Education resource personalized recommendation method and system based on deep learning

Abstract

The invention relates to the technical field of personalized recommendation, in particular to an educational resource personalized recommendation method and system based on deep learning, and the method comprises the following steps: obtaining stage scores and learning records, extracting long-term tracks, collecting short-term behaviors, generating knowledge point fusion representation, constructing a state transition structure, and generating a learning state sequence; matching resources are screened to mark a to-be-intervened node recombination path, an adaptation table is generated in combination with interaction signals, high-matching resources are screened, and a personalized recommendation task set is formed. According to the invention, through fusing long-term learning tracks and short-term behavior characteristics of students, mastering difference weights in knowledge point dimensions are constructed, individual stage learning states are archived and coded in a unified state transition structure, to-be-intervened nodes in a learning path are accurately identified, and the learning efficiency is improved. And then priority recombination is carried out in combination with the subject tag of the resource, the recommendation frequency and the path relevance, so that accurate docking between the resource recommendation content and the current learning state of the student is realized, and the pertinence, continuity and suitability of recommendation are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of personalized recommendation technology, and in particular to a method and system for personalized recommendation of educational resources based on deep learning. Background Technology

[0002] Personalized recommendation technology involves utilizing data analysis, machine learning, and artificial intelligence to provide customized product or service recommendations based on users' interests, needs, and behaviors. Core aspects of this technology include user behavior analysis, data mining, recommendation algorithms, and personalized content generation. It is primarily applied in e-commerce platforms, social networks, and media content recommendation, aiming to predict user interests and product preferences based on real-time behavior, thereby improving user experience and business efficiency. Personalized recommendation technology has become an important tool for improving service accuracy and user satisfaction. Traditional personalized recommendation methods for educational resources analyze students' learning habits, academic performance, and knowledge acquisition to push educational resources that meet their needs. These methods rely on rule matching, content-based recommendation, or collaborative filtering, using user behavior data analysis to recommend resources. However, they suffer from low recommendation accuracy when dealing with user needs and dynamically changing learning behaviors. Traditional methods use manually designed rules or simple similarity calculations to filter educational resources, but struggle to fully capture students' personalized needs and interests when handling large-scale data or non-linear learning patterns.

[0003] Existing technologies for recommending personalized educational resources to students mainly rely on rule matching, content-based recommendation, or collaborative filtering. When analyzing user behavior, they generally use static data for modeling, lacking the ability to model dynamic changes in users' learning status. This makes it difficult to identify students' actual learning obstacles and knowledge mastery deviations at specific stages. The models are insufficiently responsive to jumps or abnormal behaviors in the learning path and cannot effectively capture the deep learning intentions behind short-term behaviors. As a result, recommended resources are difficult to accurately match students' current learning status, and resource delivery lacks coherence and targeting, leading to limited coverage of recommendation results. This makes it difficult to meet the ever-changing personalized learning needs, and there are significant shortcomings in terms of service accuracy and recommendation effectiveness. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a method for personalized recommendation of educational resources based on deep learning, comprising the following steps: S1: Obtain students' periodic transcripts, course chapter completion records, and knowledge point pass records; extract long-term behavioral trajectory sets; collect practice logs, video playback interruption frequency, and interactive test scores; extract short-term behavioral feature sets; establish mastery difference weights on the knowledge point dimension by comparing the mastery of knowledge points in the long-term trajectory set with the performance in the short-term behavioral feature set; and generate a knowledge point fusion representation set. S2: Using the knowledge point fusion representation set, select the exercise question click path, video viewing start and end nodes and question type error records corresponding to the tag items under the knowledge points, and normalize and encode the execution position and type of the exercise question click path and question type error records to generate an individual stage learning state encoding sequence. S3: Based on the individual stage learning state encoding sequence, filter all resource sets that are consistent with the knowledge point labels corresponding to the encoding sequence, mark the knowledge points with continuous transfer or jump behavior in the encoding sequence as nodes to be intervened, and generate a resource path candidate set; S4: Through the resource path candidate set, collect interaction signal indicators in the real-time learning scenario, compress and map the device usage period with the interruption duration execution time, and generate an interaction matching adaptation table.

[0005] As a further aspect of the present invention, the knowledge point fusion representation set includes mastery state difference labels, behavioral feature matching vectors, fusion dimension mapping indexes, individual performance offsets, and temporal weight correction factors; the individual stage learning state encoding sequence includes behavioral encoding fragment sequences, knowledge point label sequences, state transition relationship tables, stage switching identifiers, and unified format index codes; the resource path candidate set includes resource topic index tables, matching priority matrices, path reorganization lists, intervention node marker books, and course structure alignment records; and the interaction matching adaptation table includes device usage mapping segments, interaction intensity interval tables, resource response attribute sets, time compression node sequences, and adaptation level indexes.

[0006] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Obtain the periodic grade sheet, course chapter completion records and knowledge point passing records, parse the course identifier field in the periodic grade sheet, locate and match the chapter index value in the course chapter completion record, extract the correspondence of chapter index values, aggregate the knowledge point numbers with the same chapter index value in the knowledge point passing records, and generate a long-term behavior trajectory index set. S102: Call the long-term behavior trajectory index set, collect practice logs, video playback interruption frequency and interactive test score data within one week, filter and extract the practice record timestamp field in the practice log, filter the data frames belonging to the time period within one week, count the video playback interruption frequency, and associate and match the answer accuracy field in the interactive test score with the course chapter, aggregate them into behavioral feature triplets under the same knowledge point number, and generate a short-term behavior feature set; S103: Call the short-term behavioral feature set, extract three types of data frames from the behavioral features: practice frequency, video interruption number and interactive test accuracy, calculate the numerical difference between each and the completion status value under the knowledge point number in the long-term trajectory, and normalize each difference according to the knowledge point number to generate a knowledge point fusion representation set.

