Artificial intelligence early warning intervention system for short video addiction of children

Abstract

The invention discloses an artificial intelligence early warning intervention system for short video addiction of children, and relates to the technical field of crossing of artificial intelligence and mental health. The system comprises a data collection and preprocessing module, an addiction risk assessment model module, a personalized intervention suggestion module, an intelligent early warning and feedback module and a parent education and support module. The short video use behavior of the child is monitored in real time, the addiction risk of the child is accurately evaluated, once it is detected that the child has the addiction risk, the system can automatically send early warning information to a guardian and provide personalized intervention suggestions, and the timely and effective early warning and intervention mechanism helps the guardian to take measures in time, so that the risk of addiction of the child is effectively evaluated. Children are guided to reasonably use short videos and are prevented from being addicted in the short videos, so that physical and psychological health and normal learning life of the children are protected.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of artificial intelligence and mental health, and specifically to an artificial intelligence early warning intervention system for children's short video addiction. Background Art

[0002] With the rapid development of the Internet and the popularization of smart devices, short videos have become one of the important ways for children to entertain and learn. Short video platforms have attracted a large number of child users with their rich content, vivid forms and convenient interactive methods. However, in the process of using short videos, children are prone to become addicted to them due to lack of self-control and correct guidance, leading to a series of problems such as decreased vision, affected academic performance, and deterioration of social skills.

[0003] Traditional technology has obvious shortcomings in addressing the problem of children's addiction to short videos. It is unable to continuously and effectively monitor and evaluate children's short video usage behavior. When it is detected that a child is at risk of addiction, the system cannot automatically send warning information to guardians and provide personalized intervention suggestions.

[0004] To sum up, traditional technologies have many limitations in dealing with the problem of children’s short video addiction, and cannot meet the current society’s needs for the healthy growth of children. Therefore, it is particularly important to develop an artificial intelligence early warning and intervention system for children’s short video addiction. Summary of the invention

[0005] The purpose of the present invention is to make up for the deficiencies of the prior art and to provide an artificial intelligence early warning intervention system for children's short video addiction. The system monitors children's short video usage behavior in real time, accurately assesses the risk of addiction, and provides personalized intervention suggestions and parental education and support, thereby effectively preventing children from becoming addicted to short videos and ensuring their physical and mental health and normal study and life.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: an artificial intelligence early warning intervention system for children's short video addiction, the system includes a data collection and preprocessing module, an addiction risk assessment model module, a personalized intervention suggestion module, an intelligent early warning and feedback module, and a parent education and support module;

[0007] The data collection and preprocessing module collects short video usage data through smart devices authorized by children and their guardians, including viewing time T, content type C, interactive conditions including the number of likes L, the number of comments M, the number of shares S, and eye movement data collected using eye tracking technology. The eye movement data includes the distribution of gaze points P, the gaze duration D, and the eye saccade path E. The collected data is cleaned and noise data is removed. During the data cleaning process, records with a viewing time of less than 5 seconds are deemed invalid and deleted. Manual correction is performed for content types that cannot be identified and incorrectly labeled. Outliers in the eye movement data are corrected or deleted. The eye tracking technology uses a target detection algorithm based on deep learning to identify the position and state of children's eyes. The algorithm formula is:

[0008]

[0009] Among them, O represents the detection result of eye position and state, x i represents the pixel points in the input image frame, N represents the total number of pixels in the image frame, and f k (x i ) indicates that the kth convolution kernel is at pixel x i The convolution result at ω k represents the weight of the kth convolution kernel, b represents the bias term, and σ represents the activation function;

[0010] Using the content type classification algorithm, the short video content types are divided into knowledge education C1, fun entertainment C2, creative handicraft C3, social interaction C4, and other C5. The algorithm is based on deep learning image recognition and text classification technology. For video images, a convolutional neural network is used to extract visual features. The formula is:

[0011]

[0012] Among them, V represents the visual feature vector of the video screen, P represents the set of all pixel points in the video screen, represents the pixel value of the lth channel at position (m,n), θ lmn Represents the corresponding weight parameter, φ represents the bias term, LMN represents the number of channels, height and width of the convolutional layer respectively. For the title and description text information of the video, a recurrent neural network is used for feature extraction. The formula is:

[0015]

[0016] Among them, T represents the text feature vector, r t represents the hidden state at time l, τ trepresents the corresponding weight parameter, ψ represents the bias term, T represents the length of the text sequence, and finally the visual feature vector and the text feature vector are concatenated and classified through the fully connected layer. The classification result is the content type of the short video;

[0017] The addiction risk assessment model module: builds a children's short video addiction risk assessment model based on a machine learning algorithm. The model takes into account the viewing time T, content type C, and interaction frequency. As well as multiple dimensions of eye movement data P, D, E, the short video addiction scale developed by the company is introduced as an important basis for evaluation. The short video addiction scale contains ten questions, each with five options, with scores of 1-5 from completely unacceptable to fully compliant. After the child chooses each question, the scale score Q is calculated. The model is trained using historical data, expert-annotated data, and data collected through the scale. The model training algorithm formula is:

[0018]

[0019] Among them, R represents the addiction risk assessment result, J represents the number of training data samples, and g j Represents the feature mapping function based on the support vector machine, which transforms the input feature vector (T j , C j 、F j , P j , D j 、E j , Q j ) is mapped to a high-dimensional feature space, α j represents the Lagrange multiplier of the corresponding sample, β represents the weight of the scale score in the model, h(Q j ) represents the nonlinear transformation function of the scale score;

[0020] The feature fusion algorithm is used to fuse features of different dimensions. Let the fused feature vector be X, then

[0021] Among them, γ1, γ2, …, γ7 are the weight coefficients of each feature. The fused feature vector X is input into the classifier based on the multi-layer perceptron for addiction risk assessment. The structure of MLP is input layer-hidden layer-output layer. The hidden layer uses ReLU as the activation function. The output layer uses the Sigmoid function to map the output to between 0 and 1, indicating the probability of addiction risk. The weight of MLP is trained on the training data through the back propagation algorithm, and the loss function uses the binary cross entropy loss function.

