Recommendation method and system fusing large language model reasoning and multi-source trajectory information

Abstract

The application provides a recommendation method and system for fusing large language model reasoning and multi-source trajectory information, and the method comprises the following steps: after receiving a user recommendation request, first acquiring user historical learning behavior and static attribute information, and constructing a static interest vector; combining an initial feature vector of a learned knowledge point and a graph enhanced vector of a first-order neighbor node of a knowledge graph to generate an explicit and graph expanded interest vector, and fusing to obtain a user interest vector; through similarity analysis of the user interest vector and an unlearned knowledge point vector and / or reasoning of a large language model on user associated information, screening N unlearned knowledge points to form a candidate set; verifying and screening a target knowledge point through a prerequisite knowledge point mastery degree, and outputting a recommendation result after sorting. The method improves recommendation accuracy and adaptability, and meets personalized learning needs.

Description

Technical Field

[0001] This invention relates to the field of knowledge point recommendation technology, and in particular to a recommendation method and system that integrates large language model reasoning and multi-source trajectory information. Background Technology

[0002] With the deepening development of educational informatization, online learning platforms have gradually taken on the core function of personalized delivery of teaching content, and recommendation systems play a crucial role in this process. Especially in knowledge point-level recommendation tasks, the system needs to comprehensively analyze users' historical learning behavior, mastery status, content structure, and cognitive progress to achieve accurate delivery. However, in practical applications, educational scenarios face multiple challenges, including sparse user learning data, complex knowledge point relationships, and strong cognitive path dependencies. Traditional collaborative filtering and content matching recommendation methods struggle to handle the unique prerequisite relationships, ability matching, and teaching objective constraints in the education field, often resulting in recommended content that does not match the learner's current learning ability, leading to problems such as "skipping levels in recommendations," "repeated recommendations," or "disordered learning paths," seriously affecting learning outcomes and user experience. While existing educational recommendation technologies have made some progress in basic recall, content matching, and user profile modeling, they still have significant shortcomings in in-depth personalized recommendation tasks oriented towards the teaching process. First, traditional dual-tower structures generally rely on knowledge points that users have explicitly learned for interest modeling, failing to cover users' potential, unlearned but related learning interests, resulting in overly sparse or skewed interest representations. Secondly, most recommendation methods fail to effectively incorporate information about the teaching structure, especially the prerequisite relationships, semantic similarities, and logical connections between knowledge points. They lack consideration for the order of instruction and learning curves during the recommendation process, potentially leading to chaotic content recommendation order and jumps in difficulty. Furthermore, current recommendation systems often employ manually set weighted linear scoring formulas, typically relying on single indicators such as similarity, click-through rate, and mastery, which fail to reflect the dynamic learning needs of learners at different stages. Static ranking strategies lack the ability to model the interactions between complex features and are ill-suited to the differences in learning progress, cognitive level, and interests among users, thus failing to achieve truly personalized recommendations. More critically, recommendation systems generally lack mechanisms for ensuring the rationality of learning paths and controlling cognitive load, failing to organically integrate recommendation strategies with educational goals, thus limiting their effectiveness and widespread application in teaching scenarios. Summary of the Invention

[0003] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0004] According to a first aspect of the present invention, a recommendation method integrating large language model reasoning and multi-source trajectory information is provided. The method is applied to a knowledge point recommendation scenario on an online education platform. Upon receiving a recommendation request from a user, the following steps are performed:

[0005] S100: Obtain the user's historical learning behavior information and static attribute information on the target learning platform, and construct the user's static interest vector based on the obtained user's historical learning behavior information and static attribute information.

[0006] S200: Obtain the initial knowledge point feature vector of the user's learned knowledge points, and generate an explicit interest vector through average pooling; based on the learned knowledge points, obtain the graph-enhanced knowledge point vectors of the first-order neighbor nodes of the learned knowledge points from the knowledge graph G of the target learning platform, and generate a graph-extended interest vector through average pooling; where G = (V, E, W), where V represents the set of knowledge nodes, E represents the set of edges, and W represents the edge weight matrix; the feature representation of each node in G is: the graph-enhanced knowledge point vector obtained after inputting the initial knowledge point feature vector of the corresponding knowledge point into the graph attention network.

[0007] S300, the user static interest vector, the explicit interest vector, and the graph extended interest vector are fused to generate a user interest vector.

[0008] S400, based on the similarity analysis between the user interest vector and the knowledge point vector of the unlearned knowledge points, and / or based on the analysis of user association information based on the large language model, obtain N unlearned knowledge points that the user needs to learn from the full set of unlearned knowledge points of the user, forming a candidate knowledge point set; wherein, the user association information includes at least user historical learning behavior information, static attribute information, learned knowledge points and learning objectives.

[0009] S500 selects the target knowledge point set from the candidate knowledge point set based on user interest vectors, knowledge point vectors of candidate knowledge points, knowledge point vectors of prerequisite knowledge points of candidate knowledge points, and historical behavioral characteristics of learned knowledge points.

[0010] S600 sorts the target knowledge points in the target knowledge point set and outputs the sorting result as the recommendation result.

[0011] According to a second aspect of the present invention, a recommendation system integrating large language model reasoning and multi-source trajectory information is provided, the system comprising:

[0012] The static interest vector construction module is used to obtain users' historical learning behavior information and static attribute information on the target learning platform, and to construct users' static interest vectors based on the obtained users' historical learning behavior information and static attribute information.

[0013] The explicit interest vector acquisition module is used to obtain the initial knowledge point feature vectors of the knowledge points that the user has already learned, and generate explicit interest vectors through average pooling.

[0014] The graph-enhanced interest vector acquisition module is used to obtain the graph-enhanced knowledge point vectors of the first-order neighbor nodes of the learned knowledge points from the knowledge graph G of the target learning platform, based on the learned knowledge points, and generate graph-enhanced interest vectors through average pooling; where G = (V, E, W), where V represents the set of knowledge nodes, E represents the set of edges, and W represents the edge weight matrix; the feature representation of each node in G is: the graph-enhanced knowledge point vector obtained by inputting the initial knowledge point feature vector of the corresponding knowledge point into the graph attention network.

[0015] The user interest vector acquisition module is used to obtain N unlearned knowledge points that the user needs to learn from the full set of unlearned knowledge points, based on similarity analysis between the user interest vector and the knowledge point vectors of unlearned knowledge points, and / or analysis of user association information based on a large language model, forming a candidate knowledge point set; wherein, the user association information includes at least user historical learning behavior information, static attribute information, learned knowledge points, and learning objectives.

