An online course learning intelligent system based on mobile internet

Abstract

This application discloses an intelligent online course learning system based on mobile internet, relating to the fields of artificial intelligence and education technology. The system collects multi-dimensional learning behavior data, constructs a data matrix, and uses an optimized deep learning model to extract user features. Combined with course features, it achieves accurate recommendations. In particular, it introduces a hybrid algorithm that combines scattering optimization, fitness enhancement neighborhood optimization, and restricted information mutation optimization to optimize the parameters of the deep learning model. This effectively solves the problems of traditional algorithms being prone to getting stuck in local optima and having slow convergence speed, significantly improving the accuracy of user profile construction and the matching degree of course recommendations, and realizing an intelligent and personalized online learning experience.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and education technology, and more specifically, to an intelligent system for online course learning based on the mobile Internet. Background Technology

[0002] With the rapid development of mobile internet technology, online education has become an important way for people to acquire knowledge. Traditional online education platforms often adopt a "one-size-fits-all" course display model, or recommend courses based solely on simple click-through rates and popularity, failing to meet users' personalized and differentiated learning needs.

[0003] Currently, while some technical solutions attempt to utilize collaborative filtering or basic machine learning algorithms for course recommendation, they suffer from the following significant drawbacks: Limited data collection dimensions: Existing technologies often focus only on the final learning outcome (e.g., scores), neglecting behavioral characteristics during the learning process (e.g., learning time distribution, interaction frequency, viewing duration), leading to inaccurate user profile construction. Insufficient feature extraction capabilities: Traditional shallow models struggle to uncover the deep nonlinear features hidden within user behavior data, failing to accurately represent user learning preferences. Limitations of model optimization algorithms: When constructing deep learning models, traditional gradient descent or conventional genetic algorithms and particle swarm optimization are often used for parameter optimization. These algorithms are prone to getting trapped in local optima and have slow convergence speeds, limiting the model's prediction accuracy and generalization ability. Especially for high-dimensional, complex user behavior data, conventional optimization algorithms struggle to find globally optimal model parameters within a reasonable timeframe, thus affecting the accuracy of the recommendation system. Summary of the Invention

[0004] This application aims to provide an intelligent online course learning system based on mobile internet, which seeks to solve the problem of low recommendation accuracy in existing technologies.

[0005] An intelligent online course learning system based on mobile internet includes:

[0006] The user data collection module is used to collect users' learning behavior data through mobile internet terminals. The learning behavior data includes course viewing time, answer accuracy rate, learning frequency, interaction behavior records, and learning time distribution.

[0007] The user profile building module is used to construct multi-dimensional user feature vectors based on the learning behavior data using a deep learning model.

[0008] The course feature extraction module is used to extract multimodal features from courses in the course library and generate course feature vectors corresponding to the courses.

[0009] The deep learning recommendation module is used to input the user feature vector and the course feature vector into a deep neural network model, calculate the matching degree between the user and the course, generate a personalized course recommendation list, and push the personalized course recommendation list to the user to achieve intelligent learning.

[0010] Furthermore, user learning behavior data is collected through mobile internet terminals, including:

[0011] Based on a preset data collection cycle, after each data collection cycle, the user's course viewing time, answer accuracy, learning frequency, interaction behavior records, and learning time distribution are collected through mobile Internet terminals to obtain single-cycle feature data.

[0012] A data matrix is ​​constructed from multiple consecutive periodic feature data to obtain the user's learning behavior data; each row in the data matrix is ​​a periodic feature data, and if a feature in the periodic feature data is empty, it is filled with zeros.

[0013] Furthermore, based on the aforementioned learning behavior data, a multi-dimensional user feature vector is constructed using a deep learning model, including:

[0014] The learning behavior data is normalized to obtain normalized learning behavior data.

[0015] The normalized learning behavior data is used as input to a deep learning model to obtain the features output by the deep learning model, resulting in a multi-dimensional user feature vector.

[0016] Furthermore, the deep learning model also includes the following before use:

[0017] Deep learning models are constructed using CNN-LSTM or CNN models;

[0018] Multiple historical learning behavior data are collected, and the historical learning behavior data are preprocessed using the K-means clustering algorithm to obtain deep optimization data; the deep optimization data includes deep optimization samples and deep optimization labels corresponding to the deep optimization samples;

[0019] An optimization population is initialized based on the parameters of the deep learning model to be optimized; each individual in the optimization population includes all the parameters of the model to be optimized.

[0020] The fitness of each individual in the optimized population is obtained using deep optimization data, and the globally optimal individual is determined based on the fitness.

[0021] A scattering optimization algorithm is used to perform full neighborhood development on each individual in the optimized population to obtain the first target individual;

[0022] Based on the globally optimal individual, the fitness-enhanced neighborhood optimization algorithm is used to develop the inter-individual region of the first target individual to obtain the second target individual;

[0023] A restricted information mutation optimization algorithm is used to perform global adaptive development on the second target individual to obtain the third target individual;

[0024] Before the optimization termination condition is met, the third target individual is used as the input for the next optimization, and the next optimization is performed.

