Course commodity personalized recommendation method and device, terminal and storage medium

Abstract

The invention discloses a curriculum commodity personalized recommendation method and device, a terminal and a storage medium, and relates to the field of commodity recommendation. The method comprises the following steps: acquiring client target data and multi-dimensional current situation data to perform state quantification to obtain a quantized state set, constructing a course logic relation graph, mapping knowledge level differences to graph effective nodes, performing topological sorting and time resource filtering to obtain candidate learning paths, and performing learning on the candidate learning paths. And calculating the comprehensive gain value of the candidate learning path, and selecting the path corresponding to the highest value as a personalized recommendation path. Meanwhile, the method also comprises a rapid recommendation step for new customers or data sparse customers. The technical effects of providing a personalized course commodity recommendation scheme according to the actual condition of the customer, improving the accuracy and effectiveness of recommendation, saving the learning time for the customer, improving the learning effect and facilitating the new customer to quickly obtain proper recommendation are achieved.

Description

Technical Field

[0001] This application relates to the field of product recommendation, and in particular to a method, device, terminal, and storage medium for personalized recommendation of course products. Background Technology

[0002] In the current online education market, especially in areas with high customer spending and long decision-making cycles such as postgraduate entrance exams and professional qualifications, traditional course recommendation systems face fundamental bottlenecks.

[0003] Existing technologies primarily rely on collaborative filtering and content tag matching, which are essentially static similarity calculations between "users" and "products." These systems can only output discrete, unordered course lists, failing to understand the inherent logical dependencies between courses, and unable to comprehensively consider the user's complex current situation, goals, and limited time and budget. As a result, the recommended lists often fail to align with the user's real, deep-seated need for path planning from their current situation to their goal. Users still need to piece together learning plans themselves when faced with recommendations, leading to high decision-making costs and low trust levels.

[0004] Meanwhile, existing technologies lack interpretability for recommendation results (i.e., "black box" recommendations), and can only output a simple list of course products. They cannot provide a complete diagnostic report and solution plan similar to a B2B enterprise-level pre-sales solution, nor can they systematically analyze user needs, identify gaps, or provide explanations of matching logic, making the recommendation results unconvincing and unprofessional.

[0005] Therefore, the market urgently needs a personalized recommendation technology that can deeply integrate educational principles with individual user constraints, enabling a leap from product recommendation to intelligent planning. Summary of the Invention

[0006] The purpose of this application is to overcome the above-mentioned technical problems and provide a method, device, terminal and storage medium for personalized recommendation of course products.

[0007] Firstly, a personalized recommendation method for course products employs the following technical means: A personalized recommendation method for course products includes the following steps: Based on the collected user target data and multidimensional current status data, the status is quantified to obtain a quantified status set that includes knowledge level gaps and time resource constraints. Construct a course logical relationship graph with teaching units as nodes. The nodes are defined with constraint logical relationships based on educational laws and optimization logical relationships representing learning gains. The knowledge level gap is mapped to the effective nodes in the course logical relationship graph. Based on the constraint logical relationship, the effective nodes are topologically sorted to obtain an initial path set. Based on the time resource constraint, the initial path set is filtered to obtain candidate learning paths. The gain weight associated with the optimization logic relationship is invoked to calculate the comprehensive gain value of each candidate learning path. The candidate learning path with the highest comprehensive gain value is taken as the personalized recommendation path. An interactive recommendation scheme is generated and output based on the personalized recommendation path.

[0008] By adopting the above technical solutions, and through the technical chain of "quantifying user status, constructing logical graphs, constraining sorting, and optimizing selection," a paradigm upgrade from static product recommendation to dynamic learning planning has been achieved. Pedagogical principles are encoded into computable constraints and optimization relationships, and user time and ability constraints are integrated. Algorithms such as topological sorting are used to automatically generate logically sound and realistically feasible personalized course sequences. Finally, an interactive solution with embedded computational logic is output, ensuring that the recommendation results are rigorous, feasible, and highly interpretable, providing reliable intelligent planning support for educational decision-making.

[0009] Preferably, the step of quantifying the state based on the collected user target data and multidimensional current status data to obtain a quantified state set including knowledge level gaps and time resource constraints specifically includes the following steps: Collect users' target data and multidimensional current status data. The target data includes career planning direction, and the multidimensional current status data includes static academic data and dynamic learning behavior data. Based on the target data, a preset industry job database is called to perform a target refinement analysis to obtain target refinement data. The initial knowledge level gap value is obtained by combining the static academic data. The initial knowledge level gap value is then calibrated by combining the dynamic learning behavior data to obtain the final knowledge level gap value. Based on the multidimensional current status data, the number of available learning days is determined. Combined with the static academic data, a distribution model of available learning time is statistically analyzed, and a time resource constraint value is output. The knowledge level gap value and the time resource constraint value are encapsulated into a structured quantitative state set.

[0010] By adopting the above technical solution, the target data is used to call the industry job database for detailed target analysis. The initial knowledge level gap value is obtained by combining static academic data and then calibrated with dynamic learning behavior data, which can accurately determine the user's knowledge level gap. Based on the multidimensional current situation data statistics, the available learning time distribution model can be used to obtain the time resource constraint value, which can clarify the time limit for users' learning. The knowledge level gap value and the time resource constraint value directly support the business logic of target analysis and gap analysis, which helps to improve the accuracy of course recommendation.

