Intelligent learning recommendation method and device, computer device and storage medium

Abstract

The application discloses a smart learning recommendation method and device, computer equipment and a storage medium. The method comprises: if it is determined to be a video learning scene based on request information, calculating the number of questions that each knowledge point of a teaching video should be allocated; based on the total number of questions obtained by querying, classifying and judging the user to obtain the current user category of the user, the current user category comprising a new user or an old user; based on the answer record of the user under the current user category, performing prediction calculation on the question type branch; and based on the prediction calculation result, performing sorting to obtain a final recommendation result. In the application, the prediction calculation is performed based on the real-time answer record branch, and the subsequent prediction is driven in real time with the changing state of the answer process, instead of fixing the user's ability value, thereby fully fitting and improving the learning recommendation accuracy, and solving the technical problem of poor recommendation effect when the existing technology makes intelligent learning recommendation for the user.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of educational measurement and artificial intelligence, and in particular to intelligent learning recommendation methods, devices, computer equipment, and storage media. Background Technology

[0002] When making intelligent learning recommendations, it is necessary to first assess students' abilities. Current technologies employ student ability assessment methods based on Item Response Theory (IRT). In the field of educational measurement, IRT describes the relationship between student abilities and item attributes through a model. Using a large amount of historical student response data, maximum likelihood estimation or Bayesian estimation methods are used to calibrate parameters such as difficulty, discrimination, and guessing coefficient for each item. For a student, based on their historical response records, the student's ability value is calculated through expected posterior estimation. Based on the student's current ability estimate, the item with the highest information function is selected and recommended to the student, achieving computerized adaptive testing.

[0003] However, IRT models are essentially linear or generalized linear models, assuming that item parameters (difficulty, discrimination, guessing coefficient) are constant and that students' abilities remain unchanged or change according to simple rules during the answering process. This oversimplification leads to the model's inability to adequately fit the complex student-item interactions. The prediction upper limit is approximately 75%, resulting in low prediction accuracy. Therefore, intelligent learning recommendations based on these student ability assessments are ineffective. Consequently, existing technologies suffer from poor recommendation performance when providing intelligent learning recommendations to users. Summary of the Invention

[0004] This application provides an intelligent learning recommendation method, apparatus, computer device, and storage medium, which can solve the technical problem of poor recommendation effect when making intelligent learning recommendations to users in the prior art.

[0005] In a first aspect, embodiments of this application provide an intelligent learning recommendation method, which includes:

[0006] In response to a user-initiated recommendation request, the request information is determined, including the user ID, subject ID, request background, and current topic list;

[0007] If the request information indicates a video learning scenario, then calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video.

[0008] Based on the total number of questions answered, users are categorized to determine their current user category, which includes new users or returning users.

[0009] Based on the answer records of users in the current user category, predict and calculate the question type branches;

[0010] The final recommendation results are obtained by sorting and filtering based on the prediction calculation results.

[0011] In some embodiments, based on the answer records of users in the current user category, the prediction calculation of question type branches is performed, including:

[0012] Retrieve the historical answer records of users under the current user category, including correct answers, incorrect answers, or no answers;

[0013] The candidate questions in the historical answer records are sorted according to question type, which includes objective questions and subjective questions;

[0014] For subjective questions, subjective unique features are assembled; for objective questions, objective unique features are assembled. The subjective unique features include scoring point settings, grading standards, historical scoring rate, and hit status of each scoring point. The objective unique features include basic question information, real-time statistical data, and global default feature mean.

[0015] Based on the subjective and objective unique features of the assembly, the model is invoked to perform inference and prediction to obtain the prediction accuracy.

[0016] In some embodiments, after obtaining the prediction accuracy by invoking the model to perform inference and prediction based on the assembled subjective and objective specific features, the process includes:

[0017] The recommended score is calculated by weighting the correct, incorrect, and unanswered questions in the historical answer records with the prediction accuracy. The historical answer records include the first answer record for new users and the second answer record for old users.

[0018] If the user is a new user, a first recommendation score is obtained by weighted calculation based on the first response record; if the user is an existing user, a second recommendation score is obtained by weighted calculation based on the second response record. The first response record directly reuses the group characteristics of the new user by default, and the second response record directly uses its own true statistical characteristics by default.

[0019] In some embodiments, based on the total number of questions answered, users are categorized to determine their current user category, including:

[0020] Query the total number of questions completed by the user ID under the subject ID;

[0021] The total number of questions answered is compared with a preset threshold. If the total number of questions answered is less than the threshold, the current user is determined to be a new user.

[0022] If the total number of questions answered is greater than the threshold, then the current user is determined to be a returning user.

[0023] In some embodiments, after calculating the recommended score by weighting the corresponding weights of correct, incorrect, and unanswered questions in historical response records with the prediction accuracy, the process includes:

[0024] For questions in the historical answer record that have been answered by the user within the current preset time period, the recommended score corresponding to the question is multiplied by a preset first decay factor.

