Course selection recommendation method based on hybrid algorithm

Abstract

The invention discloses a course selection recommendation method based on a hybrid algorithm, and the method comprises the steps: obtaining a course candidate recommendation sub-list through a parallel hybrid recommendation strategy in combination with a multi-source data set, and collecting and obtaining the dynamic weight data of the course candidate recommendation sub-list based on a user portrait according to a preset user target demand; taking the user portrait and the user target demand as feature data, taking the dynamic weight data as label data to obtain a data sample set, and training a preset lightweight weight prediction model according to the data sample set to obtain a course recommendation model so as to obtain dynamic weight data of a course candidate recommendation sub-list; and obtaining an individualized course selection recommendation list based on the score value and the dynamic weight data so as to realize recommended course selection for the user. The problems that in the prior art, deviation exists in single-dimension recommendation, comprehensive course recommendation logic cannot be considered in multiple aspects, recommendation precision is suddenly lowered under a new course or interdisciplinary scene, and interpretability is lacked are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart campuses, and particularly to a course selection recommendation method based on a hybrid algorithm. Background Art

[0002] Traditional course selection recommendation systems mostly rely on manual rules (such as credit limits and prerequisite requirements) or single recommendation algorithms, resulting in defects such as serious homogenization of recommendation results, prominent cold start problems, and poor dynamic adaptability. In recent years, recommendation system technology has been introduced into the education field, but still faces core challenges such as insufficient multi-dimensional data fusion and single recommendation logic in actual course selection scenarios.

[0003] For example, content-based recommendation extracts semantic features of course description texts through natural language processing (NLP) to recommend courses with similar content. However, it ignores the differences in user behavior and grades, and the recommendation results are disconnected from academic goals. In the collaborative filtering algorithm based on grades, there are problems such as high data sparsity, difficulty in constructing a reliable scoring matrix, and deviation in similarity calculation caused by abnormal grades (such as failing or getting full marks). Generally speaking, the main defects of the existing technologies are: single-dimensional recommendation deviation, inability to construct a comprehensive recommendation logic, sharp drop in recommendation accuracy in the scenario of new courses or interdisciplinary courses, and lack of interpretability, etc. Summary of the Invention

[0004] The present invention provides a course selection recommendation method based on a hybrid algorithm to overcome the above technical problems.

[0005] To achieve the above object, the technical solution of the present invention is:

[0006] A course selection recommendation method based on a hybrid algorithm specifically includes the following steps:

[0007] S1: Obtain a multi-source data set;

[0008] The multi-source data set includes a user database, a course and grade database, and a software engineering professional knowledge base;

[0009] The user database includes a user interest text vector and a user ability matrix vector;

[0010] The user interest text vector is used to represent the user's interested courses according to the user's historical course selection records and course browsing durations; the user ability matrix vector is used to represent the user's learning ability according to the user's grade data and behavior logs; and the behavior logs at least include learning progress;

[0011] The course and grade database includes a course feature vector used to represent the user's historical learning;

[0012] The course feature vector includes a course text keyword vector and a course scoring matrix vector; the course text keyword vector is used to represent different course text keywords extracted through NLP; the course scoring matrix vector is used to represent a course scoring matrix that statistically calculates the course popularity and the difficulty level of user feedback based on the course text keywords; and the course popularity is the number of times users select courses.

[0013] The software engineering professional knowledge base includes a potential relationship vector of course capabilities.

[0014] The potential relationship vector of course capabilities is used to represent a professional knowledge graph obtained based on the known professional training plan, course prerequisite relationship graph, and industry skill tree, with elective or compulsory courses as graph nodes and the dependence relationship between courses as edges.

[0015] S2: Based on the user's target requirements, obtain a multi-dimensional dynamic knowledge requirement vector for user target requirement modeling according to the large language model LLM combined with the software engineering professional knowledge base.

[0016] S3: Based on the multi-dimensional dynamic knowledge requirement vector, obtain a course candidate recommendation sub-list according to the parallel hybrid recommendation strategy combined with multi-source data sets.

[0017] The course candidate recommendation sub-list includes an ability-driven recommended course sub-list, a content-driven recommended course sub-list, a performance-driven recommended course sub-list, and a group behavior-driven recommended course sub-list.