[0011] Preferably, the construction of a curriculum logical relationship graph with teaching units as nodes, wherein the nodes are defined by binding logical relationships based on educational principles and optimization logical relationships representing learning gains, specifically includes the following steps: Each individual course product, as well as the knowledge modules with teaching objectives in each course, are entered into the graph database as independent teaching unit nodes; Analyze the teaching materials of all courses. When it is confirmed that the first node is a mandatory prerequisite for learning the second node, establish a directed edge representing a binding logical relationship between the first node and the second node. The direction of the directed edge of the binding logical relationship is from the first node to the second node. Multiple learning sequence patterns are identified based on historical learning data, which includes user learning logs and performance information. Directed edges representing optimization logical relationships are established based on the multiple learning sequence patterns, and each directed edge of optimization logical relationship is associated with a gain weight value. All the teaching unit nodes, the directed edges of the constraint logical relations, and the directed edges of the optimization logical relations are stored in the form of a graph data structure to form a course logical relation graph.

[0012] By adopting the above technical solutions, directed edges representing binding logical relationships can be established, which can standardize the order of learning, conform to the laws of pedagogy, and ensure the scientific and systematic nature of learning. Directed edges representing optimization logical relationships can be established based on historical learning data and associated with gain weight values, which can be used to refer to actual learning outcomes and provide a basis for recommending better learning paths. The graph data structure facilitates reasoning and data analysis. Combined with the user's quantified status and subsequent path selection and recommendation steps, personalized recommended paths and interactive recommendation schemes can be generated more accurately, better meeting the user's learning needs.

[0013] Preferably, the establishment of directed edges for optimizing logical relationships specifically includes the following steps: Extract all users' historical learning sequences from the user learning logs in the historical learning data, and obtain the learning performance score corresponding to each historical learning sequence from the performance information; Statistically analyze all historical learning sequences containing sequential node pairs and calculate the average performance score. Compare the average performance score with a preset baseline score threshold. If the average performance score exceeds the baseline score threshold, the node order in the sequential node pair is determined to be an effective gain mode. For each sequential node pair determined to be an effective gain pattern, an optimized logical relation directed edge is established that conforms to the node order. The magnitude by which the average performance score exceeds the baseline score threshold is stored as the gain weight of the current optimized logical relation directed edge.

[0014] By adopting the above technical solution, it is possible to determine whether the node order is an effective gain pattern. For nodes in an effective gain pattern, directed edges are established to form an optimized logical relationship and gain weights are stored. This allows for the integration of optimization logic based on historical learning outcomes when constructing the course logical relationship graph. Consequently, in subsequent course recommendations, this optimized logical relationship and its gain weights can be called to calculate the comprehensive gain value of candidate learning paths. This helps to select the personalized recommendation path with the highest comprehensive gain value, improves the accuracy and effectiveness of course recommendations, and provides users with a course recommendation solution that better meets their learning needs and achieves better learning outcomes.

[0015] Preferably, the step of mapping the knowledge level gap to valid nodes in the course logical relationship graph, performing topological sorting on the valid nodes based on the constraint logical relationship to obtain an initial path set, and filtering the initial path set based on the time resource constraint to obtain candidate learning paths specifically includes the following steps: Each knowledge level gap in the quantified state set is matched with all nodes in the course logical relationship graph, and nodes that can make up for the knowledge level gap are selected and marked as valid nodes. Based on all the valid nodes and the directed edges containing all the constraint logical relationships between the valid nodes, a course dependency subgraph is constructed; In the course dependency subgraph, key gap paths are identified based on the knowledge level gap value. A preset topology sorting algorithm is then executed along the key gap paths to generate a linear sequence of all nodes that satisfy the constraint logical relationship, which serves as the initial path set. The estimated total time for each initial path in the initial path set is calculated and compared with the time resource constraint value. All initial paths whose estimated total time does not exceed the time resource constraint value are selected as candidate learning paths.

[0016] By adopting the above technical solution, knowledge level gaps are mapped onto a course logical relationship graph to filter out effective nodes, enabling precise identification of teaching units that can bridge knowledge gaps. A course dependency subgraph is constructed and key gap paths are identified. An initial path set is generated using a topological sorting algorithm to ensure that the paths conform to the binding logical relationships of educational principles. The initial path set is then filtered based on time resource constraints to obtain candidate learning paths, ensuring that the recommended learning paths meet both knowledge enhancement needs and are within the user's available time frame. This improves the accuracy and feasibility of course recommendations and better meets the user's personalized learning needs.

[0017] Preferably, the step of calling the gain weight associated with the optimization logic relationship, calculating the comprehensive gain value of each candidate learning path, and selecting the candidate learning path with the highest comprehensive gain value as the personalized recommendation path specifically includes the following steps: Read the teaching unit nodes contained in each of the candidate learning paths in sequence; Based on the course logical relationship graph, query whether the optimized logical relationship exists between every two adjacent teaching unit nodes. If it exists, obtain the gain weight value associated with the optimization logic relationship, and sum all adjacent gain weight values ​​on the current candidate learning path to obtain the original comprehensive gain value of the current candidate learning path; The original comprehensive gain values ​​of all the candidate learning paths are normalized to obtain the final comprehensive gain value used for comparison. The candidate learning path corresponding to the highest comprehensive gain value is selected and used as the personalized recommendation path.

[0018] By employing the above technical solution, the comprehensive gain value of each candidate learning path can be accurately calculated. This allows for the selection of the candidate learning path with the highest comprehensive gain value as the personalized recommendation path, providing users with a recommendation scheme that better meets their learning needs and yields higher learning gains. This solution transforms path selection into an objective numerical comparison, eliminating the black box effect in the final recommendation. The superiority of a path can be traced through publicly available rules, greatly enhancing the credibility and persuasiveness of the results.