[0022] The personalized intervention recommendation module: Based on the addiction risk assessment results, it provides personalized intervention recommendations for children and their guardians. If the addiction risk assessment result R is less than the threshold R1 (R1 represents the mild risk stage), based on the statistical analysis of children's short video usage data, a boundary value at a relatively low level is determined for the comprehensive viewing duration factor, and basic time management strategies are provided, setting the daily short video viewing time not exceeding 30 minutes. If R1 ≤ R < R2 (R2 represents the moderate risk stage), according to data statistics and analysis, when the short video usage behavior indicators of children reach this range, simple time management is not sufficient to prevent addiction, and content intervention is required. In addition to time management, content screening recommendations are also provided, recommending educational and popular science short videos according to children's ages and interests. If R ≥ R2, family interaction activity recommendations are added on this basis, with at least 3 joint reading or outdoor activities per week. At the same time, based on the eye movement tracking analysis results, if it is found that the proportion of the fixation duration of children on specific types of content exceeds 60%, such content will be mainly blocked in the content screening recommendations, and relevant alternative content will be recommended;

[0023] Adopt a personalized recommendation algorithm to recommend short video content according to the age A, interest preferences of children, and the addiction risk assessment result R. The recommendation algorithm formula is:

[0024]

[0025] Among them, C rec represents the recommended short video content, Z is the normalization constant, K represents the number of candidate short videos, A k、 I k、 R k respectively represent the appropriate age, target interest preference, and corresponding addiction risk level of the k-th candidate short video. S k represents the distance metric between the feature vector of the k-th candidate short video and the target interest preference of children's interest preferences, and μ, λ1, λ2, λ3 are weight parameters;

[0026] The intelligent early warning and feedback module: when it is monitored that there is a risk of addiction in the short video use behavior of children, that is, the addiction risk assessment result R is greater than or equal to the warning threshold R3, R3 is a high risk stage, an in-depth study of addiction cases and related data is conducted to determine that when various evaluation indicators reach this level, the short video use behavior of children has had a relatively serious impact on their physical and mental health and normal life, and the guardian must be notified in time and a variety of comprehensive intervention measures must be taken to prevent the further deepening of the addiction level, or the scale assessment result Q shows that the child is highly addicted, or the eye movement data is abnormal, the system automatically sends an early warning message to the guardian, the early warning message includes the child’s current short video use status, addiction risk assessment results and existing problem prompts, and provides regular feedback reports, which are generated every two weeks, showing the child’s short video use, addiction risk assessment results and intervention effects, and adding eye movement data related analysis charts to the feedback report;

[0027] Adopting the early warning trigger mechanism, taking into account the dynamic changes of multiple factors, assuming that the early warning index is W, the formula is:

[0028]

[0029] Among them, R max represents the maximum value of the addiction risk assessment result, Q max Indicates the maximum value of the scale score, N a represents the number of abnormal fixations or saccade paths in the eye movement data, N total represents the total number of eye movement data points, ΔD represents the rate of change of fixation duration, and D max represents the maximum value of the rate of change of fixation duration, ω1, ω2, ω3, and ω4 are weight coefficients. When W is greater than or equal to the warning threshold W0, the warning is triggered;

[0030] The parent education and support module provides parent education courses, including the harm of short video addiction to children's physical and mental health, learning ability, and social development, as well as methods to prevent children from being addicted to short videos. The course also includes guidance on how to interpret and use eye tracking data to assist in determining the state of children's addiction, establish a parent support community where parents can share parenting experiences and provide psychological support, and build a parent capacity assessment model to assess parents' ability to deal with children's short video addiction. The model is based on parents' learning behavior in education courses, their participation in the support community, and the intervention effect in the actual parenting process. Assuming the parent capacity assessment value is P, the formula is:

[0031]

[0032] Among them, B i represents the i-th learning behavior indicator, π iRepresents the corresponding weight coefficient, C j represents the jth community participation index, ρ j Represents the corresponding weight coefficient, E k represents the kth intervention effect indicator, σ k Represents the corresponding weight coefficient. Based on the evaluation results, personalized learning suggestions and parenting guidance are provided to parents to help them improve their ability to deal with children’s addiction to short videos.

[0033] Furthermore, in the data collection and preprocessing module, the sampling frequency of the eye tracking device is 120Hz to ensure that the subtle changes of children's eye movements can be accurately captured. The eye tracking device needs to be calibrated before each use. The calibration process adopts a nine-point calibration method, that is, nine calibration points at fixed positions are displayed on the screen, and the children look at these points in turn. The device records the corresponding eye movement data, and adjusts the device parameters according to these data to improve the accuracy of eye movement data collection. For short interruptions or abnormalities in eye movement data caused by children's head movement or blinking, the system uses an interpolation algorithm to repair it. The value of the intermediate missing or abnormal point is calculated based on the valid eye movement data points before and after the interruption or abnormality to ensure the continuity and integrity of the eye movement data for subsequent analysis and processing. The eye tracking device adopts multi-light source compensation technology to reduce the interference of ambient light on eye movement data collection. By arranging multiple light sources of different wavelengths around the device, the brightness of the light source is automatically adjusted according to the ambient light intensity, so that the intensity of the reflected light from the eye is kept within a stable range, thereby improving the accuracy of eye tracking. The algorithm formula of the multi-light source compensation technology is:

[0034]

[0035] Among them, I comp represents the intensity of the reflected light from the eye after compensation, S represents the number of light sources, and I s represents the original reflected light intensity of the s-th light source, ω s Represents the weight coefficient of the sth light source, weight coefficient ω s Through a large number of experimental tests under different ambient light conditions, the weight of each light source is determined according to the stability and accuracy of the reflected light intensity, so that eye movement data can be accurately collected under various ambient light conditions.