[0018] The ability-driven recommended course sub-list is used to represent the courses that are short boards for users recommended based on the user ability matrix vector according to the gap between the matching course ability requirements and the user's current ability.

[0019] The content-driven recommended course sub-list is used to represent the interest-matching courses recommended based on the cosine similarity according to the user interest text vector and the course text keyword vector.

[0020] The performance-driven recommended course sub-list is used to represent the courses recommended by the user group scoring based on the course scoring matrix vector and the user's performance distribution.

[0021] The group behavior-driven recommended course sub-list is used to represent the courses with high-frequency course selection by similar user groups obtained by clustering the user ability matrix vector and recommended for the current user who has not selected them.

[0022] S4: Obtain the user portrait according to the user database, course and grade database.

[0023] Set the preset user target requirements, and collect and obtain the dynamic weight data of the course candidate recommendation sub-list obtained based on the parallel hybrid recommendation strategy based on the user portrait.

[0024] Using the user profile and the user's target requirements as feature data, and using the dynamic weight data as label data to obtain a data sample set;

[0025] S5: Training a preset lightweight weight prediction model according to the data sample set to obtain a course recommendation model; and the lightweight weight prediction model is a dynamic weight prediction model obtained using a machine learning model or a regression model of a given function form;

[0026] Based on the course recommendation model, according to the user profile and the multi-dimensional dynamic knowledge requirement vector, predicting and obtaining the dynamic weight data of the course candidate recommendation sub-list;

[0027] S6: Setting a score value for each recommended course in the course candidate recommendation sub-list according to the empirical value;

[0028] And obtaining an individualized course selection recommendation list based on the score value and the dynamic weight data, and realizing the recommended course selection for the user according to the individualized course selection recommendation list.

[0029] Further, step S3 specifically includes the following steps:

[0030] S31: Based on the multi-dimensional dynamic knowledge requirement vector, obtaining a course candidate recommendation sub-list according to the parallel hybrid recommendation strategy in combination with the multi-source data set;

[0031] And the parallel hybrid recommendation strategy includes a recommendation strategy based on professional ability, a recommendation strategy based on course content, a recommendation strategy based on user grades, and a recommendation strategy based on group behavior clustering;

[0032] S32: Obtaining course information based on the course text keyword vector;

[0033] And the course information at least includes the course category, the course author, and the course description text;

[0034] Through the recommendation strategy based on professional ability, according to the user ability matrix vector and the course information, obtaining the ability gap between the matching course ability requirements for the user and the user's current ability, so as to recommend the user's short-board courses according to the ability gap, that is, the ability-driven recommended course sub-list;

[0035] S33: Through the recommendation strategy based on course content, obtaining a content-driven recommended course sub-list according to the user interest text vector and the course information;

[0036] S34: Through the recommendation strategy based on user scores, recommend according to the course rating matrix vector and the user score distribution. Use the Pearson correlation coefficient to obtain the correlation index measuring the relationship between the user's course ratings and the score distribution. Based on the correlation index, obtain the courses recommended by the user group ratings, and then obtain the sub-list of courses recommended by score drive.

[0037] S35: Through the recommendation strategy based on group behavior clustering, that is, according to the user ability matrix vector, use the K-means clustering algorithm to obtain the similar user groups of the courses. To recommend the courses with high-frequency course selection by the similar user groups and not selected by the current user, and obtain the sub-list of courses recommended by group behavior drive.

[0038] Furthermore, the recommendation strategy based on professional ability in S32 specifically includes the following steps:

[0039] S321: Perform knowledge extraction on the course information based on the course text keyword vector to obtain the course information, and define the course information as a knowledge entity.

[0040] And the course information includes at least the course category, the course author, and the course description text.

[0041] And based on the potential relationship vector of course capabilities, perform entity connection on each knowledge entity to obtain the entity connection relationship.

[0042] S322: Define the course entity of the user data structure and the course entity, and based on the entity connection relationship, obtain the user course data structure model according to the user ability matrix vector.

[0043] And the user course data structure model is described as:

[0044] Data structure course entity → Entity connection relationship →...... → Course entity →...... → Entity connection relationship →...... → Course entity →......;

[0045] S323: Based on the user course data structure model, store the course information in the storage format of a relational database, obtain the gap between the data structure course and the user's current ability according to the user ability matrix vector, and obtain the sub-list of courses recommended by ability drive according to the current ability gap.