[0019] Preferably, the method further includes a quick recommendation step for new users, specifically including the following steps: Collect the historical quantitative state set of historical users and the corresponding course learning path adopted by the historical users to form a training dataset. After performing cluster analysis on all the historical users in the training dataset, they are divided into multiple clusters, and the typical state features of each cluster are extracted. The historical learning paths adopted by all historical users in each cluster are counted, and the adoption frequency and average positive feedback score of each historical learning path in its respective cluster are calculated. The historical learning paths whose adoption frequency and average positive feedback score are both higher than a preset threshold are selected and used as the high confidence recommendation path library of the cluster. When a recommendation is initiated for a new user, the basic target data and current status data of the new user are collected, and a basic quantitative state subset is calculated. The basic quantitative state subset is compared with the typical state features of each cluster to determine the preferred cluster corresponding to the current new user. From the high-confidence recommendation path library corresponding to the matched preferred cluster, the historical learning path with the highest adoption frequency or the highest average score is selected as the personalized recommendation path and directly output.

[0020] By adopting the above technical solutions and using the "clustering-statistics-matching" mechanism, high-quality default solutions validated by historical groups are provided to new users, achieving an instantly usable recommendation experience. This systematically solves the industry-wide problem of "cold start" and addresses the difficulty of recommendation due to the lack of sufficient historical data for new users or users with sparse data. It enables fast and effective personalized recommendations, improves recommendation efficiency and accuracy, and enhances product coverage and user satisfaction.

[0021] Secondly, this application provides a personalized course product recommendation device, which adopts the following technical solution: A personalized course recommendation device includes the following modules: The user status quantification module is used to quantify the status of users based on the collected target data and multidimensional current status data, and obtain a quantified status set that includes knowledge level gaps and time resource constraints. The course graph construction module is used to construct a course logical relationship graph with teaching units as nodes. The nodes are defined with constraint logical relationships based on educational principles and optimization logical relationships representing learning gains. The candidate path generation module is used to map the knowledge level gap to valid nodes in the course logical relationship graph, perform topological sorting on the valid nodes based on the constraint logical relationship to obtain an initial path set, and filter the initial path set based on the time resource constraint to obtain candidate learning paths. The recommendation result output module is used to call the gain weight associated with the optimization logic relationship, calculate the comprehensive gain value of each candidate learning path, take the candidate learning path with the highest comprehensive gain value as the personalized recommendation path, and generate and output an interactive recommendation scheme based on the personalized recommendation path.

[0022] Thirdly, this application provides a smart terminal, which adopts the following technical solution: A smart terminal includes a memory and a processor. The memory stores at least one instruction, at least one program, a code set, or an instruction set. The at least one instruction, at least one program, a code set, or an instruction set is loaded and executed by the processor to implement the personalized course product recommendation method described above.

[0023] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the personalized course product recommendation method described above.

[0024] In summary, this application includes at least one of the following beneficial technical effects: (1) This application constructs a course logical relationship graph, mines the relationship between user knowledge level and course, realizes in-depth mining of user-course relationship, improves the accuracy of human-product matching, and solves the problem of insufficient relationship mining in traditional recommendation system; (2) This application ensures the rigor of the teaching sequence through binding logical relationships and ensures the implementation of the plan through time resource constraints, thus achieving a dual guarantee of logical correctness and practical feasibility; (3) This application outputs a personalized learning path that conforms to teaching logic, meets personal constraints, and has been optimized and compared. It is no longer a disordered list of goods. It can systematically analyze user needs, identify gaps, and provide matching logic explanations, thereby enhancing the persuasiveness and professionalism of the recommendation results. Attached Figure Description

[0025] Figure 1 This is one of the flowcharts for the personalized recommendation method for course products; Figure 2 This is the second flowchart of the personalized recommendation method for course products; Figure 3 This is a diagram of the architecture of a personalized course recommendation system. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments are only possible technical implementations of the present invention, but are not limited thereto. Other embodiments obtained by those skilled in the art in conjunction with the embodiments of the present invention without creative effort are also within the protection scope of the present invention.

[0027] This application primarily uses user data to quantify user status, constructs a course relationship graph, and generates personalized course recommendation schemes. By quantitatively analyzing the user's current situation and goals, and combining this with the course logical relationship graph, it can more accurately identify the learning path suitable for the user, achieving the effect of accurately matching user needs and improving the course recommendation effect. The following is a further detailed description of this application.

[0028] The personalized course recommendation method in this embodiment includes steps such as data collection and quantification, graph construction, path generation, and interactive result output.

[0029] Specifically, a personalized recommendation method for course products, the process is as follows: Figure 1 and Figure 2 As shown, it includes the following steps: S1. Based on the collected user target data and multidimensional current status data, the status is quantified to obtain a quantified status set that includes knowledge level gaps and time resource constraints. This includes the following steps: S11. In a specific implementation method, the system collects the user's target data and multidimensional current status data through front-end interface forms, historical data import, and API docking with the learning platform.

[0030] The target data refers to the learning goals actively set by the user, including the target school, target major, and the user's career planning direction. The career planning direction is the user's expressed long-term career intention, such as "hoping to work in an internet company to engage in artificial intelligence algorithm research and development".

[0031] Multidimensional status data includes static academic data and dynamic learning behavior data, which describes the user's current objective state.

[0032] Static academic data refers to basic information that is stable over a relatively long period or obtained once, such as academic background, historical scores in various subjects, and scores in the most recent mock exams. For example: {"Current Institution": "XX University", "Current Major": "Software Engineering", "College Entrance Examination Mathematics Score": 135, "Passed CET-6": true, "Total Score of the Most Recent Mock Exam": 350}.

[0033] Dynamic learning behavior data refers to process data that is continuously generated and can be updated during recent learning, such as learning duration sequences and exercise accuracy sequences. For example: {"Average daily learning time for mathematics in the past 30 days": 1.5 hours, "Average accuracy of the past 10 sets of mathematics past exam questions": 68%, "Completion progress of the video course for Chapter 3 of Advanced Mathematics this week": 100%, "Frequency of errors in 'Function Limits' questions in the error notebook": 15 times}.

[0034] S12. Based on the collected user target data and multidimensional current status data, calculate and output a quantitative status set through a predefined quantitative status model. The quantitative status set includes quantitative values ​​for the user's knowledge level gap and time resource constraints.