[0036] Furthermore, in the addiction risk assessment model module, the kernel function of the support vector machine adopts the radial basis function, and its formula is:

[0037] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0038] Among them, xi and x j represents the feature vector of two inputs, γ is the kernel function parameter. During the training process, the five-fold cross-validation method is used to evaluate the model performance, that is, the training data is divided into five parts, four of which are used as training sets and one as validation sets in turn, and the average accuracy on the five validation sets is calculated. The model parameters are adjusted according to the average accuracy to ensure that the model has good generalization ability. The adaptive kernel function adjustment strategy is introduced. During the model training process, the kernel function parameter γ of the radial basis function is adjusted in real time according to the dynamic changes of the data distribution. Assuming the density function of the data distribution is f(x), the formula is:

[0039]

[0040] Among them, γ new represents the adjusted kernel function parameters, γ old represents the current kernel function parameters, η is the learning rate, and the density function of the data distribution f(x)

[0041] Calculated by the kernel density estimation method, the KDE formula is:

[0042]

[0043] Among them, n represents the number of samples, h represents the bandwidth parameter, and K(·) represents the kernel function. Through this adaptive adjustment strategy, the model can better adapt to the data characteristics of different stages and improve the accuracy and stability of the model.

[0044] Furthermore, in the personalized intervention suggestion module, the recommendation algorithm for content screening suggestions is based on a collaborative filtering algorithm, and the steps are as follows:

[0045] Construct a user-content matrix, where rows represent child users, columns represent short video content, and matrix elements represent children’s preference for short videos. The specific calculation method is:

[0046] For each child user u and short video content c, the preference level r u,c The calculation formula is r u,c =αT u,c +β(L u,c +M u,c +S u,c ), where T u , c represents the time that child u watches content c, L u ,cM u,c、 S u,c They represent the number of likes, comments, and shares of child u on content c, α and β are weight coefficients, and the similarity between users is calculated using the cosine similarity formula:

[0047]

[0048] Among them, sim(u1,u2) represents the similarity between users u1 and u2. and They represent the preference of users u1 and u2 for content c respectively; C represents the total number of contents;

[0049] Find the k users who are most similar to the target child, and recommend short video content that the target child has not watched and has a high recommendation degree based on the content preferences of these similar users. The recommendation degree calculation formula is:

[0050]

[0051] Among them, rec c represents the recommendation degree of content c, U k represents the set of k users most similar to the target child, sim(u,t) represents the similarity between user u and target child t, r u ,c represents the preference of user u for content c;

[0052] The time decay factor is introduced to optimize the recommendation algorithm. Taking into account the characteristics of children's interests changing over time, different weights are given to historical viewing and interaction data according to the time interval. The time decay function is set to λ(t) = e -δt , where t represents the time interval and δ is the attenuation coefficient, then the optimized preference degree calculation formula is r u,c =αT u,c λ(t u,c )+β(L u,c +M u,c +S u,c )λ(t u,c ), where t u , c represents the interval between the last operation time of child u on content c and the current time. This can make the recommendation method pay more attention to the user behavior in the past and improve the timeliness and relevance of the recommended content.

[0053] Furthermore, in the intelligent early warning and feedback module, in addition to notifying the guardian through text messages and in-app push notifications, the early warning information can also send voice reminders to the guardian's smart device with the guardian's authorization. The voice reminder content is customized according to the child's specific addiction situation. The voice reminder is implemented using text-to-speech technology. The system pre-stores a variety of voice tones for guardians to choose from to meet the personalized needs of different guardians. Regular feedback reports are sent to guardians in the form of emails. The email title contains the child's name, report time period, and key information summary, which is convenient for guardians to quickly identify the importance of emails. At the same time, the feedback report history is retained in the system, and the guardian can log in to view it at any time. A warning information priority classification model is established to comprehensively determine the priority of the warning information based on the addiction risk assessment results, the degree of abnormality of the eye movement data, and the scale score factors. Assuming the priority of the warning information is Prio, the formula is:

[0054]

[0055] Among them, R man represents the maximum value of the addiction risk assessment result, Q max Indicates the maximum value of the scale score, N a represents the number of abnormal fixations or saccade paths in the eye movement data, N total represents the total number of eye movement data points, ω5, ω6, and ω7 are weight coefficients, and weights are assigned according to the degree of influence of each factor on the urgency of the warning. Different notification methods and frequencies are used according to different priorities. For example, high-priority warning information is immediately notified to the guardian via SMS, voice reminder, and in-app push notification, and the reminder is repeated every 15 minutes until the guardian confirms receipt; medium-priority warning information is notified via SMS and in-app push notification, and the reminder is repeated every 30 minutes; low-priority warning information is only notified via in-app push notification, and the reminder is once a day.

[0056] Furthermore, in the parent education and support module, the parent education courses are presented in the form of video courses, online live lectures, and graphic materials to meet the learning habits and time arrangements of different parents. Online live lectures are held regularly, and graphic materials cover theoretical knowledge, research reports, and operation guides related to short video addiction, which are convenient for parents to consult at any time. The parent support community sets up different sections, such as experience sharing area, problem help area, and expert Q&A area. In the experience sharing area, parents can post their own successful experiences and accompany them with pictures or video descriptions. In the problem help area, parents can describe the problems they encounter, and other parents and experts can reply to provide suggestions. In the expert Q&A area, professional psychological counselors and education experts regularly publish professional answers and guidance articles on children's short video addiction problems. A parent learning progress tracking and evaluation system is developed to track parents' learning progress and mastery in education courses. The system records the length of time parents watch video courses, the degree of participation in online live lectures, and the reading and feedback of graphic materials. Assuming the parent learning progress evaluation value is Prog, the formula is:

[0057]

[0058] Among them, V i represents the index related to the i-th video course, μ i Represents the corresponding weight coefficient, L j represents the index related to the jth online live lecture, ν j Represents the corresponding weight coefficient, T k represents the kth graphic data related index, ξ k Represents the corresponding weight coefficient. Based on the evaluation results, it provides parents with personalized learning path recommendations and learning reminders to help parents better complete educational courses.

