Second classroom student demand management method

Abstract

The invention provides a second classroom student demand management method. The second classroom student demand management method comprises the steps of calling a student dynamic demand portrait model to push activity resources during login; collecting a multi-modal interaction behavior sequence of the student in real time, and extracting multi-granularity spatial-temporal characteristics to generate an instant interest snapshot; if an interest evolution signal appears between the snapshot and the long-term interest file, inputting the interest evolution signal into a dynamic demand awareness network for reasoning to obtain a demand evolution vector and an intensity value; selecting aggressive or progressive strategies to update the portrait model; when the resource list is refreshed, the updated model is used for generating a personalized recommendation sequence and pushing the personalized recommendation sequence to the terminal, so that dynamic and accurate recommendation is realized. According to the invention, the pushing accuracy of the second classroom activity resources and the student participation degree can be improved, the cold start problem of newly registered or low-activity users is solved, and the adaptability of the system and the user experience are enhanced.

Description

Technical Field

[0001] This invention relates to the field of educational informatization technology, and more specifically to a method for managing student needs in extracurricular activities. Background Technology

[0002] With the continuous development of educational informatization, extracurricular activities, as an important supplement to traditional classroom teaching, have gradually become a crucial platform for cultivating students' comprehensive qualities, expanding their interests, and enhancing their practical abilities. In the process of conducting extracurricular activities, students face numerous different types of activity resources. How to accurately recommend activity resources that match students' interests and needs has become a key issue in improving the participation and effectiveness of extracurricular activities. Traditional methods of recommending activity resources are mainly based on students' static information, such as major and grade level. This method is relatively simple and cannot accurately capture the dynamic changes in students' interests at different stages. Furthermore, most existing recommendation systems only focus on students' explicit operational behaviors, neglecting their implicit attention behaviors during browsing, resulting in an incomplete understanding of students' interests.

[0003] The existing technology has at least the following problems or defects: it cannot perceive the dynamic evolution of students' interests in real time, it is difficult to accurately push activity resources based on students' immediate interests, and it lacks an effective cold start strategy when facing newly registered users or low-activity users, which makes it impossible for these users to obtain personalized activity resource recommendations. Summary of the Invention

[0004] This invention provides a method for managing student needs in extracurricular activities, including: When a student's login activity is detected on the second classroom management platform, second classroom activity resources are pushed to the student on the platform based on a pre-built dynamic student needs profile model. The system collects real-time sequences of students' multimodal interaction behaviors on the second classroom management platform interface, wherein the multimodal interaction behavior sequences include explicit operational behavior sequences and implicit attention behavior sequences. Multi-granular spatiotemporal features are extracted from the multimodal interaction behavior sequence and the corresponding second classroom activity resources to generate real-time interest snapshots of students; When it is determined that there is an interest evolution signal between the student's real-time interest snapshot and the student's long-term interest file corresponding to the student's dynamic demand profile model, the student's real-time interest snapshot is input into the dynamic demand perception network to perform demand evolution reasoning, and the student demand evolution vector and evolution intensity value are obtained. Based on the comparison between the evolution intensity value and the preset interest drift intensity threshold, according to the comparison result, the corresponding model update strategy in the hierarchical fusion update strategy library is selectively activated to update the student dynamic demand profile model, and the updated student dynamic demand profile model is obtained. When a resource list refresh command is detected on the second classroom management platform, a personalized recommendation sequence for second classroom activity resources is generated using the updated student dynamic demand profile model. The personalized recommendation sequence of activity resources for the second classroom is sent to the display terminal of the second classroom management platform to push the activity resources.

[0005] Furthermore, the step of extracting multi-granular spatiotemporal features from the multimodal interaction behavior sequence and the corresponding second-classroom activity resources to generate a snapshot of students' real-time interests includes: Extract operation type, operation frequency, and timing information from the explicit operation behavior sequence to generate explicit behavior feature vectors; Extract page dwell time, scroll speed change, and mouse movement trajectory entropy from the implicit attention behavior sequence to generate implicit attention feature vector; The second classroom activity resources are analyzed in a structured manner to obtain resource attribute feature vectors that include activity type, activity level, activity duration, and organizer. The explicit behavior feature vector, the implicit attention feature vector, and the resource attribute feature vector are aligned and concatenated according to the timestamp of the behavior to generate the student's real-time interest snapshot.

[0006] Furthermore, before determining that there is an interest evolution signal between the student's immediate interest snapshot and the student's long-term interest profile corresponding to the student's dynamic needs profile model, the method further includes: The student's real-time interest snapshot is input into the short-term interest fit evaluation channel in the dynamic demand perception network, and the short-term interest fit score is output. The short-term interest fit evaluation channel is constructed based on a gated recurrent unit network. The student's real-time interest snapshot is input into the long-term interest deviation evaluation channel in the dynamic demand perception network, and the long-term interest deviation score and potential interest direction vector are output. The short-term interest fit score is compared with a first dynamic threshold, and the long-term interest deviation score is compared with a second dynamic threshold. When the first determination condition is met, it is determined that there is an interest evolution signal, wherein the first determination condition is that the short-term interest fit score is lower than the first dynamic threshold and the long-term interest deviation score is higher than the second dynamic threshold.

[0007] Further, the step of comparing the evolution intensity value with a preset interest drift intensity threshold, and selectively activating the corresponding model update strategy in the hierarchical fusion update strategy library based on the comparison result, includes: When the evolution intensity value is greater than or equal to the interest drift intensity threshold, the aggressive interest fusion strategy in the hierarchical fusion update strategy library is activated, wherein the aggressive interest fusion strategy is used to fuse the potential interest direction vector into the student's long-term interest profile with a higher weight. When the evolution intensity value is less than the interest drift intensity threshold, the progressive interest calibration strategy in the hierarchical fusion update strategy library is activated, wherein the progressive interest calibration strategy is used to fine-tune the relevant feature weights in the student's long-term interest profile.