[0035] In a specific implementation, the quantitative state model calls external data sources to refine the target based on data such as career planning direction of the target data, and integrates static and dynamic current status data to calculate and output quantitative values ​​of knowledge level gap and time resource constraints.

[0036] Specifically, S121. Based on the target data, a pre-defined industry job database is invoked to perform a detailed analysis of the target, resulting in detailed target data. The industry job database stores information such as skill requirements and frequently recruited universities for various industry positions. Through querying and matching, the associated target industry characteristics and job skill requirements are extracted to refine the target dimensions.

[0037] For example, {“Recommended Related Majors”: [“Artificial Intelligence”, “Computer Application Technology”], “Core Skill Requirements”: [“Python Programming”, “Machine Learning Fundamentals”, “Deep Learning Frameworks”], “Typical Target Companies”: [“Company A”, “Research Institute B”]}. Detailed target data is used to assist in validating or fine-tuning the target universities and majors initially entered by the user.

[0038] S122. Combine static academic data to obtain the initial knowledge level gap value, and combine dynamic learning behavior data to calibrate the initial knowledge level gap value to obtain the final knowledge level gap value.

[0039] In a specific implementation, taking mathematics as an example, the system determines the average score of mathematics in the target university's major over the years (e.g., 140 points) from the target refined data or user input. At the same time, it extracts the most recent full-subject mock exam mathematics score (e.g., 105 points) from static academic data, which best reflects the current stable level. The initial knowledge level gap is the difference between the two: 140 - 105 = 35 points.

[0040] Furthermore, the initial gap value is a static snapshot based on a single exam result. To reflect changes resulting from recent efforts, dynamic learning behavior data is introduced for calibration. For example, in a specific analysis, a student scored 105 on a mock exam, but their recent accuracy rate on real math questions has stabilized at 75%, higher than the mock exam level, and sufficient study time has been allocated. The system adjusts the initial gap value of 35 points downward according to predefined calibration rules, resulting in a final knowledge level gap value, for example, 30 points. Calibration rules include: for every 5% increase in recent accuracy rate compared to the mock exam accuracy rate, it is considered a 2-point improvement in ability.

[0041] S123. Extract the exam date and current date from the multidimensional current data, and directly calculate the total number of remaining days on the calendar as the number of available study days.

[0042] S124. Combining static academic data and dynamic learning behavior data, a distribution model of available learning time is statistically derived. For example, by analyzing users' daily learning time and available learning days over the past 60 days, the following distribution model is obtained: {"Average daily available time on weekdays": 2 hours, "Average daily available time on weekends": 6 hours, "Regularity of learning time periods": High}.

[0043] The time resource constraint value is output based on the distribution model of available learning time. The time resource constraint value is a structured constraint condition, including the user's total preparation days and the maximum available learning time per day, such as {"Total preparation days": 180, "Predicted total available learning hours": 540, "Recommended maximum daily load": 4 hours}.

[0044] S125. Encapsulate the knowledge level gap value and time resource constraint value into a structured quantitative state set. The knowledge level gap value is divided by subject, and the quantitative state set is as follows: {"Mathematics": 30, "English": 20, "Professional Courses": 40, Remaining Days: 180, Daily Available Time: 2}.

[0045] S2. Construct a logical relationship graph of the curriculum with teaching units as nodes. The nodes are defined by binding logical relationships based on educational principles and optimization logical relationships representing learning gains. This includes the following steps: S21. Each independent course product, as well as the knowledge modules with teaching objectives in each course, are entered into the graph database as independent teaching unit nodes.

[0046] In practice, independent course products and knowledge modules cannot coexist in the same graph.

[0047] S22. Analyze the teaching materials for all courses. In one possible implementation, the teaching materials include the official syllabus, prerequisite requirements, or teaching rules set by domain experts for all courses or chapters.

[0048] When it is confirmed that the first node is a mandatory prerequisite for learning the second node, a directed edge representing a binding logical relationship is established between the first and second nodes, with the direction of the directed edge pointing from the first node to the second node. That is, in a specific implementation, when it is confirmed that mastering the content of node A is a mandatory prerequisite for learning the content of node B, a directed edge representing a binding logical relationship is established between node A and node B, with the direction pointing from A to B.

[0049] In one specific implementation, if the node is a course product, the sequence of binding logical relations is "Course A, Course B, Course C". In this case, the binding logical relations describe the macro-level dependencies between the entire course, such as the inability to learn probability theory without completing advanced mathematics.

[0050] If the nodes are knowledge modules, the sequence is "Module A1, Module B2, Module C3", and the nodes may come from different courses. In this case, the constraint logic relationship describes the micro-dependencies between knowledge points; for example, without understanding limits, one cannot learn derivatives.

[0051] Constraint logical relationships include prerequisite dependencies defined based on the course syllabus or expert rules, which represent the necessary preconditions between nodes.

[0052] S23. Based on historical learning data, multiple learning order patterns are identified. The historical learning data includes user learning logs and performance information. Directed edges representing optimization logical relationships are established according to the multiple learning order patterns. Each directed edge of optimization logical relationship is associated with a gain weight value.

[0053] Optimization logical relationships are learning gain relationships determined based on historical learning performance data, associated with specific gain weight values, and representing quantifiable efficiency improvements brought about by the learning order. The process of establishing optimization logical relationships specifically includes the following steps: S231. Extract all users' historical learning sequences from the user learning logs in the historical learning data, and obtain the learning performance score corresponding to each historical learning sequence from the performance information in the historical learning data.

[0054] S232. Count all historical learning sequences containing sequential node pairs and calculate the average performance score. For example, sequential node pairs<X,Y> That is, learn X first and then learn Y.