[0059] Furthermore, the system also includes a data encryption and security storage module, which is used to encrypt the collected child data and store it securely in the cloud server. The data encryption adopts the AES encryption algorithm. The encryption key is generated by processing the password set by the child guardian through the hash function to ensure the security and uniqueness of the key. The encryption process encrypts all collected data block by block to prevent the data from being stolen or altered during transmission and storage. The cloud server adopts a distributed storage architecture to store data in multiple nodes. The reliability and integrity of the data are guaranteed through redundant backup and data verification mechanisms. At the same time, the access rights to cloud data are strictly restricted. Only authorized system administrators and data analysts can access the data in a specific security environment to ensure data security.

[0060] Data anonymization technology is introduced to anonymize some sensitive information before data storage, and the data records are grouped using the k-anonymity method so that at least k records in each group have the same quasi-identifier attribute value, and individual records cannot be distinguished by these quasi-identifier attribute values. This further protects children's privacy and security while ensuring data availability.

[0061] Furthermore, the system also includes an adaptive learning and model updating module, which can continuously optimize the model according to the newly collected data, regularly collect new children's short video usage data, eye movement data and scale data, add new data to the training set every month, and retrain the addiction risk assessment model. During the training process, an incremental learning algorithm is used to update only the part of the model that is most affected by the new data, avoiding retraining the entire model and improving the model update efficiency. After the model is updated, the performance of the new model is evaluated, and the same evaluation indicators as in the model training process are used to compare the evaluation results with the old model. If the performance of the new model is significantly improved (such as an increase of more than 5% in accuracy), the old model is officially replaced; otherwise, the new data is further analyzed to find existing problems (such as data anomalies or labeling errors), and the model update attempt is made again after adjustment;

[0062] Establish a model performance monitoring and early warning submodule to monitor the performance of the model in actual applications in real time. By calculating the prediction error and confusion matrix index of the model on new data, the stability and accuracy of the model are evaluated. Assuming the model performance evaluation index is Perf, the formula is:

[0063] Perf=ω8Error+ω9FPR+ω 10 FNR

[0064] Among them, Error represents the prediction error, FPR represents the false positive rate, FNR represents the false negative rate, ω8, ω9, ω 10 It is the weight coefficient. When Perf exceeds the set threshold, an early warning is triggered, prompting system administrators and data analysts to check and optimize the model to ensure that the model always maintains good performance.

[0065] Compared with existing technologies, the artificial intelligence early warning intervention system for children's short video addiction has the following beneficial effects:

[0066] 1. This system monitors children's short video usage behavior in real time and accurately assesses their addiction risk. Once it detects that a child is at risk of addiction, the system will automatically send an early warning message to the guardian and provide personalized intervention suggestions. This timely and effective early warning and intervention mechanism helps guardians take timely measures to guide children to use short videos reasonably and avoid addiction, thereby protecting children's physical and mental health and normal study and life.

[0067] 2. Through the parent education and support module, the system provides parents with a wealth of educational courses and learning resources to help parents understand the dangers of short video addiction, prevention methods, and how to interpret and use eye tracking data to assist in judging children's addiction status. At the same time, the system also establishes a parent support community where parents can share parenting experiences, seek psychological support, and receive guidance from professional psychological counselors and education experts. These measures will help improve parents' ability to deal with children's short video addiction problems and promote family harmony and the healthy growth of children. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0069] Figure 1 Operational flow chart of the artificial intelligence early warning and intervention system for children’s short video addiction. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] Embodiment 1

[0072] This embodiment describes a typical dual-income family where the parents are often unable to accompany their 8-year-old child Xiao Ming due to their busy work. After school, Xiao Ming often uses a tablet computer to watch short videos alone, which leads to his addiction to short videos, affecting his study and rest time.

[0073] Xiao Ming's parents installed a short video application with eye tracking function for Xiao Ming. The application can monitor Xiao Ming's viewing behavior in real time, including viewing time, content type, interaction and eye movement data. Through smart device authorization, the system begins to collect Xiao Ming's short video usage data, and cleans and pre-processes it to remove noise data, such as short and invalid viewing records, to ensure the accuracy and validity of the data.

[0074] The system uses an addiction risk assessment model built based on machine learning algorithms.

[0075]

[0076] Among them, R represents the addiction risk assessment result, J represents the number of training data samples, and g j Represents the feature mapping function based on the support vector machine, which transforms the input feature vector (T j , C j 、F j , P j , D j 、E j , Q j ) is mapped to a high-dimensional feature space, α j represents the Lagrange multiplier of the corresponding sample, β represents the weight of the scale score in the model, h(Q j ) represents the nonlinear transformation function of the scale score, which comprehensively evaluates Xiao Ming’s viewing time, content preference, interaction frequency, and eye movement data, and conducts a quantitative assessment through the short video addiction scale. The evaluation results show that Xiao Ming has a high risk of addiction and needs to take intervention measures.

[0077] Based on the addiction risk assessment results, the system provides personalized intervention suggestions for Xiao Ming and his parents.