[0008] Furthermore, the implementation of the radical interest fusion strategy includes: Obtain the potential interest direction vector and the current core interest vector in the student's long-term interest profile; Calculate the orthogonal decomposition components and overlapping components between the potential interest direction vector and the current core interest vector; According to the formula Update the core interest vector, where, This represents the updated core interest vector. This represents the current core interest vector. This represents the orthogonal decomposition components. The first attenuation coefficient, The first gain coefficient, and .

[0009] Furthermore, the implementation of the progressive interest calibration strategy includes: Obtain the student demand evolution vector and the corresponding interest dimension weight distribution in the student's long-term interest profile; Calculate the projection components of the student demand evolution vector on each interest dimension; According to the formula Adjust the weights of the interest dimension, where, This represents the adjusted weight of the i-th interest dimension. Indicates the weights before adjustment. To calibrate the learning rate, This represents the projected component in dimension i. This is the adjustment range calculated based on the evolution intensity value.

[0010] Furthermore, the method also includes a cold start initialization process, which includes: When a target student is identified as a newly registered user or a low-activity user, the cold start initialization process is triggered. Obtain the structured academic records and unstructured personal profile text of the target student; Based on the structured student record information, a collaborative filtering algorithm is used to retrieve neighboring samples in similar student groups to generate an initial interest preference vector based on the group. Based on the unstructured personal profile text, embedding vectors of interest keywords are extracted using a natural language processing model; By integrating the initial interest preference vector with the interest keyword embedding vector, an initial dynamic student demand profile model for the target student is generated.

[0011] Further, the step of fusing the initial interest preference vector and the interest keyword embedding vector to generate the initial dynamic student demand profile model for the target student includes: The initial interest preference vector is standardized to obtain a first standardized vector; The embedding vector of the interest keywords is standardized to obtain a second standardized vector; Based on credibility weight and The first standardized vector and the second standardized vector are weighted and summed, where the confidence weights are... Dynamic calculations are performed based on the sample size and quality of the aforementioned similar student groups; The weighted summation result vector is used as the initial value of the long-term interest profile of the initial student dynamic demand profile model.

[0012] Furthermore, the first dynamic threshold and the second dynamic threshold are dynamically adjusted in the following manner: Obtain the historical update frequency of the student dynamic needs profile model; Obtain the current teaching cycle stage of the student in the current semester; Based on the historical update frequency and the teaching cycle stage, the first dynamic threshold and the second dynamic threshold are found and set through a predefined threshold mapping table.

[0013] Furthermore, the generation of personalized recommendation sequences for extracurricular activity resources includes: Based on the updated student dynamic needs profile model, calculate the initial recommendation score for all extracurricular activity resources in the platform's candidate pool; The initial recommendation score is input into the diversity protection module, which performs diversity rearrangement of the activity resource type and organizer dimension based on the maximum boundary correlation algorithm. Based on the rearranged order, select the top K second-classroom activity resources to form a personalized recommendation sequence for second-classroom activity resources.

[0014] The embodiments of the present invention have at least the following beneficial effects: 1. By collecting students' multimodal interaction behavior sequences in real time and combining them with multi-granular spatiotemporal feature extraction to generate instant interest snapshots, and comparing and analyzing them with long-term interest archives, the system effectively identifies interest evolution signals, thereby improving the timeliness and accuracy of recommendations.

[0015] 2. Based on the comparison between the evolution intensity value and the preset threshold, the demand profile model is updated by selectively activating either aggressive or gradual interest fusion strategies. The aggressive strategy can quickly respond to significant shifts in students' interests and efficiently integrate new potential interest directions into the long-term interest profile. The gradual strategy improves the adaptability and accuracy of the recommendation system by fine-tuning the weights of relevant features.

[0016] 3. By introducing a cold start initialization process, structured student records and unstructured personal profile text are used, combined with collaborative filtering algorithms and natural language processing technology to generate initial interest preference vectors and interest keyword embedding vectors, and then integrated to generate an initial demand profile model, thereby improving user experience and the system's attractiveness to new users. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The following drawings are not intentionally drawn to scale to actual size; their focus is on illustrating the main points of this disclosure.

[0018] Figure 1 This is a flowchart illustrating a second-classroom student needs management method according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. like Figure 1 As shown, this application proposes a method for managing student needs in extracurricular activities, including: S1. When a student's login operation is detected on the second classroom management platform, second classroom activity resources are pushed to the student on the second classroom management platform according to the pre-built student dynamic demand profile model. S2. Real-time collection of students' multimodal interaction behavior sequences on the second classroom management platform interface, wherein the multimodal interaction behavior sequences include explicit operation behavior sequences and implicit attention behavior sequences; S3. Extract multi-granularity spatiotemporal features from the multimodal interaction behavior sequence and the corresponding second classroom activity resources to generate a snapshot of students' real-time interests; S4. When it is determined that there is an interest evolution signal between the student's real-time interest snapshot and the student's long-term interest file corresponding to the student's dynamic demand profile model, the student's real-time interest snapshot is input into the dynamic demand perception network to perform demand evolution reasoning, and the student demand evolution vector and evolution intensity value are obtained. S5. Based on the comparison between the evolution intensity value and the preset interest drift intensity threshold, according to the comparison result, the corresponding model update strategy in the hierarchical fusion update strategy library is selectively activated to update the student dynamic demand profile model, and the updated student dynamic demand profile model is obtained. S6. When a resource list refresh instruction is detected on the second classroom management platform, a personalized recommendation sequence for second classroom activity resources is generated using the updated student dynamic demand profile model. S7. Send the personalized recommendation sequence of the second classroom activity resources to the display terminal of the second classroom management platform to push the activity resources.

[0020] The Second Classroom Management Platform is an online system designed to provide students with various activity resources beyond traditional classroom teaching, such as club activities, lectures, competitions, and volunteer services. The platform is responsible for publishing activities, student registration, displaying resources, and collecting data on student interactions.

[0021] A dynamic student needs profile model is a collection of data structures or algorithms that comprehensively and dynamically describe students' interests, preferences, abilities, and needs. It not only includes students' long-term stable interests but also reflects the short-term fluctuations and evolutionary trends of those interests in real time, serving as the core basis for personalized recommendations.