[0007] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Call the knowledge point fusion representation set, select the practice question click path data frame, video viewing start and end node record and question type error record data frame associated with the tag item under the corresponding knowledge point number, match and verify the click node index field in the practice question click path and the error question type identifier field in the question type error record respectively, and perform index position normalization operation and error type number normalization operation on the entries with the same knowledge point number in both, and generate path and error code information; S202: Based on the path and error coding information, extract the start and end timestamps of the corresponding nodes, and perform cross-comparison and compensation merging operations on the start and end timestamps and time coverage label intervals. Based on the comparison results, reconstruct the start and end node intervals to generate a video time interval comparison set. S203: Based on the video time interval reference set, construct a one-to-one mapping structure between knowledge point number and normalized encoding field, define the state transition node as a triplet of path position, video time point and error type encoding, and perform position encoding and archiving annotation operations according to the order of transition nodes to generate an individual stage learning state encoding sequence.

[0008] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the individual stage learning state coding sequence, perform a tag matching operation, filter all educational resource items that match the tags, extract the topic tag field, recommendation frequency parameter value and course path correlation level identifier from the matched resource items, and generate a resource content attribute set; S302: Call the resource content attribute set to retrieve and identify continuous state transition nodes and jump behavior nodes. During the identification process, the time sequence field and position jump field in the state code are used to make judgments. Knowledge points with a continuous number of transitions greater than the specified jump identification threshold are marked as items that need to be processed, and nodes whose position jumps exceed the normal path length range are marked as abnormal items, generating a set of nodes to be intervened. S303: Based on the set of nodes to be intervened, perform tag matching degree determination and descending order of recommendation frequency value respectively, and perform level filtering processing in combination with course path relevance level identifier. Aggregate the sorted results according to knowledge point number to generate resource path candidate set.

[0009] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Through the resource path candidate set, collect the corresponding interactive signal indicator data frames in the real-time learning scenario, including the start and end time period in the device usage record and the interruption duration field in the question-answering interruption behavior record, identify the frequency change data points in the interactive operation log, perform time axis compression mapping operation on the device usage start and end time period and interruption duration field, and generate a compressed time behavior mapping group. S402: Call the compressed time behavior mapping group, calculate the frequency fluctuation gradient value of each curve segment, perform threshold division processing on the fluctuation gradient, perform multi-segment step-level segmentation and classification, attach multi-segment hierarchical labels to the original frequency point sequence, and generate a step-segment interactive frequency set. S403: Based on the tiered segmented interaction frequency set, call the type attribute field of the resource item, perform a mapping consistency comparison operation between the resource type field and the frequency classification label, align the matching position according to the time period of the compressed mapping, and generate an interaction matching adaptation table.

[0010] As a further aspect of the present invention, in the operation of thresholding the fluctuation gradient, at least three frequency fluctuation gradient threshold intervals are preset, each corresponding to a segmented classification label of a different level. The frequency fluctuation gradient threshold intervals include a first frequency fluctuation gradient threshold interval, a second frequency fluctuation gradient threshold interval, and a third frequency fluctuation gradient threshold interval. The first frequency fluctuation gradient threshold interval is used to mark the low frequency fluctuation state, and is correspondingly labeled as the first segmentation classification label; The second frequency fluctuation gradient threshold interval is used to mark the medium frequency fluctuation state, and is correspondingly labeled as the second segmentation classification label; The third frequency fluctuation gradient threshold interval is used to mark high frequency fluctuation states, and is correspondingly labeled as the third segmentation classification label. In the process of adding multiple segmented hierarchical labels to the original frequency point sequence, the corresponding segmented classification labels are determined based on the threshold range into which the frequency fluctuation gradient value falls, thus forming the stepped segmented interactive frequency set.

[0011] As a further aspect of the present invention, the method further includes step S5: S5: Based on the interaction matching adaptation table, extract resource items with matching levels higher than the main adaptation threshold, mark the path order and the coverage of the real-time learning state, retain resources with priority in path order and overlapping coverage with the state sequence, and combine the knowledge point coverage and learning time tags of the resources to obtain a personalized resource recommendation task set. The personalized resource recommendation task set includes a path number mapping table, a list of recommended resources, learning status coverage tags, a priority recommendation sequence, and a coverage overlap index.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the interaction matching adaptation table, extract resource entries whose matching level field is higher than the main adaptation threshold, and retrieve the path order field corresponding to each resource and the time period mapping range of the state node in the individual stage learning state encoding sequence. Perform the overlap calculation operation based on the overlap of the time periods between the two, mark the resource path order position and the degree of overlap of state coverage, and generate a resource path state mapping table. S502: Call the path order field and status coverage value in the resource path status mapping table, perform a filtering operation on resource items that are at the beginning of the path order and whose coverage and the number of times the nodes in the status sequence appear repeatedly, store the index information of the filtered resource items in a corresponding relationship with the path order, and generate a path coverage resource set. S503: Based on the path-covered resource set, call the knowledge point coverage field and the learning time tag field, perform knowledge point number merging operation and time tag standardization processing on each resource, and rearrange and combine the two types of information according to the path order structure to generate a personalized resource recommendation task set.