[0025] The questions in the same category in the historical answer records are sorted according to the recommended scores. The recommended score of the last question in the sort is multiplied by a preset second attenuation factor. The first attenuation factor and the second attenuation factor are used to avoid the question from appearing.

[0026] In some embodiments, the final recommendation results are obtained by sorting and filtering based on the prediction calculation results, including:

[0027] All candidate questions are sorted according to their recommended scores and prediction accuracy to obtain a recommended list;

[0028] Select a specified number of items from the recommendation list as the final recommendation result;

[0029] Display the reasons for the recommendation and the packaged recommendation list;

[0030] Asynchronously update user statistics in the cache.

[0031] In some embodiments, based on the answer records of users in the current user category, the prediction calculation of question type branches is performed, including:

[0032] Based on the answer records of users in the current user category, predictive calculations are performed for question type branches. At the same time, a classification and statistical task is performed to obtain video knowledge points.

[0033] Secondly, embodiments of this application also provide an intelligent learning recommendation device, which includes:

[0034] The determining unit is used to respond to a recommendation request initiated by a user and determine the request information, which includes user ID, subject ID, request background, and current topic list;

[0035] The calculation unit is used to calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video if the request information determines that it is a video learning scenario.

[0036] The classification unit is used to classify users based on the total number of questions answered in the query to determine the current user category, which includes new users or returning users.

[0037] The prediction unit is used to predict question types and branches based on the answer records of users in the current user category.

[0038] The filtering unit is used to sort and filter the results based on the prediction calculation to obtain the final recommendation results.

[0039] Thirdly, embodiments of this application also provide an intelligent learning recommendation computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0040] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, can implement the above-described method.

[0041] This application provides an intelligent learning recommendation method, apparatus, computer device, and storage medium. The method includes: responding to a user-initiated recommendation request, determining request information including a user ID, subject ID, request background, and a current list of selected topics; if the request information indicates a video learning scenario, calculating the number of questions to be assigned to each knowledge point corresponding to the teaching video; classifying the user based on the total number of questions answered, determining the user's current user category, which may include new or returning users; performing predictive calculations of question types based on the user's answer records within the current user category; and sorting and filtering the prediction results to obtain the final recommendation result. In this application, predictive calculations are performed based on the user's real-time answer records, driving subsequent predictions with the changing answer process rather than fixing the user's ability values. This provides a better fit, improving the accuracy of learning recommendations and solving the technical problem of poor recommendation performance in existing intelligent learning recommendations. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating the intelligent learning recommendation method provided in this application embodiment;

[0044] Figure 2 A schematic block diagram of the intelligent learning recommendation device provided in the embodiments of this application;

[0045] Figure 3 A schematic block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0048] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0049] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0050] When making intelligent learning recommendations, it is necessary to first assess students' abilities. Current technologies employ student ability assessment methods based on Item Response Theory (IRT). Within the framework of IRT in educational measurement, this approach describes the relationship between student abilities and item attributes through a model. Using extensive historical student response data, maximum likelihood estimation or Bayesian estimation methods are employed to calibrate parameters such as difficulty, discrimination, and guessing coefficient for each item. For a given student, their ability value is calculated using expected posterior estimation based on their historical response records. Based on the student's current ability estimate, items with the highest information function are selected and recommended to the student, thus achieving computerized adaptive testing.

[0051] However, IRT models are essentially linear or generalized linear models, assuming that item parameters (difficulty, discrimination, guessing coefficient) are constant and that students' abilities remain unchanged or change according to simple rules during the answering process. This oversimplification leads to the model's inability to adequately fit the complex student-item interactions. The prediction upper limit is approximately 75%, resulting in low prediction accuracy. Therefore, intelligent learning recommendations based on these student ability assessments are ineffective. Consequently, existing technologies suffer from poor recommendation performance when providing intelligent learning recommendations to users.

[0052] In some embodiments, the intelligent learning recommendation method is applied to an intelligent learning recommendation device, which is deployed on an intelligent learning recommendation equipment. The equipment is a computer device, which can be a terminal or a server. The terminal can be a smartphone, tablet computer, PDA, or laptop computer, etc.

[0053] Figure 1 This is a flowchart illustrating the intelligent learning recommendation method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps S110-S150:

[0054] S110. In response to the recommendation request initiated by the user, determine the request information, which includes the user ID, subject ID, request background, and current topic list;

[0055] The user ID is a unique identifier for a student, referring to the student's login account or unique identity code.

[0056] User IDs can be used to determine whether a user is new or returning, or to extract personal characteristics, as well as to query the student's historical answer status (not answered / answered correctly / answered incorrectly) for each question in the current question list.

[0057] As an example, user IDs can be used in memory decay to detect whether a particular question has been answered within 30 minutes.

[0058] Subject ID is an identifier for a subject or course package, referring to a specific subject classification code. For example: Mathematics_MATH101, Accounting_ACC01, Legal Professional Qualification Examination_Criminal Law, etc.

