Online Training Session Interaction Method and System Based on User Behavior Mining
By collecting and analyzing users' online training behavior data, a dynamic learning state model is constructed, and an adaptive conversational interaction strategy is generated. This solves the problem of rigid and inflexible interaction modes in existing technologies, realizes personalized and precise online training conversational interaction, and improves training effectiveness.
Patent Information
- Application Number
- CN202611391744.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing online training session interaction technologies lack the ability to deeply mine and accurately analyze user learning behavior data throughout the entire process. The interaction mode is rigid and inflexible, and cannot match the personalized learning status and dynamic learning needs of different users, resulting in low accuracy and insufficient effectiveness of training session interaction.
By collecting user learning behavior data throughout the entire online training process, multi-dimensional fusion analysis and feature extraction are performed to construct a dynamic learning state model for users, generate adaptive conversational interaction strategies, track and evaluate the interaction effects, and iteratively optimize the interaction strategies.
It enables dynamic adjustment of interaction strategies based on users' real-time learning status, accurately identifies user needs, improves the accuracy and effectiveness of training session interaction, forms a closed-loop interaction process, and adapts to the personalized learning needs of different users.
Smart Images

Figure CN123115341A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of training session interaction technology, and in particular to an online training session interaction method and system based on user behavior mining. Background Technology
[0002] With the popularization of digital training models, online training, with its advantages of being unrestricted by time and space, having wide resource coverage, and being able to conduct teaching in batches, has become the mainstream method for corporate employee training, vocational skills teaching, and social lifelong learning. Online training conversational interaction is a core element in ensuring training effectiveness, improving user learning participation, and achieving two-way feedback in teaching. A high-quality conversational interaction mechanism can effectively break down information barriers in online teaching and make up for the shortcomings of traditional offline training scenarios. Currently, the conversational interaction modes of existing online training platforms are mostly concentrated on basic interactive forms such as fixed question-and-answer templates, human Q&A, timed pop-up interactions, and standardized course pushes. These can achieve basic teacher-student and human-computer conversational communication, basically meeting the basic interactive needs of general online training.
[0003] However, in practical applications, existing online training session interaction technologies still have many substantial shortcomings. The core problem lies in the lack of in-depth mining and precise analysis capabilities of user learning behavior data throughout the entire process. The interaction modes are rigid and have poor adaptability, failing to match the personalized learning status and dynamic learning needs of different users. This results in low accuracy and insufficient effectiveness of training session interaction, severely restricting the teaching quality and user learning experience of online training. Specific shortcomings are as follows: First, existing technological interaction models are rigid and inflexible, primarily relying on standardized passive interaction and lacking dynamic adaptability. The conversational interaction logic of current mainstream online training platforms all follows preset, fixed rules. The interaction format, conversation content, and triggering timing are all pre-set by the system backend, resulting in uniform passive interaction modes such as course reminders, exercise pushes, and fixed Q&A scripts. Regardless of the user's learning stage (preview, study, review, practice questions, etc.) or their level of mastery of the knowledge points, the platform outputs homogenized conversational interaction content, failing to dynamically adjust interaction strategies based on the user's real-time learning status. This "one-size-fits-all" interaction method completely ignores individual learning differences among users. It cannot provide targeted Q&A guidance for users with weak foundations or lagging learning progress, while for users with good mastery and advanced progress, it easily leads to redundant interactions and information interference, resulting in a severe disconnect between conversational interaction and user learning needs, and extremely poor interaction effectiveness.
[0004] Secondly, existing technologies lack the ability to mine user behavior across all dimensions, resulting in a lack of precise data support for interactive decisions. Current online training interaction systems can only collect simple user operation data, such as course clicks, learning duration, and exercise accuracy—superficial structured data. They cannot capture and mine the refined behavioral characteristics of users during the learning process, including deeper behavioral data such as page dwell time, repeated review points, hesitation time when answering questions, conversation question preferences, repeated errors in knowledge points, and frequency of learning interruptions. Furthermore, they lack mechanisms for the fusion analysis, feature extraction, and status assessment of multi-dimensional user behavior data, making it impossible to accurately determine users' knowledge weaknesses, learning difficulties, learning habits, and cognitive shortcomings through behavioral data. This leads to the system's inability to accurately identify users' true interactive needs, making it difficult to proactively initiate targeted Q&A, knowledge point reinforcement, and explanation of key and difficult points. It remains in a passive service state of "responding only when users ask questions, and no interaction without questions," failing to achieve proactive and predictive intelligent conversational interaction.
[0005] Third, existing technologies lack a closed-loop interactive process, hindering iterative optimization of interaction effectiveness. Current online training interactive systems only complete a single interaction flow of "problem response - simple reply," without establishing mechanisms for retrospective analysis of interactive behavior data, effectiveness evaluation, and strategy iteration. After a single interaction, the system does not track changes in user learning behavior, improvement in knowledge mastery, or the emergence of subsequent learning questions, making it impossible to determine whether the interaction truly resolved the user's learning problems. Furthermore, it cannot continuously optimize dialogue scripts, timing, and content based on historical interaction data, leading to rigid interaction strategies that fail to adapt to dynamically changing user learning states and hinder the formation of a complete interactive closed loop encompassing "behavior collection - needs assessment - intelligent interaction - effectiveness evaluation - strategy optimization."
[0006] Fourth, existing technologies suffer from poor personalized interaction adaptability, resulting in unsatisfactory user learning experience and training effectiveness. Current conversational interactions are not linked to individual user profiles and learning characteristics, failing to provide differentiated interaction tailored to users with varying backgrounds, job roles, and learning abilities. In scenarios such as corporate-oriented training and specialized vocational skills training, different users have significantly different job requirements, knowledge reserves, and learning goals. Standardized conversational interaction content cannot match users' personalized training needs, easily leading to problems such as overly simplistic Q&A content, inaccurate explanations of key points and difficulties, and content that is out of touch with job training requirements. Furthermore, ineffective, redundant, and homogenized interactive content can easily distract users, reduce their learning enthusiasm, and ultimately render online training interactions merely a formality, failing to effectively help users identify and fill knowledge gaps and improve learning outcomes. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an online training session interaction method and system based on user behavior mining, which effectively solves the deficiencies of the prior art.