[0022] Multimodal interactive behavior sequences refer to the various forms of interactive data streams recorded in real time by the system when students use the extracurricular management platform. These include explicit operational behavior sequences, such as clicking, saving, and registering—behaviors that clearly express intent; and implicit attention behavior sequences, such as page dwell time, mouse movement trajectory, and scrolling speed—behaviors that indirectly reflect interest and attention.

[0023] A student instant interest snapshot is an instantaneous representation of a student's current interest state, formed by feature extraction and integration based on a sequence of multimodal interaction behaviors within a short period, such as a single conversation or browsing cycle. It reflects a student's interest in specific activity resources at a particular point in time.

[0024] A student's long-term interest profile is a component of a dynamic student needs profiling model, used to store and represent students' relatively stable and continuous interests and preferences. It is typically formed through long-term accumulation and learning of students' historical behavioral data and serves as a fundamental description of students' interests.

[0025] Interest evolution signals refer to indications that suggest a potential shift in a student's interests, identified by the system when significant differences or trend changes occur between a student's immediate interest snapshot and long-term interest profile. Detecting these signals is crucial for triggering subsequent demand evolution inference and model updates.

[0026] Dynamic Demand Aware Network is a specially designed machine learning model or algorithm framework that receives snapshots of students' immediate interests and analyzes the relationship between these snapshots and students' long-term interest profiles to infer the direction and intensity of the evolution of students' interests.

[0027] The student demand evolution vector is one of the outputs of the dynamic demand perception network. It represents the direction of the student's interest from the current state to possible future states in vector form.

[0028] The evolution intensity value is another output of the dynamic demand-aware network, which quantifies the intensity or importance of the evolution of students' interests. The higher the value, the more significant the change in interest.

[0029] The interest drift intensity threshold is a preset value used to compare with the evolution intensity value. By comparing, the system can determine whether the evolution of students' interests is a slight adjustment or a significant drift, and thus select different model update strategies.

[0030] The hierarchical fusion update strategy library contains a variety of strategies for updating student dynamic needs profile models. These strategies employ different fusion or calibration methods based on the intensity and nature of interest evolution to ensure the accuracy and adaptability of model updates.

[0031] The model update strategy is a specific algorithm or rule in the hierarchical fusion update strategy library, which is used to adjust and optimize the dynamic demand profile model of students based on the evolving information of students' needs.

[0032] The personalized recommendation sequence for extracurricular activity resources is a list of customized extracurricular activity resources generated by the system based on an updated dynamic student needs profile model. This sequence aims to match students' current interests and needs to the greatest extent possible.

[0033] When the system detects a student logging into the extracurricular management platform, it will push extracurricular activity resources to the student on the platform based on a pre-built dynamic student needs profile model. For example, the login could be a successful access to the platform using a student's account and password. This dynamic student needs profile model can be initially constructed during the student's first registration through a simple questionnaire or preset general interest tags. Based on this model, the system can push some general or initial interest-based activity resources to the student; for example, if the student selected sports as an interest during registration, some sports-related activities will be pushed.

[0034] The system collects real-time multimodal interaction sequences of students on the extracurricular management platform interface. These sequences include explicit operational behaviors and implicit attention behaviors. For example, explicit operational behaviors can be obtained by recording students' explicit interactive actions such as clicking, saving, and registering on the platform. Implicit attention behaviors can be obtained by recording behaviors that do not directly express interest, such as the time students spend on activity details pages and the mouse movement trajectory in specific areas. This behavioral data is continuously collected to reflect students' real-time activities on the platform.

[0035] The system extracts multi-granularity spatiotemporal features from the multimodal interaction sequence and corresponding extracurricular activity resources to generate a snapshot of students' immediate interests. For example, the system can simply count which activity types a student clicked on and which organizers' activities they viewed in their most recent session, and use this statistical information as part of the snapshot of their immediate interests. This snapshot is an instantaneous summary of the student's current interests.

[0036] When an interest evolution signal is determined to exist between a student's immediate interest snapshot and the student's long-term interest profile corresponding to the student's dynamic needs profile model, the immediate interest snapshot is input into the dynamic needs perception network for needs evolution inference, obtaining the student's needs evolution vector and evolution intensity value. For example, the system can simply compare the activity types appearing in the immediate interest snapshot with the activity types recorded in the student's long-term interest profile. If a new activity type appears in the immediate interest snapshot that has never appeared in the long-term interest profile, it can be preliminarily determined that an interest evolution signal exists. The dynamic needs perception network can be a simple rule engine that determines the direction and intensity of interest evolution based on preset rules.

[0037] Based on a comparison between the evolution intensity value and a preset interest drift intensity threshold, the system selectively activates the corresponding model update strategy from the hierarchical fusion update strategy library to update the student dynamic demand profile model, resulting in an updated student dynamic demand profile model. For example, if the evolution intensity value is higher than the preset threshold, indicating a significant change in interests, the system can activate a comprehensive update strategy to directly add the new interests from the immediate interest snapshot to the long-term interest profile. If the evolution intensity value is lower than the preset threshold, a fine-tuning strategy is activated, making only minor adjustments to the weights in the long-term interest profile related to the immediate interest snapshot.

[0038] When a resource list refresh command is detected on the extracurricular activity management platform, the updated student dynamic needs profile model is used to generate a personalized recommendation sequence of extracurricular activity resources. For example, the refresh command could be initiated by a student clicking the refresh button, or it could be automatically triggered by the system after browsing for a period of time. Based on the updated student dynamic needs profile model, the system will filter out the activities that best match the student's current interests from all the platform's activity resources and sort them according to their matching degree to form a recommendation sequence.

[0039] The personalized recommendation sequence of extracurricular activity resources is sent to the display terminal of the extracurricular management platform for activity resource push. This display terminal can be a student's computer screen, mobile phone screen, etc. Students will see a list of activities customized according to their latest interests and needs.