[0059] Subject IDs are used to isolate data from different subjects because student performance is subject-specific. Based on the subject ID, the question parameters, the total number of questions completed by the student in that subject, and the mean characteristics of all students in that subject are calculated separately.

[0060] The request context refers to the source of the request scenario, specifically the business context in which the user initiates the recommendation request, and it determines the subsequent calculation branches. There are mainly two types of contexts:

[0061] Normal Practice Mode: Students enter the practice / question-solving page. The system will directly calculate recommended scores for all questions in the current question list and return them sorted by ability matching.

[0062] Video learning mode: The student is watching an instructional video course. The system will then launch an additional video knowledge tag recommendation subprocess. This involves obtaining the time allocation for each knowledge point in the video and proportionally distributing the number of questions corresponding to that video, achieving a learning-practice integration where students see and practice what they see.

[0063] The current topic selection list is not the entire topic bank, but rather a list of candidate topic IDs that have been selected by the upstream system before the system enters the fine-grained scoring stage. It is a finite set of topic IDs.

[0064] The predictive scoring model only scores each question in this list, instead of scanning the entire database, which greatly reduces computational latency.

[0065] S120. If the request information determines that it is a video learning scenario, then calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video.

[0066] This application addresses the problem of disconnect between learning and practice in traditional recommendation systems within video scenarios: it avoids concentrating all recommended questions on the knowledge points covered in the first 10 minutes of the video; it ensures that a corresponding proportion of practice questions are also presented to users for the knowledge points covered in the last 40 minutes; and it uses intelligent tagging for secondary distribution to prevent repeated recommendations of the same knowledge point from the same examination angle, thereby improving recommendation diversity.

[0067] The prerequisite is that the system parses the request background field in the request information and confirms that it is a video learning scenario. The system has already obtained the complete metadata of the teaching video through concurrent tasks.

[0068] The system needs to retrieve the following basic data from the database or cache: the total playback time of the current teaching video, for example, 60 minutes; the start and end time ranges of each knowledge point in the video, such as knowledge point A corresponding to 0~20 minutes, knowledge point B corresponding to 20~45 minutes, and knowledge point C corresponding to 45~60 minutes.

[0069] The number of intelligent tag types associated with each knowledge point is used to refine the weighting. For example, knowledge point A has three tags: concept, formula, and case study. The total number of recommended questions to be returned to the user for this request is also considered. These data collectively serve as the basis for subsequent calculations of the number of recommended questions for each knowledge point.

[0070] The first step is to calculate the time allocation weight for each knowledge point. For example, if knowledge point A is 20 minutes long and the total video length is 60 minutes, then Weight_Time_A = 20 / 60 ≈ 0.333. The second step is to calculate the basic allocation quota for this knowledge point. The third step is to perform a secondary allocation based on the number of smart tags, distributing the weight evenly according to the number of smart tags under each knowledge point. This means that questions within a knowledge point cannot be clustered under the same tag; further refinement is required. The fourth step is rounding and total number correction. The result is rounded to the nearest integer to obtain the specific number of questions to be recommended under that tag.

[0071] Because rounding may cause the sum of the recommended numbers for each tag to differ from the original, a final fine-tuning is required. This includes sorting by the decimal part and padding the difference to ensure that the final total list is exactly equal to the number requested by the user.

[0072] S130. Based on the total number of questions answered, the user is classified to determine the current user category, which includes new users or old users;

[0073] S130 includes S1301-S1303:

[0074] S1301. Query the total number of questions completed by the user ID under the subject ID;

[0075] In the cache or database, using the user ID and subject ID as a composite primary key, query the total number of historical questions answered by the user in the current subject. This statistic is updated in real time with each answer and cached in Redis to avoid triggering an aggregate query in the database with each recommendation request, thus reducing response latency.

[0076] S1302. Based on the comparison between the total number of questions completed and the preset number threshold, if the total number of questions completed is less than the number threshold, the current user category is determined to be a new user.

[0077] The preset threshold for the number of questions is fixed at 20. This value is not chosen arbitrarily, but is based on the principle of statistical stability in educational measurement: when the number of questions is less than 20, the sample size is too small, and the variance of statistical characteristics such as the calculated accuracy rate is extremely large, which does not have statistical reliability; after reaching 20 questions, individual characteristics begin to stabilize, which is sufficient to support personalized prediction.

[0078] A user is considered a new user if the total number of questions they have answered in a subject is less than 20.

[0079] For new users, the 7-day / 15-day / 30-day accuracy of their student features uses the global default value, and the feature mean of all students in the subject is standardized to avoid feature anomalies caused by insufficient historical data.

[0080] S1303. If the total number of questions answered is greater than the number threshold, then the current user category is determined to be an old user.

[0081] When a user has completed more than 20 questions in a particular subject, they are considered a returning user. For returning users, their feature vectors are calculated in real time for each request, and the model is invoked for inference to ensure that the prediction results reflect the user's latest learning status.