[0008] The objective of this invention is achieved through the following technical solution: an online training session interaction method based on user behavior mining, comprising the following steps: S1: Collect user learning behavior data throughout the entire online training process. The learning behavior data includes at least course click data, learning duration data, exercise answer data, page dwell trajectory data, video playback node data, answer hesitation duration data, conversation question record data, knowledge point repeated error data, and learning interruption frequency data. S2: Perform multi-dimensional fusion analysis and feature extraction on the collected learning behavior data to mine refined behavioral characteristics of users. The refined behavioral characteristics include at least the characteristics of knowledge point mastery, learning habits, cognitive shortcomings, and learning difficulties. S3: Based on the extracted refined behavioral features, construct a user dynamic learning state model to judge the user's current learning state in real time. The current learning state includes at least the user's learning stage, distribution of knowledge weaknesses, current level of focus, and learning progress. S4: Generate an adaptive conversational interaction strategy based on the user dynamic learning state model and the current learning state. The adaptive conversational interaction strategy includes at least the interaction triggering timing decision, the interaction content selection decision, and the interaction speech style decision. S5: Based on the adaptive conversational interaction strategy, proactively generate targeted conversational interaction content and initiate intelligent conversational interaction with the user; S6: Track and collect data on changes in the user's subsequent learning behavior after the user completes the conversational interaction. The data on changes in subsequent learning behavior includes at least the improvement in knowledge mastery, the emergence of subsequent questions, and changes in learning behavior patterns. S7: Based on the subsequent learning behavior change data, evaluate the effectiveness of this conversation interaction and generate an interaction effect evaluation result; S8: Based on the interaction effect evaluation results and accumulated historical interaction data, iteratively optimize the adaptive conversation interaction strategy, and update the conversation script library, interaction trigger rules and interaction content push logic.
[0009] Preferably, in step S2, the multi-dimensional fusion analysis includes at least: Perform time-series analysis on page dwell time data to identify the distribution of user dwell time on each knowledge point page and abnormal dwell nodes; Cluster analysis was performed on video playback node data to extract knowledge points that users repeatedly watched and the frequency of playback. Statistical analysis of the time spent hesitating to answer questions is conducted, and combined with the accuracy rate of answering exercises, the critical state of users' mastery of knowledge points is determined. Semantic analysis is performed on the conversation question record data to extract the high-frequency knowledge point domains and question depth levels of users' questions; Time series analysis was performed on the frequency of learning interruptions to identify patterns in users' attention decay and periods when learning was easily interrupted.
[0010] Preferably, in step S3, the user dynamic learning state model adopts a multimodal state representation method, which includes at least a joint representation of the knowledge mastery vector, the learning behavior pattern vector, and the cognitive load vector. The real-time analysis includes: The user's mastery level in each knowledge point is determined based on the knowledge mastery vector. Identify the user's current learning stage based on learning behavior pattern vectors; The user's current cognitive load level is determined based on the cognitive load vector, and the timing of session interaction and the complexity of interaction content are adjusted according to the cognitive load level.
[0011] Preferably, in any of the above schemes, the interaction triggering timing decision in step S4 includes: When it is detected that the time a user spends on any knowledge point page exceeds a preset threshold, a knowledge point reinforcement session interaction is triggered. When it is detected that the user hesitates for more than a preset threshold when answering any exercise question, a knowledge point Q&A session is triggered. When it is detected that the number of times a user has watched the video content of the same knowledge point exceeds a preset threshold, a key point interpretation session interaction is triggered. When a user logs back in after their learning has been interrupted, a learning progress review and reconnection session is triggered. The interactive content selection decision is based on the distribution of the user's knowledge gaps and learning stage, matching the corresponding interactive content template from the conversation script library, and filling the interactive content template with personalized parameters.
[0012] Preferably, in step S5, the proactive generation of targeted conversational interaction content includes: Based on the distribution of users' knowledge gaps, generate Q&A conversation content that includes knowledge point analysis, explanation of typical examples, and related knowledge links; Based on the user's learning progress status, generate guided conversation content that includes a summary of the current learning stage, a preview of the key points for the next stage of learning, and learning suggestions; Based on the user's cognitive gap characteristics, generate consolidation conversation content that includes personalized practice push and targeted reinforcement training suggestions; The intelligent conversational interaction adopts a multi-turn dialogue format, dynamically adjusting subsequent conversation branches based on the user's real-time responses.
[0013] Preferably, in any of the above schemes, the subsequent learning behavior change data in step S6 shall at least include: The trend of user accuracy rate in knowledge point exercises within a preset time window after conversation interaction; Changes in user viewing behavior regarding relevant knowledge point video content after conversational interaction, including whether to reduce replays or speed up viewing; Changes in the frequency and depth of user-asked questions after conversational interaction; Changes in the duration of continuous learning and the frequency of learning interruptions after user interaction.
[0014] Preferably, in any of the above schemes, the interaction effect evaluation in step S7 includes: The subsequent learning behavior change data is compared and analyzed with the historical behavior data before the conversation interaction, and behavior improvement indicators of each dimension are calculated. Based on the aforementioned behavior improvement index, it is determined whether the interaction in this session effectively solved the user's learning problem. If the current session interaction is determined to be invalid or its effect is below a preset threshold, the current session interaction is marked as an interaction event to be optimized, and the interaction context information is recorded for subsequent strategy optimization.
[0015] Preferably, in any of the above schemes, the iterative optimization in step S8 includes: Based on the accumulated interaction effect evaluation results, reinforcement learning method is used to update the effect score of each dialogue template in the conversation dialogue database; Based on the tagged interactive events to be optimized, the conditional parameters of the interaction triggering rules are adaptively adjusted; Based on the differences in user responses to different interactive content and styles in historical interaction data, update the individual user interaction preference profile to achieve user-level personalized adaptation of interaction strategies.
[0016] An online training session interaction system based on user behavior mining, used to implement the above methods, includes: The data acquisition module is used to collect user learning behavior data throughout the entire online training process. The learning behavior data includes at least course click data, learning duration data, exercise answer data, page dwell trajectory data, video playback node data, answer hesitation duration data, conversation question record data, knowledge point repeated error data, and learning interruption frequency data. The behavioral feature extraction module is used to perform multi-dimensional fusion analysis and feature extraction on the learning behavior data collected by the data acquisition module, and to mine the refined behavioral features of users. The refined behavioral features include at least the features of knowledge point mastery, learning habits, cognitive shortcomings, and learning difficulties. The user status assessment module is used to construct a dynamic learning status model of the user based on the refined behavioral features extracted by the behavioral feature extraction module, and to assess the user's current learning status in real time. The current learning status includes at least the user's learning stage, the distribution of knowledge weaknesses, the current level of focus, and the learning progress status. An interaction strategy generation module is used to generate an adaptive conversational interaction strategy based on the user dynamic learning state model and current learning state output by the user state judgment module. The adaptive conversational interaction strategy includes at least interaction trigger timing decision, interaction content selection decision, and interaction speech style decision. The conversation content generation module is used to proactively generate targeted conversation interaction content based on the adaptive conversation interaction strategy output by the interaction strategy generation module. The interaction execution module is used to initiate intelligent conversational interactions with users, receive real-time responses from users, and drive multi-round dialogues. The behavior tracking module is used to track and collect data on changes in the user's subsequent learning behavior after the user completes the conversational interaction. The data on changes in subsequent learning behavior includes at least the improvement in knowledge point mastery, the emergence of subsequent questions, and changes in learning behavior patterns. The effectiveness evaluation module is used to evaluate the effectiveness of this session interaction based on the subsequent learning behavior change data collected by the behavior tracking module, and generate an interaction effectiveness evaluation result. The strategy optimization module is used to iteratively optimize the adaptive conversational interaction strategy based on the interaction effect evaluation results output by the effect evaluation module and the accumulated historical interaction data, and to update the conversational script library, interaction triggering rules and interaction content push logic.
