Method for displaying age adaptability of display content of teenagers

Abstract

The invention discloses a method for displaying age adaptability of teenager display content, and belongs to the field of age adaptability display, and the method comprises the following steps: S1, obtaining age information of a user; the content on the platform is marked with grades and marked labels, and age levels are subdivided; s2, constructing a content filtering model, automatically identifying the content by using natural language processing, image identification and video analysis technologies, and labeling an identification tag; s3, performing multi-level content screening according to the age of the user; s4, dynamically adjusting a content recommendation algorithm by using a real-time monitoring technology; s5, adjusting and optimizing a filtering mechanism by combining manpower and an algorithm; and S6, setting a transparent content adaptation rule on the platform. The method has the beneficial effects that (1) the adaptability and the timeliness are improved; and (2) the content display accuracy is improved. And (3) the content security and the user experience are balanced. And (4) the problem of rigidity of traditional filtering is avoided. And (5) the transparency and interactivity of content management are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of age-appropriate display, and more particularly, to a method for displaying the age appropriateness of content for teenagers. Background Art

[0002] The age appropriateness of content for teenagers has always been an important topic in digital technology and social culture. In the early days of the Internet, age appropriateness of content display was mainly achieved through simple filtering techniques, such as keyword filtering or age restrictions declared by users themselves. The historical technical background of these methods mainly originated from the need for static content censorship, with the focus on protecting teenagers from inappropriate content. With the popularity of social media, short video platforms, and interactive content, traditional filtering methods have gradually become unable to cope with the complex content ecosystem. Technology has started to turn to more complex algorithms such as machine learning, natural language processing (NLP), and image recognition for dynamic screening and real-time adaptation of content. In this process, platforms have started to introduce artificial intelligence (AI) to analyze text, images, and videos, automatically identify violent, pornographic, or other content inappropriate for teenagers, and decide whether to display specific content to users of certain age groups based on parameters such as content type, context, and mood.

[0003] Despite continuous technological progress, there is a significant shortcoming in this content adaptation method: algorithms often lack precision in understanding the complex context and intent of content, especially in dealing with implicit or ambiguous expressions. This defect may lead to "misjudgments", such as mislabeling educational or artistic content as inappropriate for display, or "missed judgments", where some content inappropriate for teenagers is not correctly filtered. Furthermore, algorithms may be affected by biased data during the design and training process, resulting in unfair adaptation judgments. Therefore, although modern technology enables age appropriateness to be achieved, its lack of accuracy remains an urgent problem to be solved. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for displaying the age appropriateness of content for teenagers to solve the problems raised in the above background art: Despite continuous technological progress, there is a significant shortcoming in this content adaptation method: algorithms often lack precision in understanding the complex context and intent of content, especially in dealing with implicit or ambiguous expressions. This defect may lead to "misjudgments", such as mislabeling educational or artistic content as inappropriate for display, or "missed judgments", where some content inappropriate for teenagers is not correctly filtered. Furthermore, algorithms may be affected by biased data during the design and training process, resulting in unfair adaptation judgments.

[0005] Technical solution: 1. A display method for the age adaptability of adolescent display content, characterized in that the display method for the age adaptability of adolescent display content includes the following steps:

[0006] S1. When a user registers and enters the platform, obtain the user's age information through birthday verification;

[0007] According to the nature of the content, pre-label the content on the platform with themes, sensitivities, risk levels, and label "education", "entertainment", "violence", "adult" tags, and subdivide into age levels of 7+, 13+, and 18+;

[0008] S2. Use machine learning algorithms to build a content filtering model; the content filtering model uses natural language processing, image recognition, and video analysis technologies to automatically identify the content, determines the displayed content according to the matching of the user's age and content tags, and marks the violent and pornographic elements in the images and videos that are not suitable for adolescents with recognition tags;

[0009] S3. Conduct multi-level content screening according to the user's age, classify the content into three types: suitable, suspicious, and unsuitable, and only display the "suitable" type of content for adolescent users.