[0082] S140. Based on the answer records of users under the current user category, perform prediction calculations for question type branches;

[0083] The system relies on the user category (new / old) determined in the previous step to decide whether to use the individual's real value or the default value for a near-new user when assembling features; it also relies on real-time statistics provided by concurrent tasks.

[0084] Downstream output generates a corresponding prediction accuracy for each candidate question in the current question list and passes it to subsequent steps as a basic factor for calculating the recommendation score.

[0085] By using differentiated modeling, we can achieve personalized prediction for objective and subjective questions, thus solving the accuracy bottleneck of traditional IRT models and collaborative filtering methods in mixed question types.

[0086] In some embodiments, the prediction calculation of question type branches is performed based on the answer records of users under the current user category, wherein a classification statistics task and a task of obtaining video knowledge points are performed simultaneously.

[0087] S140 includes S1401-S1404:

[0088] S1401. Obtain the historical answer records of users under the current user category, wherein the historical answer records include correct answers, incorrect answers, or no answers;

[0089] Retrieve historical answer records, using user ID and subject ID as dimensions, and pull the user's answer status for each candidate question in the current subject and the current question list, or the global history, from the database.

[0090] The criterion for judging objective questions is that the user's submitted answer completely matches the standard answer, while the criterion for judging subjective questions is that the score reaches the full score threshold.

[0091] "Answered incorrectly" means the answer is wrong or the score is below the passing threshold; "Not answered" means the user has never answered the question, or the answer has expired / been overwritten. Data is first read from the Redis cache, which is valid for half a day. If the cache is not found, the source database is queried.

[0092] S1402. The candidate questions in the historical answer records are sorted according to question type, which includes objective questions and subjective questions;

[0093] The routing is based on the question type field in the question metadata. Objective questions include multiple choice, true / false, and fill-in-the-blank questions, with unique answers and not open-ended. Subjective questions include short answer, calculation, essay, and case analysis questions, with non-unique answers, and scoring depends on manual / semi-automatic grading, resulting in partial credit.

[0094] The purpose of splitting the question types is that the score distribution and information density of the two types of questions are different, and using independent modeling can avoid the loss of prediction accuracy of a single model for a certain type of question.

[0095] S1403. Assemble subjective-specific features for subjective questions and objective-specific features for objective questions. The subjective-specific features include scoring point settings, grading standards, historical scoring rates, and hit rates for each scoring point. The objective-specific features include basic question information, real-time statistical data, and the average value of global default features.

[0096] The unique characteristics of subjective questions include the following dimensions: scoring points, grading criteria, historical score rates, and the hit rate of each scoring point.

[0097] The scoring points are defined as follows: the total number of scoring points in a question (e.g., a full score of 10 points is divided into 5 scoring points, each worth 2 points); the grading criteria refer to the specific scoring rules for each scoring point, such as keyword hits, correct formulas, and logical completeness; the historical score rate is the user's historical average score rate for this question type or knowledge point, calculated by dividing the total score of all subjective questions by the total full score; the hit rate for each scoring point reflects the user's cumulative hit count and hit rate for each scoring point in their historical answers, demonstrating the user's mastery down to the granular level of scoring points; and the knowledge point / category score rate statistics are the user's average score rate for subjective questions within the knowledge point to which the current question belongs, data obtained from real-time statistics of concurrent tasks.

[0098] The unique features of objective questions include the following dimensions: basic question information, real-time statistical data, and default feature values.

[0099] The basic information of the question includes the question's historical accuracy rate, discrimination index, knowledge point association tags (such as quadratic equations), and category associations (such as algebra). This part of the data is cached for 2 days. The user's personal question-solving characteristics are the user's global accuracy rate over the past 7, 15, and 30 days, which is cached for half a day. Real-time statistics cover the user's question-solving statistics for each knowledge point and category (including the number of correct answers, the number of incorrect answers, and the number of questions not answered), which comes from concurrent tasks. Default feature values ​​are used as a fallback for cold starts. These are the mean and standard deviation of all students in the subject across each feature dimension, which are cached for 3 hours. However, if the user is a new user, they will be replaced with the mean and standard deviation of users who registered in the past 7 days.

[0100] S1404. Based on the subjective and objective unique features of the assembly, the model is invoked to perform inference and prediction to obtain the prediction accuracy.

[0101] The objective question branch inputs the assembled objective question feature vector into the DNN model for forward inference and outputs the user's prediction accuracy for that objective question.

[0102] The subjective question branch inputs the assembled subjective question feature vector into the DNN model for inference and outputs the predicted score that the user may achieve on the subjective question.

[0103] As an example, the prediction accuracy for objective questions reached 85.9%, and the prediction accuracy for subjective questions reached 82.3%, significantly higher than the 75% of the traditional IRT model.

[0104] If there are many candidate questions, such as in a wrong question set scenario, the feature vectors of multiple questions can be assembled into a batch request, and the GPU can be called once to complete all inference, reducing IO overhead.