[0055] S233. Compare the average performance score with the preset baseline score threshold. If it exceeds the baseline score threshold, the node order in the sequential node pair is determined to be an effective gain mode. That is, the order mode from node X to node Y is determined to be an effective gain mode. In other words, the mode "learn node Y after learning node X" can bring efficiency or effect improvement.

[0056] S234, for each sequential node pair determined to be an effective gain mode<X,Y> Establish a directed edge that conforms to the node order (from X to Y) and represents the optimization logical relationship. Store the magnitude of the average performance score exceeding the baseline score threshold as the gain weight of the current optimization logical relationship directed edge.

[0057] S24. Store all teaching unit nodes, directed edges of constraint logical relations, and directed edges of optimization logical relations with attached gain weight values ​​in the form of a graph data structure to form a course logical relation graph that can be used for graph traversal and reasoning.

[0058] S3. Map the knowledge level gap to valid nodes in the course logical relationship graph. Perform topological sorting on the valid nodes based on the constraint logical relationships to obtain an initial path set. Filter the initial path set based on time resource constraints to obtain candidate learning paths. Specifically, this includes the following steps: S31. Match each knowledge level gap in the quantified state set with all nodes in the course logic relationship graph, filter out the nodes that can make up for the knowledge level gap and mark them as valid nodes.

[0059] In one specific implementation, the system maintains a teaching unit-ability dimension mapping table. This table records the specific knowledge and skill dimensions (such as advanced mathematics computation ability) that each teaching unit node (e.g., basic calculus course) can cultivate or improve, along with their expected improvement values. When the user's knowledge level gap vector is obtained, such as -30 for advanced mathematics and -15 for English reading, the system iterates through all nodes in the course logic graph and queries the aforementioned mapping table. For each gap dimension, all nodes in the mapping table marked as capable of improving that dimension, and whose expected improvement direction is opposite to the gap direction, are selected and marked as valid nodes. More specifically, for a gap of -30 in advanced mathematics, the system will select all nodes in the mapping table marked "improve advanced mathematics ability," such as "basic calculus course" and "linear algebra intensive course."

[0060] S32. Construct a course dependency subgraph based on all valid nodes and directed edges containing all constraint logical relationships between valid nodes.

[0061] S33. In the course dependency subgraph, the knowledge level gap value is quantified into the weight of the node, the key gap path is identified based on the knowledge level gap value, and the preset topology sorting algorithm is executed along the key gap path.

[0062] Specifically, in one implementation, in the course dependency subgraph, starting from the effective node corresponding to the dimension with the largest knowledge level gap (e.g., multiple nodes corresponding to "Advanced Mathematics" with the largest gap), the Depth-First Search (DFS) algorithm is used to find the connected path that covers all key gap nodes (i.e., passes through at least one effective node in each dimension) and has the largest gap value. On this connected path, a topological sorting algorithm is executed according to the dependency direction defined by the directed edge of the constraint logical relationship between nodes. In this embodiment, the Kahn algorithm is used.

[0063] S34. Generate one or more valid linear sequences of nodes that satisfy the constraint logical relationship, ensuring that the recommended course order conforms to the teaching rules. Each linear sequence of nodes is an initial path. Summarize all linear sequences of nodes to obtain the initial path set.

[0064] S35. Calculate the estimated total time for each initial path in the initial path set and compare it with the time resource constraint value. Select all initial paths whose estimated total time does not exceed the time resource constraint value as candidate learning paths.

[0065] The system maintains a teaching unit-standard time mapping table, which records the estimated learning hours of each teaching unit node under the standard learning intensity. For each initial path in the initial path set, the estimated learning hours of all the nodes it contains are summed to obtain the estimated total time of the initial path.

[0066] The estimated total time for each path is compared with the time resource constraint (e.g., remaining preparation days × daily available hours). Paths that meet the condition that the estimated total time is no greater than the time resource constraint are selected as candidate learning paths.

[0067] In another specific possible implementation, In step S2, the acquired multidimensional status data also includes the user's course procurement budget range.

[0068] For example, the system can directly collect the user's input of the course procurement budget range, such as 15,000 to 20,000 yuan, through a user interface form. To simplify calculations, the lower limit of 15,000 yuan can be used as the budget resource constraint value. Alternatively, the system can infer the user's price sensitivity and budget range based on historical purchase records from multidimensional current data.

[0069] In step S2, the output constraint values ​​also include budget resource constraint values.

[0070] In the course logic relationship graph, each teaching unit node is associated with a specific market price. After generating the initial path set, the prices of all nodes are added up for each initial path to obtain the estimated total cost of the path.

[0071] The system will check both filter conditions simultaneously: A. Time constraint: The estimated total time for the path is less than or equal to the time resource constraint value.

[0072] B. Budget constraint: The estimated total cost of the route is less than or equal to the budget resource constraint value.

[0073] Only paths that meet both of the above conditions will be retained as candidate learning paths. Any path that exceeds the time limit or budget will be eliminated.

[0074] After obtaining the initial path set in step S4, the initial path set is filtered based on time resource constraints and budget resource constraints. The filtering condition is: the sum of the total selling price of all course nodes in the path does not exceed the budget resource constraint value.

[0075] This embodiment ensures that the final recommended personalized path is not only logically correct and time-efficient, but also cost-effective.

[0076] S4. Invoke the gain weights associated with the optimization logic relationship, calculate the comprehensive gain value of each candidate learning path, and take the candidate learning path with the highest comprehensive gain value as the personalized recommendation path. Generate and output an interactive recommendation scheme based on the personalized recommendation path. The specific steps include the following: S41. Read the teaching unit nodes contained in each candidate learning path sequentially. In a specific implementation, a candidate learning path is regarded as a node sequence [Node1, Node2, Node3, ..., NodeN].