[0025] When the optimization termination condition is met, the global optimal individual is re-determined based on the third target individual, and the parameters in the global optimal individual are used as the target parameters of the deep learning model to achieve optimization of the deep learning model.

[0026] Furthermore, a scattering optimization algorithm is used to perform full neighborhood development on each individual in the optimized population to obtain the first target individual, including:

[0027] Based on the fitness of each individual in the optimized population, the degree of optimization of the solution is determined as follows:

[0028] ;

[0029] ;

[0030] In the formula, Let i be the i-th individual, where i = 1, 2, ..., L, and L is the total number of individuals. Let j be the j-th individual, j = 1, 2, ..., L, and t be the number of optimization iterations. The optimization level of the solution corresponding to the i-th individual. For the intermediate parameter corresponding to the i-th individual, Let be the fitness of the i-th individual. Let be the fitness of the j-th individual, max be the function to maximize the fitness, and min be the function to minimize the fitness. To train the population; For positive integer terms less than or equal to 0.01;

[0031] Based on the degree of solution optimization, the adaptive neighborhood development range is obtained as follows:

[0032] ;

[0033] ;

[0034] In the formula, Let be the d-th dimension parameter of the i-th individual, where d = 1, 2, ..., K, and K is the total number of parameters to be trained. Let d be the parameter of the j-th individual. Let d be the adaptive neighborhood development range corresponding to the d-th dimension parameter of the i-th individual. The first random number between (0,1) The neighborhood influence coefficient. This is the base value for the neighborhood influence coefficient. The degree of optimization of the solution corresponding to the j-th individual. Let be the Euclidean distance between the i-th individual and the i-th individual, and T be the preset maximum number of optimization attempts;

[0035] Based on the adaptive neighborhood development range, the neighborhood information corresponding to an individual is obtained as follows:

[0036] ;

[0037] In the formula, Let d be the d-th dimension parameter of the neighborhood information corresponding to the individual. for Random numbers between;

[0038] The number of scattering points obtained is:

[0039] ;

[0040] In the formula, Let i be the number of scattering ... This is the base value for the number of scattering events;

[0041] Repeatedly acquire neighborhood information corresponding to an individual until the number of acquired neighborhood information reaches the number of scattering events, thus obtaining... The neighborhood information is used to determine the neighborhood information with the highest fitness as the first target individual.

[0042] Furthermore, based on the globally optimal individual, the fitness-enhancing neighborhood optimization algorithm is used to develop inter-individual regions for the first target individual to obtain the second target individual, including:

[0043] Based on the number of optimization iterations, the adaptive inertia weight and the non-linear decreasing search factor are determined as follows:

[0044] ;

[0045] ;

[0046] In the formula, For adaptive inertia weights, This is the upper limit of the adaptive inertia weight. This is the lower limit of the adaptive inertia weight. The search factor is a non-linear decreasing factor. The first adjustment coefficient, This is the second adjustment coefficient;

[0047] Obtain the fitness value of the first target individual and sort them in descending order of fitness value to obtain the sorted first target individual;

[0048] For any first target individual, match one other first target individual to the first half of the first target individuals after the arrangement, and match another other first target individual to the second half of the first target individuals, to obtain the first vertex individual and the second vertex individual respectively.

[0049] The first enhancement development factor is obtained based on the fitness of the first vertex individual, and the second enhancement development factor is obtained based on the fitness of the second vertex individual as follows:

[0050] ;

[0051] ;

[0052] In the formula, As the first enhancement development factor, As the second enhancing development factor, It is an exponential function with the natural constant e as its base. For the m-th first target individual, This refers to the first vertex individual corresponding to the m-th first target individual. This refers to the second vertex individual corresponding to the m-th first target individual. Let be the fitness of the m-th individual with the first target. The fitness of the first vertex individual. The fitness of the second-vertex individual;

[0053] Based on the adaptive inertia weight, nonlinear decreasing search factor, first enhancement development factor, and second enhancement development factor, inter-individual region development is performed on the first target individual to obtain the second target individual:

[0054] ;

[0055] In the formula, Let d be the d-th dimension parameter of the m-th first target individual. Let d be the d-th dimension parameter of the globally optimal individual. Let d be the d-th dimension parameter of the m-th second target individual. The second random number between (0,1) The third random number between (0,1) Let d be the d-th dimension parameter of the first vertex individual. Let d be the d-th dimension parameter of the second vertex individual, and sin be the sine function.

[0056] Furthermore, a restricted information mutation optimization algorithm is used to perform global adaptive development on the second target individual to obtain the third target individual, including:

[0057] Based on the upper and lower limit values ​​of each parameter in the second target individual, the first mutation parameter and the second mutation parameter are determined as follows:

[0058] ;

[0059] ;

[0060] In the formula, Let d be the d-th dimension parameter of the nth second target individual. for The corresponding first mutation parameter, for The corresponding second mutation parameter, This provides information on the lower limit value of the d-th dimension parameter. This provides information on the upper limit of the d-th dimension parameter. The number of mutations;

[0061] Based on the first mutation parameter and the second mutation parameter, the second target individual is subjected to global adaptive development to obtain the third target individual:

[0062] ;

[0063] In the formula, Let d be the d-th dimension parameter of the nth third target individual. It is a cosine function. Random or- k is the independent variable. A newly generated individual.