[0046] Furthermore, the recommendation strategy based on course content in S33 specifically includes the following steps:

[0047] S331: Obtain the attribute text vector about the course information according to the course information.

[0048] S332: Based on the TF-IDF text processing method, perform keyword extraction on the user interest text vector and the attribute text vector to obtain the user preference vector and the course attribute vector;

[0049] Calculate and obtain the cosine similarity between the user preference vector and the course attribute vector;

[0050] The course sequences sorted in descending order of cosine similarity are used as the content-driven recommended course sublist.

[0051] Furthermore, S6 also includes a method for optimizing a course recommendation model based on user behavior feedback, which specifically includes the following steps:

[0052] S61: Count and obtain user behavior data;

[0053] The user behavior data is used to characterize the number of clicks / course selection behaviors or the adoption rate of the personalized course selection recommendation list;

[0054] S62: Determine whether the acquired user behavior data meets a preset behavior data threshold;

[0055] If satisfied, the current course recommendation model is confirmed to meet the user's needs and is used as the optimal course recommendation model;

[0056] Otherwise, based on historical user target needs and user portraits, the dynamic weight data of the course candidate recommendation sublist obtained based on the parallel hybrid recommendation strategy is obtained as a new data sample set, and the current course recommendation model is retrained to obtain the model weight parameters of the current course recommendation model to achieve feedback optimization of the course recommendation model.

[0057] Beneficial effects: The present invention provides a course recommendation method based on a hybrid algorithm, which obtains a candidate course recommendation sublist by combining a parallel hybrid recommendation strategy with a multi-source data set; and sets a preset user target requirement, and collects and obtains the dynamic weight data of the candidate course recommendation sublist obtained based on the parallel hybrid recommendation strategy based on the user portrait; and uses the user portrait and the user target requirement as feature data, and the dynamic weight data as label data to obtain a data sample set, and trains a preset lightweight weight prediction model according to the data sample set to obtain a course recommendation model to obtain the dynamic weight data of the candidate course recommendation sublist; and obtains an individualized course recommendation list based on the score value and the dynamic weight data, and implements the recommended course selection for users according to the individualized course recommendation list. It solves the problem that the single-dimensional recommendation of the prior art has deviations and cannot consider the recommendation logic of comprehensive courses from multiple aspects, and greatly improves the accuracy and interpretability of course recommendations in new courses or interdisciplinary scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0059] Figure 1 It is a flowchart of the course selection recommendation method based on the hybrid algorithm of the present invention;

[0060] Figure 2 It is a technical process block diagram of the course selection recommendation method based on the hybrid algorithm in this embodiment;

[0061] Figure 3 It is a schematic diagram of the process of the parallel hybrid recommendation strategy in this embodiment. Detailed implementation manners

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0063] This embodiment provides a course selection recommendation method based on a hybrid algorithm, as Figures 1 to 2 shown, and specifically includes the following steps:

[0064] S1: Obtain a multi-source data set;

[0065] The multi-source data set includes a user database, a course and grade database, and a software engineering professional knowledge base;

[0066] The user database includes a user interest text vector and a user ability matrix vector;

[0067] The user interest text vector is used to represent the user's interested courses according to the user's historical course selection records and course browsing durations; the user ability matrix vector is used to represent the user's learning ability according to the user's grade data and behavior logs; and the behavior logs at least include the learning progress;

[0068] Specifically, the historical course selection records, grade data, and behavior logs (such as course browsing durations) are obtained by inputting the user ID through the existing course selection system, and the abnormal data (such as invalid course selection records) is cleaned and processed, and structured tags (such as the interest tag "artificial intelligence preference") are extracted to obtain the user interest text vector and the user ability matrix vector;

[0069] The course and grade database includes a course feature vector used to characterize the user's historical learning.

[0070] The course feature vector includes a course text keyword vector and a course scoring matrix vector; the course text keyword vector is used to characterize different course text keywords extracted through NLP; the course scoring matrix vector is used to characterize a course scoring matrix that statistically calculates the course popularity and the difficulty level of user feedback based on the course text keywords; and the course popularity is the number of times users select the course.