[0105] If the inference time exceeds the 100ms threshold, the system will automatically downgrade to return a prediction result based on rules such as the global average accuracy to ensure interface availability.

[0106] Prediction results for new users are cached for 1 day, while those for existing users are not cached to ensure real-time performance.

[0107] Following S1404, the process includes: calculating a recommended score by weighting the predicted accuracy with the corresponding weights of correct, incorrect, and unanswered questions in historical answer records. The historical answer records include a first answer record for new users and a second answer record for existing users. Specifically, if the user is new, a first recommended score is calculated based on the first answer record; if the user is an existing user, a second recommended score is calculated based on the second answer record. The first answer record directly reuses the group characteristics of the relatively new user, while the second answer record directly uses its own actual statistical characteristics.

[0108] As an example, the recommended score calculation is as follows: For each selected question, the recommended score is calculated based on the user's answer status. If the question was not answered, the recommended score = prediction accuracy × 1.0, where 1.0 is the maximum weight. If the question was answered correctly, the recommended score = prediction accuracy × decay coefficient α, where α < 1.0 indicates mastery and the priority is reduced. If the question was answered incorrectly, the recommended score = prediction accuracy × reinforcement coefficient β, where β > 1.0 indicates lack of mastery and the priority is increased.

[0109] For new users, the first response record will directly reuse the group characteristics of new users by default; for existing users, the second response record will directly use their own actual statistical characteristics by default.

[0110] The recommended score is calculated by weighting the correct, incorrect, and unanswered questions from historical response records with the predicted accuracy. This process includes steps B1-B2.

[0111] B1. For questions in the historical answer record that have been answered by the user within the current preset time period, multiply the recommended score corresponding to the question by a preset first decay factor.

[0112] Step B1 is the first level of attenuation, namely, recent question attenuation, which operates on the time dimension. Its trigger condition is that the user has attempted the question within the current preset time period (the last 30 minutes), regardless of whether the answer was correct or incorrect. Once triggered, the system multiplies the question's current recommended score by a preset first attenuation factor of 0.1^10 (1e-10), an extremely small value that compresses the question's recommended score to almost zero, causing it to naturally sink to the bottom of the subsequent ranking and almost impossible to appear in the final recommended list. The purpose of this mechanism is to prevent users from repeatedly encountering questions they have just attempted within a short period, thereby reducing ineffective question-solving experiences and improving user satisfaction.

[0113] B2. Sort the questions in the same category from the historical answer records according to the recommended scores, and multiply the recommended score of the last question in the sorted list by a preset second attenuation factor. The first attenuation factor and the second attenuation factor are used to prevent questions from appearing repeatedly.

[0114] Step B2 is the second level of decay, namely intra-category ranking decay, which applies to the diversity dimension. The prerequisite for performing this step is to treat all questions in the same category as a knowledge point label, such as all questions under quadratic equations, and sort them in descending order according to the current recommended score (after applying the first level of decay).

[0115] Subsequently, the system applies a preset second decay factor γ = 0.7 to each sorted question. The specific calculation formula is as follows: if the sorting position of a question under this category is N, starting from 1, then the recommended score of the question is updated to the current score multiplied by γ^(N-1).

[0116] The following is an example of the attenuation effect:

[0117] The recommendation score for the first ranked item is multiplied by 0.7^0 = 1.0;

[0118] The second step is multiplied by 0.7^1 = 0.7;

[0119] The third question multiplied by 0.7^2 equals 0.49;

[0120] The fourth question is multiplied by 0.7^3 = 0.343, and so on. The purpose of this mechanism is to ensure that only the top 1-2 questions that best match the user's ability level for the same knowledge point receive a higher ranking, while the recommendation scores for subsequent questions of the same type are significantly suppressed. This forces the recommendation list to be more diverse in terms of knowledge point coverage, effectively preventing the recommendation results from all being concentrated on the same type of questions or the same knowledge point.

[0121] S150. Based on the prediction calculation results, sort and filter to obtain the final recommendation results.

[0122] S150 includes S1501-S1504:

[0123] S1501. Sort all candidate questions according to their recommended scores and prediction accuracy to obtain a recommended list;

[0124] A global ranking process generates a recommendation list, arranging all candidate questions in descending order of their final recommendation scores to form an ordered list. During ranking, the prediction accuracy of each question is also retained as supporting data for the recommendation rationale and as a basis for subsequent manual review. If the number of candidate questions is large, the system uses multi-threaded parallel ranking to ensure that the ranking time is within milliseconds.

[0125] S1502. Extract a specified number of items from the recommendation list as the final recommendation result;

[0126] The Top-N items are selected as the final recommendation results. The selection rule is to select a specified number of items (e.g., Ntotal=10 passed in by the business layer) from the sorted recommendation list as the final recommendation results.

[0127] If the total number of candidate questions is less than Ntotal, all available questions are returned; if the total number of candidate questions is 0, an empty list is returned and an exception log is recorded, triggering a degradation strategy, such as returning popular questions under that subject.

