrait of the service push node corresponding to the service push training sample and the first service acceptance estimation portrait of each service push node corresponding to the service push training sample, the second service acceptance estimation portrait of each service push node corresponding to the service push training sample, and the third service acceptance estimation portrait of each service push node corresponding to the service push training sample.

[0071] In this embodiment, in the online course learning service scenario of the smart campus, for the case where the first network function unit is a multi-dimensional fusion output unit, the second network function unit is a sequence response unit or a feature interaction relationship unit, and the estimation network also includes a third network function unit (feature interaction relationship unit), the following is a detailed example of each step.

[0072] First, in the stage of generating a service acceptance estimation portrait, take the advanced mathematics course push node as an example. The multi-dimensional fusion output unit estimates the service acceptance portrait based on multiple target weighted knowledge vectors corresponding to the service push training sample to generate a first service acceptance estimation portrait. Assume that the service push training sample contains online course learning-related data of multiple students. For example, the target weighted knowledge vector of student A contains his participation in the advanced mathematics course (such as weekly study time, homework completion status, etc.) and course resource utilization (such as the number of course video views, handout downloads, etc.). The multi-dimensional fusion output unit comprehensively considers this multi-dimensional information. For the advanced mathematics course push node, it may analyze the first service acceptance estimation portrait at this node based on the similar target weighted knowledge vectors of multiple students. For example, the portrait shows that 25% of students are expected to register for the course and the average learning progress is 45%.

[0073] The sequence response unit estimates the service acceptance profile based on multiple target weighted knowledge vectors corresponding to the same service push training examples to generate a second service acceptance estimation profile. Since the sequence response unit processes the target weighted knowledge vectors in a certain order, such as the chronological order of course learning. For advanced mathematics courses, it will take into account students' responses to course push at different learning stages. For example, students' reactions in different time sequences such as the early course promotion stage, the mid-term course learning in-depth stage, and the later review stage. Based on these sequence-related analyses, the second service acceptance estimation profile generated for the advanced mathematics course push node may show that 28% of students are expected to register for the course and the average learning progress is 48%.

[0074] The third network function unit (feature interaction relationship unit) also estimates the service acceptance portrait based on multiple target weighted knowledge vectors corresponding to the service push training examples to generate a third service acceptance estimation portrait. In the smart campus online course learning service, the feature interaction relationship unit mainly focuses on the impact of the interaction between different features on service acceptance. Taking advanced mathematics courses as an example, it will analyze how the interaction between students' learning foundation (such as previous grades in related courses) and course resources (such as course difficulty, richness of course materials, etc.) affects students' acceptance of course push. By analyzing the relevant data of many students, the third service acceptance estimation portrait generated for the advanced mathematics course push node may show that 26% of students are expected to register for the course and the average learning progress is 46%.

[0075] Then, in the network training error calculation phase, the network training error is calculated based on the service acceptance profile corresponding to the service push node (here, the advanced mathematics course push node) for the service push training example and the first, second, and third service acceptance estimated profiles corresponding to each service push node for the service push training example. The service acceptance profile includes actual student acceptance, for example, 30% of students actually registered for the course and their average learning progress reached 50%. For the enrollment ratio dimension, the first service acceptance estimated profile estimates 25%, which deviates from the actual 30%; the second service acceptance estimated profile estimates 28%, which also deviates from the actual 30%; and the third service acceptance estimated profile estimates 26%, which also has an error. A similar situation occurs for the learning progress dimension. Using specific error calculation methods such as mean squared error, the error for each dimension is calculated separately. These errors are then combined to obtain the network training error for the advanced mathematics course push node. The same calculation is performed for other course push nodes, such as English and physics, to obtain the network training error for each service push node for the entire service push training example. By comprehensively considering the comparison between multiple service acceptance estimation portraits and actual service acceptance description portraits to calculate the network training error, it is possible to more accurately reflect the accuracy of the estimation network in different aspects and provide an accurate basis for the subsequent optimization of neuron weight information.