[0071] Specifically, the existing course selection system inputs the course ID, statistically obtains the course description text, credit requirements, and historical user grade distribution, extracts course keywords (such as "machine learning", "Java programming") through NLP, constructs a course feature vector; statistically calculates the course popularity and difficulty score, and outputs a course feature library (content vector, scoring matrix).

[0072] The software engineering professional knowledge base includes a potential relationship vector of course capabilities.

[0073] The potential relationship vector of course capabilities is used to characterize a professional knowledge graph obtained based on the known professional training plan, course prerequisite relationship graph, and industry skill tree, with elective or required courses as graph nodes and the dependence relationship between courses as edges.

[0074] Specifically, through the existing knowledge graph construction system, by inputting the professional training plan, course prerequisite relationship graph, and industry skill tree (such as "skill requirements for software test engineers"), a knowledge graph (course → ability mapping) is obtained, and the ability weight is defined (such as the weight of "algorithm design ability" is 0.3) to obtain a professional knowledge graph (node = course / ability, edge = dependence relationship).

[0075] S2: Based on the user's target requirements, a multi-dimensional dynamic knowledge requirement vector for user target requirement modeling is obtained according to the large language model LLM combined with the software engineering professional knowledge base.

[0076] Specifically, in this embodiment, the large language model LLM dynamically analyzes the user's interest and ability requirements to obtain a personalized knowledge requirement vector, that is, the user's natural language question (such as "I want to become a full-stack engineer. Which courses should I choose?") + user portrait (obtaining the user portrait according to the user database, course and grade database) is used as input, and the large language model LLM performs interactive processing on the input, which includes the LLM parsing the semantics of the question (such as "target position = full-stack engineer"), and at the same time combining the professional knowledge base to match the skill requirements (such as "need to master front-end + back-end development") to obtain a multi-dimensional dynamic knowledge requirement vector (such as [front-end development: 0.8, back-end development: 0.7, algorithm foundation: 0.5]).

[0077] S3: Based on the multi-dimensional dynamic knowledge demand vector, obtain the course candidate recommendation sub-list according to the parallel hybrid recommendation strategy by combining multi-source data sets;

[0078] The course candidate recommendation sub-list includes an ability-driven recommended course sub-list, a content-driven recommended course sub-list, a performance-driven recommended course sub-list, and a group behavior-driven recommended course sub-list;

[0079] The ability-driven recommended course sub-list is used to represent the courses with user's weaknesses recommended based on the user ability matrix vector according to the gap between the matching course ability requirements and the user's current ability;

[0080] The content-driven recommended course sub-list is used to represent the interest-matching courses recommended based on the cosine similarity according to the user interest text vector and the course text keyword vector;

[0081] The performance-driven recommended course sub-list is used to represent the user group rating recommended courses recommended based on the course rating matrix vector and the user performance distribution;

[0082] The group behavior-driven recommended course sub-list is used to represent the courses with high-frequency course selections recommended by similar user groups obtained by clustering the user ability matrix vector and not selected by the current user;

[0083] Such as Figure 3 shown, specifically including the following steps:

[0084] S31: Based on the multi-dimensional dynamic knowledge demand vector, obtain the course candidate recommendation sub-list according to the parallel hybrid recommendation strategy by combining multi-source data sets;

[0085] And the parallel hybrid recommendation strategy includes a recommendation strategy based on professional ability, a recommendation strategy based on course content, a recommendation strategy based on user performance, and a recommendation strategy based on group behavior clustering;

[0086] S32: Obtain course information based on the course text keyword vector;

[0087] And the course information includes at least course category, course author, and course description text;

[0088] Through the recommendation strategy based on professional ability, according to the user ability matrix vector and the course information, obtain the gap between the matching course ability requirements and the user's current ability for the user, so as to recommend the courses with user's weaknesses according to the ability gap, that is, the ability-driven recommended course sub-list;

[0089] In a specific embodiment, the recommendation strategy based on professional ability is:

[0090] S321: Extract knowledge from the course information based on the keyword vectors of the course text to obtain the course information, and define the course information as a knowledge entity;

[0091] And the course information includes at least the course category, course author, course description text, and course form;

[0092] And based on the potential relationship vectors of the course capabilities, connect the knowledge entities to obtain the entity connection relationships; for example, the relationship between a course and a course type, the relationship between a course and a user, etc.