[0017] Preferably, as described in any of the above schemes, the interaction strategy generation module includes: The trigger timing decision submodule is used to dynamically determine the trigger timing of the conversation interaction based on the user's current learning status. The trigger timing includes at least the following: triggering when the user stays on the knowledge point page for too long, triggering when the user hesitates to answer the exercise for too long, triggering when the user repeatedly watches the video content, and triggering when the user logs in again after the learning is interrupted. The content selection submodule is used to match corresponding interactive content templates from the conversation script library based on the distribution of the user's knowledge gaps and learning stage, and to fill in the interactive content templates with personalized parameters. The conversation style submodule is used to dynamically adjust the conversation style parameters based on the user's cognitive load level and historical interaction preferences. The system also includes a conversation script library update module, which updates the script templates and their weight parameters in the conversation script library in real time based on the iterative optimization results of the strategy optimization module.
[0018] The present invention has the following advantages: 1. This online training session interaction method and system based on user behavior mining collects full-process learning behavior data in step S1, including page dwell trajectory data, video replay node data, answer hesitation time data, session question record data, repeated knowledge point error data, and learning interruption frequency data. Step S2 performs multi-dimensional fusion analysis on the above multi-source heterogeneous data, conducts time-series analysis on page dwell trajectory data to identify abnormal dwell nodes, performs cluster analysis on video replay node data to extract repeatedly replayed knowledge point segments and replay frequency, and performs statistical analysis on answer hesitation time data to determine the user's critical state of knowledge point mastery. Semantic analysis of conversation question records is performed to extract high-frequency knowledge points and question depth levels. Time series analysis of learning interruption frequency data is performed to identify attention decay patterns. This invention can comprehensively capture the deep cognitive behavioral characteristics of users during the learning process. Compared with existing technologies that only collect surface-level structured data such as course clicks, learning time, and exercise accuracy, this invention establishes a quantitative mapping relationship between refined behavioral characteristics and user cognitive states, providing accurate data support for subsequent interaction decisions. It solves the technical problem that existing technologies cannot accurately identify users' real interaction needs due to the lack of full-dimensional behavioral mining capabilities, and are always in a passive response state.
[0019] 2. This online training conversation interaction method and system based on user behavior mining constructs a dynamic learning state model for users in step S3, which includes a knowledge mastery vector, a learning behavior pattern vector, and a cognitive load vector. Based on the knowledge mastery vector, it determines the user's mastery level for each knowledge point; based on the learning behavior pattern vector, it identifies the user's current learning stage; and based on the cognitive load vector, it determines the user's current cognitive load level. Step S4 generates an adaptive conversation interaction strategy based on this dynamic learning state model, which includes decisions on interaction trigger timing, interaction content selection, and interaction style. Different types of conversation interactions are triggered when page dwell timeouts, question hesitation timeouts, repeated video replays, or re-login after learning interruptions are detected. This achieves dynamic adaptation of the interaction strategy to the user's real-time learning state, solving the technical problems of existing technologies where rigid and inflexible interaction patterns and a "one-size-fits-all" passive interaction approach fail to differentiate interaction content based on the user's learning stage and knowledge mastery level, leading to redundancy or insufficient targeting.
[0020] 3. This online training session interaction method and system based on user behavior mining tracks and collects subsequent learning behavior data within a preset time window after user session interaction, including trends in the accuracy of knowledge point exercises, changes in viewing behavior of related knowledge point video content, changes in the frequency and depth of session questions, and changes in continuous learning duration and frequency of learning interruptions. Step S7 compares and analyzes the subsequent learning behavior change data with historical behavior data before the interaction to calculate the improvement index of each dimension of behavior. This comprehensively determines whether the session interaction effectively solved the user's learning problem. Invalid or less effective interaction events are marked and the interaction context information is recorded. This establishes a quantifiable evaluation mechanism for the effect of a single session interaction, solving the technical problems of existing technologies that only complete a single "problem response - simple reply" process, cannot track changes in user learning behavior and improvement in knowledge point mastery after interaction, and cannot determine whether the interaction truly solves the learning problem.
[0021] 4. This online training conversation interaction method and system based on user behavior mining, through step S8, uses reinforcement learning to update the effect scores of each dialogue template in the conversation script library based on the accumulated interaction effect evaluation results. It adaptively adjusts the condition parameters of the interaction triggering rules based on the marked interaction events to be optimized, and updates the individual user interaction preference profile based on the differences in user responses to different interaction content and interaction styles in historical interaction data. In addition, the cascading linkage between the strategy optimization module and the conversation script library update module in the system architecture feeds back the quantitative evaluation results of historical interaction effects to each decision sub-module in the interaction strategy generation stage, driving the dynamic iterative update of conversation script template weights, triggering rule thresholds and user preference profiles. This forms a closed-loop interaction process of "behavior collection - demand analysis - intelligent interaction - effect evaluation - strategy optimization". It solves the technical problem of existing technologies that, due to the lack of interaction effect backtracking analysis and strategy iteration mechanisms, result in long-term solidification of interaction strategies and inability to adapt to the dynamic changes in users' learning states.
[0022] 5. This online training conversation interaction method and system based on user behavior mining, in step S4, matches corresponding interaction content templates from the conversation script library according to the distribution of users' knowledge weaknesses and learning stages, and fills in personalized parameters. In step S5, it generates Q&A conversation content based on the distribution of knowledge weaknesses, including knowledge point analysis, explanation of typical examples, and related knowledge links; it generates guiding conversation content based on the learning progress status, including stage summary, key point preview, and learning suggestions; and it generates consolidation conversation content based on cognitive shortcomings, including personalized practice push and reinforcement training suggestions. At the same time, it combines the user's individual interaction preference profile updated based on historical interaction data in step S8 to perform differentiated speech style adaptation for different users. Personalized interaction adaptation is achieved from two dimensions: matching the knowledge level of the interaction content and matching the expression style of the interaction script. This solves the technical problems of existing technologies that cannot perform differentiated interaction adaptation for users with different foundations and learning abilities due to the lack of association between the conversation interaction and the user's individual profile and learning characteristics, resulting in inappropriate Q&A content depth and the disconnect between the pushed content and training needs. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation
[0024] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.
[0025] Example 1: Online Training Session Interaction Method Based on User Behavior Mining like Figure 1 As shown, this embodiment provides an online training session interaction method based on user behavior mining. The method includes the following steps: Step S1: Collect user learning behavior data throughout the entire online training process.
[0026] Data collection is accomplished collaboratively through collection programs deployed on both the server and client sides of the online training platform. On the server side, an API gateway interceptor records the request URL, request time, and response time for each user's course page, generating a course click log; a scheduled task polls the user's learning progress table, recording the cumulative learning time for each course chapter; and the input parameters and return values of the exercise submission interface record exercise answer data, including exercise ID, user answer, whether it is correct, and submission timestamp. On the client side, an embedded JavaScript tracking SDK collects page dwell trajectory data: each time a user navigates to a page, the current page URL, entry timestamp, and exit timestamp are recorded, generating a page dwell sequence.