[0064] Furthermore, multimodal feature extraction is performed on the courses in the course database to generate course feature vectors, including:

[0065] Obtain the knowledge point tags, difficulty level, course type and / or prerequisite relationships corresponding to the course to obtain multimodal features;

[0066] The multimodal features are encoded using one-hot encoding to obtain the course feature vector corresponding to the course.

[0067] Furthermore, the user feature vector and the course feature vector are input into a deep neural network model to calculate the matching degree between the user and the course, generating a personalized course recommendation list, including:

[0068] The user feature vector and the course feature vector are weighted and fused or concatenated to obtain a fused feature vector;

[0069] The fused feature vector is used as the input to the deep neural network model to obtain the matching degree between the user and the course output by the deep neural network model.

[0070] Based on the matching degree between the user and the course, a TOP-N recommendation is generated to obtain a personalized course recommendation list.

[0071] Furthermore, it also includes:

[0072] The learning path planning module is used to generate the optimal learning path sequence based on the knowledge graph and the personalized course recommendation list, using a reinforcement learning algorithm.

[0073] The real-time feedback and adjustment module is used to monitor the user's real-time learning status and dynamically adjust the recommendation strategy and learning path.

[0074] This application provides an intelligent online course learning system based on mobile internet. By collecting multi-dimensional learning behavior data, a data matrix is ​​constructed, and an optimized deep learning model is used to extract user features. Combined with course features, accurate recommendations are achieved. In particular, a hybrid algorithm combining scattering optimization, fitness enhancement neighborhood optimization, and restricted information mutation optimization is introduced to optimize the parameters of the deep learning model. This effectively solves the problems of traditional algorithms being prone to getting stuck in local optima and having slow convergence speed, significantly improving the accuracy of user profile construction and the matching degree of course recommendations, and realizing an intelligent and personalized online learning experience. Attached Figure Description

[0075] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a schematic diagram of the structure of an intelligent online course learning system based on mobile Internet proposed in one embodiment of this application;

[0077] Among them, 101-User Data Collection Module, 102-User Profile Construction Module, 103-Course Feature Extraction Module, 104-Deep Learning Recommendation Module, 105-Learning Path Planning Module, and 106-Real-time Feedback Adjustment Module. Detailed Implementation

[0078] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0079] like Figure 1 As shown in the figure, this application provides an intelligent online course learning system based on mobile Internet, including:

[0080] The user data collection module 101 is used to collect users' learning behavior data through mobile Internet terminals. The learning behavior data includes course viewing time, answer accuracy rate, learning frequency, interaction behavior records, and learning time distribution.

[0081] User profile building module 102 is used to build a multi-dimensional user feature vector based on the learning behavior data using a deep learning model;

[0082] The course feature extraction module 103 is used to extract multimodal features from courses in the course database and generate course feature vectors corresponding to the courses.

[0083] The deep learning recommendation module 104 is used to input the user feature vector and the course feature vector into a deep neural network model, calculate the matching degree between the user and the course, generate a personalized course recommendation list, and push the personalized course recommendation list to the user to realize intelligent learning.

[0084] In this embodiment of the application, collecting user learning behavior data through a mobile internet terminal includes:

[0085] Based on a preset data collection cycle, after each data collection cycle, the user's course viewing time, answer accuracy, learning frequency, interaction behavior records, and learning time distribution are collected through mobile Internet terminals to obtain single-cycle feature data.

[0086] The user's learning behavior data is obtained by constructing a data matrix from multiple consecutive periodic feature data (the number of periods in each data matrix construction is the same); each row in the data matrix is ​​a periodic feature data, and if a feature in the periodic feature data is empty, it is filled with zeros.

[0087] Course viewing time reflects learning engagement; quiz accuracy reflects knowledge mastery; learning frequency reflects learning habits; interactive behavior records, including questions, notes, and discussions, reflect learning activity; and learning time distribution reflects optimal learning periods. Data from multiple consecutive periods is constructed into a matrix, with zeros filled in the corresponding row if a user did not study in a given period. This matrix format preserves time-series characteristics, facilitating subsequent processing by models such as LSTM.

[0088] This application collects multi-dimensional data from mobile internet terminals, including course viewing time, quiz accuracy, learning frequency, interaction records, and learning time distribution, breaking through the limitations of traditional technologies that only focus on a single result. In particular, the use of a "preset data collection cycle" to construct a data matrix introduces the time dimension into the data collection process, enabling a complete record of the dynamic changes in users' learning habits. The zero-filling method for missing features preserves the matrix structure of the data for processing by deep learning models while avoiding calculation errors caused by missing values. This end-to-end, periodic data collection method provides a solid data foundation for building high-precision user profiles, allowing the system to capture subtle fluctuations in users' learning status and generate recommendations that are more relevant to the user's current learning stage.