[0093] S322: In this embodiment, the knowledge entity representation abstracts things, concepts, relationships, etc. in the real world into forms that can be understood and processed by a computer. In the knowledge entity representation, the knowledge entities, attributes, and relationships are transformed into data structures that can be processed by a computer;

[0094] In this implementation, define the user data structure course entity and the course entity, and based on the entity connection relationship, obtain the user course data structure model according to the user ability matrix vector;

[0095] And the user course data structure model is described as:

[0096] Data structure course entity → Entity connection relationship →...... → Course entity →...... → Entity connection relationship →...... → Course entity →......;

[0097] For example: Entity (Data structure course) → Relationship (Belongs to) → Entity (Compulsory course), where the attribute of the data structure course as the entity is grade = 73, and the attribute of the compulsory course as the entity is the second semester of sophomore year;

[0098] S323: Based on the user course data structure model, store the course information in the storage format of the relational database mysql, obtain the gap between the data structure course and the user's current ability according to the user ability matrix vector, and obtain the ability-driven recommended course sub-list according to the user's current ability gap;

[0099] For example: Based on the professional knowledge graph and the user ability matrix, obtain the gap between the matching course ability requirements and the user's current ability, and recommend courses to fill the ability shortcoming (such as "Lack of algorithm design ability → Recommend 'Data Structure'") to obtain the ability-driven recommended sub-list;

[0100] S33: Through the recommendation strategy based on the course content, obtain the content-driven recommended course sub-list according to the user interest text vector and the course information;

[0101] In a specific embodiment, the recommendation strategy based on the course content is:

[0102] S331: Obtain the attribute text vector of the course information according to the course information;

[0103] For example: Obtain the feature representation of the course, that is, the attribute text vector, according to the relevant information of the course (course category, course description, etc.);

[0104] S332: Based on the TF-IDF text processing method, extract keywords from the user interest text vector and the attribute text vector to obtain the user preference vector and the course attribute vector;

[0105] Specifically, since the user interest text vector and the attribute text vector mentioned in this embodiment are usually represented by words, that is, keywords are used to represent the features of the project, the weights of each project feature can be determined according to TF-IDF (term frequency-inverse document frequency), and then their similarity can be calculated;

[0106] Among them, the term frequency-inverse document frequency (TF-IDF) in this embodiment is a statistical method commonly used in text processing, which can evaluate the importance of a word in a document. Simply put, it can be used for keyword extraction of documents. The basic idea of TF-IDF is: If a certain word appears frequently in an article and rarely appears in other articles, it is considered that the word is probably a keyword;

[0107] Term Frequency (TF): The frequency of the word w appearing in the document d. The idea of Inverse Document Frequency (IDF) is: If a word appears in many documents, it means that the importance of the word is not high; on the contrary, it means that the importance of the word is very high. Mainly considering the importance of the word, the IDF calculation method of the word w is as follows:

[0108]

[0109] In the formula: N represents the total number of documents in the corpus; N(w) represents the number of documents in which the word w appears; the larger the number of documents, and the fewer documents in which the word appears, the larger the IDF value, indicating that the word is more important; With the above definitions of TF and IDF in this embodiment, TF-IDF is the product of the above two quantities. The TF-IDF calculation method of the word w is as follows:

[0110] TF-IDF = TF(d, w) × IDF(w)

[0111] The larger the TF-IDF value of a word, the more important the word is. A high term frequency within a specific document and a low document frequency of the term across the entire document collection can result in a high-weight TF-IDF. Therefore, TF-IDF tends to filter out common words and retain important ones;

[0112] Based on the vector representation of the subject matter calculated by TF-IDF, it is very easy in this embodiment to obtain the similarity between the user interest preference vector and the course attribute vector. Next is to make personalized recommendations for the user. Assume the user interest preference vector is u and the course feature vector is p. The similarity between the user interest preference vector and the course attribute vector can be obtained using a similarity calculation formula such as cosine similarity;

[0113] And the expression for the similarity cos(u, p) between the user interest preference vector and the course attribute vector is

[0114]

[0115] Calculate and obtain the cosine similarity between the user preference vector and the course attribute vector;

[0116] Take the course sequence sorted in descending order of cosine similarity as the content-driven recommended course sub-list; By calculating and comparing the similarity between the user interest preference vector and the attribute vectors of each unevaluated course, generate a project prediction score or top-N recommendation for the target user. By calculating and comparing the similarity between the user interest preference vector and the attribute vectors of each unevaluated course, generate a project prediction score or top-N recommendation for the target user.