[0077] S42. Based on the course logical relationship graph, for each pair of adjacent nodes (Node1, Node2), (Node2, Node3), ... in the sequence, query whether there is an optimal logical relationship between each pair of adjacent teaching unit nodes: If it exists, obtain the gain weight value w_i associated with the optimization logical relationship on the directed edge; if it does not exist, treat the gain weight value of the pair of nodes as 0 or a preset base value.

[0078] S43. Sum the gain weights of all adjacent candidates on the current candidate learning path to obtain the original comprehensive gain value G_raw=Σw_i of the current candidate learning path.

[0079] S44. To avoid unfair comparisons caused by differences in path lengths, this embodiment uses the average gain per unit node as a unified comparison metric: the original comprehensive gain values ​​of all candidate learning paths are normalized and mapped to the same comparable interval, and the normalized comprehensive gain value of each path is calculated. G_final = G_raw / N; Where G_final is the overall gain value, and N is the total number of teaching unit nodes included in the candidate learning path.

[0080] S45. Compare the overall gain values ​​of all candidate learning paths, select the candidate learning path with the highest overall gain value, and use it as the personalized recommendation path.

[0081] S46. Generate and output an interactive visualization solution, which includes the following steps: S461. Based on the personalized recommended path, automatically generate an HTML5 webpage report. In a specific implementable scenario, this report includes: a knowledge level gap analysis presented as a radar chart, a path course node sequence and time planning presented as a Gantt chart, and a table explaining the matching logic between each module and the gap points. In the Gantt chart, node connections with high gain weight values ​​are highlighted, and their specific weights are labeled.

[0082] S462. In the webpage report, configure adjustable interactive components for the target school, major, and daily study time.

[0083] S463. Listen for adjustment events of interactive components. When a user submits modifications through an interactive component, the event is triggered. The adjusted parameters are then submitted to the backend system as new target data and time resource constraints.

[0084] S464. After receiving the new parameters, the backend system re-executes all steps from step S1 and pushes the newly generated personalized recommendation path and report content to the frontend webpage in real time for a refresh-free update.

[0085] To better address the cold start problem faced by new users or users with sparse data, this invention also provides a parallel fast recommendation mechanism.

[0086] S5. The quick recommendation process for new users or users with sparse data is independent of the above steps S1-S4, and specifically includes the following steps: S51. Collect the historical quantitative state set of historical users and the corresponding course learning paths adopted by historical users to form a training dataset.

[0087] In one specific implementation, the system background runs an offline computing task that collects the final historical quantitative status set (e.g., including math gap value, English gap value, total available days) of 10,000 historical users who have completed learning in the past year, as well as the course path (e.g., course A, course C, course F sequence) they ultimately purchased and completed.

[0088] S52. After performing cluster analysis on all historical users in the training dataset, the dataset is divided into multiple clusters, and the typical state features of each cluster are extracted.

[0089] In a specific implementation, different clustering algorithms are used to analyze the historical quantitative state sets of historical users. As a preferred option, when the business requires a clearly defined and distinct typical group, the K-means clustering algorithm can be used. For example, setting the number of clusters K=5, the algorithm will automatically divide the data into 5 clusters, such as "weak in math", "time-pressured", and "balanced sprint", and calculate the central feature vector for each cluster, i.e., the typical state feature.

[0090] As another preferred approach, the DBSCAN density clustering algorithm can be used to more automatically discover naturally formed clusters and identify noise points in the data. By setting an appropriate neighborhood radius and minimum number of samples, the algorithm can divide users into multiple core clusters and mark sparsely distributed users as noise points. These noise points can be excluded from the subsequent path library construction or processed separately.

[0091] Based on the distribution characteristics of historical data or the configuration of maintenance personnel, any of the above clustering algorithms can be flexibly selected to perform cluster analysis, thereby obtaining customer groups with similar state characteristics and their typical features.

[0092] S53. Statistically analyze the historical learning paths adopted by all historical users in each cluster, and calculate the adoption frequency and average positive feedback score of each historical learning path in its respective cluster. Select historical learning paths whose adoption frequency and average positive feedback score are both higher than the preset threshold, and use them as the high-confidence recommendation path library for the cluster.

[0093] In one specific implementation, for the "weak in math" cluster, the system collects historical learning paths adopted by all users within that cluster. Assuming the path "Course A (Basic Math) to Course M (Intensive Math)" is chosen by 60% of users in this cluster, and these users' average course rating is as high as 4.8 (out of 5), then this path will be stored in the high-confidence recommendation path library for the "weak in math" cluster. The system records the adoption rate and average rating for each path in the library.

[0094] S54. When starting a recommendation for a new user, collect the new user's basic target data and current status data, such as poor math skills and 6 months of preparation time.

[0095] The basic quantized state subset is calculated as {Mathematical gap: High, Time: 6 months}.

[0096] S55. Compare the basic quantified state subset with the typical state features of each cluster to determine the preferred cluster corresponding to the current new user. From the high-confidence recommendation path library corresponding to the matched preferred cluster, select the historical learning path with the highest adoption frequency or the highest average score as the personalized recommendation path and output it directly.

[0097] In one specific implementation, the system calculates the cosine similarity between the subset of states and the feature vectors of the five group centers, finding that it best matches the cluster of "weak in mathematics". Subsequently, it directly retrieves the path with the highest adoption rate (i.e., from course A to course M) from the path library of this cluster as a personalized recommendation path, along with the explanation "This recommendation is based on the choices of many successful students with similar situations to yours" and presents it to the user.

[0098] In another specific implementation, a button is provided on the recommendation page, such as "Get a more accurate personalized plan," to guide users to fill in more detailed test questions and goals, thereby transitioning to the core process of step S1.

[0099] In another specific implementation, this application provides a method for quantitatively matching a user's micro-level skill gaps with the specific teaching value points of a course and automatically generating explanatory descriptions, which specifically includes the following steps.