[0013] A deep learning-based personalized recommendation system for educational resources includes: The trajectory extraction module obtains students' periodic report cards, course chapter completion records, and knowledge point pass records. It also collects students' practice logs, video playback interruption frequency, and interactive test scores within a week. The module calls upon the completion rate of knowledge points in the long-term behavioral trajectory set and the interaction performance in the short-term behavioral feature set. It performs absolute value calculation on the difference between the two dimensions of data under the same knowledge point label to generate a mastery difference fusion value set. Based on the mastery difference fusion value set, the status coding module obtains the click path of the practice questions, the start and end nodes of video viewing, and the question type error record under the corresponding knowledge point tag. It calls the time index value in the click path of the practice questions and the position number of the question type error record to perform unified position coding. It performs interval cross-judgment on the playback time segment in the start and end nodes of video viewing and the knowledge point coverage time tag to generate an individual stage learning status coding sequence. The path filtering module calls the individual stage learning state encoding sequence, filters all resource sets that match the knowledge point tags, extracts the topic tags, recommendation frequency and course path relevance level of each resource, sets knowledge points with multiple consecutive jump behaviors in the transition state as nodes to be intervened, and generates a resource path candidate set. The interaction adaptation module collects the device usage time, answer interruption duration and interaction frequency curves of students in real-time learning scenarios through the resource path candidate set, calls the time parameters of device usage time and answer interruption duration to perform time period compression mapping processing, and generates an interaction matching adaptation table. The resource recommendation module extracts resource items with a matching degree higher than the main matching threshold based on the interaction matching adaptation table, marks the number of overlapping labels between the sequential position in the original resource path and the learning state sequence, and generates a personalized resource recommendation task set.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by integrating students' long-term learning trajectories and short-term behavioral characteristics, a weighted average of mastery differences in knowledge points is constructed. Individual learning states at different stages are archived and encoded in a unified state transition structure, accurately identifying nodes in the learning path that require intervention. Then, resources are prioritized and reorganized based on their topic tags, recommendation frequency, and path relevance to form a candidate set of resources that deeply matches the learning state at each stage. Simultaneously, real-time interactive signals are introduced to refine the resource suitability, and highly matched resources are filtered and sorted according to path order and state coverage. This achieves precise alignment between recommended resource content and students' current learning state, effectively improving the targeting, continuity, and suitability of recommendations. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a method for personalized recommendation of educational resources based on deep learning, comprising the following steps: S1: Obtain students' periodic report cards, course chapter completion records, and knowledge point pass records; extract long-term behavioral trajectory sets; collect practice logs, video playback interruption frequency, and interactive test scores within a week; extract short-term behavioral feature sets; establish mastery difference weights on the knowledge point dimension by comparing the mastery of knowledge points in the long-term trajectory set with the performance in the short-term behavioral feature set; and generate a knowledge point fusion representation set. S2: Using a knowledge point fusion representation set, select the exercise click path, video viewing start and end nodes, and question type error records corresponding to the tag items under the knowledge point. Normalize the execution position and type of the exercise click path and question type error records. Reconstruct the interval between the video viewing start and end nodes and the knowledge point time coverage tags. Establish a unified state transition structure for the knowledge point. Encode and archive in the transition structure to generate an individual stage learning state coding sequence. S3: Based on the individual stage learning state coding sequence, filter all resource sets that are consistent with the knowledge point tags corresponding to the coding sequence, extract the topic tags, recommendation frequency and course path relevance level of the resource content, mark the knowledge points with continuous transfer or jump behavior in the coding sequence as nodes to be intervened, and reorganize the nodes to be intervened with the resource topic tags after matching and prioritizing, and generate a resource path candidate set. S4: Collect interaction signal indicators in real-time learning scenarios through the resource path candidate set, including device usage time, answer interruption duration and interaction frequency curve. Compress and map the device usage time and interruption duration execution time, perform step-level segmentation of the interaction frequency curve, and compare it with resource type attributes to generate an interaction matching adaptation table. S5: Based on the interaction matching adaptation table, extract resource items with matching levels higher than the main adaptation threshold, mark the path order and the coverage of the real-time learning state, retain resources with priority in path order and overlapping coverage with the state sequence, and combine the knowledge point coverage and learning time tags of the resources to obtain a personalized resource recommendation task set. The knowledge point fusion representation set includes mastery state difference labels, behavioral feature matching vectors, fusion dimension mapping indexes, individual performance offsets, and temporal weight correction factors. The individual stage learning state encoding sequence includes behavioral encoding fragment sequences, knowledge point label sequences, state transition relationship tables, stage switching identifiers, and unified format index codes. The resource path candidate set includes a resource topic index table, matching priority matrix, path reorganization list, intervention node tag book, and course structure alignment records. The interaction matching adaptation table includes device usage mapping segments, interaction intensity interval tables, resource response attribute sets, time compression node sequences, and adaptation level indexes. The personalized resource recommendation task set includes a path number mapping table, a recommended resource list, learning state coverage labels, priority recommendation sequences, and coverage overlap indicators.

[0023] Please see Figure 2 The specific steps of S1 are as follows: S101: Obtain the periodic grade sheet, course chapter completion records and knowledge point passing records, parse the course identifier field in the periodic grade sheet, locate and match the chapter index value in the course chapter completion record, extract the correspondence of chapter index values, aggregate the knowledge point numbers with the same chapter index value in the knowledge point passing records, and generate a long-term behavior trajectory index set. The course identifier field for each student is extracted, consisting of the course number and class code. The parsing process requires calling the course information table to read the correspondence between the course number and the course chapter. The course identifier is standardized, for example, "2024 Course-A1" is uniformly converted to "Course Number 2024, Chapter Number A1". Then, it is matched with the chapter number field in the course chapter completion record. The student's learning chapter under the course is located by field association. After successful matching, the chapter index value is extracted and recorded in a temporary table to store the correspondence between student ID and chapter number. The knowledge point is called through the record, and the knowledge point number field is aggregated. In the aggregation operation, it is necessary to locate the chapter to which the knowledge point belongs to, to ensure that it is consistent with the aforementioned chapter index. The corresponding knowledge point numbers are summarized under each chapter. For example, if the knowledge point numbers passed by the student under chapter A1 are K01, K02, and K05, the numbers are uniformly collected under chapter A1 and recorded in the long-term behavior trajectory index set. The whole process involves multiple association matching operations between chapter numbers and knowledge point numbers to ensure that the relationship between data is accurate and to facilitate the extraction and analysis of behavioral features, generating a long-term behavior trajectory index set.