[0016] The candidate knowledge point acquisition module is used to obtain N unlearned knowledge points that the user needs to learn from all unlearned knowledge points based on the user's interest vector and the knowledge point vector of the user's unlearned knowledge points, or based on the user's interest vector, the knowledge point vector of the user's unlearned knowledge points and the large language model, forming a candidate knowledge point set. The knowledge point vector of a certain knowledge point is obtained based on the corresponding initial knowledge point feature vector, which is constructed based on the basic information of the knowledge point.

[0017] The target knowledge point acquisition module is used to filter the target knowledge point set from the candidate knowledge point set based on the user interest vector, the knowledge point vector of the candidate knowledge point, the knowledge point vector of the prerequisite knowledge point of the candidate knowledge point, and the historical behavioral characteristics of the learned knowledge point.

[0018] The recommendation result generation module is used to sort the target knowledge points in the target knowledge point set and output the sorting results as the recommendation results.

[0019] The present invention has at least the following beneficial effects:

[0020] Accurately depict user needs: Integrate static interest vectors, explicit interest vectors, and graph-extended interest vectors to construct user profiles from multiple dimensions, including basic user attributes, direct learning preferences, and knowledge-related needs. This overcomes the limitations of traditional recommendations that rely on single-dimensional user characteristics and improves the accuracy of capturing user learning needs.

[0021] Adaptable to diverse recommendation scenarios: Through the optional and combined paths of similarity analysis and large language model inference, it can achieve efficient recommendation based on vector matching, and also use large language models to deeply analyze complex information such as users' learning goals and basic levels, adapting to the recommendation needs of different users (such as those with clear goals and those exploring interests).

[0022] Ensuring the rationality and effectiveness of learning: Target knowledge points are selected based on the user's mastery of prerequisite knowledge points to avoid recommending content beyond the user's current ability; through multi-dimensional feature sorting, knowledge points with high matching degree and suitable for the learning progress are prioritized to improve learning efficiency and experience, and help implement personalized teaching in online education.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating the recommendation method that integrates large language model reasoning and multi-source trajectory information, provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0029] This invention aims to provide a personalized education recommendation method that integrates a knowledge graph expansion mechanism, a user mastery prediction model, and a learnable ranking strategy, in order to solve problems such as insufficient expression of interest, disordered content recommendation paths, rigid ranking logic, and insufficient cognitive adaptability in existing education recommendation systems.

[0030] This invention provides a recommendation method that integrates large language model reasoning and multi-source trajectory information. The method is applied to knowledge point recommendation scenarios on online education platforms. Upon receiving a recommendation request from a user, it executes the following steps: Figure 1 The following steps are shown:

[0031] S100: Obtain the user's historical learning behavior information and static attribute information on the target learning platform, and construct the user's static interest vector based on the obtained user's historical learning behavior information and static attribute information.

[0032] In this embodiment of the invention, the target learning platform refers to an online education platform that provides users with knowledge point learning resources.

[0033] In this embodiment of the invention, historical learning behavior information includes quantifiable behavioral data such as the user's course completion progress, chapter access frequency, quiz scores, practice accuracy, and answering time. Static attribute information includes inherent attributes such as user ID, user's educational level, professional field, and learning style tags (e.g., visual, auditory).

[0034] In this embodiment of the invention, the steps for constructing the user's static interest vector are as follows:

[0035] The historical learning behavior information is sorted in reverse chronological order (with recent behavior given priority), and continuous feature vectors are constructed through standardization (such as normalization to the [0, 1] interval).

[0036] Static attribute information is encoded to generate a dense vector of static attributes, wherein discrete attributes are mapped to a low-dimensional dense vector through an embedding layer;

[0037] Continuous feature vectors and dense static attribute vectors are concatenated by dimension to form user static interest vectors, which serve as a unified expression of user basic features.

[0038] S200: Obtain the initial knowledge point feature vector of the user's learned knowledge points, and generate an explicit interest vector through average pooling; Based on the learned knowledge points, obtain the graph-enhanced knowledge point vector of the first-order neighbor nodes of the learned knowledge points from the knowledge graph G of the target learning platform, and generate a graph-extended interest vector through average pooling.

[0039] Where G = (V, E, W), where V represents the set of knowledge nodes, E represents the set of edges, and W represents the edge weight matrix; the feature representation of each node in G is: the graph-enhanced knowledge point vector obtained by inputting the initial knowledge point feature vector of the corresponding knowledge point into the graph attention network.

[0040] In this embodiment of the invention, the knowledge points that a user has already learned can be obtained through historical learning behavior information.

[0041] In this embodiment of the invention, the initial knowledge point feature vector is constructed based on the basic information of the knowledge point. The basic information of the knowledge point may include the knowledge point ID, knowledge point text information (title, description), tag system (such as mathematics, algebra, geometry), difficulty level, and course information. The initial knowledge point feature vector is constructed as follows:

[0042] The textual information of knowledge points is encoded into a fixed-dimensional dense semantic vector using Sentence-BERT, and the tags and categories are converted into vectors after multi-hot encoding and embedding mapping.

[0043] All features are concatenated and uniformly projected to 128 dimensions to obtain the initial knowledge point feature vector, which serves as the input to the knowledge point tower.

[0044] The initial feature vectors of all knowledge points on the target learning platform are pre-stored in an initial feature vector knowledge base for efficient use in the recommendation process. The determination of knowledge points on the platform is based primarily on course content, with the following specific rules:

[0045] All knowledge points on the platform are generated based on the breakdown of the course content system: each course is divided into several chapters according to the teaching logic, and each chapter contains multiple teaching videos;

[0046] The core knowledge content carried by a single teaching video is defined as an independent knowledge point. For example, the video "reduction formula of trigonometric functions" corresponds to the knowledge point of the same name. This ensures that the knowledge points correspond one-to-one with the video teaching units, avoiding both overly coarse knowledge granularity, such as merging the entire chapter into a knowledge point, and over-splitting, such as breaking down the derivation steps of a formula into multiple knowledge points.

[0047] The knowledge points must cover the core concepts, principles, or skills in the video (such as the "basic formulas of calculus" explained in the video and the corresponding application examples, which all belong to this knowledge point), and be manually verified to ensure that the boundaries of each knowledge point are clear, the content is complete, and it is consistent with the teaching objectives of the course.

[0048] In this embodiment of the invention, the edge types in the knowledge graph may include pre-relational edges, semantically similar edges, and reinforcement edges. Each edge is assigned a fixed initial weight in W according to its semantics; for example, the weight of a pre-relational edge is set to 1.0, a similar edge to 0.8, and a reinforcement edge to 0.5. The knowledge graph is stored in the form of an adjacency list.