[0100] Maintain a basic gap point library that covers all key knowledge points and ability dimensions within the target exam scope. For example, for postgraduate entrance exam mathematics, it includes hundreds of entries such as "Advanced Mathematics - Function Limits" and "Linear Algebra - Matrix Operations".

[0101] The knowledge level gap is broken down into a structured list of gap points. Specifically, based on the subject to which the knowledge level gap belongs, all gap point entries for that subject are retrieved from the basic gap point database. The system calculates a current mastery score for each relevant gap point based on the user's answers to preset ability diagnostic questions, ranging from 0 to 100. A target mastery level is preset, and the difference between this and the current mastery score is calculated to obtain the quantified gap value for each gap point. A structured list of user gap points is output, with each entry formatted as {Gap Point ID: "G001", Gap Point Name: "Function Limit", Quantified Gap Value: 15}.

[0102] When constructing the logical relationship graph of the course, a value point attribute is added to each teaching unit node. Specifically, each teaching unit node is labeled with the ID of one or more core gap points that it aims to solve. Based on the historical teaching effect data of the teaching unit, the average improvement value for each labeled gap point is calculated and the node attributes are stored. For example, the attributes of the node "Calculus Intensive Course" include: {Solve gap points: ["G001", "G002"], expected improvement value: [12, 18]}.

[0103] When filtering valid nodes, for each candidate node that can resolve the user's gap, calculate its matching score with the user, using the following formula: Match score = Σ(the expected improvement value of the node to a certain gap point / the quantitative difference value of the user at that gap point × weight coefficient); The weighting coefficient is determined by the importance of the gap point in the overall gap.

[0104] Finally, when calculating the overall gain value of the path, this embodiment incorporates the calculated matching score.

[0105] The overall gain of the path is obtained by weighted summation of two parts: Overall gain = α × sum of path gain weights + β × average path matching score; Wherein, α and β are adjustable hybrid parameters. The path gain weights are derived from the optimization logic relationship and are used to reflect learning efficiency; the average path matching score is derived from the matching calculation in this embodiment and is used to reflect the accuracy of solving user shortcomings.

[0106] When generating the final recommendation, the system automatically generates explanatory text for each key node in the path. For the top Q nodes with the highest matching scores in the path, the corresponding template is selected from a pre-set explanatory template library and populated with specific data.

[0107] For example, we recommend the "Intensive Calculus Course", which mainly targets your weakness in the area of ​​function limits (current gap: 15). According to historical data, studying this part can improve your score by an average of 12 points, effectively making up for your weakness in this area.

[0108] Based on the same inventive concept described above, this application also discloses a personalized course product recommendation device, the structure of which is as follows: Figure 3 As shown, the device includes the following modules: The user status quantification module is used to quantify the status of users based on the collected target data and multidimensional current status data, and obtain a quantified status set that includes knowledge level gaps and time resource constraints. The course graph construction module is used to construct a course logical relationship graph with teaching units as nodes. The nodes are defined with constraint logical relationships based on educational principles and optimization logical relationships representing learning gains. The candidate path generation module maps knowledge level gaps to valid nodes in the course logical relationship graph, performs topological sorting on the valid nodes based on the constraint logical relationship to obtain an initial path set, and filters the initial path set based on time resource constraints to obtain candidate learning paths. The recommendation result output module is used to call the gain weight associated with the optimization logic relationship, calculate the comprehensive gain value of each candidate learning path, select the candidate learning path with the highest comprehensive gain value as the personalized recommendation path, and generate and output an interactive recommendation scheme based on the personalized recommendation path.

[0109] Based on the same inventive concept described above, this application also discloses a smart terminal, including a memory and a processor. The memory stores at least one instruction, at least one program, code set, or instruction set. The processor loads and executes the at least one instruction, at least one program, code set, or instruction set to implement the personalized course product recommendation method described above.

[0110] Based on the same inventive concept described above, this application also discloses a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the personalized course product recommendation method described above.

[0111] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code.

[0112] Those skilled in the art will understand that the step numbers of the above methods or processes are only used to distinguish different steps and do not constitute an absolute restriction on the execution order. Some steps may be executed simultaneously or in a different order than the numbers.

[0113] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for personalized recommendation of course products, characterized in that, Includes the following steps: Based on the collected user target data and multidimensional current status data, the status is quantified to obtain a quantified status set that includes knowledge level gaps and time resource constraints. Construct a course logical relationship graph with teaching units as nodes. The nodes are defined with constraint logical relationships based on educational laws and optimization logical relationships representing learning gains. The knowledge level gap is mapped to valid nodes in the course logical relationship graph. Based on the constraint logical relationship, the valid nodes are topologically sorted to obtain an initial path set. Based on the time resource constraint, the initial path set is filtered to obtain candidate learning paths. The gain weight associated with the optimization logic relationship is invoked to calculate the comprehensive gain value of each candidate learning path. The candidate learning path with the highest comprehensive gain value is taken as the personalized recommendation path. An interactive recommendation scheme is generated and output based on the personalized recommendation path.

2. The personalized course recommendation method according to claim 1, characterized in that, The step of quantifying the state based on the collected user target data and multidimensional current status data to obtain a quantified state set that includes knowledge level gaps and time resource constraints specifically includes the following steps: Collect users' target data and multidimensional current status data. The target data includes career planning direction, and the multidimensional current status data includes static academic data and dynamic learning behavior data. Based on the target data, a preset industry job database is called to perform a target refinement analysis to obtain target refinement data. The initial knowledge level gap value is obtained by combining the static academic data. The initial knowledge level gap value is then calibrated by combining the dynamic learning behavior data to obtain the final knowledge level gap value. Based on the multidimensional current status data, the number of available learning days is determined. Combining the static academic data and the dynamic learning behavior data, a distribution model of available learning time is statistically analyzed, and a time resource constraint value is output. The knowledge level gap value and the time resource constraint value are encapsulated into a structured quantitative state set.