[0024] S102: Call the long-term behavior trajectory index set, collect practice logs, video playback interruption frequency and interactive test score data within a week, filter and extract the practice record timestamp field in the practice log, select data frames belonging to the time period within a week, count the video playback interruption frequency, and associate and match the answer accuracy field in the interactive test score with the course chapter, aggregate them into behavioral feature triplets under the same knowledge point number, and generate a short-term behavior feature set; Relevant practice log data from the past week was filtered out. The logs included a practice time field. By setting a time range, only records from the most recent 7 days were retained. The data was categorized and statistically analyzed according to the knowledge point number to form practice frequency data for each knowledge point. Simultaneously, video playback records were retrieved to check for interruptions during playback. The identification method was generally based on user pause, skip, or exit operation markers. The number of interruptions in the video playback records of specific chapters was accumulated to obtain the interruption frequency of videos associated with each knowledge point. Then, the interactive test score data was processed to extract students' answers to questions under relevant knowledge points, including the number of questions and the number of correct answers in each test. This was summarized as the test accuracy rate and uniformly aggregated according to the knowledge point number. The practice frequency, video interruption frequency, and test accuracy rate were combined to construct behavioral feature triples. These were then categorized and organized according to the knowledge point number to generate a short-term behavioral feature set.

[0025] S103: Call the short-term behavioral feature set, extract three types of data frames from the behavioral features: practice frequency, number of video interruptions and accuracy of interactive tests, calculate the numerical difference between them and the completion status value under the knowledge point number in the long-term trajectory, and normalize each difference according to the knowledge point number to generate a knowledge point fusion representation set. The system compares each knowledge point number with the completion status records in the long-term trajectory, calculating the numerical difference between the two. For example, if a student's past practice frequency for a certain knowledge point is significantly lower than the number of practice sessions within a week, the difference is positive. The difference data is then organized and archived by knowledge point number. The difference data is normalized to map the original values ​​to a unified interval for easy representation and analysis. In this process, each data item needs to be normalized separately. For example, before normalizing the difference in the number of video interruptions, the interruption difference data under the knowledge point needs to be collected first, and processed based on the minimum and maximum values. The difference in the accuracy rate of interactive tests is also normalized using the same method. Finally, the data from the three dimensions are combined as the fusion feature representation of the same knowledge point to generate a knowledge point fusion representation set.

[0026] Please see Figure 3 The specific steps of S2 are as follows: S201: Call the knowledge point fusion representation set, select the practice question click path data frame, video viewing start and end node record and question type error record data frame associated with the tag item under the corresponding knowledge point number, match and verify the click node index field in the practice question click path and the error question type identifier field in the question type error record respectively, and perform index position normalization operation and error type number normalization operation on the entries with the same knowledge point number in both, and generate path and error code information; The click node index field is extracted from the click path data of practice questions. The corresponding structure contains the sequential number of the practice questions clicked by the student on the page. For example, a student clicks in the order [1, 3, 5, 6] under knowledge point K101, indicating that some questions were skipped. Then, the error question type identifier field is extracted from the question type error record. Fields such as "fill in the blank", "selection", and "true" are used to record which type of question the student made a mistake in. The two types of data are cross-filtered by the knowledge point number field, and records with the same knowledge point number are retained. A uniform standard normalization operation is performed on the click node index in the practice question path. For example, the maximum number of click steps is normalized to 1, and the rest are scaled according to the ratio. The original path [1, 3, 5, 6] is mapped to [0.2, 0.6, 1.0, 1.2]. Similarly, for different question types in the error type identifier field, number normalization is also performed. "Fill in the blank", "selection", and "true" are uniformly mapped to fixed numbers such as [0.1, 0.5, 1.0] to generate path and error code information.

[0027] S202: Based on the path and error coding information, extract the start and end timestamps of the corresponding nodes, and perform cross-comparison and compensation merging operations on the start and end timestamps and time coverage label intervals. Based on the comparison results, reconstruct the start and end node intervals to generate a video time interval comparison set. The start and end time information is cross-referenced with a preset time coverage label interval. The time coverage label interval generally refers to the time period considered to be effective behavior in the knowledge point learning task. For example, 08:30 to 09:00 is a high-frequency learning period. The cross-reference operation is completed by judging whether the start and end times are within the label interval. Then, the times that are not within the interval but are close to the boundary are compensated and merged. For example, if the start time of a path is 08:28 and the end time is 09:02, which has a slight deviation from the label interval, it is integrated into the coverage interval and redefined as 08:30 to 09:00. This process is achieved by setting a compensation interval for the time difference range. The compensation rule is generally set to a merging tolerance interval of no more than 5 minutes in advance and no more than 10 minutes in delay. After completing the cross-reference and compensation operations, a video time interval comparison set is generated.

[0028] S203: Based on the video time interval reference set, construct a one-to-one mapping structure between knowledge point number and normalized coding field, define the state transition node as a triplet of path position, video time point and error type code, and perform position coding and archiving annotation operations according to the order of transition nodes to generate an individual stage learning state coding sequence. The system retrieves the path location code, video time interval, and error type code from each record, combining these three elements to define a state transition node. This forms a complete triplet structure, representing the error type encountered by the student at a specific time and path location. This state transition node is stored as an important recording unit of the behavioral trajectory. The state transition nodes are then sorted and numbered according to the order of the start time of the video time interval, forming an ordered sequence. Each node is assigned a location code, such as S1 for the first node, S2 for the second, and so on. Next, archiving and annotation are performed, categorizing the ordered sequence into the individual stage learning record table according to the knowledge point number. Each student has an independent coding sequence under each knowledge point. The sequence records the complete path information of the behavioral path, time point, and error manifestation, generating an individual stage learning state coding sequence.