[0049] In this embodiment of the invention, the first-order neighbor nodes of the learned knowledge point are specifically the first-order neighbor nodes whose edge type with the learned knowledge point is a pre-relational edge or a semantically similar edge. The graph expansion interest vector can be obtained by average pooling the graph enhancement knowledge point vectors of the first-order neighbor nodes of the learned knowledge point.

[0050] S300, the user static interest vector, the explicit interest vector, and the graph extended interest vector are fused to generate a user interest vector.

[0051] In this embodiment of the invention, the user static interest vector, the explicit interest vector, and the graph extended interest vector are fused using a gating fusion mechanism. The user interest vector satisfies the following condition: u = g × u base +(1-g)×(u graph +u static ); where u is the user interest vector, which can be 128-dimensional, comprehensively reflecting the user's learned preferences, potential knowledge needs, and static attribute characteristics. base For the explicit interest vector, u graph To extend the interest vector in the graph, u static The vectors are user static interest vectors, all 128-dimensional. g is the fusion gating parameter, g = sigmoid(W). g [u base ;u graph ;u static ]+b g ), sigmoid() is the activation function, [;] indicates vector concatenation, W g b is the weight parameter of the user tower fusion layer. g These are the bias terms for the fusion layer, all of which are optimized through model training.

[0052] In another embodiment of the present invention, u = g × u base +(1-g)×(α×u graph +(1-α)×u static α is an adjustment coefficient, 0 < α < 1, used to dynamically balance the influence of knowledge association expansion and user inherent attributes, in order to avoid feature dilution caused by the direct superposition of graph expansion interests and static interests, so that the weights of the two can be dynamically optimized according to training.

[0053] In this embodiment of the invention, the adjustment coefficient can be regarded as a trainable parameter of the model. During the model training process, it is updated together with other parameters through backpropagation and eventually converges to the optimal value. Specifically, α is initialized to 0.5 (or randomly initialized in the range of 0 to 1). In each iteration, the value of α is adjusted according to the gradient of the loss function (such as cross-entropy loss, mean squared error, etc. in a ranking task) to make it eventually adapt to the data distribution.

[0054] S400, based on the similarity analysis between the user's interest vector and the knowledge point vector of the unlearned knowledge points, and / or based on the analysis of user association information based on the large language model, obtain N unlearned knowledge points that the user needs to learn from the full set of unlearned knowledge points of the user, forming a candidate knowledge point set; wherein, the user association information includes at least the user's historical learning behavior information, static attribute information, learned knowledge points and learning objectives, and N is a preset positive integer, the value of which is determined according to the actual scenario.

[0055] (Example 1)

[0056] In this embodiment, the candidate knowledge point set is obtained solely based on the similarity analysis between the user's interest vector and the knowledge point vectors of unlearned knowledge points. Further, S400 may specifically include:

[0057] S401, obtain the knowledge point vector corresponding to each of the unlearned knowledge points for all users.

[0058] S402, calculate the similarity between the user's interest vector and the knowledge point vector of each unlearned knowledge point, and obtain the similarity result corresponding to each unlearned knowledge point.

[0059] S403, Sort all unlearned knowledge points according to the similarity results, and select the top N unlearned knowledge points after sorting to form a candidate knowledge point set.

[0060] In this embodiment of the invention, "knowledge points not learned by all users" refers to all knowledge points that have not been learned by all users. The knowledge point vector for each knowledge point is obtained based on the corresponding initial knowledge point feature vector.

[0061] Specifically, the system retrieves the knowledge point vectors of all unlearned knowledge points from a pre-defined knowledge point vector library within the platform, calculates their cosine similarity with the user's interest vector, and selects N knowledge points as candidate knowledge points in descending order of similarity. The number of N can be determined based on the actual situation. Common methods for determining N are as follows:

[0062] 1. Experience-based setup based on business scenarios (basic solution)

[0063] Based on the knowledge density and display scenario settings of the target learning platform:

[0064] For scenarios with low knowledge density (such as general education courses and introductory knowledge points): N can be 20~50 to provide a wider range of choices and meet users' exploration needs;

[0065] For scenarios with high knowledge density (such as professional courses and advanced knowledge points): N should be 5~15 to avoid information overload and focus on core matching knowledge points;

[0066] Mobile / Lightweight Scenarios: N is 5~10, adapted for small screen display, reducing user decision-making costs.

[0067] This type of approach is simple and direct, and is suitable for rapid implementation. Usually, the optimal empirical value can be determined first through small-scale testing (such as A / B testing to compare the click-through rates of users with N=10 and N=20).

[0068] 2. Dynamic adjustment based on user characteristics (personalized solutions)

[0069] By associating the N value with user attributes or behavioral characteristics, the number of candidates suitable for different users can be dynamically generated, for example:

[0070] New users (limited behavioral data): Take a larger value for N (e.g., 30~50) to help the system quickly explore user preferences through broader candidate coverage;

[0071] Active users (rich in behavioral data): N is taken as a small value (e.g., 5~15), and accurate recommendations are made based on stable interests to reduce redundancy;

[0072] For users who learn quickly: N should be a larger value (e.g., 20~30) to match their need for rapid advancement;

[0073] For users with slow learning progress: take a smaller value for N (e.g., 5~10) and focus on consolidating the core content of the current knowledge.

[0074] 3. Constraint settings based on system performance (engineering solution)

[0075] The value of N is determined by considering the platform's computing resources and response speed limitations:

[0076] If the platform has a large user base and high real-time requirements (such as millions of users online at the same time): N needs to be controlled within a small range (such as 10~20) to avoid excessive time consumption in similarity calculation and subsequent processing (such as path compliance screening), which would affect the response speed;

[0077] If the platform has a small user base but sufficient resources, the N value can be appropriately increased (e.g., 30-50) to enhance the richness of the candidate pool.

[0078] 4. Adaptive optimization based on recommendation performance (data-driven approach)

[0079] N is treated as an optimizable parameter and dynamically adjusted through model training or reinforcement learning:

[0080] Offline optimization: During the model training phase, test the recommendation metrics (such as click-through rate, learning completion rate, and dwell time) corresponding to different N values ​​(such as 5, 10, 15, and 20), and select the N value that optimizes the core metrics.

[0081] Online learning: Through reinforcement learning, we track user feedback on recommendations with different N values ​​in real time (such as clicks, skips, and completion of learning), and dynamically update the N value strategy (for example, if the click rate of recommendations with N=10 is consistently higher than that with N=20, the N value will be automatically lowered).