[0089] In this embodiment of the application, based on the learning behavior data, a multi-dimensional user feature vector is constructed using a deep learning model, including:

[0090] The learning behavior data is normalized to obtain normalized learning behavior data.

[0091] The normalized learning behavior data is used as input to a deep learning model to obtain the features output by the deep learning model, resulting in a multi-dimensional user feature vector.

[0092] In this embodiment of the application, the deep learning model further includes the following before use:

[0093] Deep learning models are constructed using CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) or CNN models;

[0094] Multiple historical learning behavior data are collected, and the historical learning behavior data are preprocessed using the K-means clustering algorithm to obtain deep optimization data; the deep optimization data includes deep optimization samples and deep optimization labels corresponding to the deep optimization samples;

[0095] For example, historical learning behavior data can be clustered using K-means to obtain multiple data clusters. Each data cluster can be assigned a label to obtain a deep-optimized label for each historical learning behavior data. After normalizing the historical learning behavior data, it can be used as a deep-optimized sample.

[0096] An optimization population is initialized based on the parameters of the deep learning model to be optimized; each individual in the optimization population includes all the parameters of the model to be optimized.

[0097] For example, the upper and lower limits of the parameters of the model to be optimized can be randomized and encoded into a vector to obtain an individual. Multiple different individuals can be obtained repeatedly to form an optimization population. The parameters of the model to be optimized can be all the parameters of the deep learning model or some of the parameters (such as connection weights and thresholds).

[0098] The fitness of each individual in the optimized population is obtained using deep optimization data, and the globally optimal individual is determined based on the fitness.

[0099] For example, you can use deeply optimized samples as input and their corresponding deeply optimized labels as the desired output to obtain the cross-entropy loss function value or the root mean square loss function value. Then, take the reciprocal of the loss function value to obtain the fitness. You can add a positive constant less than 0.01 to the loss function value in the denominator to avoid the denominator being zero.

[0100] A scattering optimization algorithm is used to perform full neighborhood development on each individual in the optimized population to obtain the first target individual;

[0101] Based on the globally optimal individual, the fitness-enhanced neighborhood optimization algorithm is used to develop the inter-individual region of the first target individual to obtain the second target individual;

[0102] A restricted information mutation optimization algorithm is used to perform global adaptive development on the second target individual to obtain the third target individual;

[0103] Before the optimization termination condition is met (such as before the preset maximum number of optimizations has been reached), the third target individual is used as the input for the next optimization, and the next optimization is performed.

[0104] When the optimization termination condition is met, the global optimal individual is re-determined based on the third target individual, and the parameters in the global optimal individual are used as the target parameters of the deep learning model to achieve optimization of the deep learning model.

[0105] In this embodiment of the application, a scattering optimization algorithm is used to perform full neighborhood development on each individual in the optimized population to obtain a first target individual, including:

[0106] Based on the fitness of each individual in the optimized population, the degree of optimization of the solution is determined as follows:

[0107] ;

[0108] ;

[0109] In the formula, Let i be the i-th individual, where i = 1, 2, ..., L, and L is the total number of individuals. Let j be the j-th individual, j = 1, 2, ..., L, and t be the number of optimization iterations. The optimization level of the solution corresponding to the i-th individual. For the intermediate parameter corresponding to the i-th individual, Let be the fitness of the i-th individual. Let be the fitness of the j-th individual, max be the function to maximize the fitness, and min be the function to minimize the fitness. To train the population; For positive integer terms less than or equal to 0.01;

[0110] Based on the degree of solution optimization, the adaptive neighborhood development range is obtained as follows:

[0111] ;

[0112] ;

[0113] In the formula, Let be the d-th dimension parameter of the i-th individual, where d = 1, 2, ..., K, and K is the total number of parameters to be trained. Let d be the parameter of the j-th individual. Let d be the adaptive neighborhood development range corresponding to the d-th dimension parameter of the i-th individual. The first random number between (0,1) The neighborhood influence coefficient. The baseline value for the neighborhood influence coefficient can be set to a constant between 5 and 20. The degree of optimization of the solution corresponding to the j-th individual. Let be the Euclidean distance between the i-th individual and the i-th individual, and T be the preset maximum number of optimization attempts;

[0114] Based on the adaptive neighborhood development range, the neighborhood information corresponding to an individual is obtained as follows:

[0115] ;

[0116] In the formula, Let d be the d-th dimension parameter of the neighborhood information corresponding to the individual. for Random numbers between;

[0117] The number of scattering points obtained is:

[0118] ;

[0119] In the formula, Let i be the number of scattering ... This is the base value for the number of scattering points and can be set to a constant between 50 and 100.

[0120] Repeatedly acquire neighborhood information corresponding to an individual until the number of acquired neighborhood information reaches the number of scattering events, thus obtaining... The neighborhood information is used to determine the neighborhood information with the highest fitness as the first target individual.