[0029] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the individual stage learning state coding sequence, perform a label matching operation, filter all educational resource items that match the labels, extract the topic label field, recommendation frequency parameter value and course path relevance level identifier from the matched resource items, and generate a set of resource content attributes; The system extracts the knowledge point number and error type code corresponding to each node and compares them one by one with the tags already set in the educational resource library. Each educational resource entry contains a knowledge point topic tag, a recommendation frequency parameter, and a course path relevance level field. Matching is achieved through the knowledge point tag field. For example, if a student has multiple error indicators in the status code under the K102 knowledge point, educational resource entries labeled K102 will be filtered out from the resource library. After a successful match, the topic tag field is extracted from the resource entry, such as "function application" or "probability basics". The recommendation frequency parameter is then read from it. The recommendation frequency parameter is set according to the resource's past usage frequency, such as 3 times / week or 1 time / day. The course path relevance level identifier, such as "high relevance", "medium relevance", or "low relevance", is read to generate a set of resource content attributes.

[0030] S302: Call the resource content attribute set to retrieve and identify continuous state transition nodes and jump behavior nodes. During the identification process, the time sequence field and position jump field in the state code are used to make judgments. Knowledge points with a continuous number of transitions greater than the specified jump identification threshold are marked as items that need to be processed, and nodes whose position jumps exceed the normal path length range are marked as abnormal items, generating a set of nodes to be intervened. The system retrieves and identifies nodes exhibiting continuous state transitions and jump behaviors. The identification process primarily relies on analyzing the time sequence and location number fields of each node in the encoded sequence to check for continuous state transitions. If multiple consecutive nodes appear under a given knowledge point, and the time interval between adjacent nodes is less than the set minimum learning interval standard, the node is considered a continuous behavior chain. The number of transitions in each chain is then counted. If the number of consecutive transitions exceeds the set jump identification threshold, for example, more than 5 times, the knowledge point is marked as an item requiring processing. On the other hand, the system checks for position jump behaviors, identifying whether there are abnormal spans in the path location numbers of nodes. If the difference in path location numbers between two adjacent nodes significantly exceeds the upper limit of the normal path length, such as more than 20 steps, it is determined to be a jump behavior. These nodes are marked as anomalies, generating a set of nodes to be intervened.

[0031] S303: Based on the set of nodes to be intervened, perform tag matching degree judgment and recommendation frequency value descending sorting operations respectively, and perform level filtering processing in combination with course path relevance level identifier. Aggregate the sorted results by knowledge point number to generate resource path candidate set; For each node, a tag matching degree judgment operation is performed to check the consistency between the corresponding knowledge point tags and the topic tag fields in the resource content attribute set. The matching degree is evaluated based on tag similarity. If they are completely consistent, it is considered a 100% match. If only some keywords match, it is divided into levels such as 80% and 60%. A preset tag matching algorithm is used to score and record each matching result. After completing the matching degree scoring, the resource items associated with the node are sorted in descending order according to the recommendation frequency parameter value, and resources with higher recommendation frequency are placed in a priority position. For example, if there are three resources under a certain knowledge point with recommendation frequencies of 5 times / week, 2 times / week, and 1 time / week, the order is 5>2>1. Based on the course path relevance level identifier, a level filter is performed, retaining only items with a relevance level of "high" or "medium" and deleting "low" related resources. The filtered items are aggregated by knowledge point number to generate a resource path candidate set.

[0032] Please see Figure 5 The specific steps of S4 are as follows: S401: Collect interactive signal indicator data frames corresponding to the real-time learning scenario through the resource path candidate set, including the start and end time period in the device usage record and the interruption duration field in the question-answering interruption behavior record, identify frequency change data points in the interactive operation log, perform time axis compression mapping operation on the device usage start and end time period and interruption duration field, and generate compressed time behavior mapping group. The system collects various interactive signal data from students' current learning environment, primarily including device usage records, question interruption behaviors, and interactive operation logs. Device usage records show the time points for each device startup and shutdown, identifying the actual time period of the learning activity. For example, if a student starts their tablet at 08:00 and shuts it down at 09:15, the usage period is 75 minutes. The system extracts the interruption duration field from the question interruption behavior records, reflecting the duration of user pauses, departures, or prolonged periods of unresponsiveness during the question-answering process. For instance, if two interruptions occur during a question-answering session, lasting 90 seconds and 120 seconds respectively, the system synchronizes the time periods and interruption times, compressing the complete device usage timeline onto the effective learning behavior timeline. A mapping method is used to remove interruption times from the main timeline, retaining only continuous effective time periods. After timeline compression during the mapping process, if 10 minutes of the original 75-minute timeline are invalid interruptions, the remaining 65 minutes are used as the compressed effective timeline. Simultaneously, frequency change data points extracted from the interactive operation logs, such as page click frequency and question submission frequency, are marked at the corresponding positions on the compressed timeline, generating compressed time behavior mapping groups.