[0076] In a possible implementation, step S150 includes:

[0077] If the campus service resource label in the service push training sample is the target campus service resource label, based on the network training error, the neuron weight information in the service acceptance portrait estimation network used to estimate the service push node of the service push training sample as the target service push node is optimized.

[0078] If the campus service resource label in the service push training example is not the target campus service resource label, at least one neuron weight information of the service acceptance portrait estimation network is optimized based on the network training error.

[0079] In this embodiment, taking the advanced mathematics course as the target service push node, it is assumed that there are numerous service push training examples, each of which contains relevant information such as a campus service resource tag. When the campus service resource tag in the service push training example is the target campus service resource tag (such as a tag directly related to the advanced mathematics course, such as a specific textbook version or a specific teaching video series), the neuron weight information used in the service acceptance profile estimation network to estimate that the service push node for the service push training example is the advanced mathematics course push node is optimized based on the network training error.

[0080] For example, in the attention network, certain neurons are responsible for calculating service knowledge vectors related to advanced mathematics courses. For example, when calculating the global similarity matrix, if it is found that the setting of certain neuron weights for the service knowledge vector related to advanced mathematics courses results in a large error when comparing its similarity with other service knowledge vectors (such as vectors related to other courses), this will affect the overall service acceptance estimate. For example, when comparing the similarity of the service knowledge vector of advanced mathematics courses with the service knowledge vector of physics courses, due to the unreasonable setting of a certain neuron weight, two vectors that should have low similarity are judged to have high similarity, which in turn affects the accuracy of the enhanced service knowledge vector obtained by the subsequent weighted summation that integrates global context information. Based on the network training error feedback, the weights of these neurons related to advanced mathematics courses are optimized and adjusted to improve the calculation accuracy in this regard.

[0081] In the estimation network, if the multi-dimensional fusion output unit of the first network functional unit improperly sets the neuron weights of certain key dimensions (such as the dimension of the impact of course difficulty on service acceptance) when fusing multi-dimensional information to generate the first service acceptance estimation portrait, it will lead to a large deviation from the actual service acceptance description portrait (such as the proportion of people who actually registered to study advanced mathematics courses, the actual learning progress of students in the course, and other multi-dimensional accurate information). For example, in the multi-dimensional fusion output unit, the dimension of course difficulty originally had a greater impact on service acceptance, but because the neuron weight was set too low, the impact of this factor was not accurately reflected in the generated first service acceptance estimation portrait. For example, in actual situations, the high difficulty of the course leads to a low proportion of registered people, but this trend is not reflected in the estimated portrait. Based on the network training error, the neuron weight information of the advanced mathematics course push node in the multi-dimensional fusion output unit used to estimate the service push training examples is optimized.

[0082] Similarly, for the second network functional unit (such as the sequence response unit or the feature interaction relationship unit), if there is an error when sequentially processing the target weighted knowledge vector to generate the second or third service acceptance estimation portrait, for example, when processing the target weighted knowledge vector related to the advanced mathematics course in the sequence response unit according to the chronological order of course learning, some neuron weights are set inaccurately, resulting in a discrepancy between the actual situation and the impact of different learning stages on service acceptance. For example, in the early promotion stage of the course, due to insufficient course guidance information, the number of registrations was small, but this was not reflected in the estimated portrait. These neuron weights are adjusted based on the network training error.

[0083] When the campus service resource label in the service push training example is not the target campus service resource label (for example, a resource label related to an English course, while the target is an advanced mathematics course push node), the weight information of at least one neuron in the service acceptance profile estimation network is optimized based on the network training error. Because the entire service acceptance profile estimation network is a complex whole, each part is interrelated. Although the campus service resource label is not related to the target service push node, some parts of the network may indirectly affect the estimation of the target service push node.

[0084] For example, in an attention network, some general neurons are responsible for calculating basic relationships between different courses. Improper weighting of these neurons can affect overall network performance and, in turn, the estimation of the node that pushes advanced mathematics courses. Even if the campus service resource labels of the training examples currently being pushed by the service are unrelated to advanced mathematics courses, improper weighting of these neurons can lead to deviations in the calculation of the global similarity matrix and other steps, thereby affecting the entire network's estimation of the service acceptance of the node that pushes advanced mathematics courses. Based on the network training error, these neuron weights that may affect overall network performance are optimized.