[0082] (Example 2)

[0083] In this embodiment, the candidate knowledge point set can be obtained solely based on the analysis of user-related information using a large language model. Further, S400 may specifically include:

[0084] S410, the user's historical learning behavior information, static attribute information, list of learned knowledge points and learning objectives are concatenated into structured input text. The structured input text is combined with preset prompt words and then input into the large language model to obtain an initial knowledge point set consisting of K unlearned knowledge points; where K is a preset positive integer, K≥N, which is reserved space for subsequent filtering.

[0085] In this embodiment of the invention, the large language model can be Qwen3-14B. The preset prompt words can be determined based on actual needs. In one illustrative embodiment, exemplary preset prompt words may be:

[0086] "You are a learning path recommendation assistant. Based on the following user information, infer the K unlearned knowledge points that need to be learned next:"

[0087] The knowledge points must be relevant to the user's goals, not be on the completed list, and their difficulty level must match the user's current skill level.

[0088] Output only a list of knowledge point names.

[0089] Input information:

[0090] User attributes: {u_static, corresponding to user static attribute information}

[0091] Learning trajectory: {u_traj, corresponding to the user's historical learning behavior information}

[0092] Current level: {u_level, determined based on mastery of learned knowledge points}

[0093] Learning objective: {u_goal, the user's explicitly stated learning objective}

[0094] Completed knowledge points: {u_mastered, list of knowledge points learned by the user}

[0095] Recommended quantity: {K}.

[0096] S411, each initial knowledge point in the initial knowledge point set is feature-encoded to generate an initial knowledge point vector corresponding to each initial knowledge point.

[0097] In this embodiment of the invention, a sentence-BERT encoder can be used to perform feature encoding on each initial knowledge point in the initial knowledge point set.

[0098] S412, calculate the similarity between each initial knowledge point vector and each knowledge point vector in the preset knowledge point vector library, and determine the real unlearned knowledge points matched by each initial knowledge point; select N knowledge points from the matching results to form a candidate knowledge point set.

[0099] S412 specifically includes five stages: vector standardization, similarity calculation and initial screening, candidate vector selection, multi-dimensional verification, and matching result determination. This ensures accurate matching of the initial knowledge points output by the large language model with the actual unlearned knowledge points in the pre-defined knowledge point vector library, as detailed below:

[0100] Step 1: Vector Standardization

[0101] The vectors generated from the initial knowledge points output by the large language model (such as "application of the quadratic equation root-finding formula" and "Python function parameter passing mechanism") are standardized through feature encoding to ensure consistency with the calculation benchmark of vectors in the preset knowledge point vector library. The specific operation is as follows:

[0102] (1) Dimensional consistency verification

[0103] Read the dimensions of the initial knowledge point vector (e.g., a 128-dimensional vector generated by sentence-BERT) and compare it with the standard dimensions (e.g., 128-dimensional) of the vectors in the preset knowledge point vector library. If the dimensions are inconsistent, adjust them to the standard dimensions through zero-padding (low-dimensional zero filling) or principal component analysis (PCA, high-dimensional dimensionality reduction) to avoid distortion of similarity calculation due to dimensional differences.

[0104] (2) Vector normalization processing

[0105] If cosine similarity is used: L2 normalization is performed on the initial knowledge point vectors to make the normalized vectors have a magnitude of 1, eliminating the interference of vector magnitude on the similarity results (such as avoiding misjudgment of semantic relevance due to vector numerical scaling). If Euclidean distance is used: Min-Max normalization is uniformly performed on the initial knowledge point vectors and vectors in the preset library (mapping vector elements to the [0,1] interval) to ensure that the vector numerical range is consistent and improve the comparability of distance calculation.

[0106] Step 2: Similarity Calculation and Initial Screening

[0107] Iterate through the preset knowledge point vector library, calculate the similarity between the initial knowledge point vector and all unlearned knowledge point vectors in the library, and filter out highly relevant vectors to form a candidate list, specifically including:

[0108] (1) Selection of similarity calculation method

[0109] Select the appropriate method based on the characteristics of the knowledge points:

[0110] For semantic knowledge points (such as concepts and principle descriptions): cosine similarity should be used first, focusing on the semantic relevance of vector directions (such as the vector directions of "finding the root of a quadratic equation" and "formula for finding the root of a quadratic equation in one variable" are highly consistent).

[0111] Skill-based knowledge points (such as operating procedures and formula applications): Euclidean distance is used to supplement the verification of the similarity of vector values ​​(such as the smaller difference in vector values ​​between "Python function parameter passing" and "Python position parameter passing").

[0112] (2) Dynamic threshold screening

[0113] A similarity threshold is set, and vectors that meet the criteria are selected to form a "candidate matching vector list". The threshold is determined as follows:

[0114] Data-driven threshold: Based on historical matching data statistics, select the minimum threshold that makes the "matching accuracy ≥ 90%" (e.g., by analyzing the past 100,000 matching records, it was determined that the accuracy is 92% when the cosine similarity threshold is 0.6, so 0.6 is adopted).

[0115] Scene adaptation threshold: dynamically adjusted based on knowledge point attributes.

[0116] For specialized fields (such as advanced mathematics, programming languages): the threshold must be ≥0.7 (strict matching of specialized terminology is required);

[0117] General knowledge (such as popular science knowledge and basic concepts): threshold ≥ 0.5 (flexibility in expression is allowed);

[0118] Newly added knowledge point (few vector data): Threshold reduced by 0.1 (to avoid missed matches).

[0119] Example: The vector of the initial knowledge point "Python function parameter passing mechanism" meets the threshold of ≥0.6 with K002 (Python function positional parameter passing, cosine similarity 0.82) and K005 (Python function keyword parameter passing, cosine similarity 0.75) in the library, and is included in the candidate list; it does not meet the threshold with K010 (Java method parameter passing, cosine similarity 0.32), and is excluded.

[0120] Step 3: Selection of candidate vectors

[0121] The "candidate matching vector list" is sorted in descending order of similarity value, and the top M vectors (M is a preset positive integer, usually 3 to 5) are selected as "high-priority candidate vectors" to reduce the computational load in subsequent verification stages. The specific rules are as follows:

[0122] If the number of vectors in the candidate list is less than or equal to M: keep all vectors (e.g., if there are 2 vectors in the candidate list, M=3, then keep all of them).

[0123] If the number of candidate vectors is greater than M: take the first M vectors (e.g., if there are 10 candidate vectors and M=3, then take the top 3 vectors based on similarity).

[0124] If the number of candidate list vectors is 0: mark the initial knowledge point as "awaiting manual matching" and trigger the supplementary processing flow (such as automatically recording it to the platform backend and prompting the operations staff to match it manually).