[0033] S402: Call the compressed time behavior mapping group, calculate the frequency fluctuation gradient value of each curve segment, perform threshold division processing on the fluctuation gradient, perform multi-segment step-level segmentation and classification, attach multi-segment hierarchical labels to the original frequency point sequence, and generate a step-segment interactive frequency set. The fluctuation gradient value is calculated for the frequency change within each continuous time period. The process relies on the comparison of the frequency change rate. For example, if a user's click frequency increases from 1 time per minute to 5 times per minute between the 10th and 15th minutes, it is identified as an upward fluctuation gradient. The upward fluctuation gradient is calculated by the ratio of the change before and after, and is divided according to the preset gradient range. For example, gradient 0 to 1 is stable, 1 to 3 is moderate fluctuation, and above 3 is severe fluctuation. Based on this, a multi-segment step-like classification is performed on each frequency curve, dividing the original frequency points into multiple level segments, and the level label is directly attached to the original frequency point sequence. For example, the click frequency at the 12th minute is marked as "moderate fluctuation", and the frequency change at the 18th minute is large, so it is marked as "severe fluctuation". Each point in the entire sequence is given a clear step label, generating a step-segmented interactive frequency set.

[0034] S403: Based on the tiered segmented interaction frequency set, call the type attribute field of the resource item, perform a mapping consistency comparison operation between the resource type field and the frequency classification label, align the matching position according to the time period of the compressed mapping, and generate an interaction matching adaptation table. The system calls the resource entry type field, which identifies the resource type, such as "video explanation," "practice reinforcement," and "concept demonstration." A frequency adaptation level range is defined for each resource type; for example, "video explanation" adapts to the low-frequency range, "practice reinforcement" to the mid-frequency range, and "concept demonstration" to the high-frequency range. By comparing the frequency classification label with the preset adaptation rules between the resource types, the system checks whether the current frequency level matches the resource type. If they match, it is marked as a match. For example, if a student interaction frequency is marked as "medium amplitude fluctuation" at the 15th minute, the corresponding recommended resource type should be "practice reinforcement." This comparison operation is performed on frequency points to ensure that the resources allocated at each time point match the actual behavior intensity. Based on the positional relationship of the compressed time periods in the time behavior mapping group, the matching resources are allocated to the corresponding time axis segments, generating an interaction matching adaptation table.

[0035] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the interaction matching adaptation table, extract resource entries whose matching level field is higher than the main adaptation threshold, and retrieve the path order field corresponding to each resource and the time period mapping range of the state node in the individual stage learning state coding sequence. Perform the overlap calculation operation based on the overlap of the time periods between the two, mark the resource path order position and the degree of overlap of state coverage, and generate a resource path state mapping table. The threshold is set based on learning behavior and resource matching accuracy. For example, if it is set to 80%, the matching level must reach 80% or higher to enter the filtering list. After extraction, the path sequence field is called for each resource. The field indicates the sequential position of the resource in the original recommended path. At the same time, the time period mapping data of the corresponding knowledge point in the individual stage learning status coding sequence is retrieved. The overlap between the resource recommendation time and the actual learning time of the student status node is calculated. By comparing the ratio of the intersection length of the two time periods to the length of the resource time period, the overlap of time periods is determined. If the overlap exceeds 50% of the original resource time period, it is marked as high overlap. At the same time, the path sequence position of the resource and the coverage of the corresponding status node are recorded. The resource path position, status time overlap, resource number and matching level are assembled into a resource path status mapping table.

[0036] S502: Call the path order field and status coverage value in the resource path status mapping table, perform a filtering operation on resource items that are at the beginning of the path order and have a high degree of coverage and a high frequency of repetition of nodes in the status sequence, store the index information of the filtered resource items in a corresponding relationship with the path order, and generate a path coverage resource set. Prioritize resource items that appear earlier in the path order, i.e., earlier in the recommendation structure. Perform state coverage analysis on resource items to identify whether the corresponding knowledge point nodes in the learning state encoding sequence of the individual stage appear repeatedly. If the number of times a state node covered by a resource appears in the sequence is greater than a specified repetition threshold, such as more than 3 times, it is considered that the corresponding behavior is frequent. Based on this, resource items with high coverage and behavior repetition are selected. After selection, the unique identifier field of the resource item is retained, and it is organized into a structured index group that corresponds one-to-one with the original path order field to ensure that each resource still retains its relative position order in the original recommendation path, thus generating a path-covered resource set.

[0037] S503: Based on the path-covered resource set, call the knowledge point coverage field and the learning time tag field, perform knowledge point number merging operation and time tag standardization processing on each resource, and rearrange and combine the two types of information according to the path order structure to generate a personalized resource recommendation task set; Extract the IDs of multiple knowledge points involved in the resource and cluster them. For example, if a resource covers three knowledge points K101, K102, and K105, it will be aggregated into the same knowledge point cluster for representation. The merged knowledge point field is used to uniformly evaluate the breadth of knowledge points in the resource. The learning time tag field is standardized. Text-based time tags such as "morning study" and "after-class review" are uniformly converted into standardized time period codes. For example, "morning study" is uniformly mapped to "08:00-12:00". The standardization process is performed according to the set time period template to ensure that tags from different sources have a unified time expression structure. The knowledge point merging results and the standardized time period information are recombined according to the path order. Each resource is rearranged according to the order in the original path, and structured resource recommendation task entries are output. Each task includes the resource ID, merged knowledge point cluster, learning time period and its path order tag, forming a personalized resource recommendation task set.