[0040] In other implementations, upon detecting a student's login activity on the extracurricular management platform, extracurricular activity resources are pushed to the student on the platform based on a pre-built dynamic student needs profile model. The method also collects real-time sequences of students' multimodal interaction behaviors on the extracurricular management platform interface, including explicit operational behavior sequences and implicit attention behavior sequences.

[0041] Multi-granular spatiotemporal features are extracted from multimodal interaction behavior sequences and corresponding extracurricular activity resources to generate student instant interest snapshots. When it is determined that there is an interest evolution signal between the student's instant interest snapshot and the student's long-term interest profile corresponding to the student's dynamic needs profile model, the student's instant interest snapshot is input into the dynamic needs perception network for needs evolution inference to obtain the student's needs evolution vector and evolution intensity value.

[0042] Based on a comparison between the evolution intensity value and a preset interest drift intensity threshold, and according to the comparison results, the corresponding model update strategy in the hierarchical fusion update strategy library is selectively activated to update the student dynamic demand profile model, resulting in an updated student dynamic demand profile model. When a resource list refresh command is detected from the second classroom management platform, a personalized recommendation sequence of second classroom activity resources is generated using the updated student dynamic demand profile model, and this personalized recommendation sequence is sent to the display terminal of the second classroom management platform for activity resource push.

[0043] Extracting operation type, frequency, and timing information from explicit operational behavior sequences to generate explicit behavioral feature vectors refers to analyzing the records of students' direct and explicit interactive actions on the second-classroom management platform interface. Operation type aims to identify the specific actions students performed, such as clicking, saving, registering, and searching, reflecting their direct intentions. Operation frequency quantifies the repetition of specific actions, indicating the intensity of interest. Timing information records the time sequence and intervals of operations, revealing dynamic changes in interest. By comprehensively extracting this information, an explicit behavioral feature vector representing students' active interest preferences can be constructed. This process can be achieved by embedding tracking points on the front-end interface to record the type, timestamp, and related parameters of each user's click, input, and submission, and transmitting this log data to the back-end for parsing and aggregation to form a structured explicit behavioral feature vector. Alternatively, browser plugins or client SDKs can be used to capture user interaction events in specific areas in real time, such as accessing an activity details page or clicking a registration button, and then classifying, statistically analyzing, and performing time-series analysis on these events to generate explicit behavioral feature vectors.

[0044] Extracting page dwell time, scroll speed changes, and mouse trajectory entropy from implicit attention behavior sequences to generate implicit attention feature vectors refers to analyzing students' indirect and implicit interactive behaviors on the second-classroom management platform interface. Page dwell time measures students' attention to specific content; scroll speed changes indicate students' browsing pace and points of interest; mouse trajectory entropy reflects the breadth and uncertainty of students' exploration on the page, with higher entropy values ​​potentially indicating stronger exploration or more dispersed interests. Extracting these implicit behavioral features helps capture students' deeper, unexpressed interests. This process involves using front-end JavaScript to listen to the `visibilitychange` and `scroll` events to record page entry and exit times and scroll bar position changes, and using the `mousemove` event to record mouse coordinates. This data is then processed to calculate page dwell time, scroll speed, and mouse trajectory entropy, thereby generating the implicit attention feature vector. Alternatively, eye-tracking technology or visual attention-based models can be used to estimate the user's gaze points and gaze duration on the screen. Combined with mouse behavior data, this can more accurately assess the user's implicit attention and quantify it into an implicit attention feature vector.

[0045] Structured analysis of extracurricular activity resources yields resource attribute feature vectors containing activity type, activity level, activity duration, and organizer. This involves converting various activity information provided by the platform, such as lectures, competitions, practical activities, and volunteer services, into structured data that can be processed by computers through predefined data models and parsing rules. Activity type (e.g., academic lectures, sports competitions), activity level (e.g., school-level, college-level, national-level), activity duration (e.g., 1 hour, half-day, week), and organizer (e.g., academic affairs office, a specific college, student association) are key dimensions describing the core attributes of the activities. These attributes are crucial for understanding the nature of the activities and matching them with student interests. Activity information on the platform can be categorized and tagged using manual annotation or semi-automated methods. Key entities, such as time, location, and organizer, can be extracted from the activity description text. This structured data is then stored in a database for subsequent querying and construction of resource attribute feature vectors. Alternatively, Natural Language Processing (NLP) techniques, such as Named Entity Recognition (NER) and text classification models, can be used to automatically identify and extract information such as event type, level, duration, and organizer from texts such as event titles and descriptions, and encode it into resource attribute feature vectors.

[0046] Aligning and splicing explicit behavior feature vectors, implicit attention feature vectors, and resource attribute feature vectors according to the timestamps of the behavior occurrence to generate a snapshot of students' real-time interests means synchronizing feature vectors from different sources and modalities according to the time points of their occurrence or association, and combining them into a unified, high-dimensional feature representation.

[0047] When generating a snapshot, a time window is selected, such as the last 5 minutes or the last 10 interactions. All relevant feature vectors within this window are arranged in chronological order, and then the features are concatenated, for example, through vector concatenation or weighted averaging, to form a snapshot representing the student's immediate interest within that time window. Alternatively, a sliding window mechanism can be used, updating the feature set within the current time window whenever a new interaction occurs, and recalculating or updating the student's immediate interest snapshot. During concatenation, feature fusion techniques, such as multimodal fusion networks, can be used to weightedly fuse features from different modalities to generate a more expressive snapshot of the student's immediate interest.

[0048] Furthermore, the short-term interest fit assessment channel is used by the dynamic demand-aware network to evaluate the degree of match between students' immediate interest snapshots and their current short-term interest patterns. This channel can capture the degree of consistency between students' recent behavior reflecting their immediate interests and the short-term interests that the system has learned, thereby identifying immediate changes in students' interests.