[0027]
[0028] in Simultaneously, the page URL is mapped to the knowledge point identifiers in the course knowledge graph to obtain the dwell time of each knowledge point page. The video replay node data is collected as follows: In the HTML5 video player, the `timeupdate` event (triggered every 250ms) and the `seeked` event (triggered after the user drags the progress bar) are listened to. When a `seeked` event is detected and the playback time after the jump is more than 2 seconds earlier than the time before the jump, it is determined as a replay operation. The replay start timestamp, replay end timestamp, and corresponding video segment identifier are recorded, and the replay segment is mapped to the corresponding knowledge point according to the preset video-knowledge point mapping table. The data collection method for the hesitation time is as follows: When the user enters the exercise page, the timestamp of the page DOM loading completion is recorded; when the user clicks any option for the first time, the click timestamp is recorded. The difference between the two is the hesitation time for that exercise. If the user exits the page without answering, this data is not recorded. Data on learning interruption frequency is collected by listening to page visibility change events. When the page becomes hidden, the start time of the interruption is recorded; when the page becomes visible again, the end time is recorded. If the interruption duration exceeds a preset threshold (e.g., 5 minutes), it is counted as a valid interruption event. The number of interruptions and the total interruption duration are calculated per unit time (e.g., 1 hour). Conversation question records are obtained by listening to the submission event of the chat input box, acquiring the text of each submitted question, the question time, and the current chapter ID. Regular expressions and named entities are used to extract key knowledge points from the questions. All collected data is structured and stored according to user ID, course ID, and collection time, forming the original behavioral dataset.
[0029] Step S2: Perform multi-dimensional fusion analysis and feature extraction on the collected learning behavior data to uncover the refined behavioral characteristics of users. This step is performed by the behavioral feature extraction module and includes the following sub-analysis operations: (1) Perform time-series analysis on page dwell trajectory data: Arrange the dwell time of the same user on each knowledge point page into a time-series sequence according to the learning order, and use the sliding window method (window size is 5 knowledge points) to calculate the local mean and standard deviation. Mark the nodes whose current dwell time exceeds the local mean plus twice the standard deviation as abnormal dwell nodes, and mark the knowledge points corresponding to the abnormal dwell nodes as potential points of difficulty in understanding; at the same time, sort all dwell times in descending order, and take the top 20% of knowledge point pages as the pages that users focus on.
[0030] (2) The DBSCAN clustering algorithm was used to analyze the video playback node data: the timestamp of each playback event was used as a one-dimensional data point, and the clustering radius was set. seconds, minimum number of samples The system clusters the rewatch points that are concentrated in time into clusters, with each cluster corresponding to a knowledge point segment that is repeatedly rewatched. The number of rewatch points in each cluster is counted as the rewatch frequency of that knowledge point. Clusters with a rewatch frequency exceeding the global average rewatch frequency plus one standard deviation are marked as high rewatch hotspots, and the corresponding knowledge points are identified as the user's comprehension difficulties.
[0031] (3) Statistical analysis of the time spent hesitating during answering questions: For all questions under each knowledge point, calculate the average correct answer rate of users on that knowledge point. and average time spent hesitating when answering questions Construct two-dimensional decision rules: If and ( If the median time spent hesitating over all the exercises for a user is taken as the average time spent on a particular knowledge point, then that knowledge point is considered to be in a critical state of mastery, indicating that the user is in a vague cognitive range of "understanding but not quite understanding". At this time, targeted Q&A intervention is most needed.
[0032] (4) Use BERT pre-trained language model to perform semantic encoding and text classification on the conversation question record data: Input the question text into the BERT model, extract the 768-dimensional vector at the [CLS] position as the sentence vector, and map the sentence vector to the multi-label classification result of the knowledge point domain through a fully connected classification layer. At the same time, use a cognitive verb dictionary (containing memory layer verbs such as "what" and "define", understanding layer verbs such as "why" and "explain", application layer verbs such as "how" and "how to apply", analysis layer verbs such as "analyze" and "compare", and creation layer verbs such as "design" and "create") to match the question text, determine the cognitive depth level of the question, count the frequency of questions at each level, and extract the knowledge point domain and high-frequency question level of high-frequency questions.
[0033] (5) Time series analysis of learning interruption frequency data: Using 10-minute intervals as a time bucket, the number of interruptions within each bucket is counted to form an interruption frequency time series. An ARIMA model (order of...) is then used. The sequence is fitted, trend and periodic components are extracted, and time intervals with significantly higher interruption frequency than the upper bound of the confidence interval of the fitted value are identified as high-incidence periods of user attention decay.
[0034] Through the above multi-dimensional integrated analysis, refined behavioral characteristics are output, including: the mastery score of each knowledge point (combined with a comprehensive weighted average of accuracy, dwell time, revisit frequency, and question frequency), user learning habit tags (such as "morning type", "night type", and "scattered type" based on the distribution of learning time), a list of cognitive weaknesses (knowledge points with a mastery score below 0.4), and a list of learning difficulties (highly revisited hot topics and critical knowledge points).
[0035] Step S3: Based on the extracted refined behavioral features, construct a dynamic learning state model for users and assess their current learning status in real time.
[0036] A multimodal state representation method is used to construct the user's current learning state as a joint form of three-dimensional vectors: First, construct the knowledge mastery vector:
[0037] in This represents the total number of knowledge points in the course knowledge graph. For the first The assessment of mastery of each knowledge point. The value range is [0,1]. The calculation formula is:
[0038] in This represents the average accuracy rate of the exercises for this knowledge point. The normalized quantile of the time spent on the page for this knowledge point (the longer the stay, the more attention it receives). This is a challenging indicator value based on the normalization of replay frequency (the more replays, the higher the difficulty). (smaller) This is the normalized value for the frequency of questions. , , , The preset weights are set to 0.40, 0.25, 0.20, and 0.15 respectively. According to... The value will be divided into the following levels: <0.4 indicates a "weak" value, 0.4≤ <0.7 is considered "normal". A score of ≥0.7 indicates "proficient".
[0039] Secondly, construct learning behavior pattern vectors:
[0040] The probability values correspond to four categories: "Preview Stage," "New Lesson Learning Stage," "Review Stage," and "Practice and Consolidation Stage." User learning behavior features from the past 30 minutes are extracted: new page visit frequency, number of completed exercises, video viewing completion rate, and graph structure depth of page jumps. These features are input into a pre-trained softmax classifier, which outputs the probability distribution of the four stages. The stage with the highest probability is taken as the current learning stage.