[0085] In the estimation network, if the multi-dimensional fusion output unit of the first network functional unit has problems with setting the neuron weights when processing training samples pushed by other services, it may affect the estimation of the advanced mathematics course push node through the internal association mechanism of the network. For example, when the multi-dimensional fusion output unit processes service push training samples related to English courses, due to the unreasonable setting of a certain neuron weight, an error occurs when fusing multi-dimensional information. This error may affect the estimation of the advanced mathematics course push node through the information transmission and interaction mechanism within the network. Based on the network training error, the neuron weight information in the multi-dimensional fusion output unit that may affect the estimation of the advanced mathematics course push node is optimized.

[0086] Similarly, for the second network functional unit (such as the sequence response unit or the feature interaction relationship unit), the error caused by improper setting of neuron weights when processing other service push training samples may affect the estimation of the push node of the advanced mathematics course through the association between networks. For example, when the sequence response unit processes the target weighted knowledge vector related to the English course in sequence, there is an error in setting the neuron weights, which leads to a deviation in analyzing the impact of the learning sequence on service acceptance. This deviation may indirectly affect the estimation of the push node of the advanced mathematics course through the internal association of the network. Based on the network training error, the information of these potentially affected neuron weights is optimized, thereby improving the accuracy of the entire service acceptance portrait estimation network in estimating the push node of the advanced mathematics course.

[0087] In a possible implementation, step S150 includes:

[0088] Step S151: Obtain a service usage trajectory vector of a target push object and a plurality of campus service resource semantic vectors of campus service resource data to be pushed.

[0089] Step S152: Generate a service acceptance estimation portrait of each service push node corresponding to the campus service resource data to be pushed through the service acceptance portrait estimation network based on the service usage trajectory vector and the campus service resource semantic vector of each campus service resource data to be pushed.

[0090] Step S153, based on the service acceptance estimation portrait of each campus service resource data to be pushed corresponding to each service push node, determine the target campus service resource data and the service push node corresponding to the target campus service resource data from the multiple campus service resource data to be pushed.

[0091] Step S154: Based on the target campus service resource data and the service push node corresponding to the target campus service resource data, the smart campus service is pushed to the target push object.

[0092] In a possible implementation, step S151 includes:

[0093] Step S1511: Acquire a service push task, wherein the service push task is marked with object index information of a target push object and service requirement knowledge points.

[0094] Step S1512: Acquire a service usage trajectory vector of the target push object based on the service push task.

[0095] Step S1513: acquiring a plurality of campus service resource data to be pushed based on the service demand knowledge point, and extracting a campus service resource semantic vector of each campus service resource data to be pushed.

[0096] In a possible implementation, step S153 includes: based on the service acceptance estimation portraits corresponding to the service push nodes of each campus service resource data to be pushed, extracting the campus service resource data to be pushed corresponding to the service acceptance estimation portraits whose probability values ​​meet the set conditions as the target campus service resource data, and using the service push nodes corresponding to the service acceptance estimation portraits whose probability values ​​meet the set conditions in the target campus service resource data as the service push nodes corresponding to the target campus service resource data.

[0097] In this embodiment, when a service push task is acquired, it is annotated with the object index information and service requirement knowledge points of the target push object. For example, if the target push object in a service push task is a group of students in a specific grade of a certain major, the object index information clearly identifies the characteristics of this student group, such as the major code and grade number. The service requirement knowledge points are knowledge needs related to students' academic development or campus life, such as the need for in-depth study of advanced mathematics courses or knowledge needs for organizing campus cultural activities.

[0098] Taking online course learning services as an example, the service usage trajectory vector of the target push recipient contains information about the various course-related operations of the student group on the smart campus platform. For example, information such as the student group's login time on different online course platforms, the length of time spent on the advanced mathematics course page, the frequency of interaction in course-related discussion forums, and the time and quality of homework submissions. This information comprehensively reflects the target push recipient's past behavior patterns and level of participation in online course learning.