[0038] Please see Figure 7 A personalized educational resource recommendation system based on deep learning includes: The trajectory extraction module obtains students' periodic report cards, course chapter completion records, and knowledge point pass records. It also collects students' practice logs, video playback interruption frequency, and interactive test scores within a week. The module calls upon the completion rate of knowledge points in the long-term behavioral trajectory set and the interaction performance in the short-term behavioral feature set. It performs absolute value calculation on the difference between the two dimensions of data under the same knowledge point label to generate a mastery difference fusion value set. The status coding module, based on the mastery difference fusion value set, obtains the click path of the practice questions under the corresponding knowledge point tags, the start and end nodes of video viewing, and the question type error records. It calls the time index value in the click path of the practice questions and the position number of the question type error record to perform unified position coding. It performs interval cross-judgment on the playback time segment in the start and end nodes of video viewing and the knowledge point coverage time tag to generate an individual stage learning status coding sequence. The path filtering module calls the individual stage learning state encoding sequence, filters all resource sets that match the knowledge point tags, extracts the topic tags, recommendation frequency and course path relevance level of each resource, sets the knowledge points with multiple consecutive jump behaviors in the transition state as nodes to be intervened, and generates a resource path candidate set. The interaction adaptation module collects the device usage time, answer interruption duration and interaction frequency curves of students in real-time learning scenarios through the resource path candidate set, calls the time parameters of device usage time and answer interruption duration to perform time period compression mapping processing, and generates an interaction matching adaptation table. The resource recommendation module extracts resource items with a matching degree higher than the main matching threshold based on the interaction matching adaptation table, marks the number of overlapping labels between the sequential position in the original resource path and the learning state sequence, and generates a personalized resource recommendation task set.

[0039] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A personalized recommendation method for educational resources based on deep learning, characterized in that, Includes the following steps: S1: Obtain students' periodic transcripts, course chapter completion records, and knowledge point pass records; extract long-term behavioral trajectory sets; collect practice logs, video playback interruption frequency, and interactive test scores; extract short-term behavioral feature sets; establish mastery difference weights on the knowledge point dimension by comparing the mastery of knowledge points in the long-term trajectory set with the performance in the short-term behavioral feature set; and generate a knowledge point fusion representation set. S2: Using the knowledge point fusion representation set, select the exercise question click path, video viewing start and end nodes and question type error records corresponding to the tag items under the knowledge points, and normalize and encode the execution position and type of the exercise question click path and question type error records to generate an individual stage learning state encoding sequence. S3: Based on the individual stage learning state encoding sequence, filter all resource sets that are consistent with the knowledge point labels corresponding to the encoding sequence, mark the knowledge points with continuous transfer or jump behavior in the encoding sequence as nodes to be intervened, and generate a resource path candidate set; S4: Through the resource path candidate set, collect interaction signal indicators in the real-time learning scenario, compress and map the device usage period with the interruption duration execution time, and generate an interaction matching adaptation table.

2. The method for personalized recommendation of educational resources based on deep learning according to claim 1, characterized in that, The knowledge point fusion representation set includes mastery state difference labels, behavioral feature matching vectors, fusion dimension mapping indexes, individual performance offsets, and temporal weight correction factors. The individual stage learning state encoding sequence includes behavioral encoding fragment sequences, knowledge point label sequences, state transition relationship tables, stage switching identifiers, and unified format index codes. The resource path candidate set includes a resource topic index table, matching priority matrix, path reorganization list, intervention node tag book, and course structure alignment records. The interaction matching adaptation table includes device usage mapping segments, interaction intensity interval tables, resource response attribute sets, time compression node sequences, and adaptation level indexes.

3. The method for personalized recommendation of educational resources based on deep learning according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Obtain the periodic grade sheet, course chapter completion records and knowledge point passing records, parse the course identifier field in the periodic grade sheet, locate and match the chapter index value in the course chapter completion record, extract the correspondence of chapter index values, aggregate the knowledge point numbers with the same chapter index value in the knowledge point passing records, and generate a long-term behavior trajectory index set. S102: Call the long-term behavior trajectory index set, collect practice logs, video playback interruption frequency and interactive test score data within one week, filter and extract the practice record timestamp field in the practice log, filter the data frames belonging to the time period within one week, count the video playback interruption frequency, and associate and match the answer accuracy field in the interactive test score with the course chapter, aggregate them into behavioral feature triplets under the same knowledge point number, and generate a short-term behavior feature set; S103: Call the short-term behavioral feature set, extract three types of data frames from the behavioral features: practice frequency, video interruption number and interactive test accuracy, calculate the numerical difference between each and the completion status value under the knowledge point number in the long-term trajectory, and normalize each difference according to the knowledge point number to generate a knowledge point fusion representation set.

4. The method for personalized recommendation of educational resources based on deep learning according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Call the knowledge point fusion representation set, select the practice question click path data frame, video viewing start and end node record and question type error record data frame associated with the tag item under the corresponding knowledge point number, match and verify the click node index field in the practice question click path and the error question type identifier field in the question type error record respectively, and perform index position normalization operation and error type number normalization operation on the entries with the same knowledge point number in both, and generate path and error code information; S202: Based on the path and error coding information, extract the start and end timestamps of the corresponding nodes, and perform cross-comparison and compensation merging operations on the start and end timestamps and time coverage label intervals. Based on the comparison results, reconstruct the start and end node intervals to generate a video time interval comparison set. S203: Based on the video time interval reference set, construct a one-to-one mapping structure between knowledge point number and normalized encoding field, define the state transition node as a triplet of path position, video time point and error type encoding, and perform position encoding and archiving annotation operations according to the order of transition nodes to generate an individual stage learning state encoding sequence.

5. The method for personalized recommendation of educational resources based on deep learning according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Based on the individual stage learning state coding sequence, perform a tag matching operation, filter all educational resource items that match the tags, extract the topic tag field, recommendation frequency parameter value and course path correlation level identifier from the matched resource items, and generate a resource content attribute set; S302: Call the resource content attribute set to retrieve and identify continuous state transition nodes and jump behavior nodes. During the identification process, the time sequence field and position jump field in the state code are used to make judgments. Knowledge points with a continuous number of transitions greater than the specified jump identification threshold are marked as items that need to be processed, and nodes whose position jumps exceed the normal path length range are marked as abnormal items, generating a set of nodes to be intervened. S303: Based on the set of nodes to be intervened, perform tag matching degree determination and descending order of recommendation frequency value respectively, and perform level filtering processing in combination with course path relevance level identifier. Aggregate the sorted results according to knowledge point number to generate resource path candidate set.