[0128] S1503, Show the reasons for recommendation and the recommended package list;

[0129] Display the reasons for the recommendation and package the recommendation list.

[0130] For each question in the final recommendation list, generate an explainable recommendation reason. Examples include "Reinforcement based on wrong answers": the question was answered incorrectly, β>1.0, so the priority is increased; "Based on ability matching": the prediction accuracy is moderate, no state is set, so the weight is full.

[0131] To avoid duplicate recommendations, if a question triggers time decay but is still within the truncation range, a note will be added indicating that it has been recently completed. The recommendation list is encapsulated into a standardized JSON / Protobuf response structure, including the question ID, final recommendation score, prediction accuracy, recommendation reason, and basic question information. The encapsulated result is returned to the client for rendering and display via HTTP / HTTPS protocol.

[0132] S1504. Asynchronously update user statistics in the cache.

[0133] The system asynchronously updates user statistics in the cache, specifically including the total number of questions the user has answered. If the current recommendation involves questions newly answered by the user, their answer count is incremented by 1, and the user's 7-day / 15-day / 30-day accuracy rates are updated. If new answers are generated, the accuracy statistics for each time window and the new / old user status markers are recalculated and updated. If the user was originally a new user and had answered fewer than 20 questions, their status is switched to old user after their answer count reaches 20. Subsequent requests will automatically use the user's real characteristics to replace the default characteristics for new users, recording the question IDs and answer times actually answered by the user in this request for use in determining the "30-minute window" in subsequent steps.

[0134] In terms of technical implementation, the system uses message queues or asynchronous task thread pools to decouple the above update operations, ensuring that the main link response latency is not affected.

[0135] This application provides an intelligent learning recommendation method, apparatus, computer device, and storage medium. The method includes: responding to a user-initiated recommendation request, determining request information including a user ID, subject ID, request background, and a current list of selected topics; if the request information indicates a video learning scenario, calculating the number of questions to be assigned to each knowledge point corresponding to the teaching video; classifying the user based on the total number of questions answered, determining the user's current user category, which may include new or returning users; performing predictive calculations of question types based on the user's answer records within the current user category; and sorting and filtering the prediction results to obtain the final recommendation result. In this application, predictive calculations are performed based on the user's real-time answer records, driving subsequent predictions with the changing answer process rather than fixing the user's ability values. This fully fits the user's learning situation, improving the accuracy of learning recommendations and solving the technical problem of poor recommendation performance in existing intelligent learning recommendations.

[0136] Figure 2 This is a schematic block diagram of an intelligent learning recommendation device provided in an embodiment of this application. Figure 2 As shown, corresponding to the above-described intelligent learning recommendation method, this application also provides an intelligent learning recommendation device 600. This intelligent learning recommendation device 600 includes a unit for executing the above-described intelligent learning recommendation method, and can be configured in terminals such as desktop computers, tablet computers, and laptops. Specifically, please refer to... Figure 2 The intelligent learning recommendation device 600 includes a determination unit 601, a calculation unit 602, a classification unit 603, a prediction unit 604, and a filtering unit 605, wherein:

[0137] The determining unit 601 is used to respond to the recommendation request initiated by the user and determine the request information, which includes the user ID, subject ID, request background and current topic list;

[0138] The calculation unit 602 is used to calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video if the request information determines that it is a video learning scenario.

[0139] Classification unit 603 is used to classify users based on the total number of questions answered and to obtain the current user category, which includes new users or old users;

[0140] Prediction unit 604 is used to predict and calculate question type branches based on the answer records of users under the current user category;

[0141] The filtering unit 605 is used to sort and filter based on the prediction calculation results to obtain the final recommendation results. In some embodiments, the prediction unit 604 performs prediction calculations for question type branches based on the answer records of users under the current user category, specifically for:

[0142] Retrieve the historical answer records of users under the current user category, including correct answers, incorrect answers, or no answers;

[0143] The candidate questions in the historical answer records are sorted according to question type, which includes objective questions and subjective questions;

[0144] For subjective questions, subjective unique features are assembled; for objective questions, objective unique features are assembled. The subjective unique features include scoring point settings, grading standards, historical scoring rate, and hit status of each scoring point. The objective unique features include basic question information, real-time statistical data, and global default feature mean.

[0145] Based on the subjective and objective unique features of the assembly, the model is invoked to perform inference and prediction to obtain the prediction accuracy.

[0146] In some embodiments, after obtaining the prediction accuracy by invoking the model to perform inference and prediction based on the assembled subjective and objective unique features, the prediction unit 604 is specifically used for:

[0147] The recommended score is calculated by weighting the correct, incorrect, and unanswered questions in the historical answer records with the prediction accuracy. The historical answer records include the first answer record for new users and the second answer record for old users.

[0148] If the user is a new user, a first recommendation score is obtained by weighted calculation based on the first response record; if the user is an existing user, a second recommendation score is obtained by weighted calculation based on the second response record. The first response record directly reuses the group characteristics of the new user by default, and the second response record directly uses its own true statistical characteristics by default.