[0041] Next, construct a cognitive load vector, using the current cognitive load level value. express. The calculation uses the following formula:
[0042] in This is the normalized value of the average hesitation time for the most recent 10 practice questions. The coefficient of variation of page dwell time in the last 20 minutes (the larger the coefficient of variation, the less stable the focus). This is the normalized value of the interruption frequency over the past hour. This is the offset of the ratio between the current video playback speed and the default speed (this will increase if the playback speed is too slow or if the speed is frequently adjusted). ), These are weighting coefficients, with values of 0.3, 0.3, 0.25, and 0.15 respectively. The time was determined to be high cognitive load. The time was judged as low cognitive load.
[0043] Real-time analysis is triggered every 5 minutes or when the user performs a new key action (such as submitting exercises, dragging the video progress bar, or asking a question), outputting the current learning stage, the distribution of weaknesses across all knowledge points, and the current level of focus (as determined by...). The values are mapped to three levels: "high focus", "normal" and "low focus" and the learning progress status (number of chapters completed / total number of chapters).
[0044] Step S4: Generate an adaptive conversational interaction strategy based on the user's dynamic learning state model and current learning state. The strategy includes at least interaction trigger timing decision, interaction content selection decision, and interaction speech style decision.
[0045] (1) Trigger timing decision: Set four types of trigger conditions, which correspond to four types of interaction scenarios.
[0046] Knowledge point page dwell timeout trigger: When the system detects that a user's current dwell time on any knowledge point page exceeds 2.5 times the historical average dwell time for that knowledge point, and the user's mastery score for that knowledge point is below 0.5, a knowledge point reinforcement session interaction is triggered. The threshold is stored in the system configuration table and can be dynamically adjusted by the strategy optimization module.
[0047] Exercise answering hesitation timeout trigger: When a user's hesitation time for answering a certain exercise exceeds the 75th percentile of the user's historical hesitation time, and the mastery score of the corresponding knowledge point for that exercise is lower than 0.6, the corresponding knowledge point Q&A session interaction is triggered.
[0048] Video replay trigger: When the system counts that a user has replayed the same knowledge point video content more than 3 times (within the last 24 hours), a key point interpretation session will be triggered.
[0049] Learning interruption triggers upon re-login: When the system detects that the user's current login time is more than 30 minutes from the last logout time, and the user was in the process of learning an unfinished chapter when the user last logged out, the learning progress review and connection session interaction will be automatically triggered within 2 minutes of logging in.
[0050] (2) Interactive Content Selection Decision: Based on the distribution of users' knowledge weaknesses and learning stages, corresponding interactive content templates are matched from the conversation script database. The conversation script database adopts a three-level index structure: the first level is divided by learning stage (preview / new lesson / review / practice), the second level is indexed by knowledge point ID, and the third level is indexed by interaction type (Q&A, consolidation, guidance, reminder). After matching, the current user's personalized parameters (username, mastery score, list of weaknesses, progress percentage, etc.) are filled into the placeholders in the template to generate the main framework of personalized interactive content.
[0051] (3) Interaction style decision: based on the user's current cognitive load level The L value and the user's historical interaction preference profile (stored in the user profile database) are used to dynamically adjust the communication style parameters. If L>0.7, a concise sentence length (no more than 30 characters per message), formality, and reduced encouragement frequency are selected; if L<0.3 and the user preference profile is marked as "prefers encouragement", a detailed sentence length, conversational expression, and high encouragement frequency are selected.
[0052] Step S5: Based on the adaptive conversational interaction strategy, proactively generate targeted conversational interaction content and initiate intelligent conversational interaction with the user.
[0053] After receiving the interaction strategy, the conversation content generation module extracts the corresponding content from the knowledge resource base and exercise base and assembles it into output conversation text based on the interaction type, content template identifier and speech style parameters in the strategy.
[0054] If the interaction type is Q&A, then extract the ID of the weak knowledge point specified in the strategy, obtain the structured explanation text, typical examples (difficulty level matched with the user's mastery level; if the mastery level < 0.4, select basic questions; if 0.4 ≤ mastery level < 0.7, select medium-level questions) and their step-by-step solution process from the knowledge point parsing library, and attach the links to the preceding and succeeding knowledge points of the knowledge point, fill them into the Q&A template to generate complete content.
[0055] If the interaction type is guided, the percentage of completed chapters and the estimated learning time for the remaining chapters are calculated based on the user's current learning stage and progress status. The summary statement and learning suggestion template corresponding to the current stage are extracted from the guidance strategy library, and the guided session content is generated after filling in the personalized data.
[0056] If the interaction type is reinforcement, then based on the user's cognitive weaknesses (list of weak knowledge points), select practice questions (up to 5 questions) related to the weak knowledge points from the question bank, with a difficulty coefficient within ±0.15 of the user's current mastery level. At the same time, extract review suggestions from the reinforcement training strategy library and combine them to generate reinforcement conversation content.
[0057] The generated content undergoes smoothness optimization using a Transformer-based text polishing model before being pushed to the user's client interface via a WebSocket persistent connection. The interaction employs a multi-turn dialogue format, providing preset quick response options (such as "Understood," "Still have questions?", "Explain again"), and dynamically adjusts subsequent conversation branches based on user responses.
[0058] Step S6: Track and collect data on changes in the user's learning behavior after completing the conversational interaction.
[0059] The interaction execution module starts a behavior tracking timer immediately after the session ends. The preset tracking time window is 24 hours after the interaction is completed, or until the user completes the learning of the next 5 knowledge points (whichever comes first). Within this time window, the behavior tracking module continuously collects the following data: The trend of user accuracy in answering questions related to the knowledge points involved in the interaction: The cumulative accuracy of the knowledge point is calculated every 2 hours to form a time series.
[0060] Changes in user viewing behavior for relevant knowledge-point videos: Monitor whether users rewatch the corresponding video segments, and record the number of rewatches and the total duration.
[0061] Changes in the frequency and depth of user questions after interaction: Statistics on the number of questions and the distribution of cognitive levels within the window.
[0062] Changes in users' continuous learning duration and interruption frequency: Average continuous duration and interruption frequency are calculated per hour.
[0063] All tracking data is stored in association with user ID and session interaction ID.
[0064] Step S7: Based on subsequent learning behavior change data, evaluate the effectiveness of this conversation interaction and generate interaction effect evaluation results.
[0065] The effectiveness evaluation module reads behavior tracking records and retrieves baseline behavior data for the same user on the same knowledge point from the historical behavior database, specifically from the 24 hours prior to the interaction. Improvement metrics are calculated for each dimension. Improvement in mastery of knowledge points:
[0066] in To track the average accuracy rate of this knowledge point within the window, This represents the average accuracy rate over the 24 hours prior to the interaction.
[0067] Improvement in video viewing behavior:
[0068] in , These represent the number of times the video for that knowledge point was watched before and after the interaction. To prevent division by zero for extremely small constants, This indicates positive improvement.
[0069] Improvement in Question Depth: The difference between the average cognitive level score of questions asked after interaction (memory level=1, comprehension level=2, application level=3, analysis level=4, creation level=5) and the average score before interaction is denoted as . .
[0070] Improvement in focus:
[0071] in This represents the average continuous learning time (in minutes). This indicates an increase in focus.