[0117] For example: Based on the course feature library and the user interest vector, calculate the course content similarity (cosine similarity) to obtain recommended interest-matching courses (such as "preference for Python → recommend 'Advanced Python Programming'") to obtain the content-driven recommended sub-list;

[0118] S34: Through a recommendation strategy based on user grades, recommend according to the course rating matrix vector and the user grade distribution. Use the Pearson correlation coefficient to obtain a correlation index measuring the relationship between the user's course rating and the grade distribution. Based on the correlation index, obtain the courses recommended by the user group rating, and then obtain the grade-driven recommended course sub-list;

[0119] And the expression for obtaining the correlation index measuring the relationship between the user's course rating and the user grade distribution using the Pearson correlation coefficient is

[0120]

[0121] In the formula: sim(X, Y) represents the correlation index between the user's course rating and the user grade distribution; It represents the product of the covariance of the user's course ratings and the covariance of the user's grade distribution; It represents the product of the standard deviation of the user's course ratings and the standard deviation of the user's grade distribution;

[0122] For example: Based on the course rating matrix vector and the user's grade distribution, the Pearson correlation coefficient is used to obtain a correlation index measuring the relationship between the user's course ratings and the grade distribution, construct a user-course rating matrix to fill in sparse data (matrix factorization), and recommend high-rated courses for similar user groups based on the user-course rating matrix to obtain a grade-driven recommended course sub-list;

[0123] S35: Through a recommendation strategy based on group behavior clustering, that is, according to the user ability matrix vector, the K-means clustering algorithm is used to obtain similar user groups for the courses, and based on the frequently selected courses recommended by the similar user groups and not selected by the current user, obtain a group behavior-driven recommended course sub-list;

[0124] For example: Based on the user behavior log and the clustered similar user groups (K-means), potential interest associations are mined, and courses that are frequently selected by the group but not selected by the current user are recommended to obtain a group behavior-driven recommended course sub-list;

[0125] S4: Obtain a user portrait based on the user database, course, and grade database;

[0126] Set a preset user target requirement, and collect and obtain dynamic weight data of the course candidate recommended sub-list obtained based on the parallel hybrid recommendation strategy based on the user portrait;

[0127] And use the user portrait and the user target requirement as feature data, and the dynamic weight data as label data to obtain a data sample set;

[0128] S5: And train a preset lightweight weight prediction model according to the data sample set to obtain a course recommendation model; and the lightweight weight prediction model is a dynamic weight prediction model obtained using a machine learning model or a regression model of a given function type;

[0129] And based on the course recommendation model, according to the user portrait and the multi-dimensional dynamic knowledge requirement vector, predict and obtain the dynamic weight data of the course candidate recommended sub-list;

[0130] For example, the proportion of the dynamic weight data of the ability-driven recommended course sub-list obtained by the recommendation strategy based on professional ability: the content-driven recommended course sub-list obtained by the recommendation strategy based on course content: the grade-driven recommended course sub-list obtained by the recommendation strategy based on user grades: the group behavior-driven recommended course sub-list obtained by the recommendation strategy based on group behavior clustering = 0.2:0.2:0.3:0.3;

[0131] S6: Set a score value for each recommended course in the candidate recommended course sub - list according to the empirical value;

[0132] And obtain an individualized course selection recommendation list based on the score value and the dynamic weight data, and implement the recommended course selection for the user according to the individualized course selection recommendation list;

[0133] In a specific embodiment, S6 further includes a method for optimizing the course recommendation model based on user behavior feedback, which specifically includes the following steps:

[0134] S61: Statistically obtain user behavior data;

[0135] The user behavior data is used to characterize the number of click / course selection behaviors or the adoption rate of the individualized course selection recommendation list;

[0136] S62: Determine whether the obtained user behavior data meets a preset behavior data threshold;

[0137] If it is satisfied, confirm that the current course recommendation model meets the user's needs and use it as the optimal course recommendation model;

[0138] Otherwise, based on the historical user target needs, obtain the dynamic weight data of the course candidate recommended sub - list obtained according to the user portrait based on the parallel hybrid recommendation strategy as a new data sample set, and retrain the current course recommendation model to obtain the model weight parameters of the current course recommendation model to achieve the feedback optimization of the course recommendation model.