[0078]

[0079] Among them, C rec represents the recommended short video content, Z is a normalization constant, K represents the number of candidate short videos, and A k、 I k、 R k They represent the appropriate age, target interest preference, and corresponding addiction risk level of the kth candidate short video, respectively. k The feature vector representing the kth candidate short video represents the distance measure between the child's interest preference and the target interest preference of the candidate short video. μ, λ1, λ2, and λ3 are weight parameters. It is recommended that Xiao Ming set the daily short video viewing time to no more than 30 minutes, and recommend educational and popular science short video content to enrich Xiao Ming's knowledge horizon. At the same time, the system also recommends that parents read or do outdoor sports with Xiao Ming at least three times a week to increase family interaction and cultivate Xiao Ming's interests and social skills.

[0080] When the system detects that Xiao Ming's short video usage behavior is at risk of addiction, it automatically sends a warning message to Xiao Ming's parents. Assuming the warning index is W, the formula is:

[0081]

[0082] Among them, R max represents the maximum value of the addiction risk assessment result, Q max Indicates the maximum value of the scale score, N a represents the number of abnormal fixations or saccade paths in the eye movement data, Ntotal represents the total number of eye movement data points, ΔD represents the rate of change of fixation duration, and D max It represents the maximum value of the rate of change of gaze duration, ω1, ω2, ω3, and ω4 are weight coefficients. When W is greater than or equal to the warning threshold W0, an early warning is triggered, which includes the current short video usage status, addiction risk assessment results, and existing problem prompts. The system generates a feedback report every two weeks, showing Xiao Ming's short video usage, addiction risk assessment results, and intervention effects. The report also adds eye movement data related analysis charts to help parents understand Xiao Ming's viewing behavior more intuitively.

[0083] The system provides parent education courses, including the harm of short video addiction to children's physical and mental health and prevention methods, to help parents recognize the seriousness of the problem and master effective intervention strategies. Through course learning, parents have mastered how to interpret and use eye tracking data to assist in judging Xiao Ming's addiction state, thereby improving the targetedness and effectiveness of intervention. At the same time, parents join the parent support community to share parenting experiences with other parents, obtain psychological support, and jointly deal with the problem of children being addicted to short videos.

[0084] The system continuously monitors Xiao Ming's short video usage behavior and adjusts the intervention strategy according to the actual situation. If Xiao Ming's addiction risk decreases, the system will gradually relax the viewing time limit and recommend more beneficial short video content. If Xiao Ming's addiction risk continues to be high, the system will strengthen early warning and intervention efforts to ensure that Xiao Ming can gradually get rid of the trouble of short video addiction.

[0085] Parents strengthen their interaction and supervision with Xiao Ming in daily life, pay attention to Xiao Ming's psychological changes and interests and hobbies, and cultivate Xiao Ming's self-discipline and sense of responsibility by participating in family activities together and formulating family rules. At the same time, parents also encourage Xiao Ming to actively participate in school and social activities to broaden his horizons and enhance his social skills.

[0086] Embodiment 2

[0087] This embodiment describes the observation of a primary school in recent years that with the popularity of smart phones, students frequently use their mobile phones to watch short videos during breaks and after school. While this phenomenon brings entertainment and relaxation, it also causes addiction problems for some students. Watching short videos for a long time not only affects students' academic performance, but also interferes with their normal rest and physical and mental health development.

[0088] The school cooperates with well-known smart device manufacturers to equip all students with smart glasses or smart watches with high-precision eye tracking functions. These devices can not only record students' short video viewing behavior, but also capture their eye movement trajectories while watching.

[0089] Among them, O represents the detection result of eye position and state, x i represents the pixel points in the input image frame, N represents the total number of pixels in the image frame, and f k (x i ) indicates that the kth convolution kernel is at pixel x i The convolution result at ω k represents the weight of the kth convolution kernel, b represents the bias term, and σ represents the activation function, which provides important data for the subsequent addiction risk assessment. Before collecting data, the school signs a detailed data use agreement with parents to ensure that all information collection is based on parental consent and strictly abides by relevant laws and regulations to protect students' personal privacy. The system background cleans the large amount of raw data collected, removes invalid, abnormal or duplicate data, and ensures the accuracy and completeness of the data used for analysis.

[0090] Based on the multi-dimensional features of students’ viewing time, content preferences, interaction frequency, and eye movement data, we construct feature vectors and use ensemble learning methods, such as random forests or gradient boosting trees, combined with deep learning models to train a model that can accurately identify addiction risks.

[0091] Among them, R represents the addiction risk assessment result, J represents the number of training data samples, and g j Represents the feature mapping function based on the support vector machine, which transforms the input feature vector (T j , C j 、F j , P j , D j 、E j , Q j ) is mapped to a high-dimensional feature space, α j represents the Lagrange multiplier of the corresponding sample, β represents the weight of the scale score in the model, h(Q j ) represents the nonlinear transformation function of the scale score. The model performance is evaluated through cross-validation method, and it is continuously optimized to ensure that the model can accurately identify addiction tendencies and avoid false positives and negatives.

[0092] According to the age, gender, interests and learning needs of students, educational and inspiring short video content is intelligently recommended. Short video content is recommended based on the child’s age A, interest preference and addiction risk assessment result R. The recommendation algorithm formula is:

[0093]

[0094] Among them, C rec represents the recommended short video content, Z is a normalization constant, K represents the number of candidate short videos, and A k、 I k、 Rk They represent the appropriate age, target interest preference, and corresponding addiction risk level of the kth candidate short video, respectively. k The feature vector representing the kth candidate short video represents the distance measure between the child's interest preference and the target interest preference of the candidate short video. μ, λ1, λ2, and λ3 are weight parameters, such as scientific experiments, historical culture, and art appreciation. The system guides students to watch high-quality content. The system automatically sets a reasonable viewing time window based on the students' daily study and life patterns, and automatically locks the short video application after the time is exceeded. Students are encouraged to participate in outdoor activities or rest, and cooperate with schools to plan a series of colorful recess activities, such as sports competitions, art creation, and small scientific experiments, to enhance students' collective participation and interests and hobbies, and reduce their dependence on short videos.