[0121] Traditional algorithms tend to cluster in local regions during initialization or early search phases, leading to missed global optima. This application introduces a scattering optimization algorithm, designing a mechanism to determine the degree of optimization and an adaptive neighborhood development range. By calculating the fitness difference between an individual and other individuals in the population, the search step size (i.e., the adaptive neighborhood development range) is dynamically adjusted. Individuals with lower fitness have a larger search radius, enabling them to escape current inferior regions; individuals with higher fitness perform a more refined search. Simultaneously, a scattering number mechanism is introduced to generate multiple trial solutions within the neighborhood, selecting the optimal one. This mechanism ensures population diversity in the early search phase, effectively preventing the algorithm from prematurely converging to local optima, and improving the robustness of both global and local optimization.

[0122] In this embodiment of the application, based on the globally optimal individual, a fitness-enhancing neighborhood optimization algorithm is used to develop inter-individual regions for the first target individual to obtain a second target individual, including:

[0123] Based on the number of optimization iterations, the adaptive inertia weight and the non-linear decreasing search factor are determined as follows:

[0124] ;

[0125] ;

[0126] In the formula, For adaptive inertia weights, The upper limit of the adaptive inertia weight can be set to 0.95; The lower limit of the adaptive inertia weight can be set to 0.001; The search factor is a non-linear decreasing factor. The first adjustment coefficient can be set to 1; The second adjustment factor can be set to 3;

[0127] Obtain the fitness value of the first target individual and sort them in descending order of fitness value to obtain the sorted first target individual;

[0128] For any first target individual, match one other first target individual to the first half of the first target individuals after the arrangement, and match another other first target individual to the second half of the first target individuals, to obtain the first vertex individual and the second vertex individual respectively.

[0129] The first enhancement development factor is obtained based on the fitness of the first vertex individual, and the second enhancement development factor is obtained based on the fitness of the second vertex individual as follows:

[0130] ;

[0131] ;

[0132] In the formula, As the first enhancement development factor, As the second enhancing development factor, It is an exponential function with the natural constant e as its base. For the m-th first target individual, This refers to the first vertex individual corresponding to the m-th first target individual. This refers to the second vertex individual corresponding to the m-th first target individual. Let be the fitness of the m-th individual with the first target. The fitness of the first vertex individual. The fitness of the second-vertex individual;

[0133] Based on the adaptive inertia weight, nonlinear decreasing search factor, first enhancement development factor, and second enhancement development factor, inter-individual region development is performed on the first target individual to obtain the second target individual:

[0134] ;

[0135] In the formula, Let d be the d-th dimension parameter of the m-th first target individual. Let d be the d-th dimension parameter of the globally optimal individual. Let d be the d-th dimension parameter of the m-th second target individual. The second random number between (0,1) The third random number between (0,1) Let d be the d-th dimension parameter of the first vertex individual. Let d be the d-th dimension parameter of the second vertex individual, and sin be the sine function.

[0136] After full neighborhood development, fine-grained exploration is needed in high-potential regions. This application employs adaptive inertia weights and nonlinear decreasing search factors to balance convergence speed and accuracy during the development process. More importantly, the algorithm divides individuals into two halves based on fitness ranking and matches different vertex individuals as guiding directions for each half, dividing the solution space into multiple smaller spaces for searching, thus improving the convenience and searchability of the solution space. By introducing a first and a second enhancement development factor, the guiding role of excellent individuals is amplified using an exponential function and fitness differences. This fitness-difference-based directional development strategy enables the population to quickly converge towards the globally optimal individual, significantly accelerating the convergence speed and improving the accuracy of the final solution.

[0137] In this embodiment of the application, a restricted information mutation optimization algorithm is used to perform global adaptive development on the second target individual to obtain a third target individual, including:

[0138] Based on the upper and lower limit values ​​of each parameter in the second target individual, the first mutation parameter and the second mutation parameter are determined as follows:

[0139] ;

[0140] ;

[0141] In the formula, Let d be the d-th dimension parameter of the nth second target individual. for The corresponding first mutation parameter, for The corresponding second mutation parameter, This provides information on the lower limit value of the d-th dimension parameter. This provides information on the upper limit of the d-th dimension parameter. This represents the number of mutations and can be set to 3 or 4.

[0142] Based on the first mutation parameter and the second mutation parameter, the second target individual is subjected to global adaptive development to obtain the third target individual:

[0143] ;

[0144] In the formula, Let d be the d-th dimension parameter of the nth third target individual. It is a cosine function. It is a random factor, and random is or- k is the independent variable. A newly generated individual.

[0145] To address the stagnation caused by population diversity loss in the later stages of optimization, this application introduces a constrained information mutation optimization algorithm. This algorithm dynamically generates the mutation parameter range based on the upper and lower bounds of the parameters and the number of mutations. It employs a combination of cosine functions and random perturbations to adaptively mutate individuals. This mutation is not a blind random walk but is constrained by parameter boundaries and historical information, ensuring both the effectiveness of the search and giving individuals the ability to escape current local traps. This allows the model to maintain its optimization capability even when facing complex, non-convex objective functions, ultimately achieving better model parameters.