[0125] Step 4: Multi-dimensional verification

[0126] The knowledge points corresponding to the "high-priority candidate vectors" are subject to dual verification based on attributes and unlearned status to exclude mismatched or already learned knowledge points, as follows:

[0127] (1) Attribute validation

[0128] Based on user-related information (learning objectives, static attributes, and learned knowledge points), verify the core attributes of candidate knowledge points:

[0129] Subject Consistency: If the user's learning goal is "high school math exam preparation", the candidate knowledge points must belong to the "mathematics" subject, excluding knowledge points from other subjects such as "physics" and "computer science".

[0130] Difficulty suitability: Based on the user's static attributes (such as "first-year high school student"), the difficulty of the candidate knowledge points should be "introductory high school" or "advanced junior high school", excluding out-of-syllabus content such as "university calculus";

[0131] Knowledge Relevance: The correlation strength between candidate knowledge points and user-learned knowledge points is calculated through the knowledge graph (e.g., the sum of edge weights ≥ 0.5) to ensure that there is a prior or extended relationship (e.g., if the user has already learned "definition of a quadratic equation", the correlation strength of the candidate "quadratic equation root-finding formula" is 0.8, which meets the requirements).

[0132] (2) Unlearned verification

[0133] Based on the user's list of learned knowledge points, verify whether the candidate knowledge point is content that has not been learned:

[0134] Exact match verification: Perform an exact string match between the candidate knowledge point name and the learned list (e.g., if the candidate "solution method of quadratic equation in one variable" is completely consistent with the learned "solution method of quadratic equation in one variable", it will be excluded).

[0135] Semantic similarity verification: Calculate the cosine similarity between the candidate knowledge point vector and the learned knowledge point vector. If the similarity is ≥0.8 (judged as semantic duplication), exclude it (e.g., the candidate "method of finding the roots of a quadratic equation" and the learned "application of the formula for finding the roots of a quadratic equation in one variable" have a similarity of 0.85, so exclude it).

[0136] Step 5: Determine the matching result.

[0137] Based on the verification results, the actual unlearned knowledge points corresponding to each initial knowledge point are determined according to the following rules:

[0138] (1) Unique matching: If only one candidate vector passes all the verifications, its corresponding knowledge point is directly taken as the real unlearned knowledge point;

[0139] (2) Multiple matching: If multiple candidate vectors pass the verification, the optimal one is selected according to priority.

[0140] First priority: the highest similarity score (e.g., candidate A has a similarity of 0.82, candidate B has a similarity of 0.75, select A);

[0141] Second priority: the strongest knowledge association (if the similarity is the same, choose the one with a stronger association with the already learned knowledge points).

[0142] Third priority: User preference adaptation (if the first two are the same, select the knowledge point type that the user has the highest historical learning frequency).

[0143] (3) No match: If all candidate vectors fail the validation, perform supplementary processing:

[0144] Lower the similarity threshold (e.g., from 0.6 to 0.5), re-traverse the vector library to generate a candidate list, and repeat the verification.

[0145] The initial knowledge point names are standardized (e.g., redundant words such as "mechanism" and "method" are removed), and the vectors are re-encoded before matching again.

[0146] If no results are found, the initial knowledge point is marked as "pending confirmation," and the corresponding knowledge point is manually matched or added in the preset knowledge point base, while the vector base is updated synchronously.

[0147] Step 6: Generation of candidate knowledge point set

[0148] From all the actual unlearned knowledge points that match the initial knowledge points, N knowledge points are selected to form a candidate set, and the selection rules are as follows:

[0149] If the total number of matching results is ≥ N: sort them by weighted order of "large language model inference priority (output order) + matching similarity", and select the top N results;

[0150] If the total number of matching results is less than N: Using the similarity analysis method in Example 1, supplement and filter (N - number of matching results) knowledge points from the full set of unlearned knowledge points to ensure that the total number is N.

[0151] The above process enables precise matching between initial knowledge points and actual unlearned knowledge points, ensuring semantic consistency and adapting to user learning needs, thus laying the foundation for subsequent selection of target knowledge points.

[0152] (Example 3)

[0153] In this embodiment, the candidate knowledge point set is obtained based on the similarity analysis between the user interest vector and the knowledge point vector of unlearned knowledge points, and the analysis of user association information based on a large language model. Further, S400 specifically includes:

[0154] S420, based on the similarity between the user interest vector and the knowledge point vector of the user's unlearned knowledge points, obtain a first knowledge point set containing N unlearned knowledge points from the full set of user unlearned knowledge points.

[0155] S421, the user's historical learning behavior information, static attribute information, list of learned knowledge points and learning objectives are concatenated into structured input text. The structured input text is combined with preset prompt words and then input into the large language model to obtain an initial knowledge point set consisting of K unlearned knowledge points.

[0156] S422, each initial knowledge point in the initial knowledge point set is feature-encoded to generate a knowledge point vector corresponding to each initial knowledge point.

[0157] S423, calculate the similarity between each initial knowledge point vector and each knowledge point vector in the preset knowledge point vector library, determine the real unlearned knowledge points matched by each initial knowledge point in the preset knowledge point vector library, and form a second knowledge point set.

[0158] S424, take the union of the first knowledge point set and the second knowledge point set and remove duplicate knowledge points, and use the deduplicated set as the candidate knowledge point set.

[0159] If the number of unions is greater than N, the top N knowledge points are selected by weighted sorting based on "similarity (first set) + large language model inference priority (second set)". If the number of unions is less than N, the unlearned knowledge points are supplemented and filtered to N based on similarity, and finally a candidate knowledge point set is formed.

[0160] This embodiment combines two approaches: relying on vector similarity to ensure recommendation accuracy, and using a large language model to uncover potential learning needs, thereby improving recommendation diversity and coverage.

[0161] The S400 process in Example 1 can be implemented using a dual-tower model, which consists of a user tower, a knowledge point tower, and an output module. The specific structure and working principle are as follows:

[0162] 1. User Tower

[0163] Architecture: It adopts a gated recurrent network (GRU) combined with a feature fusion layer to adapt to the characteristics of temporal behavior and static attributes.

[0164] Input: User's original feature data, including dynamic features (user's historical learning behavior sequence, such as the order of accessing knowledge points, practice records, viewing time, etc.) and static features (user's static attribute information, such as education level, major, learning style, etc.).

[0165] Processing flow:

[0166] ①GRU extracts temporal vectors of dynamic features, while the embedding layer encodes static features into dense vectors;

[0167] ②Based on the S300-based gating fusion mechanism, dynamic feature vectors, user static interest vectors, explicit interest vectors, and graph-extended interest vectors are fused to output a 128-dimensional user interest vector u.