[0095] The system monitors students' short video usage in real time. Once the risk of addiction is detected, it immediately sends warning information to the class teacher and parents through SMS and APP push. Suppose the warning index is W, then the formula is:

[0096]

[0097] Among them, R max represents the maximum value of the addiction risk assessment result, Q max Indicates the maximum value of the scale score, N a represents the number of abnormal fixations or saccade paths in the eye movement data, N total represents the total number of eye movement data points, ΔD represents the rate of change of fixation duration, and D max represents the maximum value of the rate of change of gaze duration, ω1, ω2, ω3, and ω4 are weight coefficients. When W is greater than or equal to the warning threshold W0, an early warning is triggered and detailed feedback reports are generated regularly, including analysis of students’ viewing behavior, addiction risk level, implementation of intervention measures, and effect evaluation, for reference by parents and schools.

[0098] Invite psychology experts and educational technology experts to hold online or offline lectures to educate parents about the dangers, preventive measures and family management methods of short video addiction. Establish a home-school communication APP or WeChat group to facilitate instant communication between parents, students and teachers, share successful cases, and jointly solve problems encountered by students in using short videos. Organize parents to participate in workshop activities, and enhance parents' ability to identify signs of children's addiction through role-playing and scenario simulation, and learn effective intervention techniques.

[0099] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An artificial intelligence early warning and intervention system for children's short video addiction, characterized by: The system includes a data collection and preprocessing module, an addiction risk assessment model module, a personalized intervention suggestion module, an intelligent early warning and feedback module, and a parent education and support module: The data collection and preprocessing module collects short video usage data through smart devices authorized by children and their guardians, including viewing time T, content type C, interactive conditions including the number of likes L, the number of comments M, the number of shares S, and eye movement data collected using eye tracking technology. The eye movement data includes the distribution of gaze points P, the gaze duration D, and the eye saccade path E. The collected data is cleaned and noise data is removed. During the data cleaning process, records with a viewing time of less than 5 seconds are deemed invalid and deleted. Manual correction is performed for content types that cannot be identified and incorrectly labeled. Outliers in the eye movement data are corrected or deleted. The eye tracking technology uses a target detection algorithm based on deep learning to identify the position and state of children's eyes. The algorithm formula is: Among them, O represents the detection result of eye position and state, x i represents the pixel points in the input image frame, N represents the total number of pixels in the image frame, and f k (x i ) indicates that the kth convolution kernel is at pixel x i The convolution result at ω k represents the weight of the kth convolution kernel, b represents the bias term, and σ represents the activation function; Using the content type classification algorithm, the short video content types are divided into knowledge education C1, fun entertainment C2, creative handicraft C3, social interaction C4, and other C5. The algorithm is based on deep learning image recognition and text classification technology. For video images, a convolutional neural network is used to extract visual features. The formula is: Among them, V represents the visual feature vector of the video screen, P represents the set of all pixel points in the video screen, represents the pixel value of the lth channel at position (m,n), θ lmn represents the corresponding weight parameter, φ represents the bias term, LMN represents the number of channels, height and width of the convolutional layer, and for the video labels, Title, description text information, and feature extraction using a recurrent neural network. The formula is: Among them, T represents the text feature vector, r t represents the hidden state at time l, τ t represents the corresponding weight parameter, ψ represents the bias term, T represents the length of the text sequence, and finally the visual feature vector and the text feature vector are concatenated and classified through the fully connected layer. The classification result is the content type of the short video; The addiction risk assessment model module: builds a children's short video addiction risk assessment model based on a machine learning algorithm. The model takes into account the viewing time T, content type C, and interaction frequency. As well as multiple dimensions of eye movement data P, D, E, the short video addiction scale developed by the company is introduced as an important basis for evaluation. The short video addiction scale contains ten questions, each with five options, with scores of 1-5 from completely unacceptable to fully compliant. After the child chooses each question, the scale score Q is calculated. The model is trained using historical data, expert-annotated data, and data collected through the scale. The model training algorithm formula is: Among them, R represents the addiction risk assessment result, J represents the number of training data samples, and g j Represents the feature mapping function based on the support vector machine, which transforms the input feature vector (T j , C j 、F j , P j , D j 、E j , Q j ) is mapped to a high-dimensional feature space, α j represents the Lagrange multiplier of the corresponding sample, β represents the weight of the scale score in the model, h(Q j ) represents the nonlinear transformation function of the scale score; The feature fusion algorithm is used to fuse features of different dimensions. Let the fused feature vector be X, then Among them, γ1, γ2, …, γ7 are the weight coefficients of each feature. The fused feature vector X is input into the classifier based on the multi-layer perceptron for addiction risk assessment. The structure of MLP is input layer-hidden layer-output layer. The hidden layer uses ReLU as the activation function. The output layer uses the Sigmoid function to map the output to between 0 and 1, indicating the probability of addiction risk. The weight of MLP is trained on the training data through the back propagation algorithm, and the loss function uses the binary cross entropy loss function. The personalized intervention recommendation module: According to the addiction risk assessment results, it provides personalized intervention recommendations for children and their guardians. If the addiction risk assessment result R is less than the threshold R1 (R1 is the mild risk stage), based on the statistical analysis of children's short video usage data, a boundary value at a relatively low level is determined for the comprehensive viewing duration factor, and basic time management strategies are provided, setting the daily short video viewing time not to exceed 30 minutes. If R1 ≤ R < R2 (R2 is the moderate risk stage), according to data statistics and analysis, when the short video usage behavior indicators of children reach this range, simple time management is not sufficient to prevent addiction, and content intervention is required. In addition to time management, content screening recommendations are also provided, recommending educational and popular science short videos according to children's age and interests. If R ≥ R2, on the above basis, family interaction activity suggestions are added, with at least 3 joint reading or outdoor sports activities per week. At the same time, based on the eye movement tracking analysis results, if the fixation duration ratio of children for specific types of content exceeds 60%, such content will be mainly blocked in the content screening recommendations, and relevant alternative content will be recommended; Adopt a personalized recommendation algorithm to recommend short video content according to the age A, interest preferences of children, and the addiction risk assessment result R. The recommendation algorithm formula is: Among them, C rec represents the recommended short video content, Z is a normalization constant, K represents the number of candidate short videos, and A k、 I k、 R k They represent the appropriate age, target interest preference, and corresponding addiction risk level of the kth candidate short video, respectively. k The feature vector representing the kth candidate short video represents the distance measure between the child's interest preference and the target interest preference of the candidate short video, and μ, λ1, λ2, and λ3 are weight parameters; The intelligent early warning and feedback module: When it is detected that there is an addiction risk in children's short video usage behavior, that is, the addiction risk assessment result R is greater than or equal to the early warning threshold R3 (R3 is the high risk stage), through in-depth research on addiction cases and relevant data, it is determined that when various evaluation indicators reach this level, children's short video usage behavior has had a relatively serious impact on their physical and mental health and normal life, and guardians must be notified in a timely manner and various comprehensive intervention measures must be taken to prevent the further deepening of the addiction level. Or when the scale evaluation result Q shows that the child's addiction level is relatively high, or when abnormal eye movement data appears, the system automatically sends an early warning message to the guardian. The early warning message includes the child's current short video usage status, addiction risk assessment result, and problem prompts, and provides a regular feedback report, which is generated once every two weeks, showing the child's short video usage situation, addiction risk assessment result, and intervention effect, and adding an eye movement data related analysis chart in the feedback report; Adopt an early warning trigger mechanism, comprehensively considering the dynamic changes of various factors. Let the early warning index be W, then the formula is: Among them, R max represents the maximum value of the addiction risk assessment result, Q max Indicates the maximum value of the scale score, N a represents the number of abnormal fixations or saccade paths in the eye movement data, N total represents the total number of eye movement data points, ΔD represents the rate of change of fixation duration, and D max represents the maximum value of the rate of change of fixation duration, ω1, ω2, ω3, and ω4 are weight coefficients. When W is greater than or equal to the warning threshold W0, the warning is triggered; The parent education and support module: Provide parent education courses, the course content includes the harms of short video addiction to children's physical and mental health, learning ability, and social development, as well as methods for preventing children's short video addiction. The course also includes guidance on how to interpret and utilize eye movement tracking data to assist in judging the child's addiction status. Establish a parent support community where parents can share parenting experiences, and at the same time provide psychological support. Construct a parent ability evaluation model to evaluate the ability level of parents in dealing with children's short video addiction problems. This model is based on parents' learning behaviors in education courses, participation in the support community, and intervention effects in actual parenting processes. Let the parent ability evaluation value be P, then the formula is: Among them, B i represents the i-th learning behavior indicator, π i Represents the corresponding weight coefficient, C j represents the jth community participation index, ρ j Represents the corresponding weight coefficient, E k represents the kth intervention effect indicator, σ k Represents the corresponding weight coefficient. Based on the evaluation results, personalized learning suggestions and parenting guidance are provided to parents to help them improve their ability to deal with children’s addiction to short videos.