[0099] At the same time, if the service demand knowledge point is in-depth learning of advanced mathematics courses, then the campus service resource data to be pushed may include electronic materials supporting different versions of advanced mathematics textbooks, advanced mathematics course videos recorded by different teachers, online tutoring courses, etc. For each piece of campus service resource data to be pushed, its campus service resource semantic vector is extracted. Taking advanced mathematics course videos as an example, by recognizing the voice content in the video and converting it into text, and then analyzing the semantic information in these texts, concepts closely related to the core content of the course, such as limits, derivatives, and integrals, are extracted as keywords. These keywords are then converted into vector form. The dimension of this vector is related to the knowledge system of the entire smart campus. Each element in the vector represents the characteristic value of this keyword at different knowledge dimensions, thereby obtaining the campus service resource semantic vector of the course video. For other campus service resource data to be pushed, such as electronic textbooks and online tutoring courses, their campus service resource semantic vectors are also extracted in a similar manner.

[0100] Next, taking the campus service resource data to be pushed related to advanced mathematics courses as an example, the attention network in the service acceptance profile estimation network first captures the relationship between the service usage trajectory vector and the campus service resource semantic vector. For example, when calculating the global similarity matrix, the historical learning behavior of the target push object in the advanced mathematics course (part of the information in the service usage trajectory vector) is compared with the campus service resource semantic vector of the advanced mathematics course video to be pushed, and the similarity between them is determined, reflecting the importance of the course video to the target push object in the global context. In this way, each campus service resource data to be pushed is processed to obtain an enhanced service knowledge vector that integrates global context information.

[0101] Then, for example, for an enhanced service knowledge vector for a high-level mathematics course video, it is segmented according to dimensions such as course content difficulty, learning time requirements, and expected learning outcomes. A local attention mechanism is used to determine the local focus of each sub-vector. The final service knowledge vector is then obtained through multi-level attention fusion (e.g., semantic attention to the mathematical concept relationships within the course content, behavioral attention to the learning habits of the target audience, and resource attention to how the course resources are utilized). The attention coefficient is adjusted according to pre-defined dynamic adjustment rules (e.g., time decay-based adjustment rules: if the target audience last studied high-level mathematics a long time ago, the attention coefficient of the relevant service knowledge vector will be reduced; user preference-based adjustment rules: if the target audience's previous learning behavior shows no interest in mathematics, the attention coefficient of the relevant service knowledge vector will also be reduced; service popularity-based adjustment rules: if the current high-level mathematics course is very popular on campus, the attention coefficient of the relevant service knowledge vector will be increased). This results in a normalized attention coefficient.

[0102] Next, the first network functional unit (multi-dimensional fusion output unit), the second network functional unit (e.g., the sequence response unit), and the third network functional unit (e.g., the feature interaction unit) in the estimation network estimate service acceptance profiles based on multiple target weighted knowledge vectors corresponding to the service push training examples, generating service acceptance estimation profiles for each service push node (e.g., push nodes at different teaching stages of a higher mathematics course) corresponding to the campus service resource data to be pushed. The multi-dimensional fusion output unit comprehensively considers the multi-dimensional information in the target weighted knowledge vectors, such as the target push recipient's learning behavior in different courses, course resource characteristics, and the importance weights of this information. For different push nodes in the higher mathematics course (e.g., the course pre-study stage, the course mid-study stage, the course review stage, etc.), it fuses the information in these vectors and outputs service acceptance estimation profiles for these nodes. The sequence response unit processes the target weighted knowledge vectors in a specific order (e.g., by course learning time or by increasing course difficulty). For different push nodes in the higher mathematics course, it considers the target push recipient's response to the course push at different learning stages or different course difficulty levels to generate service acceptance estimation profiles for these nodes. The feature interaction relationship unit focuses on the impact of the interaction between different features on service acceptance. It analyzes how the interaction between the learning foundation of the target push object (such as the grades of previous related courses) and course resources (such as course difficulty, richness of course materials, etc.) affects the target push object's acceptance of different push nodes of advanced mathematics courses, thereby generating a corresponding service acceptance estimation portrait.