[0139] Compared with the prior art, the beneficial effects of the method in this embodiment are as follows: Obtain the course candidate recommended sub - list through the parallel hybrid recommendation strategy combined with multi - source data sets; set the preset user target needs, and collect and obtain the dynamic weight data of the course candidate recommended sub - list obtained according to the user portrait based on the parallel hybrid recommendation strategy; and use the user portrait and user target needs as feature data, and the dynamic weight data as label data to obtain a data sample set, train the preset lightweight weight prediction model according to the data sample set to obtain a course recommendation model to obtain the dynamic weight data of the course candidate recommended sub - list; and obtain an individualized course selection recommendation list based on the score value and the dynamic weight data, and implement the recommended course selection for the user according to the individualized course selection recommendation list. It solves the problem that the single - dimension recommendation in the prior art has deviations and cannot consider the recommendation logic of comprehensive courses in multiple aspects, and greatly improves the accuracy and interpretability of course recommendation in new course or interdisciplinary scenarios.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A course selection recommendation method based on a hybrid algorithm, characterized in that, Specifically, it includes the following steps: S1: Obtain a multi-source data set; The multi-source data set includes a user database, a course and grade database, and a software engineering professional knowledge base; The user database includes a user interest text vector and a user ability matrix vector; The user interest text vector is used to represent the user's interested courses based on the user's historical course selection records and course browsing durations; the user ability matrix vector is used to represent the user's learning ability based on the user's grade data and behavior logs; and the behavior logs at least include the learning progress; The course and grade database includes a course feature vector used to represent the user's historical learning; The course feature vector includes a course text keyword vector and a course rating matrix vector; the course text keyword vector is used to represent different course text keywords extracted through NLP; the course rating matrix vector is used to represent a course rating matrix for statistically calculating the course popularity and the user feedback difficulty level based on the course text keywords; and the course popularity is the number of times the user selects a course; The software engineering professional knowledge base includes a course ability potential relationship vector; The course ability potential relationship vector is used to represent a professional knowledge graph obtained based on the known professional training plan, course prerequisite relationship graph, and industry skill tree, with elective or compulsory courses as graph nodes and the dependence relationship between courses as edges; S2: Based on the user's target requirements, obtain a multi-dimensional dynamic knowledge requirement vector for user target requirement modeling according to the large language model LLM combined with the software engineering professional knowledge base; S3: Based on the multi-dimensional dynamic knowledge requirement vector, obtain a course candidate recommendation sub-list according to the parallel hybrid recommendation strategy combined with the multi-source data set; The course candidate recommendation sub-list includes an ability-driven recommended course sub-list, a content-driven recommended course sub-list, a grade-driven recommended course sub-list, and a group behavior-driven recommended course sub-list; The ability-driven recommended course sub-list is used to represent the user's short-board courses recommended based on the user ability matrix vector according to the matching of the course ability requirements and the user's current ability gap; The content-driven recommended course sub-list is used to represent the interest-matching courses recommended based on the cosine similarity according to the user interest text vector and the course text keyword vector; The grade-driven recommended course sub-list is used to represent the user group rating recommended courses recommended based on the course rating matrix vector and the user grade distribution; The group behavior-driven recommended course sub-list is used to represent the courses that are frequently selected by similar user groups obtained by clustering the user ability matrix vector and are not selected by the current user, and recommended according to the similar user groups; S4: Obtain a user portrait according to the user database and the course and grade database; Set preset user target requirements, and collect and obtain dynamic weight data of the course candidate recommendation sub-list obtained based on the parallel hybrid recommendation strategy based on the user portrait; And use the user portrait and the user target requirements as feature data, and the dynamic weight data as label data to obtain a data sample set; S5: Train a preset lightweight weight prediction model according to the data sample set to obtain a course recommendation model; and the lightweight weight prediction model is a dynamic weight prediction model obtained by using a machine learning model or a regression model of a given function type; And based on the course recommendation model, predict and obtain the dynamic weight data of the course candidate recommendation sub-list according to the user profile and the multi-dimensional dynamic knowledge requirement vector; S6: Set a score value for each recommended course in the course candidate recommendation sub-list according to the empirical value; And obtain an individualized course selection recommendation list based on the score value and the dynamic weight data, and implement the recommended course selection for the user according to the individualized course selection recommendation list.