6. The method for personalized recommendation of educational resources based on deep learning according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Through the resource path candidate set, collect the corresponding interactive signal indicator data frames in the real-time learning scenario, including the start and end time period in the device usage record and the interruption duration field in the question-answering interruption behavior record, identify the frequency change data points in the interactive operation log, perform time axis compression mapping operation on the device usage start and end time period and interruption duration field, and generate a compressed time behavior mapping group. S402: Call the compressed time behavior mapping group, calculate the frequency fluctuation gradient value of each curve segment, perform threshold division processing on the fluctuation gradient, perform multi-segment step-level segmentation and classification, attach multi-segment hierarchical labels to the original frequency point sequence, and generate a step-segment interactive frequency set. S403: Based on the tiered segmented interaction frequency set, call the type attribute field of the resource item, perform a mapping consistency comparison operation between the resource type field and the frequency classification label, align the matching position according to the time period of the compressed mapping, and generate an interaction matching adaptation table.

7. The method for personalized recommendation of educational resources based on deep learning according to claim 6, characterized in that, In the operation of thresholding the fluctuation gradient, at least three frequency fluctuation gradient threshold intervals are preset, each corresponding to a segmented classification label of a different level. The frequency fluctuation gradient threshold intervals include a first frequency fluctuation gradient threshold interval, a second frequency fluctuation gradient threshold interval, and a third frequency fluctuation gradient threshold interval. The first frequency fluctuation gradient threshold interval is used to mark the low frequency fluctuation state, and is correspondingly labeled as the first segmentation classification label; The second frequency fluctuation gradient threshold interval is used to mark the medium frequency fluctuation state, and is correspondingly labeled as the second segmentation classification label; The third frequency fluctuation gradient threshold interval is used to mark high frequency fluctuation states, and is correspondingly labeled as the third segmentation classification label. In the process of adding multiple segmented hierarchical labels to the original frequency point sequence, the corresponding segmented classification labels are determined based on the threshold range into which the frequency fluctuation gradient value falls, thus forming the stepped segmented interactive frequency set.

8. The method for personalized recommendation of educational resources based on deep learning according to claim 1, characterized in that, The method further includes step S5: S5: Based on the interaction matching adaptation table, extract resource items with matching levels higher than the main adaptation threshold, mark the path order and the coverage of the real-time learning state, retain resources with priority in path order and overlapping coverage with the state sequence, and combine the knowledge point coverage and learning time tags of the resources to obtain a personalized resource recommendation task set. The personalized resource recommendation task set includes a path number mapping table, a list of recommended resources, learning status coverage tags, a priority recommendation sequence, and a coverage overlap index.

9. The method for personalized recommendation of educational resources based on deep learning according to claim 8, characterized in that, The specific steps of S5 are as follows: S501: Based on the interaction matching adaptation table, extract resource entries whose matching level field is higher than the main adaptation threshold, and retrieve the path order field corresponding to each resource and the time period mapping range of the state node in the individual stage learning state encoding sequence. Perform the overlap calculation operation based on the overlap of the time periods between the two, mark the resource path order position and the degree of overlap of state coverage, and generate a resource path state mapping table. S502: Call the path order field and status coverage value in the resource path status mapping table, perform a filtering operation on resource items that are at the beginning of the path order and whose coverage and the number of times the nodes in the status sequence appear repeatedly, store the index information of the filtered resource items in a corresponding relationship with the path order, and generate a path coverage resource set. S503: Based on the path-covered resource set, call the knowledge point coverage field and the learning time tag field, perform knowledge point number merging operation and time tag standardization processing on each resource, and rearrange and combine the two types of information according to the path order structure to generate a personalized resource recommendation task set.

10. A personalized educational resource recommendation system based on deep learning, characterized in that, The system is used to implement the deep learning-based personalized recommendation method for educational resources as described in any one of claims 1-9, and the system comprises: The trajectory extraction module obtains students' periodic report cards, course chapter completion records, and knowledge point pass records. It also collects students' practice logs, video playback interruption frequency, and interactive test scores within a week. The module calls upon the completion rate of knowledge points in the long-term behavioral trajectory set and the interaction performance in the short-term behavioral feature set. It performs absolute value calculation on the difference between the two dimensions of data under the same knowledge point label to generate a mastery difference fusion value set. Based on the mastery difference fusion value set, the status coding module obtains the click path of the practice questions, the start and end nodes of video viewing, and the question type error record under the corresponding knowledge point tag. It calls the time index value in the click path of the practice questions and the position number of the question type error record to perform unified position coding. It performs interval cross-judgment on the playback time segment in the start and end nodes of video viewing and the knowledge point coverage time tag to generate an individual stage learning status coding sequence. The path filtering module calls the individual stage learning state encoding sequence, filters all resource sets that match the knowledge point tags, extracts the topic tags, recommendation frequency and course path relevance level of each resource, sets knowledge points with multiple consecutive jump behaviors in the transition state as nodes to be intervened, and generates a resource path candidate set. The interaction adaptation module collects the device usage time, answer interruption duration and interaction frequency curves of students in real-time learning scenarios through the resource path candidate set, calls the time parameters of device usage time and answer interruption duration to perform time period compression mapping processing, and generates an interaction matching adaptation table. The resource recommendation module extracts resource items with a matching degree higher than the main matching threshold based on the interaction matching adaptation table, marks the number of overlapping labels between the sequential position in the original resource path and the learning state sequence, and generates a personalized resource recommendation task set.

Images