[0103] Finally, for the campus service resource data to be pushed for advanced mathematics courses, including different versions of electronic textbooks, course videos recorded by multiple teachers, and online tutoring courses, there are corresponding service acceptance estimation profiles at different service push nodes (such as the course preview phase, mid-course learning phase, and course review phase). Assuming that the registration probability value in the service acceptance estimation profile during the course preview phase is the highest and the learning engagement prediction value is good, then the campus service resource data to be pushed that meets this set condition (such as a course preview video recorded by a teacher) is determined as the target campus service resource data, and the service push node corresponding to this set condition (such as the course preview phase) in this target campus service resource data is determined as the service push node corresponding to this target campus service resource data.

[0104] In the above example, based on the determined target campus service resource data (a course preview video recorded by a teacher) and the corresponding service push node (course preview stage), the smart campus service push system will push the course preview video recorded by this teacher to the target push object (a group of students in a specific grade of a certain major) when they are in the course preview stage. Such push decisions are made based on the analysis of the service usage trajectory vector of the target push object, the campus service resource semantic vector of the campus service resource data to be pushed, and the service acceptance estimation profile generated by the service acceptance profile estimation network, thereby achieving the precise push of smart campus services, improving the effectiveness and pertinence of service push, meeting the service demand knowledge points of the target push object, and helping to improve the learning experience and learning effect of the target push object in the smart campus.

[0105] Figure 2 The hardware structure of the artificial intelligence-based smart campus service push system 100 for implementing the above-mentioned artificial intelligence-based smart campus service push method provided in an embodiment of the present invention is shown. Figure 2 As shown, the artificial intelligence-based smart campus service push system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130 and a communication unit 140.

[0106] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store data and / or instructions used by the artificial intelligence-based smart campus service push system 100 to execute or use to complete the exemplary method described in the present invention.

[0107] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the artificial intelligence-based smart campus service push method of the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0108] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned artificial intelligence-based smart campus service push system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0109] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based smart campus service push method is implemented.