[0168] 2. Knowledge Pyramid

[0169] Architecture: A three-layer fully connected feedforward network with the structure “FC(din→256)→ReLU→FC(256→128)→ReLU”; where din=128 (the dimension of the initial knowledge point feature vector), FC is a fully connected layer, and ReLU is an activation function (formula: ReLU(x)=max(0,x), used to introduce non-linear expressive power).

[0170] Input: Initial knowledge point feature vector.

[0171] Output: 128-dimensional knowledge point representation vector, enhancing feature discrimination.

[0172] Output module: Calls the cosine similarity formula to calculate the similarity between the user interest vector u and the knowledge point vector, quantifies the degree of "user-knowledge point" matching, and provides a basis for the screening in Example 1.

[0173] 4. Model Training

[0174] Training strategy: Joint training + triplet contrastive learning, training samples are triples (us, c) + c - ); where us is the user's original feature data, c + For positive samples (initial feature vectors of knowledge points already learned by the user), c - Negative samples (initial feature vectors of knowledge points that users have not learned, selected through random sampling or "sampling of discontinuous learning paths").

[0175] Loss function: L(u, k) + k - )=max(0,γ-score(u,k) + )+score(u,k - )); where score(u, k + ) represents u and k + The similarity score between them, score(u, k) - ) represents the user interest vectors u and k - The similarity score between them, k + For c + The vector output by the knowledge point tower, k - For c - The vector output by the knowledge point tower has γ as the boundary margin hyperparameter, which can be set to 0.2.

[0176] Training parameters: The AdamW optimizer is used, and the learning rate (example 1e-4), batch size (example 128), etc. can be adjusted as needed; during training, the GRU and Embedding layers of the user tower and the fully connected layers of the knowledge tower participate in backpropagation synchronously to update parameters.

[0177] Training objective: To force the model to learn "u and k" + Higher similarity, and with k - By following the pattern of "lower similarity", the vector space is optimized to better reflect the actual relationship, thereby improving matching accuracy.

[0178] 5. Knowledge base update: All knowledge point vectors of the target learning platform are pre-stored in the knowledge base; when a new knowledge point is added, the system automatically generates its vector through the knowledge point tower and updates the knowledge base to ensure the integrity of the candidate pool.

[0179] S500 selects the target knowledge point set from the candidate knowledge point set based on user interest vectors, knowledge point vectors of candidate knowledge points, knowledge point vectors of prerequisite knowledge points of candidate knowledge points, and historical behavioral characteristics of learned knowledge points.

[0180] In this embodiment of the invention, prerequisite knowledge points can be obtained based on a preset prerequisite path dependency matrix P. P∈{0,1} N×N , where P ij =1 indicates that knowledge point i is a prerequisite knowledge point for knowledge point j. The values ​​of i and j range from 1 to N, where N is the total number of knowledge points provided by the target learning platform.

[0181] Furthermore, the S500 specifically includes:

[0182] S510 inputs the user's interest vector, the knowledge point vector of the candidate knowledge points, and the historical behavioral features of the learned knowledge points into the trained mastery prediction model, and outputs the predicted mastery value of the user for each candidate knowledge point.

[0183] In this embodiment of the invention, the mastery prediction model can be MasteryNet, which can adopt a three-layer feedforward neural network structure (256→128→64→1). The historical behavioral characteristics of learned knowledge points can include 10 dimensions of historical behavioral characteristics such as the number of practice sessions, completion rate, first learning time, most recent learning time, learning frequency, review ratio, learning time, number of skips, learning content coverage (coverage of videos, handouts, exercises, etc. under this knowledge point), and completion rate of prerequisite knowledge points.

[0184] S520 inputs the user's interest vector, the knowledge point vector of the prerequisite knowledge points of each candidate knowledge point, and the historical behavioral features of the learned knowledge points into the trained mastery prediction model, and outputs the predicted value of the user's mastery of each prerequisite knowledge point.

[0185] S530, for a single candidate knowledge point, the following judgment is performed: if the predicted mastery values ​​of all prerequisite knowledge points of the candidate knowledge point are greater than the preset mastery threshold, then the candidate knowledge point is included in the target knowledge point set; if the predicted mastery value of any prerequisite knowledge point of the candidate knowledge point is less than or equal to the preset mastery threshold, then the candidate knowledge point is removed from the candidate knowledge point set, so as to ensure that the recommended content is consistent with the user's current knowledge base and avoid skipping learning.

[0186] In this embodiment of the invention, the mastery threshold can be set to 0.7.

[0187] S600 sorts the target knowledge points in the target knowledge point set and outputs the sorting result as the recommendation result.

[0188] Furthermore, the S600 specifically includes:

[0189] S610, construct a ranking feature vector for each target knowledge point, input the ranking feature vector into the trained ranking score prediction model, and output the ranking score of each target knowledge point; wherein, the ranking feature vector includes at least: the knowledge point vector of the target knowledge point, the user interest vector, the similarity between the knowledge point vector of the target knowledge point and the user interest vector, the mastery prediction value of the target knowledge point, the association strength between the target knowledge point and the user's learned knowledge points in the knowledge graph, the mean of the mastery prediction values ​​of all prerequisite knowledge points of the target knowledge point, and the number of times the user has clicked on knowledge points of the same type as the target knowledge point in history.

[0190] In this embodiment of the invention, the correlation strength between the knowledge point in the knowledge graph and the knowledge points already learned by the user can be, for example, the average edge weight.

[0191] In this embodiment of the invention, the ranking score prediction model can be RankNet, a ranking model based on a multilayer perceptron, with the following model structure: FC(256→128)→ReLU→FC(128→64)→ReLU→FC(64→1).

[0192] The ranking score is used to indicate the recommendation priority of the target knowledge point.

[0193] In this embodiment of the invention, the ranking and scoring prediction model is trained using RankNet's pairwise ranking loss, which improves model performance by optimizing the ranking relationship between pairs of knowledge points. Pairwise ranking loss is a commonly used loss function in ranking model training; its core idea is to learn ranking rules by optimizing the relative order of paired samples.

[0194] For each user, select one positive sample c from the knowledge points they have actually learned. +(The knowledge points that the user has actually learned), and select a negative sample c from the knowledge points that have not been learned. - (Knowledge points that users have not yet learned) constitute sample pairs (c + c - );

[0195] The loss function is defined as: L ranc =-logσ(s(c + )-s(c - ), where σ is the sigmoid function, s(c + ) and s(c - The ranking scores for positive and negative samples are respectively.

[0196] The training objective is to maximize the positive sample scores s(c + ) as large as possible the negative sample score s(c) - (i.e., s(c)) + )-s(c - () > 0), thus allowing the model to learn the rule that "knowledge points that users are more likely to learn should be listed first".