[0049] Short-term interest fit assessment channels can be constructed based on Gated Recurrent Unit (GRU) networks. The GRU network learns and memorizes behavioral sequence features from students' immediate interest snapshots, compares them with pre-trained short-term interest patterns, and ultimately outputs a quantified short-term interest fit score. Alternatively, channels can be constructed based on Long Short-Term Memory (LSTM) networks or Transformer models. These models also possess the ability to process sequential data and capture complex temporal features, effectively evaluating the fit of immediate interests.

[0050] The Long-Term Interest Deviation Assessment Channel measures the difference between a student's immediate interest snapshot and their long-term interest profile, further identifying potential directions for interest development. This channel aims to uncover deeper, trend-based changes in students' interests, rather than superficial short-term fluctuations. The Long-Term Interest Deviation Assessment Channel obtains a deviation score by calculating the cosine similarity or Euclidean distance between the vector representation of a student's immediate interest snapshot and the core interest vectors in their long-term interest profile.

[0051] Deviation scores can be calculated based on similarity or the inverse or complement of distance. Simultaneously, by analyzing the difference between the immediate interest snapshot vector and the long-term interest profile vector, potential interest direction vectors can be derived. Furthermore, this can also be achieved using techniques such as autoencoders or cluster analysis, by learning low-dimensional representations of long-term interests or classifying them into different interest clusters, thereby assessing the degree of deviation of immediate interests and identifying potential interest directions. The short-term interest fit score is compared with a first dynamic threshold to determine whether the student's current immediate interests significantly deviate from known short-term interest patterns. If the fit score is lower than the preset first dynamic threshold, it may indicate that the student has developed new interests that are not captured by the current short-term interest pattern.

[0052] The long-term interest deviation score is compared with a second dynamic threshold to determine whether a student's current immediate interest has significantly deviated from their long-term interest profile. If the deviation score is higher than the preset second dynamic threshold, it may mean that the student's interest has undergone substantial long-term evolution. The dynamic threshold can be adjusted according to factors such as system operation status, student activity level, or teaching cycle stage to adapt to different situations. The first judgment condition combines the two dimensions of short-term interest mismatch and long-term interest deviation, forming a more rigorous judgment mechanism. Only when a student's immediate interest neither conforms to their short-term pattern (i.e., the short-term interest fit score is lower than the first dynamic threshold) nor significantly deviates from their long-term profile (i.e., the long-term interest deviation score is higher than the second dynamic threshold) is a genuine interest evolution signal considered to exist.

[0053] Furthermore, the evolution intensity value is an indicator that quantifies the degree or intensity of changes in student interests, and its function is to measure the significance of interest drift. This value can be a numerical value output by the dynamic demand-aware network, for example, it can be a measure of the distance between a student's immediate interest snapshot and a student's long-term interest profile, or it can be a scalar output learned internally by the network, reflecting the confidence level or magnitude of interest deviation. The interest drift intensity threshold is a preset critical value used to distinguish different intensities of interest drift. Its function is to serve as a decision-making basis to classify interest evolution into different levels, thereby triggering different update strategies. This threshold can be a fixed value, set based on experience or historical data statistics; or it can be a dynamically adjusted value, for example, adaptively adjusted based on factors such as student activity level and learning stage.

[0054] The hierarchical fusion update strategy library is a collection of various model update strategies. These strategies are designed hierarchically according to the intensity of interest evolution, providing diverse update mechanisms to adapt to different degrees of interest drift and ensuring the flexibility and accuracy of model updates. This strategy library can be a software library storing different algorithm modules or parameter configurations; it can also be a decision tree or rule set, selecting the appropriate update logic based on input conditions. The aggressive interest fusion strategy is an update strategy designed for strong interest drift, aiming to quickly integrate new interest directions into students' long-term interest profiles. Its function is to rapidly respond and adjust the model when students' interests change significantly, avoiding a severe disconnect between recommended content and students' current interests.

[0055] Potential interest direction vectors can be quickly integrated into students' long-term interest profiles through methods such as direct replacement, high-weighted averaging, or orthogonal decomposition. These potential interest direction vectors, output by a dynamic demand-aware network, represent newly emerging or evolving interest directions in students. They serve as a primary source of information in aggressive strategies, indicating the specific direction of interest evolution. They can be vectors output from long-term interest deviation assessment channels, capturing trends and content of student interest deviations.

[0056] The student's long-term interest profile is a comprehensive representation of the student's long-term interest preferences and characteristics on the second-classroom management platform. It serves as a core component of the student dynamic needs profiling model, guiding personalized recommendations. This profile can be a high-dimensional vector, where each dimension represents an interest feature or topic, and its value indicates the student's preference for that feature; it can also be a complex data structure based on graph structures or knowledge graphs. The progressive interest calibration strategy is an update strategy designed to address slight interest drift. It aims to make subtle adjustments to the student's long-term interest profile. When changes in student interests are not significant, the existing interest weights are fine-tuned to maintain model stability and avoid inaccuracies caused by over-updates. This strategy can fine-tune the relevant feature weights in the student's long-term interest profile through small-scale weighted adjustments, gradient descent optimization, or attention-based weight allocation.

[0057] This application further proposes the implementation of a radical interest fusion strategy, including: obtaining the potential interest direction vector and the current core interest vector in the student's long-term interest profile; calculating the orthogonal decomposition components and overlapping components between the potential interest direction vector and the current core interest vector; and according to the formula... Update the core interest vector, where, This represents the updated core interest vector. This represents the current core interest vector. This represents the orthogonal decomposition components. The first attenuation coefficient, The first gain coefficient, and .

[0058] The potential interest direction vector is inferred from students' instant interest snapshots through a dynamic demand-aware network, representing new interest directions that students may be developing. It can be a high-dimensional vector, with dimensions corresponding to different interest topics or features; for example, it can be obtained by encoding students' instant interest snapshots using a deep learning model, such as the Transformer encoder. The current core interest vector in the student's long-term interest profile is a stable representation of the student's historical interests, reflecting the student's main interest preferences over a long period. This vector can be a vector aggregated by a weighted average of historical behavioral feature vectors, or an embedding vector learned from students' historical interaction sequences using a recurrent neural network, such as LSTM.