[0110] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for pushing smart campus services based on artificial intelligence, characterized in that: The method comprises: Obtaining a service push training sample for a smart campus service, wherein the service push training sample is annotated with a service acceptance description portrait corresponding to a target service push node of the service push training sample; Performing feature extraction on the service usage trajectory data and campus service resource data in the service push training example to generate a plurality of service knowledge vectors included in the service push training example; Obtaining an attention coefficient of each service knowledge vector through an attention network, and performing a service acceptance profile estimation based on the multiple service knowledge vectors included in the service push training sample and the attention coefficient of each service knowledge vector through an estimation network, thereby generating a service acceptance estimation profile of each service push node among the multiple service push nodes corresponding to the service push training sample; Calculate the network training error based on the service acceptance description portrait of the target service push node corresponding to the service push training sample and the service acceptance estimation portrait of each service push node corresponding to the service push training sample; Optimizing the neuron weight information of the service acceptance portrait estimation network used to estimate the service push training sample corresponding to the target service push node based on the network training error until the network convergence requirement is met, and making a smart campus service push decision for the target push object based on the service acceptance portrait estimation network, wherein the service acceptance portrait estimation network includes the attention network and the estimation network; The estimation network includes a first network function unit and a second network function unit. The step of extracting features from the service usage trajectory data and campus service resource data in the service push training example to generate a plurality of service knowledge vectors included in the service push training example includes: Cleaning, denoising, and normalizing the service usage trajectory data to extract key behavioral features. Simultaneously, semantic analysis, keyword extraction, and vectorization are performed on the campus service resource data to generate corresponding resource feature vectors. The key behavioral features and resource feature vectors are then fused to generate service knowledge vectors containing service information and resource semantics. The step of obtaining the attention coefficient of each service knowledge vector through the attention network includes: The global attention mechanism of the attention network is used to capture the global dependencies between the various service knowledge vectors. Specifically, the similarity between each service knowledge vector and all other service knowledge vectors is calculated to obtain a global similarity matrix. The global similarity matrix reflects the importance of each service knowledge vector in the global context. The global similarity matrix is ​​then used to perform a weighted summation of the service knowledge vectors to obtain an enhanced service knowledge vector that integrates the global context information. Segmenting each enhanced service knowledge vector to generate multiple sub-vectors, and applying a local attention function to each sub-vector based on the local attention mechanism of the attention network. The local attention function identifies a local focus by calculating the correlation between internal elements of the sub-vectors, and merges the local focus of each sub-vector to generate a local focus vector for the entire service knowledge vector; The multi-level attention fusion strategy of the attention network is adopted to define multiple levels of attention heads, each of which is used to capture information at the corresponding level; Applying the attention head to each local focus vector simultaneously to obtain multiple levels of attention representation, and fusing the multiple levels of attention representation through weighted summation to generate a final service knowledge vector that incorporates multiple levels of attention information, including semantic level, behavioral level, and resource level; Applying a predefined dynamic adjustment rule to the attention coefficient of each final service knowledge vector to obtain a dynamically adjusted attention coefficient, wherein the dynamic adjustment rule includes an adjustment rule based on time decay, an adjustment rule based on user preference, and an adjustment rule based on service popularity; Calculating the sum or maximum value of all dynamically adjusted attention coefficients, and dividing each attention coefficient by the sum or maximum value to obtain a normalized attention coefficient; The step of estimating a service acceptance profile based on a plurality of service knowledge vectors included in the service push training example and an attention coefficient of each of the service knowledge vectors through an estimation network to generate an estimated service acceptance profile of each of the plurality of service push nodes corresponding to the service push training example includes: For each service knowledge vector, perform weighted calculation on the service knowledge vector and the attention coefficient corresponding to the service knowledge vector to generate a target weighted knowledge vector corresponding to each service knowledge vector; Estimating a service acceptance profile based on a plurality of target weighted knowledge vectors corresponding to the service push training sample by the first network function unit to generate a first service acceptance estimation profile corresponding to each service push node of the service push training sample; Estimating a service acceptance profile based on a plurality of target weighted knowledge vectors corresponding to the service push training sample by the second network function unit, and generating a second service acceptance estimation profile corresponding to each service push node of the service push training sample; The step of calculating the network training error based on the service acceptance description portrait corresponding to the target service push node of the service push training sample and the service acceptance estimation portrait corresponding to each service push node of the service push training sample comprises: Calculating a network training error based on a service acceptance description portrait of the service push node corresponding to the service push training sample, a first service acceptance estimation portrait corresponding to each service push node of the service push training sample, and a second service acceptance estimation portrait corresponding to each service push node of the service push training sample; The step of optimizing the neuron weight information of the service push training example corresponding to the target service push node in the service acceptance portrait estimation network based on the network training error includes: Based on the network training error, the neuron weight information of the attention network is optimized, the neuron weight information in the first network function unit used to estimate the service push training sample corresponding to the target service push node is optimized, and the neuron weight information of the second network function unit is optimized.

2. The method for pushing smart campus services based on artificial intelligence according to claim 1, characterized in that: The step of optimizing the neuron weight information of the service push training example corresponding to the target service push node in the service acceptance portrait estimation network based on the network training error includes: Masking and marking target neuron weight information in the service acceptance profile estimation network, wherein the target neuron weight information is other than neuron weight information used in a fully connected mapping unit of the estimation network to estimate the neuron weight corresponding to the target service push node corresponding to the service push training example; Based on the network training error, the unshielded and labeled neuron weight information in the service acceptance portrait estimation network is optimized.

3. The method for pushing smart campus services based on artificial intelligence according to claim 1, characterized in that: The first network function unit is a multi-dimensional fusion output unit, and the second network function unit is a sequence response unit or a feature interaction relationship unit.