2. The course selection recommendation method based on a hybrid algorithm according to claim 1, wherein The specific steps of S3 are as follows: S31: Based on the multi-dimensional dynamic knowledge requirement vector, obtain a course candidate recommendation sub-list according to the parallel hybrid recommendation strategy combined with multi-source data sets; And the parallel hybrid recommendation strategy includes a recommendation strategy based on professional ability, a recommendation strategy based on course content, a recommendation strategy based on user grades, and a recommendation strategy based on group behavior clustering; S32: Obtain course information based on the course text keyword vector; And the course information includes at least the course category, the course author, and the course description text; Through the recommendation strategy based on professional ability, according to the user ability matrix vector and the course information, obtain the ability gap between the matching course ability requirements for the user and the user's current ability, so as to recommend the user's short-board courses according to the ability gap, that is, the ability-driven recommended course sub-list; S33: Through the recommendation strategy based on course content, obtain a content-driven recommended course sub-list according to the user interest text vector and the course information; S34: Through the recommendation strategy based on user grades, according to the course score matrix vector and the user grade distribution recommendation, use the Pearson correlation coefficient to obtain a correlation index measuring the correlation between the user's course score and the grade distribution, and obtain the courses recommended by the user group score according to the correlation index, and then obtain the grade-driven recommended course sub-list; S35: Through the recommendation strategy based on group behavior clustering, that is, according to the user ability matrix vector, use the K-means clustering algorithm to obtain the similar user group of the course, so as to recommend the high-frequency selected courses of the similar user group and the courses not selected by the current user, and obtain the group behavior-driven recommended course sub-list.

3. The course selection recommendation method based on a hybrid algorithm according to claim 2, wherein, The recommendation strategy based on professional ability in S32 specifically includes the following steps: S321: Perform knowledge extraction on the course information based on the course text keyword vector to obtain the course information, and define the course information as a knowledge entity; And the course information includes at least the course category, the course author, and the course description text; And based on the potential relationship vector of the course ability, connect each knowledge entity to obtain an entity connection relationship; S322: Define the user data structure course entity and the course entity, and based on the entity connection relationship, obtain the user course data structure model according to the user ability matrix vector; And the user course data structure model is described as: Data Structure Course Entity → Entity Connection Relationship →...... → Course Entity →...... → Entity Connection Relationship →...... → Course Entity →......; S323: Based on the user's course data structure model, store the course information in the storage format of a relational database, obtain the gap between the data structure course and the user's current ability according to the user ability matrix vector, and obtain the ability-driven recommended course sub-list according to the current ability gap.

4. The course selection recommendation method based on a hybrid algorithm according to claim 2, wherein The recommendation strategy based on course content in S33 specifically includes the following steps: S331: Obtain the attribute text vector of the course information according to the course information; S332: Based on the TF-IDF text processing method, extract keywords from the user interest text vector and the attribute text vector to obtain the user preference vector and the course attribute vector; Calculate and obtain the cosine similarity between the user preference vector and the course attribute vector; Use the course sequence with the cosine similarity sorted in descending order as the content-driven recommended course sub-list.

5. The course selection recommendation method based on a hybrid algorithm according to claim 3, characterized in that S6 also includes a method for optimizing the course recommendation model based on user behavior feedback, specifically including the following steps: S61: Statistically obtain the user behavior data; The user behavior data is used to represent the number of click / enrollment behavior times or the adoption rate of the individualized enrollment recommendation list; S62: Determine whether the obtained user behavior data meets the preset behavior data threshold; If it is satisfied, confirm that the current course recommendation model meets the user's needs and use it as the optimal course recommendation model; Otherwise, obtain the dynamic weight data of the course candidate recommendation sub-list obtained according to the parallel hybrid recommendation strategy based on the historical user target needs according to the user portrait as a new data sample set, and retrain the current course recommendation model to obtain the model weight parameters of the current course recommendation model to achieve feedback optimization of the course recommendation model.

Images