[0059] The calculation of the orthogonal decomposition components and overlapping components between the latent interest direction vector and the current core interest vector aims to decompose the latent interest direction vector into components perpendicular to the current core interest vector, orthogonal decomposition components, components with the same or similar directions, and overlapping components. The orthogonal decomposition components represent novel interest points in the student's interests that are unrelated to existing core interests; for example, they can be obtained by projecting the latent interest direction vector onto the orthogonal complement space of the current core interest vector. The overlapping components represent interest points in the student's interests that are related to or reinforce existing core interests; for example, they can be obtained by projecting the latent interest direction vector onto the direction of the current core interest vector.

[0060] First attenuation coefficient This is used to control the degree to which the original core interest vector is retained in the updated core interest vector. Its value is typically between 0 and 1, representing the degree of forgetting or attenuation of historical interests. First gain coefficient. The strength of the introduction of orthogonal decomposition components of the latent interest direction vector into the updated core interest vector is used to control the strength of these components; their values ​​are typically greater than 0, and in this aggressive policy, This indicates that higher weights are given to newly discovered parts that are orthogonal to old interests.

[0061] This application further proposes that the implementation of a progressive interest calibration strategy includes: obtaining the student demand evolution vector and the corresponding interest dimension weight distribution in the student's long-term interest profile; calculating the projection components of the student demand evolution vector on each interest dimension; and applying the formula... Adjust the weights of the interest dimension, where, This represents the adjusted weight of the i-th interest dimension. Indicates the weights before adjustment. To calibrate the learning rate, This represents the projected component in dimension i. This is the adjustment range calculated based on the evolution intensity value.

[0062] Obtaining the student demand evolution vector refers to extracting a vector from the output of a dynamic demand perception network that quantifies the specific direction and degree of changes in student interests over a short period. This vector is typically high-dimensional, with each dimension representing a specific aspect of the student's interest, such as a preference for a particular subject, activity type, or learning style. Obtaining the corresponding interest dimension weight distribution from the student's long-term interest profile involves extracting information on the intensity or importance of the student's preferences across different interest dimensions from a comprehensive profile formed by the accumulation of the student's historical behavior and preferences. This distribution is usually stored as a multi-dimensional vector or a weighted feature set. Calculating the projection components of the student demand evolution vector onto each interest dimension aims to quantify the degree of influence of this evolution vector on each specific interest dimension. This can be achieved through vector dot product operations, i.e., performing an inner product operation between the student demand evolution vector and the basis vectors of each interest dimension.

[0063] For example, the standard projection method in linear algebra can be used to project the evolved vector onto the basis vectors of each interest dimension defined by the long-term interest profile; alternatively, neural network layers can learn and output the mapping or contribution values ​​of the evolved vector on each interest dimension, and these output values ​​are the projection components. Adjusting the weights of the interest dimensions according to the formula is the core operation of the progressive interest calibration strategy, aiming to refine and progressively modify the weights of the interest dimensions in the student's long-term interest profile based on the evolved vector of the student's needs. The formula... This indicates the adjusted weights. Indicates the weights before adjustment. The learning rate is used to control the step size or magnitude of each adjustment for calibration. Its value can be set empirically or optimized through cross-validation. Indicates in dimension The projected components on, and It is a normalization term in the form of a Softmax function that transforms each projected component into a probability distribution, ensuring that the sum of the adjustment contributions across all dimensions is 1, and highlighting those dimensions with larger projected components, making the adjustment more inclined towards the dimensions of interest that are more relevant to the current evolution direction. The adjustment magnitude is calculated based on the evolution intensity value. It is a scalar determined by the evolution intensity value output by the dynamic demand-aware network. When the evolution intensity value is small, It is also relatively small, thus ensuring the gradual nature of the adjustment.

[0064] Furthermore, this process is initiated when the system identifies a target student as a newly registered user or a low-activity user. Newly registered users typically refer to students logging into the second-classroom management platform for the first time, and their identification can be determined by checking the user's registration timestamp or first login record. Low-activity users refer to students who, within a certain time period, log in less than a preset threshold, have an excessively short sequence of interaction behaviors, or whose most recent operation lasted longer than a preset duration. These identification criteria can be flexibly configured to adapt to different operational strategies.

[0065] After triggering the cold start initialization process, the system retrieves the target student's structured academic record information and unstructured personal profile text. The structured academic record information typically includes the student's student ID, major, year, class, completed courses, grades, and records of honors or disciplinary actions. This information is usually stored in the academic affairs system or student management system and can be retrieved through an interface. The unstructured personal profile text refers to the self-introduction, description of interests and hobbies, personal signature, or other free text information posted by the student during registration. This information reflects the student's personalized interests and potential needs.

[0066] Based on the acquired structured student records, the system uses a collaborative filtering algorithm to retrieve neighboring samples within similar student groups and generates an initial interest preference vector based on the group. Collaborative filtering algorithms can employ various methods, such as user-based collaborative filtering or item-based collaborative filtering. For example, by calculating the similarity of students across dimensions such as major, grade level, and completed courses, a student group with similar backgrounds to the target student can be identified. Then, the system analyzes the historical behavior of these neighboring samples on extracurricular activity resources, such as registration, browsing, and saving, aggregating their interests to generate a preliminary interest preference vector based on group commonalities for the target student.

[0067] Based on the acquired unstructured personal profile text, the system extracts interest keyword embedding vectors using a natural language processing (NLP) model. NLP models can employ word embedding models such as Word2Vec and GloVe, pre-trained language models such as BERT and GPT, or text classification models. These models can perform semantic analysis on the text, identify the interest keywords contained within, and convert them into high-dimensional vector representations. For example, from text expressing an interest in programming, artificial intelligence, and data science, keywords such as "programming," "artificial intelligence," and "data science" can be extracted, and corresponding embedding vectors can be generated. These vectors capture the interest direction expressed in the text.