[0197] This loss function can directly optimize the core objective of recommendation ranking (relative order), which is more in line with the needs of recommendation scenarios than simply predicting absolute scores, and ultimately improves the accuracy of ranking results.

[0198] S620: Sort the target knowledge points in descending order according to the ranking score, and output the sorted list of target knowledge points as the recommendation result, so that the recommendation reflects both the matching degree of user interests and the cognitive readiness and compliance of teaching order.

[0199] The recommendation method provided in this invention optimizes the entire process from interest modeling to content recommendation by integrating multi-source user features with the reasoning capabilities of a large language model. The specific technical effects are as follows:

[0200] First, the fusion of multi-dimensional interest vectors improves the accuracy of user demand characterization. By integrating user static interest vectors (basic attributes), explicit interest vectors (features of learned knowledge points), and graph-extended interest vectors (first-order neighbor features of the knowledge graph), a comprehensive user interest representation is constructed. This not only covers stable user attributes and direct learning preferences, but also mines potential related needs through the knowledge graph, solving the problem of one-sided interest characterization caused by traditional single-feature modeling, and making interest vectors more in line with real learning needs.

[0201] Secondly, the dual-path candidate generation of "similarity analysis + large language model" enhances the adaptability of recommendations. A flexible method for selecting candidate knowledge points is provided through an "AND / OR" mechanism: the vector similarity-based path can efficiently match knowledge points directly related to user interests; the large language model-based path can deeply analyze user-related information (historical behavior, static attributes, learned knowledge points, and learning goals) to uncover potential learning needs (such as knowledge points that the user has not explicitly expressed but which align with their goals). The combination of these two paths ensures recommendation accuracy while improving adaptability to complex scenarios (such as the differentiated needs of goal-oriented and exploratory users).

[0202] Third, multi-layered screening and intelligent sorting ensure the rationality and personalization of learning. The selection of target knowledge points is verified by the user's prior knowledge mastery, avoiding recommendations of content beyond the user's current capabilities. During the sorting phase, priorities are dynamically adjusted based on multiple features such as user interests and the relevance of knowledge points, ensuring that the recommended results conform to cognitive logic and align with the user's learning pace and preferences.

[0203] In summary, this invention effectively improves the accuracy (matching interests and needs), diversity (covering explicit and potential needs), and learning effectiveness (conforming to cognitive laws) of knowledge point recommendations through a collaborative mechanism of multi-source interest fusion, dual-path candidate generation, rationality screening, and intelligent ranking, providing technical support for personalized online education. Based on the same inventive concept, embodiments of this invention provide a recommendation system that integrates large language model reasoning and multi-source trajectory information, the system comprising:

[0204] The static interest vector construction module is used to obtain users' historical learning behavior information and static attribute information on the target learning platform, and to construct users' static interest vectors based on the obtained users' historical learning behavior information and static attribute information.

[0205] The explicit interest vector acquisition module is used to obtain the initial knowledge point feature vectors of the knowledge points that the user has already learned, and generate explicit interest vectors through average pooling.

[0206] The graph-enhanced interest vector acquisition module is used to obtain the graph-enhanced knowledge point vectors of the first-order neighbor nodes of the learned knowledge points from the knowledge graph G of the target learning platform, based on the learned knowledge points, and generate graph-enhanced interest vectors through average pooling; where G = (V, E, W), where V represents the set of knowledge nodes, E represents the set of edges, and W represents the edge weight matrix; the feature representation of each node in G is: the graph-enhanced knowledge point vector obtained by inputting the initial knowledge point feature vector of the corresponding knowledge point into the graph attention network.

[0207] The user interest vector acquisition module is used to generate user interest vectors from the user static interest vectors, the explicit interest vectors, and the graph extended interest vectors.

[0208] The candidate knowledge point acquisition module, based on the similarity analysis between the user's interest vector and the knowledge point vectors of unlearned knowledge points, and / or based on the analysis of user association information using a large language model, acquires N unlearned knowledge points that the user needs to learn from the full set of unlearned knowledge points, forming a candidate knowledge point set; wherein, the user association information includes at least the user's historical learning behavior information, static attribute information, learned knowledge points, and learning objectives. The target knowledge point acquisition module is used to filter the target knowledge point set from the candidate knowledge point set based on the user's interest vector, the knowledge point vectors of candidate knowledge points, the knowledge point vectors of prerequisite knowledge points for candidate knowledge points, and the historical behavioral characteristics of learned knowledge points;

[0209] The recommendation result generation module is used to sort the target knowledge points in the target knowledge point set and output the sorting results as the recommendation results.

[0210] This system can be used to perform Figure 1 The method shown in the illustrated embodiment can be used as a reference for understanding the functions that each functional module of the system can achieve. Figure 1 The embodiments shown are described in detail below.

[0211] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0212] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0213] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0214] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A recommendation method integrating large language model reasoning and multi-source trajectory information, characterized in that, The method is applied to the knowledge point recommendation scenario of an online education platform. Upon receiving a recommendation request from a user, the following steps are performed: S100: Obtain the user's historical learning behavior information and static attribute information on the target learning platform, and construct the user's static interest vector based on the obtained user's historical learning behavior information and static attribute information; S200: Obtain the initial knowledge point feature vector of the user's learned knowledge points, and generate an explicit interest vector through average pooling; based on the learned knowledge points, obtain the graph-enhanced knowledge point vectors of the first-order neighbor nodes of the learned knowledge points from the knowledge graph G of the target learning platform, and generate a graph-extended interest vector through average pooling; where G = (V, E, W), where V represents the set of knowledge nodes, E represents the set of edges, and W represents the edge weight matrix; the feature representation of each node in G is: the graph-enhanced knowledge point vector obtained by inputting the initial knowledge point feature vector of the corresponding knowledge point into the graph attention network; the first-order neighbor nodes of the learned knowledge points are specifically the first-order neighbor nodes whose edge type with the learned knowledge point is a pre-relational edge or a semantically similar edge. S300, the user static interest vector, the explicit interest vector, and the graph extended interest vector are fused to generate a user interest vector; S400, based on the similarity analysis between the user interest vector and the knowledge point vector of the unlearned knowledge points, and / or based on the analysis of user association information based on the large language model, obtain N unlearned knowledge points that the user needs to learn from the full set of unlearned knowledge points of the user, and form a candidate knowledge point set; wherein, the user association information includes at least user historical learning behavior information, static attribute information, learned knowledge points and learning objectives; S500 filters the target knowledge point set from the candidate knowledge point set based on user interest vector, knowledge point vector of candidate knowledge points, knowledge point vector of prerequisite knowledge points of candidate knowledge points, and historical behavioral characteristics of learned knowledge points. S600 sorts the target knowledge points in the target knowledge point set and outputs the sorting result as the recommendation result.