[0146] Based on the optimized deep learning model described above, the system can extract more representative user feature vectors. Combined with the multimodal features of the course, the calculated matching degree is more accurate. In summary, this application, through the synergistic improvement of data collection, model optimization algorithms, and recommendation strategies, constructs an efficient, accurate, and intelligent online course learning system, effectively solving the technical problems of low recommendation accuracy and difficulty in model optimization in existing technologies.

[0147] In this embodiment of the application, multimodal feature extraction is performed on courses in the course library to generate course feature vectors corresponding to the courses, including:

[0148] Obtain the knowledge point tags, difficulty level, course type and / or prerequisite relationships corresponding to the course to obtain multimodal features;

[0149] The multimodal features are encoded using one-hot encoding to obtain the course feature vector corresponding to the course.

[0150] In this embodiment, the user feature vector and the course feature vector are input into a deep neural network model to calculate the matching degree between the user and the course, generating a personalized course recommendation list, including:

[0151] The user feature vector and the course feature vector are weighted and fused or concatenated to obtain a fused feature vector;

[0152] The fused feature vector is used as the input to the deep neural network model to obtain the matching degree between the user and the course output by the deep neural network model.

[0153] Based on the matching degree between the user and the course, a TOP-N recommendation is generated to obtain a personalized course recommendation list.

[0154] Optionally, to achieve the matching degree, embodiments of this application provide a method for training a deep neural network model, which may include:

[0155] For each data cluster in K-means clustering, determine the future course for the next cycle corresponding to each user's historical learning behavior data in the data cluster;

[0156] The matching degree of the most frequent future course in the data cluster is set to 1, and the ratio between other future courses and the most frequent future course is used as the matching degree of other future courses. Thus, the matching degree corresponding to each historical learning behavior data is obtained, which is also the matching degree corresponding to the historical fusion feature vector.

[0157] By taking the historical fusion feature vector corresponding to the historical learning behavior data as input and the corresponding matching degree as the expected output, the deep neural network model can be trained.

[0158] In this embodiment of the application, the system further includes:

[0159] The learning path planning module 105 is used to generate the optimal learning path sequence based on the knowledge graph and the personalized course recommendation list, using a reinforcement learning algorithm.

[0160] The real-time feedback adjustment module 106 is used to monitor the user's real-time learning status and dynamically adjust the recommendation strategy and learning path.

[0161] The system's learning path planning module and real-time feedback adjustment module, utilizing knowledge graphs and reinforcement learning techniques, not only solve the problem of which courses to recommend, but also the problem of the order in which to learn them. This represents a leap from simple recommendations to systematic learning planning, greatly improving users' learning efficiency and experience.

[0162] For example, the learning path planning module employs a Deep Q-Network (DQN) reinforcement learning algorithm. This module models the learning path planning problem as a Markov Decision Process (MDP), defined as follows:

[0163] State space S: States s represent the user's learning state, including the set of knowledge points already mastered, the current knowledge level vector, the learning history sequence, and a personalized course recommendation list. The states are represented as fixed-dimensional vectors, facilitating neural network processing. All courses can be constructed as a knowledge graph, where nodes represent knowledge points and edges represent relationships between them. Based on the user's historical accuracy rate, a certain accuracy rate is set for each knowledge point in past practice questions. When the accuracy rate exceeds a preset threshold, the knowledge point is considered mastered. The knowledge mastery vector can be based on the knowledge point nodes in the knowledge graph, using knowledge tracing algorithms (such as DKT (Deep Knowledge Tracking)) to output the mastery probability of each knowledge point. Alternatively, the historical practice question accuracy rate can be used as the mastery probability, and the learning history sequence can be a sequence of multiple knowledge points recently learned by the user.

[0164] Action Space A: Action 'a' represents the course selected for study. The action space is the set of all available courses in the course library. For large-scale course libraries, a candidate set filtering strategy is used to restrict the action space to Top-N recommended courses.

[0165] Reward function R: The design of reward r takes into account three dimensions:

[0166] ;

[0167] Where ΔK represents the increase in knowledge mastery (assessed through pre- and post-test evaluations), ΔE represents the improvement in learning efficiency (the number of knowledge points mastered per unit of time), and S represents the user satisfaction score (from user feedback). α, β, and γ are weighting coefficients, set to 0.5, 0.3, and 0.2, respectively.

[0168] Q-value function: Q(s,a) represents the long-term value estimate of choosing action a in state s. Deep Q-networks use a three-layer fully connected network (hidden layer dimensions [128,64]) to approximate the Q-value function.

[0169] The training process employs experience replay and target network techniques:

[0170] (1) Experience replay: Store the experience (s,a,r,s') of each interaction into the experience pool, and randomly sample batch experiences during training to break the data correlation.

[0171] (2) Target network: Use an independent target network to calculate the target Q value and update the target network parameters regularly to improve training stability.

[0172] The loss function is:

[0173] ;

[0174] Where θ is the policy network parameter, θ⁻ is the target network parameter, γ is the discount factor (set to 0.95), E is the expected value of the data, and r is the reward. The Q-value is the output of the policy network. The Q-value output by the target network.