[0068] Furthermore, the initial interest preference vector and the interest keyword embedding vector may have different numerical ranges or dimensions. To ensure they contribute information fairly during fusion, they need to be standardized. Standardization aims to eliminate these differences, making the various dimensions of the vectors comparable. For example, Z-score standardization can be used, which adjusts the data distribution by subtracting the mean and dividing by the standard deviation; or Min-Max standardization can be used, scaling the data to a fixed interval, such as [0,1]; or L2 norm normalization can be used, making the Euclidean length of the vector 1.

[0069] Multiply the first normalized vector by Multiply the second standardized vector by The two weighted vectors are then summed to obtain a comprehensive result vector. This weighted summation method allows the system to dynamically adjust the contributions of the two vector sources to the final fusion result based on the level of trust in each source. The resulting weighted summation vector is used as the initial value for the long-term interest profile of the initial student dynamic demand profile model. The long-term interest profile is a core component of the student profile, representing the student's relatively stable and deep-seated interest preferences. Using it as the initial value lays the foundation for subsequent student interest evolution and personalized recommendations, ensuring that the profile in the cold start phase is as accurate and effective as possible.

[0070] Furthermore, dynamically adjusting the first and second dynamic thresholds refers to changing the threshold parameters used to judge the evolution of interest signals in real time or periodically based on factors such as system operating status, user behavior patterns, or changes in the external environment. This adjustment makes the system more sensitive to changes in student interests, adapting to the needs of different situations. For example, the optimal threshold can be predicted based on a machine learning model, or conditional adjustments can be made according to preset rules. The historical update frequency refers to the number or frequency at which the student dynamic demand profile model has been updated over a period of time. This frequency can reflect the stability or volatility of student interests. For example, the system can record the timestamp of each model update and calculate the number of updates in the most recent week or month; or, it can calculate the time interval between two adjacent updates and statistically analyze its distribution.

[0071] The teaching cycle stage refers to the different learning stages a student experiences within the current semester, such as the beginning of the semester, mid-term, final exam week, and holidays. Students' learning pressure, focus, and behavioral patterns may differ significantly across different teaching cycle stages. For example, the system can interface with the school's academic affairs system to obtain semester stage information corresponding to the current date; or, by analyzing students' activity data at different time periods, it can infer their current teaching cycle stage. The predefined threshold mapping table is a data structure used to store the first dynamic threshold and the second dynamic threshold corresponding to different scenarios. This mapping table can be a two-dimensional array, a hash table, or a set of conditional rules. For example, the mapping table can take the range of historical update frequency and the teaching cycle stage as input and output the corresponding threshold combinations; or, it can be a decision tree model that provides threshold suggestions based on input features.

[0072] Furthermore, the initial recommendation score of all extracurricular activity resources in the candidate pool of the computing platform aims to quantify the degree of matching between each extracurricular activity resource in the platform and the student's current interests, providing a basic relevance measure for subsequent recommendation ranking.

[0073] Specifically, content-based recommendation algorithms can be used. These algorithms analyze the text descriptions, tags, and categories of activity resources and match them with the interest preference vectors in a student dynamic needs profile model, calculating cosine similarity or dot product as the initial recommendation score. Alternatively, collaborative filtering algorithms, such as user-based collaborative filtering, can be employed to identify other student groups with similar interests to the target student. Based on the historical behavioral data of these similar students, the algorithm can predict the target student's preference for activities they haven't participated in, thus obtaining an initial recommendation score.

[0074] The diversity assurance module is a functional component that, while ensuring personalized recommendations, increases the breadth and diversity of the recommendation results. This prevents the recommendation list from becoming overly concentrated on a specific type or theme of activity, thereby broadening students' horizons and encouraging them to explore new areas of interest. This module can function as a standalone post-processing unit, receiving preliminary recommendation results and activity metadata and reordering them. Alternatively, it can be integrated into the later stages of the recommendation algorithm to optimize the diversity of the candidate set before generating the final recommendation list.

[0075] The maximum boundary relevance algorithm is a widely used reordering algorithm in information retrieval and recommendation systems, aiming to balance the relevance and diversity of recommendation results. When selecting a recommendation item, it considers not only the item's relevance to the user's interests but also its similarity to previously selected items, thus avoiding the selection of items highly similar to existing choices. The type of activity resource and the organizer dimension are attributes used to measure the differences between activity resources. The type of activity resource can include, but is not limited to, academic lectures, practical activities, volunteer services, cultural and sports competitions, and innovation and entrepreneurship, reflecting the nature and content of the activity. The organizer dimension specifies the organizer of the activity, such as various colleges within the university, student associations, the university's Youth League Committee, and external partner institutions, which helps distinguish activities from different sources and styles. During diversity reordering, the system evaluates the similarity between candidate activities and previously selected activities based on these dimensions to ensure that the recommendation list has sufficient coverage in these aspects.

[0076] The top K extracurricular activity resources are selected to determine the length of the final recommendation list presented to students, where K represents the number of activities included in the list. Its purpose is to provide a reasonable number of recommendations based on the display capabilities of the user interface and the user's attention span. The value of K can be a preset fixed value, such as 10, 20, or 30, to adapt to different application scenarios. Furthermore, the value of K can also be dynamically adjusted based on factors such as student personalization settings, the context of the current session, or device type.

[0077] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for managing student needs in extracurricular activities, characterized in that: include: When a student's login activity is detected on the second classroom management platform, second classroom activity resources are pushed to the student on the platform based on a pre-built dynamic student needs profile model. The system collects real-time sequences of students' multimodal interaction behaviors on the second classroom management platform interface, wherein the multimodal interaction behavior sequences include explicit operational behavior sequences and implicit attention behavior sequences. Multi-granular spatiotemporal features are extracted from the multimodal interaction behavior sequence and the corresponding second classroom activity resources to generate real-time interest snapshots of students; When it is determined that there is an interest evolution signal between the student's real-time interest snapshot and the student's long-term interest file corresponding to the student's dynamic demand profile model, the student's real-time interest snapshot is input into the dynamic demand perception network to perform demand evolution reasoning, and the student demand evolution vector and evolution intensity value are obtained. Based on the comparison between the evolution intensity value and the preset interest drift intensity threshold, according to the comparison result, the corresponding model update strategy in the hierarchical fusion update strategy library is selectively activated to update the student dynamic demand profile model, and the updated student dynamic demand profile model is obtained. When a resource list refresh command is detected on the second classroom management platform, a personalized recommendation sequence for second classroom activity resources is generated using the updated student dynamic demand profile model. The personalized recommendation sequence of activity resources for the second classroom is sent to the display terminal of the second classroom management platform to push the activity resources.