2. The artificial intelligence early warning and intervention system for children's short video addiction according to claim 1 is characterized in that: In the data collection and preprocessing module, the eye tracking device needs to be calibrated before each use. The calibration process adopts the nine-point calibration method, that is, nine calibration points at fixed positions are displayed on the screen, and the children look at these points in turn. The device records the corresponding eye movement data and adjusts the device parameters according to these data. For the short interruption or abnormality of the eye movement data caused by the child's head movement or blinking, the system uses an interpolation algorithm to repair it, and calculates the value of the middle missing or abnormal point according to the valid eye movement data points before and after the interruption or abnormality. The eye tracking device adopts multi-light source compensation technology, by arranging multiple light sources of different wavelengths around the device, and automatically adjusting the brightness of the light source according to the ambient light intensity, so that the intensity of the eye reflected light remains within a stable range. The algorithm formula of the multi-light source compensation technology is: Among them, I comp represents the intensity of the reflected light from the eye after compensation, S represents the number of light sources, and I s represents the original reflected light intensity of the s-th light source, ω represents the weight coefficient of the s-th light source, and the weight coefficient ω is obtained through a large number of experimental tests under different ss ambient light conditions.

3. The artificial intelligence early warning and intervention system for children's short video addiction according to claim 1 is characterized in that: In the addiction risk assessment model module, the kernel function of the support vector machine adopts the radial basis function, and its formula is: K(x i ,x j )=exp(-γ||x i -x j || 2 ) Among them, x i and x j represents the feature vector of two inputs, γ is the kernel function parameter. During the training process, the five-fold cross-validation method is used to evaluate the model performance, that is, the training data is divided into five parts, four of which are used as training sets and one as validation sets in turn, and the average accuracy on the five validation sets is calculated. The model parameters are adjusted according to the average accuracy, and the adaptive kernel function adjustment strategy is introduced. During the model training process, the kernel function parameter γ of the radial basis function is adjusted in real time according to the dynamic changes of the data distribution. Assuming the density function of the data distribution is f(x), the formula is: Among them, γ new represents the adjusted kernel function parameters, γ old represents the current kernel function parameters, η is the learning rate, and the density function of the data distribution f(x) Calculated by the kernel density estimation method, the KDE formula is: Among them, n represents the number of samples, h represents the bandwidth parameter, and K(·) represents the kernel function. Through this adaptive adjustment strategy, the model can better adapt to the data characteristics at different stages.