2. The method according to claim 1, characterized in that, The user static interest vector, the explicit interest vector, and the graph extended interest vector are fused using a gating fusion mechanism. The user interest vector satisfies the following condition: u = g × u base +(1-g)×(u graph +u static ); where u is the user interest vector, u base For the explicit interest vector, u graph To extend the interest vector in the graph, u static denoted as the user's static interest vector, and g as the fusion gating parameter.

3. The method according to claim 1, characterized in that, The user interest vector satisfies the following condition: u = g × u base +(1-g)×(α×u graph +(1-α)×u static ), where u is the user interest vector, u base For the explicit interest vector, u graph To extend the interest vector in the graph, u static Let g be the user's static interest vector, g be the fusion gating parameter, and α be the adjustment coefficient.

4. The method according to claim 2 or 3, characterized in that, The fusion gate parameter g satisfies: g = sigmoid(W g [u base ;u graph ;u static ]+b g ), sigmoid() is the activation function, [; ] indicates vector concatenation, W g b is the weight parameter of the user tower fusion layer. g This is the bias term for the fusion layer.

5. The method according to claim 1, characterized in that, The S500 specifically includes: S510: Input the user's interest vector, the knowledge point vector of the candidate knowledge points, and the historical behavioral features of the learned knowledge points into the trained mastery prediction model, and output the user's mastery prediction value for each candidate knowledge point. S520 inputs the user's interest vector, the knowledge point vector of the prerequisite knowledge points of each candidate knowledge point, and the historical behavioral features of the learned knowledge points into the trained mastery prediction model, and outputs the user's mastery prediction value of each prerequisite knowledge point. S530, perform the following judgment on a single candidate knowledge point: if the predicted mastery values ​​of all prerequisite knowledge points of the candidate knowledge point are greater than the preset mastery threshold, then include the candidate knowledge point in the target knowledge point set; if the predicted mastery value of any prerequisite knowledge point of the candidate knowledge point is less than or equal to the preset mastery threshold, then remove the candidate knowledge point from the candidate knowledge point set.

6. The method according to claim 4, characterized in that, The S600 specifically includes: S610, construct a ranking feature vector for each target knowledge point, input the ranking feature vector into the trained ranking score prediction model, and output the ranking score of each target knowledge point; wherein, the ranking feature vector includes at least: the knowledge point vector of the target knowledge point, the user interest vector, the similarity between the knowledge point vector of the target knowledge point and the user interest vector, the mastery prediction value of the target knowledge point, the association strength between the target knowledge point and the user's learned knowledge points in the knowledge graph, the mean of the mastery prediction values ​​of all prerequisite knowledge points of the target knowledge point, and the number of times the user has clicked on knowledge points of the same type as the target knowledge point in history; S620: Sort the target knowledge points in descending order according to their ranking scores, and output the sorted list of target knowledge points as the recommendation result.

7. The method according to claim 1, characterized in that, The S400 specifically includes: S401, obtain the knowledge point vector corresponding to each of the unlearned knowledge points of all users; S402, calculate the similarity between the user interest vector and the knowledge point vector of each unlearned knowledge point to obtain the similarity result corresponding to each unlearned knowledge point; S403, Sort all unlearned knowledge points according to the similarity results, and select the top N unlearned knowledge points after sorting to form a candidate knowledge point set.

8. The method according to claim 1, characterized in that, The S400 specifically includes: S420, Based on the similarity between the user interest vector and the knowledge point vector of the user's unlearned knowledge points, obtain a first knowledge point set containing N unlearned knowledge points from the full set of user unlearned knowledge points; S421, the user's historical learning behavior information, static attribute information, list of learned knowledge points and learning objectives are concatenated into structured input text. The structured input text is combined with preset prompt words and then input into the large language model to obtain an initial knowledge point set consisting of K unlearned knowledge points; where K is a preset positive integer. S422, Perform feature encoding on each knowledge point in the initial knowledge point set to generate a knowledge point vector corresponding to each knowledge point; S423, calculate the similarity between the knowledge point vector of each knowledge point in the initial knowledge point set and the knowledge point vectors in the preset knowledge point vector library, determine the real unlearned knowledge points matched by each knowledge point in the initial knowledge point set in the preset knowledge point vector library, and form a second knowledge point set; S424, take the union of the first knowledge point set and the second knowledge point set and remove duplicate knowledge points, and use the deduplicated set as the candidate knowledge point set.

9. A recommendation system that integrates large language model reasoning and multi-source trajectory information, characterized in that, The system includes: The static interest vector construction module is used to obtain users' historical learning behavior information and static attribute information on the target learning platform, and to construct users' static interest vectors based on the obtained users' historical learning behavior information and static attribute information. The explicit interest vector acquisition module is used to obtain the initial knowledge point feature vector of the user's learned knowledge points, and generate explicit interest vectors through average pooling. The graph-enhanced interest vector acquisition module is used to obtain graph-enhanced knowledge point vectors of the first-order neighbor nodes of the learned knowledge points from the knowledge graph G of the target learning platform based on the learned knowledge points, and generate graph-enhanced interest vectors through average pooling; where G = (V, E, W), where V represents the set of knowledge nodes, E represents the set of edges, and W represents the edge weight matrix; the feature representation of each node in G is: the graph-enhanced knowledge point vector obtained by inputting the initial knowledge point feature vector of the corresponding knowledge point into the graph attention network; the first-order neighbor nodes of the learned knowledge points are specifically the first-order neighbor nodes whose edge type with the learned knowledge point is a pre-relational edge or a semantically similar edge. The user interest vector acquisition module is used to generate a user interest vector from the user static interest vector, the explicit interest vector, and the graph extended interest vector; The candidate knowledge point acquisition module, based on the similarity analysis between the user's interest vector and the knowledge point vector of the unlearned knowledge points, and / or based on the analysis of user association information based on the large language model, acquires N unlearned knowledge points that the user needs to learn from the full set of unlearned knowledge points of the user, forming a candidate knowledge point set; wherein, the user association information includes at least the user's historical learning behavior information, static attribute information, learned knowledge points and learning objectives; The target knowledge point acquisition module is used to filter the target knowledge point set from the candidate knowledge point set based on the user interest vector, the knowledge point vector of the candidate knowledge point, the knowledge point vector of the prerequisite knowledge point of the candidate knowledge point, and the historical behavioral characteristics of the learned knowledge point. The recommendation result generation module is used to sort the target knowledge points in the target knowledge point set and output the sorting results as the recommendation results.

Images