3. The personalized course recommendation method according to claim 2, characterized in that, The construction of a curriculum logical relationship graph with teaching units as nodes, wherein the nodes are defined by binding logical relationships based on educational principles and optimization logical relationships representing learning gains, specifically includes the following steps: Each individual course product, as well as the knowledge modules with teaching objectives in each course, are entered into the graph database as independent teaching unit nodes; Analyze the teaching materials of all courses. When it is confirmed that the first node is a mandatory prerequisite for learning the second node, establish a directed edge representing a binding logical relationship between the first node and the second node. The direction of the directed edge of the binding logical relationship is from the first node to the second node. Multiple learning sequence patterns are identified based on historical learning data, which includes user learning logs and performance information. Directed edges representing optimization logical relationships are established based on the multiple learning sequence patterns, and each directed edge of optimization logical relationship is associated with a gain weight value. All the teaching unit nodes, the directed edges of the constraint logical relations, and the directed edges of the optimization logical relations are stored in the form of a graph data structure to form a course logical relation graph.

4. The personalized course recommendation method according to claim 3, characterized in that, The establishment of directed edges for optimizing logical relationships specifically includes the following steps: Extract all users' historical learning sequences from the user learning logs in the historical learning data, and obtain the learning performance score corresponding to each historical learning sequence from the performance information; Statistically analyze all historical learning sequences containing sequential node pairs and calculate the average performance score. Compare the average performance score with a preset baseline score threshold. If the average performance score exceeds the baseline score threshold, the node order in the sequential node pair is determined to be an effective gain mode. For each sequential node pair determined to be an effective gain pattern, an optimized logical relation directed edge is established that conforms to the node order. The magnitude by which the average performance score exceeds the baseline score threshold is stored as the gain weight of the current optimized logical relation directed edge.

5. The personalized course recommendation method according to claim 3, characterized in that, The process of mapping the knowledge level gap to valid nodes in the course logical relationship graph, performing topological sorting on the valid nodes based on the constraint logical relationship to obtain an initial path set, and filtering the initial path set based on the time resource constraint to obtain candidate learning paths specifically includes the following steps: Each knowledge level gap in the quantified state set is matched with all nodes in the course logical relationship graph, and nodes that can make up for the knowledge level gap are selected and marked as valid nodes. Based on all the valid nodes and the directed edges containing all the constraint logical relationships between the valid nodes, a course dependency subgraph is constructed; In the course dependency subgraph, key gap paths are identified based on the knowledge level gap value, and a preset topology sorting algorithm is executed along the key gap paths to generate a linear sequence of all nodes that satisfy the constraint logical relationship, which serves as the initial path set. The estimated total time for each initial path in the initial path set is calculated and compared with the time resource constraint value. All initial paths whose estimated total time does not exceed the time resource constraint value are selected as candidate learning paths.

6. The personalized course recommendation method according to claim 3, characterized in that, The step of invoking the gain weight associated with the optimization logic relationship, calculating the comprehensive gain value of each candidate learning path, and selecting the candidate learning path with the highest comprehensive gain value as the personalized recommendation path specifically includes the following steps: Read the teaching unit nodes contained in each of the candidate learning paths in sequence; Based on the course logical relationship graph, query whether the optimized logical relationship exists between every two adjacent teaching unit nodes. If it exists, obtain the gain weight value associated with the optimization logic relationship, and sum all adjacent gain weight values ​​on the current candidate learning path to obtain the original comprehensive gain value of the current candidate learning path; The original comprehensive gain values ​​of all the candidate learning paths are normalized to obtain the final comprehensive gain value used for comparison. The candidate learning path corresponding to the highest comprehensive gain value is selected and used as the personalized recommendation path.

7. The personalized course recommendation method according to claim 1, characterized in that, The method also includes a quick recommendation step for new users, specifically including the following steps: Collect the historical quantitative state set of historical users and the corresponding course learning path adopted by the historical users to form a training dataset; After performing cluster analysis on all the historical users in the training dataset, they are divided into multiple clusters, and the typical state features of each cluster are extracted. The historical learning paths adopted by all historical users in each cluster are counted, and the adoption frequency and average positive feedback score of each historical learning path in its respective cluster are calculated. The historical learning paths whose adoption frequency and average positive feedback score are both higher than a preset threshold are selected and used as the high confidence recommendation path library of the cluster. When a recommendation is initiated for a new user, the basic target data and current status data of the new user are collected, and a basic quantitative state subset is calculated. The basic quantitative state subset is compared with the typical state features of each cluster to determine the preferred cluster corresponding to the current new user. From the high-confidence recommendation path library corresponding to the matched preferred cluster, the historical learning path with the highest adoption frequency or the highest average score is selected as the personalized recommendation path and directly output.

8. A personalized course product recommendation device, characterized in that, Includes the following modules: The user status quantification module is used to quantify the status of users based on the collected target data and multidimensional current status data, and obtain a quantified status set that includes knowledge level gaps and time resource constraints. The course graph construction module is used to construct a course logical relationship graph with teaching units as nodes. The nodes are defined with constraint logical relationships based on educational principles and optimization logical relationships representing learning gains. The candidate path generation module is used to map the knowledge level gap to valid nodes in the course logical relationship graph, perform topological sorting on the valid nodes based on the constraint logical relationship to obtain an initial path set, and filter the initial path set based on the time resource constraint to obtain candidate learning paths. The recommendation result output module is used to call the gain weight associated with the optimization logic relationship, calculate the comprehensive gain value of each candidate learning path, take the candidate learning path with the highest comprehensive gain value as the personalized recommendation path, and generate and output an interactive recommendation scheme based on the personalized recommendation path.

9. A smart terminal, characterized in that, The method includes a memory and a processor, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the personalized course product recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the personalized course product recommendation method as described in any one of claims 1 to 7.

Images