4. The artificial intelligence early warning and intervention system for children's short video addiction according to claim 1 is characterized in that: In the personalized intervention suggestion module, the recommendation algorithm for content screening suggestions is based on a collaborative filtering algorithm, and the steps are as follows: Construct a user-content matrix, where rows represent child users, columns represent short video content, and matrix elements represent children’s preference for short videos. The specific calculation method is: For each child user u and short video content c, the preference level r u,c The calculation formula is r u,c =αT u,c +β(L u,c +M u,c +S u,c ), where T u , c represents the time that child u watches content c, L u ,cM u,c、 S u,c They represent the number of likes, comments, and shares of child u on content c, α and β are weight coefficients, and the similarity between users is calculated using the cosine similarity formula: Among them, sim(u1,u2) represents the similarity between users u1 and u2, r u1,c and r u2,c They represent the preference of users u1 and u2 for content c respectively; C represents the total number of contents; Find the k users who are most similar to the target child, and recommend short video content that the target child has not watched and has a high recommendation degree based on the content preferences of these similar users. The recommendation degree calculation formula is: Among them, rec c represents the recommendation degree of content c, U k represents the set of k users most similar to the target child, sim(u,t) represents the similarity between user u and target child t, r u ,c represents the preference of user u for content c; The time decay factor is introduced to optimize the recommendation algorithm. Taking into account the characteristics of children's interests changing over time, different weights are given to historical viewing and interaction data according to the time interval. The time decay function is set to λ(t) = e -δt , where t represents the time interval and δ is the attenuation coefficient, then the optimized preference degree calculation formula is r u,c =αT u,c λ(t u,c )+β(L u,c +M u,c +S u,c )λ(t u,c ), where t u ,c represents the interval between the last operation time of child u on content c and the current time.

5. According to claim 1, the artificial intelligence early warning and intervention system for children's short video addiction is characterized in that: In the intelligent early warning and feedback module, in addition to notifying the guardian through text messages and in-app push notifications, the early warning information can also send voice reminders to the guardian's smart device with the guardian's authorization. The voice reminder content is customized according to the child's specific addiction situation. The voice reminder is implemented using text-to-speech technology. The system pre-stores a variety of voice tones for guardians to choose from. Regular feedback reports are sent to guardians in the form of emails. The email title contains the child's name, report time period, and key information summary. At the same time, the feedback report history is retained in the system, and the guardian can log in to view it at any time. A warning information priority classification model is established to comprehensively determine the priority of the warning information based on the addiction risk assessment results, the degree of eye movement data abnormality, and the scale score factors. Suppose the warning information priority is Prio, then the formula is: Among them, R man represents the maximum value of the addiction risk assessment result, Q max Indicates the maximum value of the scale score, N a represents the number of abnormal fixations or saccade paths in the eye movement data, N total It represents the total number of eye movement data points. ω5, ω6, and ω7 are weight coefficients. The weights are assigned according to the influence of each factor on the urgency of the warning. Different notification methods and frequencies are adopted according to different priorities.

6. The artificial intelligence early warning and intervention system for children's short video addiction according to claim 1 is characterized in that: In the parent education and support module, the parent education courses are presented in the form of video courses, online live lectures, and graphic materials. Online live lectures are held regularly, and the graphic materials cover theoretical knowledge, research reports, and operation guides related to short video addiction. The parent support community sets up different sections and develops a parent learning progress tracking and evaluation system to track the learning progress and mastery of parents in the education courses, record the length of time parents watch video courses, the degree of participation in online live lectures, and the reading and feedback of graphic materials. Assuming that the parent learning progress evaluation value is Prog, the formula is: Among them, V i represents the index related to the i-th video course, μ i Represents the corresponding weight coefficient, L j represents the index related to the jth online live lecture, ν j Represents the corresponding weight coefficient, T k represents the kth graphic data related index, ξ k Represents the corresponding weight coefficient, and provides parents with personalized learning path recommendations and learning reminders based on the evaluation results.

7. The artificial intelligence early warning and intervention system for children's short video addiction according to claim 1 is characterized in that: The system also includes a data encryption and security storage module, which is used to encrypt the collected child data and store it securely in a cloud server. The data encryption adopts the AES encryption algorithm. The encryption key is generated by processing the password set by the child's guardian through a hash function. The encryption process encrypts all the collected data block by block. The cloud server adopts a distributed storage architecture to store the data in multiple nodes. At the same time, the access rights to the cloud data are strictly restricted. Only authorized system administrators and data analysts can access the data in a specific security environment. Data anonymization technology is introduced to anonymize some sensitive information before data storage, and the data records are grouped using the k-anonymity method, so that at least k records in each group have the same quasi-identifier attribute value, and individual records cannot be distinguished by these quasi-identifier attribute values.

8. The artificial intelligence early warning and intervention system for children's short video addiction according to claim 1 is characterized in that: The system also includes an adaptive learning and model updating module, which can continuously optimize the model according to the newly collected data, regularly collect new children's short video usage data, eye movement data and scale data, add new data to the training set every month, and retrain the addiction risk assessment model. During the training process, an incremental learning algorithm is used to update only the part of the model that is most affected by the new data. After the model is updated, the performance of the new model is evaluated, and the same evaluation indicators as in the model training process are used to compare the evaluation results with the old model; Establish a model performance monitoring and early warning submodule to monitor the performance of the model in actual applications in real time. By calculating the prediction error and confusion matrix index of the model on new data, the stability and accuracy of the model are evaluated. Assuming the model performance evaluation index is Perf, the formula is: Perf=ω8Error+ω9FPR+ω 10 FNR Among them, Error represents the prediction error, FPR represents the false positive rate, FNR represents the false negative rate, ω8, ω9, ω 10 It is the weight coefficient. When Perf exceeds the set threshold, an early warning is triggered, prompting system administrators and data analysts to check and optimize the model.

Images