[0149] In some embodiments, the classification unit 603, when performing a classification judgment based on the total number of questions answered to determine the user's current user category, specifically uses the following:

[0150] Query the total number of questions completed by the user ID under the subject ID;

[0151] The total number of questions answered is compared with a preset threshold. If the total number of questions answered is less than the threshold, the current user is determined to be a new user.

[0152] If the total number of questions answered is greater than the threshold, then the current user is determined to be a returning user.

[0153] In some embodiments, after calculating the recommendation score by weighting the corresponding weights of correct, incorrect, and unanswered questions in historical response records with the prediction accuracy, the prediction unit 604 is specifically used for:

[0154] For questions in the historical answer record that have been answered by the user within the current preset time period, the recommended score corresponding to the question is multiplied by a preset first decay factor.

[0155] The questions in the same category in the historical answer records are sorted according to the recommended scores. The recommended score of the last question in the sort is multiplied by a preset second attenuation factor. The first attenuation factor and the second attenuation factor are used to avoid the occurrence of the question.

[0156] In some embodiments, the filtering unit 605 performs sorting and filtering based on prediction calculation results to obtain a final recommendation result, specifically for:

[0157] All candidate questions are sorted according to their recommended scores and prediction accuracy to obtain a recommended list;

[0158] Select a specified number of items from the recommendation list as the final recommendation result;

[0159] Display the reasons for the recommendation and the packaged recommendation list;

[0160] Asynchronously update user statistics in the cache.

[0161] In some embodiments, when the prediction unit 604 performs prediction calculations for question type branches based on the user's answer records under the current user category, it is specifically used for:

[0162] Based on the answer records of users in the current user category, predictive calculations are performed for question type branches, while simultaneously performing classification statistics tasks and acquiring video knowledge points.

[0163] In summary, the intelligent learning recommendation device 600 in this embodiment responds to a user-initiated recommendation request by determining the request information, which includes the user ID, subject ID, request background, and current topic selection list. If the request information indicates a video learning scenario, the device calculates the number of questions to be assigned to each knowledge point corresponding to the teaching video. Based on the total number of questions answered, the device categorizes the user to obtain their current user category, which may include new or returning users. Based on the user's answer records within the current user category, the device performs prediction calculations for question type branches. The prediction calculation results are then sorted and filtered to obtain the final recommendation result. In this application, prediction calculations are performed based on the user's real-time answer records, driving subsequent predictions with the changing answer process rather than fixing the user's ability values. This fully fits the user's situation, improves the accuracy of learning recommendations, and solves the technical problem of poor recommendation performance when providing intelligent learning recommendations to users in the prior art.

[0164] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned intelligent learning recommendation device and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0165] The aforementioned intelligent learning recommendation device can be implemented as a computer program, which can, for example... Figure 3 It runs on the computer device shown.

[0166] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device 700 provided in an embodiment of this application. The computer device 700 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0167] See Figure 3 The computer device 700 includes a processor 702, a memory, and a network interface 705 connected via a system bus 701. The memory may include a non-volatile storage medium 703 and internal memory 704.

[0168] The non-volatile storage medium 703 may store an operating system 7031 and a computer program 7032. The computer program 7032 includes program instructions that, when executed, cause the processor 702 to perform an intelligent learning recommendation method.

[0169] The processor 702 provides computing and control capabilities to support the operation of the entire computer device 700.

[0170] The internal memory 704 provides an environment for the execution of the computer program 7032 in the non-volatile storage medium 703. When the computer program 7032 is executed by the processor 702, the processor 702 can execute an intelligent learning recommendation method.

[0171] This network interface 705 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 700 to which the present application is applied. The specific computer device 700 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0172] The processor 702 is used to run the computer program 7032 stored in the memory to perform the following steps:

[0173] In response to a user-initiated recommendation request, the request information is determined, including the user ID, subject ID, request background, and current topic list;

[0174] If the request information indicates a video learning scenario, then calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video.

[0175] Based on the total number of questions answered, users are categorized to determine their current user category, which includes new users or returning users.

[0176] Based on the answer records of users in the current user category, predict and calculate the question type branches;

[0177] The final recommendation results are obtained by sorting and filtering based on the prediction calculation results.

[0178] It should be understood that in the embodiments of this application, the processor 702 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0179] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0180] It should be noted that the specific embodiments are merely illustrative examples intended to aid in understanding the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. The scope of protection of this application is determined by the contents of the claims. Specific embodiments, parameter ranges, or technical features described in the specification should not be construed as undue limitation or expansion of the scope of protection of the claims. The technical effects described in the specification are only used to illustrate the innovation of the present invention. Any technical solution that does not simultaneously possess all the technical features of the present invention, even if it claims to solve the same technical problem, does not fall within the scope of protection of the present invention.

[0181] Therefore, this application also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the following steps:

[0182] In response to a user-initiated recommendation request, the request information is determined, including the user ID, subject ID, request background, and current topic list;

[0183] If the request information indicates a video learning scenario, then calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video.