4. The method for pushing smart campus services based on artificial intelligence according to claim 1, characterized in that: The estimation network further includes a third network function unit, and the step of performing, through the estimation network, service acceptance profile estimation based on a plurality of service knowledge vectors included in the service push training example and an attention coefficient of each of the service knowledge vectors, and generating a service acceptance estimation profile of each of the plurality of service push nodes corresponding to the service push training example, further includes: performing a service acceptance profile estimation based on the third network function unit according to a plurality of target weighted knowledge vectors corresponding to the service push training sample, and generating a third service acceptance estimation profile corresponding to each service push node of the service push training sample, wherein the second network function unit is a sequence response unit and the third network function unit is a feature interaction relationship unit; The calculating of the network training error based on the service acceptance description portrait of the service push node corresponding to the service push training sample, the first service acceptance estimation portrait of each service push node corresponding to the service push training sample, and the second service acceptance estimation portrait of each service push node corresponding to the service push training sample includes: The network training error is calculated based on the service acceptance description portrait of the service push node corresponding to the service push training sample, the first service acceptance estimation portrait of each service push node corresponding to the service push training sample, the second service acceptance estimation portrait of each service push node corresponding to the service push training sample, and the third service acceptance estimation portrait of each service push node corresponding to the service push training sample.

5. The method for pushing smart campus services based on artificial intelligence according to any one of claims 1 to 4, characterized in that: The step of optimizing the neuron weight information of the service push training example corresponding to the target service push node in the service acceptance portrait estimation network based on the network training error includes: If the campus service resource label in the service push training example is the target campus service resource label, optimizing the neuron weight information in the service acceptance portrait estimation network for estimating that the service push node of the service push training example is the target service push node based on the network training error; If the campus service resource label in the service push training example is not the target campus service resource label, at least one neuron weight information of the service acceptance portrait estimation network is optimized based on the network training error.

6. The method for pushing smart campus services based on artificial intelligence according to any one of claims 1 to 4, characterized in that: The step of making a smart campus service push decision for the target push object based on the service acceptance profile estimation network includes: Obtaining a service usage trajectory vector of a target push object and a plurality of campus service resource semantic vectors of campus service resource data to be pushed; Generate, by the service acceptance portrait estimation network, a service acceptance estimation portrait of each service push node corresponding to the campus service resource data to be pushed based on the service usage trajectory vector and the campus service resource semantic vector of each campus service resource data to be pushed; Determining target campus service resource data and the service push node corresponding to the target campus service resource data from the plurality of campus service resource data to be pushed based on the service acceptance estimation portrait corresponding to each service push node of each campus service resource data to be pushed; Based on the target campus service resource data and the service push node corresponding to the target campus service resource data, smart campus service is pushed to the target push object.

7. The method for pushing smart campus services based on artificial intelligence according to claim 6, characterized in that: The step of obtaining the service usage trajectory vector of the target push object and the campus service resource semantic vectors of the plurality of campus service resource data to be pushed includes: Obtain a service push task, wherein the service push task is marked with object index information of a target push object and service requirement knowledge points; Acquire a service usage trajectory vector of the target push object based on the service push task; Based on the service demand knowledge point, a plurality of campus service resource data to be pushed are acquired, and a campus service resource semantic vector of each campus service resource data to be pushed is extracted.

8. The method for pushing smart campus services based on artificial intelligence according to claim 6, characterized in that: The step of determining target campus service resource data and the service push node corresponding to the target campus service resource data from the plurality of campus service resource data to be pushed based on the service acceptance estimation portrait corresponding to each service push node of each campus service resource data to be pushed includes: Based on the service acceptance estimation portraits corresponding to the service push nodes of each campus service resource data to be pushed, the campus service resource data to be pushed corresponding to the service acceptance estimation portrait whose probability value meets the set conditions is extracted as the target campus service resource data, and the service push node corresponding to the service acceptance estimation portrait whose probability value meets the set conditions in the target campus service resource data is used as the service push node corresponding to the target campus service resource data.

9. An artificial intelligence-based smart campus service push system, characterized in that: The artificial intelligence-based smart campus service push system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based smart campus service push method described in any one of claims 1 to 8 above.

Images