[0175] Parameter determination method: The experience pool capacity is set to 100,000, the batch size is 64, the exploration rate ε decreases linearly from 1.0 to 0.01 (decay steps 100,000), and the target network is updated every 1,000 steps. The learning rate is 0.0001, and the RMSProp optimizer is used.

[0176] The real-time feedback adjustment module employs an online learning algorithm to dynamically optimize the recommendation strategy. The module's workflow is as follows: It continuously monitors user learning behavior, including video viewing progress, quiz results, course evaluations, and learning duration. Data is streamed to the server in real-time, and the learning path and recommendation results are updated every preset adjustment cycle.

[0177] Optionally, if a user skips a certain type of course consecutively, the recommendation weight of that type of course will be reduced; if a user gives a course a high rating, the recommendation probability of similar courses will be increased. The strategy adjustment can adopt a multi-armed slot machine (MAB) framework. If the system detects that the user's learning progress is behind expectations, or that the user's mastery of a certain knowledge point is not up to standard, the system will automatically adjust the learning path, insert supplementary learning content, or adjust the difficulty of subsequent courses.

[0178] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0179] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0180] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0181] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0182] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0183] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0184] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An intelligent online course learning system based on mobile internet, characterized in that, include: The user data collection module is used to collect users' learning behavior data through mobile internet terminals. The learning behavior data includes course viewing time, answer accuracy rate, learning frequency, interaction behavior records, and learning time distribution. The user profile building module is used to construct multi-dimensional user feature vectors based on the learning behavior data using a deep learning model. The course feature extraction module is used to extract multimodal features from courses in the course library and generate course feature vectors corresponding to the courses. The deep learning recommendation module is used to input the user feature vector and the course feature vector into a deep neural network model, calculate the matching degree between the user and the course, generate a personalized course recommendation list, and push the personalized course recommendation list to the user to achieve intelligent learning.

2. The intelligent online course learning system based on mobile internet according to claim 1, characterized in that, Collect user learning behavior data through mobile internet terminals, including: Based on a preset data collection cycle, after each data collection cycle, the user's course viewing time, answer accuracy, learning frequency, interaction behavior records, and learning time distribution are collected through mobile Internet terminals to obtain single-cycle feature data. A data matrix is ​​constructed from multiple consecutive periodic feature data to obtain the user's learning behavior data; each row in the data matrix is ​​a periodic feature data, and if a feature in the periodic feature data is empty, it is filled with zeros.

3. The intelligent online course learning system based on mobile internet according to claim 1, characterized in that, Based on the aforementioned learning behavior data, a multi-dimensional user feature vector is constructed using a deep learning model, including: The learning behavior data is normalized to obtain normalized learning behavior data. The normalized learning behavior data is used as input to a deep learning model to obtain the features output by the deep learning model, resulting in a multi-dimensional user feature vector.

4. The intelligent online course learning system based on mobile internet according to claim 1, characterized in that, Before the deep learning model is used, it also includes: Deep learning models are constructed using CNN-LSTM or CNN models; Multiple historical learning behavior data are collected, and the historical learning behavior data are preprocessed using the K-means clustering algorithm to obtain deep optimization data; the deep optimization data includes deep optimization samples and deep optimization labels corresponding to the deep optimization samples; An optimization population is initialized based on the parameters of the deep learning model to be optimized; each individual in the optimization population includes all the parameters of the model to be optimized. The fitness of each individual in the optimized population is obtained using deep optimization data, and the globally optimal individual is determined based on the fitness. A scattering optimization algorithm is used to perform full neighborhood development on each individual in the optimized population to obtain the first target individual; Based on the globally optimal individual, the fitness-enhanced neighborhood optimization algorithm is used to develop the inter-individual region of the first target individual to obtain the second target individual; A restricted information mutation optimization algorithm is used to perform global adaptive development on the second target individual to obtain the third target individual; Before the optimization termination condition is met, the third target individual is used as the input for the next optimization, and the next optimization is performed. When the optimization termination condition is met, the global optimal individual is re-determined based on the third target individual, and the parameters in the global optimal individual are used as the target parameters of the deep learning model to achieve optimization of the deep learning model.

5. The intelligent online course learning system based on mobile internet according to claim 4, characterized in that, A scattering optimization algorithm is used to perform full neighborhood development on each individual in the optimized population to obtain the first target individual, including: Based on the fitness of each individual in the optimized population, the degree of optimization of the solution is determined as follows: ； ； In the formula, Let i be the i-th individual, where i = 1, 2, ..., L, and L is the total number of individuals. Let j be the j-th individual, j = 1, 2, ..., L, and t be the number of optimization iterations. The optimization level of the solution corresponding to the i-th individual. For the intermediate parameter corresponding to the i-th individual, Let be the fitness of the i-th individual. Let be the fitness of the j-th individual, max be the function to maximize the fitness, and min be the function to minimize the fitness. To train the population; For positive integer terms less than or equal to 0.01; Based on the degree of solution optimization, the adaptive neighborhood development range is obtained as follows: ； ； In the formula, Let be the d-th dimension parameter of the i-th individual, where d = 1, 2, ..., K, and K is the total number of parameters to be trained. Let d be the parameter of the j-th individual. Let d be the adaptive neighborhood development range corresponding to the d-th dimension parameter of the i-th individual. The first random number between (0,1) The neighborhood influence coefficient. This is the base value for the neighborhood influence coefficient. The degree of optimization of the solution corresponding to the j-th individual. Let be the Euclidean distance between the i-th individual and the i-th individual, and T be the preset maximum number of optimization attempts; Based on the adaptive neighborhood development range, the neighborhood information corresponding to an individual is obtained as follows: ； In the formula, Let d be the d-th dimension parameter of the neighborhood information corresponding to the individual. for Random numbers between; The number of scattering points obtained is: ； In the formula, Let i be the number of scattering ... This is the base value for the number of scattering events; Repeatedly acquire neighborhood information corresponding to an individual until the number of acquired neighborhood information reaches the number of scattering events, thus obtaining... The neighborhood information is used to determine the neighborhood information with the highest fitness as the first target individual.