[0184] Based on the total number of questions answered, users are categorized to determine their current user category, which includes new users or returning users.

[0185] Based on the answer records of users in the current user category, predict and calculate the question type branches;

[0186] The final recommendation results are obtained by sorting and filtering based on the prediction calculation results.

[0187] The storage medium can be any computer-readable storage medium that can store program code, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0188] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0189] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0190] The steps in the methods of this application embodiment can be adjusted, merged, or deleted according to actual needs. The units in the apparatus of this application embodiment can be merged, divided, or deleted according to actual needs. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0191] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0192] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent learning recommendation method, characterized in that, The method includes: In response to a user-initiated recommendation request, the request information is determined, including the user ID, subject ID, request background, and current topic list; If the request information indicates a video learning scenario, then calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video. Based on the total number of questions answered, users are categorized to determine their current user category, which includes new users or returning users. Based on the answer records of users in the current user category, predict and calculate the question type branches; The final recommendation results are obtained by sorting and filtering based on the prediction calculation results.

2. The method according to claim 1, characterized in that, Based on the answer records of users in the current user category, predictive calculations are performed for question type branches, including: Retrieve the historical answer records of users under the current user category, including correct answers, incorrect answers, or no answers; The candidate questions in the historical answer records are sorted according to question type, which includes objective questions and subjective questions; For subjective questions, subjective unique features are assembled; for objective questions, objective unique features are assembled. The subjective unique features include scoring point settings, grading standards, historical scoring rate, and hit status of each scoring point. The objective unique features include basic question information, real-time statistical data, and global default feature mean. Based on the subjective and objective unique features of the assembly, the model is invoked to perform inference and prediction to obtain the prediction accuracy.

3. The method according to claim 2, characterized in that, Based on the subjective and objective unique features of the assembly, after calling the model to perform inference and prediction to obtain the prediction accuracy, the following is included: The recommended score is calculated by weighting the correct, incorrect, and unanswered questions in the historical answer records with the prediction accuracy. The historical answer records include the first answer record for new users and the second answer record for old users. If the user is a new user, a first recommendation score is obtained by weighted calculation based on the first response record; if the user is an existing user, a second recommendation score is obtained by weighted calculation based on the second response record. The first response record directly reuses the group characteristics of the new user by default, and the second response record directly uses its own true statistical characteristics by default.

4. The method according to claim 1, characterized in that, Based on the total number of questions answered, users are categorized to determine their current user category, including: Query the total number of questions completed by the user ID under the subject ID; The total number of questions answered is compared with a preset threshold. If the total number of questions answered is less than the threshold, the current user is determined to be a new user. If the total number of questions answered is greater than the threshold, then the current user is determined to be a returning user.

5. The method according to claim 4, characterized in that, The recommended score is calculated by weighting the correct, incorrect, and unanswered questions from historical response records with the prediction accuracy, and then includes: For questions in the historical answer record that have been answered by the user within the current preset time period, the recommended score corresponding to the question is multiplied by a preset first decay factor. The questions in the same category in the historical answer records are sorted according to the recommended scores. The recommended score of the last question in the sort is multiplied by a preset second attenuation factor. The first attenuation factor and the second attenuation factor are used to avoid the question from appearing.

6. The method according to claim 1, characterized in that, The final recommendation results are obtained by sorting and filtering based on the prediction calculation results, including: All candidate questions are sorted according to their recommended scores and prediction accuracy to obtain a recommended list; Select a specified number of items from the recommendation list as the final recommendation result; Display the reasons for the recommendation and the packaged recommendation list; Asynchronously update user statistics in the cache.

7. The method according to claim 1, characterized in that, Based on the answer records of users in the current user category, predictive calculations are performed for question type branches, including: Based on the answer records of users in the current user category, predictive calculations are performed for question type branches. At the same time, a classification and statistical task is performed to obtain video knowledge points.

8. An intelligent learning recommendation device, characterized in that, The device includes: The determining unit is used to respond to a recommendation request initiated by a user and determine the request information, which includes user ID, subject ID, request background, and current topic list; The calculation unit is used to calculate the number of questions that should be assigned to each knowledge point corresponding to the teaching video if the request information determines that it is a video learning scenario. The classification unit is used to classify users based on the total number of questions answered in the query to determine the current user category, which includes new users or returning users. The prediction unit is used to predict question types and branches based on the answer records of users in the current user category. The filtering unit is used to sort and filter the results based on the prediction calculation to obtain the final recommendation results.

9. A computer device for intelligent learning recommendation, characterized in that, The method includes a memory, a processor, and an intelligent learning recommendation program stored in the memory and executable on the processor, wherein the processor executes the intelligent learning recommendation program to implement the steps of the intelligent learning recommendation method according to any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a program that implements the intelligent learning recommendation method, which is executed by a processor to implement the steps of the intelligent learning recommendation method as described in any one of claims 1 to 7.

Images