[0072] After normalizing the improvement of each dimension, a weighted sum is taken to obtain the comprehensive score:
[0073] in This is the normalized improvement value (mapped to [−1,1]). If Score > 0.3, it is considered a valid interaction; if 0.1 ≤ Score ≤ 0.3, it is considered a moderate effect; if Score < 0.1, it is considered an invalid interaction. Interactions with moderate or invalid effects are marked as "interaction events to be optimized", and the relevant context is packaged and stored in the event library to be optimized.
[0074] Step S8: Based on the interaction effect evaluation results and accumulated historical interaction data, iteratively optimize the adaptive conversation interaction strategy, and update the conversation script library, interaction trigger rules, and interaction content push logic.
[0075] The strategy optimization module performs batch optimization at fixed intervals (e.g., 2 AM daily), based on accumulated historical interaction evaluation records (at least 50 records) and performs the following three iterative operations: (1) Update of the effectiveness score of the speech template: The Q-learning reinforcement learning method is adopted, and the user's state (knowledge mastery vector, learning stage, cognitive load level) is used as the state space. The choice of a script template as an action The effectiveness of the interaction will be comprehensively scored. As an instant reward For each historical interaction record, update according to the Q-value formula:
[0076] in For learning rate, As a discount factor, This reflects the user's state after interaction. The expected return scores for each script template have been adjusted after the update.
[0077] (2) Adaptive adjustment of trigger rule condition parameters: Extract all events marked as "to be optimized" from the event library to be optimized, and count the false trigger rate and missed trigger rate under each trigger condition. Use gradient descent to fine-tune the threshold parameter of each trigger condition, with each adjustment step being 5% of the current value, and verify the adjusted precision and recall on the validation set (the most recent 20% of historical records). If both improve, retain the adjustment; otherwise, revert.
[0078] (3) Update the user's individual interaction preference profile: Based on the feedback performance of the same user under different interaction content and interaction style in historical interaction data (including whether to click on links, whether to favorite, whether to like, etc.), update the user's preference profile and store it as key-value pairs (such as preference interaction frequency, preference speech length, preference formality, etc.). The update rule is to count the proportion of positive feedback from users to various styles of interaction and use an exponentially weighted moving average for smoothing.
[0079] After optimization, the updated dialogue template Q-value table is sent to the conversation dialogue library update module, the adjusted trigger threshold is synchronized to the interaction strategy generation module, and the updated preference profile is written to the user profile database.
[0080] Example 2: Online Training Session Interaction System Based on User Behavior Mining like Figure 2 As shown, this embodiment provides an online training session interaction system based on user behavior mining, used to implement the method described in Embodiment 1 above. The system includes a data acquisition module, a behavior feature extraction module, a user state analysis module, an interaction strategy generation module, a session content generation module, an interaction execution module, a behavior tracking module, an effect evaluation module, a strategy optimization module, and a session script database update module.
[0081] Data Acquisition Module: Deployed on both the server and client sides, the server sub-module collects course clicks, learning duration, and exercise answer data through the API gateway interceptor; the client sub-module collects page dwell time, video playback nodes, hesitation time for answering questions, learning interruption frequency, and conversation question records through the JavaScript SDK. The collected data is validated and deduplicated before being stored in HBase.
[0082] The behavioral feature extraction module employs a combination of offline batch processing (Spark) and real-time stream processing (Flink). The offline portion performs deep computations such as DBSCAN clustering and ARIMA time-series analysis; the real-time portion calculates lightweight features within the recent sliding window. Feature results are stored in Redis (real-time) and Elasticsearch (historical), and outputs include knowledge mastery scores, learning habit tags, and a list of cognitive weaknesses.
[0083] The user status assessment module includes a model building submodule and a real-time assessment submodule. The model building submodule calculates the knowledge mastery vector according to the formula in step S3 of Example 1. Behavioral pattern vectors And cognitive load level L. The real-time analysis submodule updates the status every 5 minutes or when it receives a new feature push, and stores it in Redis after serialization in Protobuf format (key user:state:{userId}, expiration time 2 hours).
[0084] Interaction strategy generation module: contains three sub-modules: Trigger Timing Decision Submodule: Based on the Drools rule engine, it polls user status and generates a trigger event when any trigger condition is met. Threshold parameters are stored in the Apollo configuration center and support dynamic updates.
[0085] Content selection submodule: Based on the trigger type, it maps the interaction type, matches templates from the conversation script library by three-level index, selects the optimal template by priority weight (updated by the strategy optimization module), and populates the user's personalized parameters.
[0086] The conversation style submodule is based on cognitive load levels. Based on user preference profiles, determine the three-part structure of the communication style (length, formality, and frequency of encouragement) and attach it to the content framework.
[0087] The conversation content generation module receives a content framework and style parameters, extracts specific content from the Neo4j knowledge graph and the MySQL exercise library, renders the final conversation text using the FreeMarker template engine, truncates or expands the content according to the style parameters, and outputs it to the interaction execution module.
[0088] Interactive Execution Module: Maintains the WebSocket connection pool and pushes session content to the client. Creates a dialogue context object and sets a timeout timer (60 seconds). Determines whether to proceed to the next round of dialogue based on user responses, records explicit feedback behaviors (clicks, likes, favorites, etc.), and writes them to the behavior tracking database.
[0089] Behavior Tracking Module: Creates a tracking task after the session ends, with the tracking window lasting 24 hours or until the user completes 5 knowledge points. Continuously monitors new behavioral data, aggregating and statistically analyzing it by knowledge point ID (average accuracy every 2 hours, cumulative review times, question frequency, etc.). After the window ends, the aggregated data, along with the baseline data from the 24 hours prior to the interaction, is sent to the performance evaluation module.
[0090] The performance evaluation module uses Python microservices and NumPy / Pandas to calculate the improvement in each dimension and the overall score. Score. The validity level is determined based on the threshold, and an evaluation record is generated and stored in MySQL. For events requiring optimization, the complete context is serialized into JSON and stored in the MongoDB table for events requiring optimization.
[0091] The strategy optimization module executes daily at midnight, retrieving evaluation records from MySQL for the past 7 days and historical state snapshots from Redis. It performs three optimizations: ①Q-learning updates the Q table of the script template (stored in RedisHash); ② Search the grid for the optimal trigger threshold and update it to the configuration center; ③ Update user preference profiles using exponentially weighted moving average (RedisHash).
[0092] Generate an optimization report for administrators to view.
[0093] The conversation script library update module receives the Q-value update result from the strategy optimization module, recalculates the priority weight (priority_score) of each script template (0,1), and increments the version number. It also provides a management backend for manual addition, deletion, and modification of templates. After the update is complete, a notification content selection submodule refreshes the local cache.
[0094] System workflow: 1. The data acquisition module continuously collects behavioral data; 2. Behavioral feature extraction module analyzes and extracts features; 3. The user status assessment module constructs the status in real time; 4. The interaction strategy generation module automatically generates strategies; The 5-session content generation and interaction execution module completes the session; The 6-behavior tracking and effectiveness evaluation module assesses the effectiveness; The strategy optimization and script library update modules implement closed-loop iteration. The modules are decoupled through Kafka message queues and RESTful APIs, supporting horizontal scaling and service registration and discovery (Consul).