2. The method according to claim 1, characterized in that, The step of extracting multi-granular spatiotemporal features from the multimodal interaction behavior sequence and the corresponding second classroom activity resources to generate a snapshot of students' real-time interests includes: Extract operation type, operation frequency, and timing information from the explicit operation behavior sequence to generate explicit behavior feature vectors; Extract page dwell time, scroll speed change, and mouse movement trajectory entropy from the implicit attention behavior sequence to generate implicit attention feature vector; The second classroom activity resources are analyzed in a structured manner to obtain resource attribute feature vectors that include activity type, activity level, activity duration, and organizer. The explicit behavior feature vector, the implicit attention feature vector, and the resource attribute feature vector are aligned and concatenated according to the timestamp of the behavior to generate the student's real-time interest snapshot.

3. The method according to claim 1 or 2, characterized in that, Before determining that there is an interest evolution signal between the student's immediate interest snapshot and the student's long-term interest profile corresponding to the student's dynamic needs profile model, the method further includes: The student's real-time interest snapshot is input into the short-term interest fit evaluation channel in the dynamic demand perception network, and the short-term interest fit score is output. The short-term interest fit evaluation channel is constructed based on a gated recurrent unit network. The student's real-time interest snapshot is input into the long-term interest deviation evaluation channel in the dynamic demand perception network, and the long-term interest deviation score and potential interest direction vector are output. The short-term interest fit score is compared with a first dynamic threshold, and the long-term interest deviation score is compared with a second dynamic threshold. When the first determination condition is met, it is determined that there is an interest evolution signal, wherein the first determination condition is that the short-term interest fit score is lower than the first dynamic threshold and the long-term interest deviation score is higher than the second dynamic threshold.

4. The method according to claim 3, characterized in that, The step involves comparing the evolution intensity value with a preset interest drift intensity threshold, and selectively activating the corresponding model update strategy in the hierarchical fusion update strategy library based on the comparison result, including: When the evolution intensity value is greater than or equal to the interest drift intensity threshold, the aggressive interest fusion strategy in the hierarchical fusion update strategy library is activated, wherein the aggressive interest fusion strategy is used to fuse the potential interest direction vector into the student's long-term interest profile with a higher weight. When the evolution intensity value is less than the interest drift intensity threshold, the progressive interest calibration strategy in the hierarchical fusion update strategy library is activated, wherein the progressive interest calibration strategy is used to fine-tune the relevant feature weights in the student's long-term interest profile.

5. The method according to claim 4, characterized in that, The implementation of the radical interest fusion strategy includes: Obtain the potential interest direction vector and the current core interest vector in the student's long-term interest profile; Calculate the orthogonal decomposition components and overlapping components between the potential interest direction vector and the current core interest vector; According to the formula Update core interest vectors. in, This represents the updated core interest vector. This represents the current core interest vector. This represents the orthogonal decomposition components. The first attenuation coefficient, The first gain coefficient, and .

6. The method according to claim 4, characterized in that, The implementation of the progressive interest calibration strategy includes: Obtain the student demand evolution vector and the corresponding interest dimension weight distribution in the student's long-term interest profile; Calculate the projection components of the student demand evolution vector on each interest dimension; According to the formula Adjust the weights of the interest dimension, where, This represents the adjusted weight of the i-th interest dimension. Indicates the weights before adjustment. To calibrate the learning rate, This represents the projected component on dimension i. This is the adjustment range calculated based on the evolution intensity value.

7. The method according to claim 1, characterized in that, The method further includes a cold start initialization process, which includes: When a target student is identified as a newly registered user or a low-activity user, the cold start initialization process is triggered. Obtain the structured academic records and unstructured personal profile text of the target student; Based on the structured student record information, a collaborative filtering algorithm is used to retrieve neighboring samples in similar student groups to generate an initial interest preference vector based on the group. Based on the unstructured personal profile text, embedding vectors of interest keywords are extracted using a natural language processing model; By integrating the initial interest preference vector with the interest keyword embedding vector, an initial dynamic student demand profile model for the target student is generated.

8. The method according to claim 7, characterized in that, The process of fusing the initial interest preference vector with the interest keyword embedding vector to generate the initial dynamic student needs profile model for the target student includes: The initial interest preference vector is standardized to obtain a first standardized vector; The embedding vector of the interest keywords is standardized to obtain a second standardized vector; Based on credibility weight and The first standardized vector and the second standardized vector are weighted and summed, where the confidence weights are... Dynamic calculations are performed based on the sample size and quality of the aforementioned similar student groups; The weighted summation result vector is used as the initial value of the long-term interest profile of the initial student dynamic demand profile model.

9. The method according to claim 3, characterized in that, The first dynamic threshold and the second dynamic threshold are dynamically adjusted in the following ways: Obtain the historical update frequency of the student dynamic needs profile model; Obtain the current teaching cycle stage of the student in the current semester; Based on the historical update frequency and the teaching cycle stage, the first dynamic threshold and the second dynamic threshold are found and set through a predefined threshold mapping table.

10. The method according to claim 1, characterized in that, The generation of personalized recommendation sequences for extracurricular activity resources includes: Based on the updated student dynamic needs profile model, calculate the initial recommendation score for all extracurricular activity resources in the platform's candidate pool; The maximum boundary correlation algorithm is used to perform diversity rearrangement of event resources by type and organizer dimension; Based on the rearranged order, select the top K second-classroom activity resources to form a personalized recommendation sequence for second-classroom activity resources.

Images