6. The intelligent online course learning system based on mobile internet according to claim 5, characterized in that, Based on the globally optimal individual, the fitness-enhancing neighborhood optimization algorithm is used to develop the inter-individual region of the first target individual to obtain the second target individual, including: Based on the number of optimization iterations, the adaptive inertia weight and the non-linear decreasing search factor are determined as follows: ； ； In the formula, For adaptive inertia weights, This is the upper limit of the adaptive inertia weight. This is the lower limit of the adaptive inertia weight. The search factor is a non-linear decreasing factor. The first adjustment coefficient, This is the second adjustment coefficient; Obtain the fitness value of the first target individual and sort them in descending order of fitness value to obtain the sorted first target individual; For any first target individual, match one other first target individual to the first half of the first target individuals after the arrangement, and match another other first target individual to the second half of the first target individuals, to obtain the first vertex individual and the second vertex individual respectively. The first enhancement development factor is obtained based on the fitness of the first vertex individual, and the second enhancement development factor is obtained based on the fitness of the second vertex individual as follows: ； ； In the formula, As the first enhancement development factor, As the second enhancing development factor, It is an exponential function with the natural constant e as its base. For the m-th first target individual, This refers to the first vertex individual corresponding to the m-th first target individual. This refers to the second vertex individual corresponding to the m-th first target individual. Let be the fitness of the m-th individual with the first target. The fitness of the first vertex individual. The fitness of the second-vertex individual; Based on the adaptive inertia weight, nonlinear decreasing search factor, first enhancement development factor, and second enhancement development factor, inter-individual region development is performed on the first target individual to obtain the second target individual: ； In the formula, Let d be the d-th dimension parameter of the m-th first target individual. Let d be the d-th dimension parameter of the globally optimal individual. Let d be the d-th dimension parameter of the m-th second target individual. The second random number between (0,1) The third random number between (0,1) Let d be the d-th dimension parameter of the first vertex individual. Let d be the d-th dimension parameter of the second vertex individual, and sin be the sine function.

7. The intelligent online course learning system based on mobile internet according to claim 6, characterized in that, The second target individual is globally adaptively developed using a constrained information mutation optimization algorithm to obtain the third target individual, including: Based on the upper and lower limit values ​​of each parameter in the second target individual, the first mutation parameter and the second mutation parameter are determined as follows: ； ； In the formula, Let d be the d-th dimension parameter of the nth second target individual. for The corresponding first mutation parameter, for The corresponding second mutation parameter, This provides information on the lower limit value of the d-th dimension parameter. This provides information on the upper limit of the d-th dimension parameter. The number of mutations; Based on the first mutation parameter and the second mutation parameter, the second target individual is subjected to global adaptive development to obtain the third target individual: ； In the formula, Let d be the d-th dimension parameter of the nth third target individual. It is a cosine function. Random or- k is the independent variable. A newly generated individual.

8. The intelligent online course learning system based on mobile internet according to claim 1, characterized in that, Multimodal feature extraction is performed on courses in the course database to generate course feature vectors, including: Obtain the knowledge point tags, difficulty level, course type and / or prerequisite relationships corresponding to the course to obtain multimodal features; The multimodal features are encoded using one-hot encoding to obtain the course feature vector corresponding to the course.

9. The intelligent online course learning system based on mobile internet according to claim 1, characterized in that, The user feature vector and the course feature vector are input into a deep neural network model to calculate the matching degree between the user and the course, generating a personalized course recommendation list, including: The user feature vector and the course feature vector are weighted and fused or concatenated to obtain a fused feature vector; The fused feature vector is used as the input to the deep neural network model to obtain the matching degree between the user and the course output by the deep neural network model. Based on the matching degree between the user and the course, a TOP-N recommendation is generated to obtain a personalized course recommendation list.

10. The intelligent online course learning system based on mobile internet according to claim 1, characterized in that, Also includes: The learning path planning module is used to generate the optimal learning path sequence based on the knowledge graph and the personalized course recommendation list, using a reinforcement learning algorithm. The real-time feedback and adjustment module is used to monitor the user's real-time learning status and dynamically adjust the recommendation strategy and learning path.

Images