[0095] In summary, this invention collects full-process learning behavior data, including page dwell trajectory data, video playback node data, question hesitation duration data, conversation question record data, repeated knowledge point error data, and learning interruption frequency data, through step S1. Step S2 performs multi-dimensional fusion analysis on the aforementioned multi-source heterogeneous data, conducts time-series analysis on the page dwell trajectory data to identify abnormal dwell nodes, performs cluster analysis on the video playback node data to extract repeatedly viewed knowledge point segments and playback frequency, performs statistical analysis on the question hesitation duration data to determine the user's critical state of knowledge point mastery, performs semantic analysis on the conversation question record data to extract high-frequency knowledge point domains and question depth levels, and analyzes the learning interruption frequency data. By conducting time-series analysis to identify attention decay patterns, this invention can comprehensively capture the deep cognitive behavioral characteristics of users during the learning process. Compared to existing technologies that only collect surface-level structured data such as course clicks, learning duration, and exercise accuracy, this invention establishes a quantitative mapping relationship between refined behavioral characteristics and user cognitive states, providing accurate data support for subsequent interaction decisions. It solves the technical problem of existing technologies failing to accurately identify users' true interaction needs due to the lack of full-dimensional behavioral mining capabilities, resulting in a passive response state. Step S3 constructs a dynamic user learning state model jointly represented by a knowledge mastery vector, a learning behavior pattern vector, and a cognitive load vector. Based on the knowledge mastery vector, the invention determines the user's cognitive state. Based on the mastery level of each knowledge point, the system identifies the user's current learning stage based on learning behavior pattern vectors and determines the user's current cognitive load level based on cognitive load vectors. Step S4 generates an adaptive conversational interaction strategy based on this dynamic learning state model, which includes decisions on interaction trigger timing, interaction content selection, and interaction style. Different types of conversational interactions are triggered when page dwell timeout, hesitation timeout, repeated video playback, or re-login after learning interruption are detected. This achieves dynamic adaptation of the interaction strategy to the user's real-time learning state, solving the problem that existing technologies, due to rigid and inflexible interaction modes and a "one-size-fits-all" passive interaction approach, cannot differentiate interaction output according to the user's learning stage and knowledge mastery level. Technical issues such as content redundancy or insufficient targeting in interactions are addressed through step S6, which tracks and collects subsequent learning behavior data within a preset time window after user session interaction. This data includes trends in the accuracy of knowledge point exercises, changes in viewing behavior of related knowledge point video content, changes in the frequency and depth of questions asked, and changes in continuous learning duration and frequency of learning interruptions. Step S7 compares and analyzes this subsequent learning behavior data with historical behavior data before the interaction to calculate behavior improvement indicators across various dimensions. This comprehensive assessment determines whether the session interaction effectively solved the user's learning problem. Ineffective or below-preset threshold interaction events are marked and their context information is recorded, establishing a quantifiable evaluation mechanism for the effectiveness of a single session interaction.This solution addresses the technical problems of existing technologies that only complete a single "problem response - simple reply" process, cannot track changes in user learning behavior and improvement in knowledge mastery after interaction, and cannot determine whether the interaction truly solves the learning problem. Step S8 uses reinforcement learning to update the effectiveness scores of each dialogue template in the conversation script library based on accumulated interaction effect evaluation results. It adaptively adjusts the conditional parameters of interaction trigger rules based on marked interaction events to be optimized, and updates individual user interaction preference profiles based on differences in user responses to different interaction content and styles in historical interaction data. Furthermore, the cascading linkage between the strategy optimization module and the conversation script library update module in the system architecture feeds back the quantitative evaluation results of historical interaction effects to each decision sub-module in the interaction strategy generation stage, driving dynamic iterative updates of conversation script template weights, trigger rule thresholds, and user preference profiles. This forms a complete interactive closed loop of "behavior collection - demand analysis - intelligent interaction - effect evaluation - strategy optimization," solving the problems caused by the lack of interaction effect retrospective analysis and strategy iteration mechanisms in existing technologies. The technical problem of long-term fixed content failing to adapt to the dynamic changes in users' learning states is addressed by: Step S4, matching corresponding interactive content templates from the conversation script library based on the distribution of users' knowledge weaknesses and learning stages, and filling in personalized parameters; Step S5, generating Q&A conversation content based on the distribution of knowledge weaknesses, including knowledge point analysis, typical example explanations, and related knowledge links; generating guiding conversation content based on learning progress status, including stage summaries, key point previews, and learning suggestions; and generating consolidation conversation content based on cognitive gap characteristics, including personalized practice pushes and reinforcement training suggestions. Simultaneously, Step S8, based on updated user individual interaction preference profiles using historical interaction data, performs differentiated script style adaptation for different users. Personalized interaction adaptation is achieved from two dimensions: matching the knowledge level of the interactive content and matching the expression style of the interactive scripts. This solves the technical problems of existing technologies that fail to differentiate interaction adaptation for users with different foundations and learning abilities due to the lack of association between conversation interaction and user individual profiles and learning characteristics, resulting in inappropriate Q&A content depth and a disconnect between pushed content and training needs.
[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An online training session interaction method based on user behavior mining, characterized by: Includes the following steps: S1: Collect user learning behavior data throughout the entire online training process. The learning behavior data includes at least course click data, learning duration data, exercise answer data, page dwell trajectory data, video playback node data, answer hesitation duration data, conversation question record data, knowledge point repeated error data, and learning interruption frequency data. S2: Perform multi-dimensional fusion analysis and feature extraction on the collected learning behavior data to mine refined behavioral characteristics of users. The refined behavioral characteristics include at least the characteristics of knowledge point mastery, learning habits, cognitive shortcomings, and learning difficulties. S3: Based on the extracted refined behavioral features, construct a user dynamic learning state model to judge the user's current learning state in real time. The current learning state includes at least the user's learning stage, distribution of knowledge weaknesses, current level of focus, and learning progress. S4: Generate an adaptive conversational interaction strategy based on the user dynamic learning state model and the current learning state. The adaptive conversational interaction strategy includes at least the interaction triggering timing decision, the interaction content selection decision, and the interaction speech style decision. S5: Based on the adaptive conversational interaction strategy, proactively generate targeted conversational interaction content and initiate intelligent conversational interaction with the user; S6: Track and collect data on changes in the user's subsequent learning behavior after the user completes the conversational interaction. The data on changes in subsequent learning behavior includes at least the improvement in knowledge mastery, the emergence of subsequent questions, and changes in learning behavior patterns. S7: Based on the subsequent learning behavior change data, evaluate the effectiveness of this conversation interaction and generate an interaction effect evaluation result; S8: Based on the interaction effect evaluation results and accumulated historical interaction data, iteratively optimize the adaptive conversation interaction strategy, and update the conversation script library, interaction trigger rules and interaction content push logic.
2. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: In step S2, the multi-dimensional fusion analysis includes at least: Perform time-series analysis on page dwell time data to identify the distribution of user dwell time on each knowledge point page and abnormal dwell nodes; Cluster analysis was performed on video playback node data to extract knowledge points that users repeatedly watched and the frequency of playback. Statistical analysis of the time spent hesitating to answer questions is conducted, and combined with the accuracy rate of answering exercises, the critical state of users' mastery of knowledge points is determined. Semantic analysis is performed on the conversation question record data to extract the high-frequency knowledge point domains and question depth levels of users' questions; Time series analysis was performed on the frequency of learning interruptions to identify patterns in users' attention decay and periods when learning was easily interrupted.
3. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: In step S3, the user dynamic learning state model adopts a multimodal state representation method, which includes at least a joint representation of the knowledge mastery vector, the learning behavior pattern vector, and the cognitive load vector. The real-time analysis includes: The user's mastery level in each knowledge point is determined based on the knowledge mastery vector. Identify the user's current learning stage based on learning behavior pattern vectors; The user's current cognitive load level is determined based on the cognitive load vector, and the timing of session interaction and the complexity of interaction content are adjusted according to the cognitive load level.
4. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: The interaction trigger timing decision in step S4 includes: When it is detected that the time a user spends on any knowledge point page exceeds a preset threshold, a knowledge point reinforcement session interaction is triggered. When it is detected that the user hesitates for more than a preset threshold when answering any exercise question, a knowledge point Q&A session is triggered. When it is detected that the number of times a user has watched the video content of the same knowledge point exceeds a preset threshold, a key point interpretation session interaction is triggered. When a user logs back in after their learning has been interrupted, a learning progress review and reconnection session is triggered. The interactive content selection decision is based on the distribution of the user's knowledge gaps and learning stage, matching the corresponding interactive content template from the conversation script library, and filling the interactive content template with personalized parameters.
5. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: In step S5, the proactive generation of targeted conversational interaction content includes: Based on the distribution of users' knowledge gaps, generate Q&A conversation content that includes knowledge point analysis, explanation of typical examples, and related knowledge links; Based on the user's learning progress status, generate guided conversation content that includes a summary of the current learning stage, a preview of the key points for the next stage of learning, and learning suggestions; Based on the user's cognitive gap characteristics, generate consolidation conversation content that includes personalized practice push and targeted reinforcement training suggestions; The intelligent conversational interaction adopts a multi-turn dialogue format, dynamically adjusting subsequent conversation branches based on the user's real-time responses.
6. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: The subsequent learning behavior change data in step S6 shall include at least the following: The trend of user accuracy rate in knowledge point exercises within a preset time window after conversation interaction; Changes in user viewing behavior regarding relevant knowledge point video content after conversational interaction, including whether to reduce replays or speed up viewing; Changes in the frequency and depth of user-asked questions after conversational interaction; Changes in the duration of continuous learning and the frequency of learning interruptions after user interaction.
7. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: The interaction effect evaluation in step S7 includes: The subsequent learning behavior change data is compared and analyzed with the historical behavior data before the conversation interaction, and behavior improvement indicators of each dimension are calculated. Based on the aforementioned behavior improvement index, it is determined whether the interaction in this session effectively solved the user's learning problem. If the current session interaction is determined to be invalid or its effect is below a preset threshold, the current session interaction is marked as an interaction event to be optimized, and the interaction context information is recorded for subsequent strategy optimization.
8. The online training session interaction method based on user behavior mining according to claim 1, characterized in that: The iterative optimization in step S8 includes: Based on the accumulated interaction effect evaluation results, reinforcement learning method is used to update the effect score of each dialogue template in the conversation dialogue database; Based on the tagged interactive events to be optimized, the conditional parameters of the interaction triggering rules are adaptively adjusted; Based on the differences in user responses to different interactive content and styles in historical interaction data, update the individual user interaction preference profile to achieve user-level personalized adaptation of interaction strategies.
9. An online training session interaction system based on user behavior mining, used to implement the method as described in any one of claims 1-8, characterized in that: include: The data acquisition module is used to collect user learning behavior data throughout the entire online training process. The learning behavior data includes at least course click data, learning duration data, exercise answer data, page dwell trajectory data, video playback node data, answer hesitation duration data, conversation question record data, knowledge point repeated error data, and learning interruption frequency data. The behavioral feature extraction module is used to perform multi-dimensional fusion analysis and feature extraction on the learning behavior data collected by the data acquisition module, and to mine the refined behavioral features of users. The refined behavioral features include at least the features of knowledge point mastery, learning habits, cognitive shortcomings, and learning difficulties. The user status assessment module is used to construct a dynamic learning status model of the user based on the refined behavioral features extracted by the behavioral feature extraction module, and to assess the user's current learning status in real time. The current learning status includes at least the user's learning stage, the distribution of knowledge weaknesses, the current level of focus, and the learning progress status. An interaction strategy generation module is used to generate an adaptive conversational interaction strategy based on the user dynamic learning state model and current learning state output by the user state judgment module. The adaptive conversational interaction strategy includes at least interaction trigger timing decision, interaction content selection decision, and interaction speech style decision. The conversation content generation module is used to proactively generate targeted conversation interaction content based on the adaptive conversation interaction strategy output by the interaction strategy generation module. The interaction execution module is used to initiate intelligent conversational interactions with users, receive real-time responses from users, and drive multi-round dialogues. The behavior tracking module is used to track and collect data on changes in the user's subsequent learning behavior after the user completes the conversational interaction. The data on changes in subsequent learning behavior includes at least the improvement in knowledge point mastery, the emergence of subsequent questions, and changes in learning behavior patterns. The effectiveness evaluation module is used to evaluate the effectiveness of this session interaction based on the subsequent learning behavior change data collected by the behavior tracking module, and generate an interaction effectiveness evaluation result. The strategy optimization module is used to iteratively optimize the adaptive conversational interaction strategy based on the interaction effect evaluation results output by the effect evaluation module and the accumulated historical interaction data, and to update the conversational script library, interaction triggering rules and interaction content push logic.
10. The online training session interaction system based on user behavior mining according to claim 9, characterized in that, The interaction strategy generation module includes: The trigger timing decision submodule is used to dynamically determine the trigger timing of the conversation interaction based on the user's current learning status. The trigger timing includes at least the following: triggering when the user stays on the knowledge point page for too long, triggering when the user hesitates to answer the exercise for too long, triggering when the user repeatedly watches the video content, and triggering when the user logs in again after the learning is interrupted. The content selection submodule is used to match corresponding interactive content templates from the conversation script library based on the distribution of the user's knowledge gaps and learning stage, and to fill in the interactive content templates with personalized parameters. The conversation style submodule is used to dynamically adjust the conversation style parameters based on the user's cognitive load level and historical interaction preferences. The system also includes a conversation script library update module, which updates the script templates and their weight parameters in the conversation script library in real time based on the iterative optimization results of the strategy optimization module.