[0010] Provide content permission setting options for parents and guardians, allowing parents to customize the content viewing permissions of adolescents;

[0011] S4. Use real-time monitoring technology to dynamically analyze the user's content interaction situation, detect whether there is an abnormal content browsing pattern including frequent attempts to view inappropriate content; dynamically adjust the content recommendation algorithm according to the user's age and usage;

[0012] S5. Introduce a user feedback and content review mechanism, allow users and parents to raise objections to the content, and if the content is repeatedly reported, it will enter the manual review process; combine manual and algorithm to adjust and optimize the filtering mechanism;

[0013] S6. Set transparent content adaptation rules on the platform, explain the adaptation criteria for different contents to users and parents, and regularly update the rules to reflect the changes in new technology content; provide a prompt message of "Why can't I see this content" to show the reasons for content adaptation.

[0014] Preferably, in the content filtering model, the construction of the content classification model includes the following steps:

[0015] S2-1. Collect and organize the content data on the platform, use the pre-trained BERT model to assign initial tags to each collected content to form training data;

[0016] S2-2. The content text generates semantic vectors through the embedding layer of the pre-trained BERT model, adds a classification layer, assigns the classified content to the corresponding age labels, and identifies text content inappropriate for teenagers.

[0017] Using the pre-trained BERT model, fine-tuning on a specific dataset, the pre-trained BERT model identifies sensitive content including violence and adult content, inputs the content text, including the title and description, and outputs the age level corresponding to the label.

[0018] S2-3. Build a CNN-based image classification model ResNet to identify potential sensitive elements in pictures, such as violence and nudity; for real-time processing scenarios, use the YOLO model to detect objects in pictures and video frames.

[0019] Input the image data into the ResNet to generate the feature vector of the image, and then assign it to the corresponding content labels of "violence", "education", and "risk-free".

[0020] S2-4. Use CNN combined with LSTM to process video sequences; use VisionTransformer to analyze risk elements and emotions in video frames; split the video sequence into frames, then extract features through the CNN, and the LSTM processes the sequence information to detect sensitive information in the video.

[0021] Extract key frames from the video, input them into the CNN model for frame-by-frame analysis, and generate classification results after aggregating frame information through LSTM to identify dynamic video content inappropriate for teenagers to watch.

[0022] Preferably, the content classification model integrates mimic data, combines text, image, and video features through a multi-modal neural network to generate a unified content understanding model; uses multi-modal Transformer and CLI to jointly analyze the text and visual features of the content.

[0023] Preferably, in the content filtering model, a multi-label classification model, random forest, processes comprehensive content labels, and the output content is assigned to multiple labels at the same time; the multi-label classification model matches the content features of text, image, video, and the appropriate age labels.

[0024] The content labels output by the multi-label classification model are matched with the user's age labels to screen out content suitable for the current user's age, and content display rules are set according to different labels and age levels.

[0025] Preferably, in the content filtering model, a collaborative filtering recommendation system generates personalized content recommendations suitable for the user's age based on matrix factorization.

[0026] Based on the user's browsing history and behavioral data, the user-based collaborative filtering algorithm and the item-based collaborative filtering algorithm are adopted to recommend content suitable for the user's age; combined with the user preference data, the high-age content that is not suitable is excluded;

[0027] Use the deep Q-network to optimize the content display strategy in real time; the reinforcement learning algorithm continuously learns based on the user's behavior feedback, dynamically adjusts the recommended content, and optimizes the content display suitable for users of different age groups; use the reinforcement learning strategy to continuously adjust the content display and recommendation strategy.

[0028] Preferably, the calculation formula of the user-based collaborative filtering algorithm is as follows:

[0029] Assume that the ratings of users u and v for multiple items i are r ui and r vi , then the cosine similarity formula is:

[0030]

[0031] where: I uv represents the set of users who have rated both user u and user v;

[0032] The rating prediction formula for user u and item i is:

[0033]

[0034] where, is the average rating of user u; U i is the set of users who have rated item i;

[0035] sim(u, v) represents the similarity between users u and v.

[0036] Preferably, the calculation formula of the item-based rating prediction formula is as follows:

[0037] Assume that the ratings of items i and j are r ui and r vj , then the similarity formula of items i and j is:

[0038]

[0039] where, U ij represents the set of users who have rated both items i and j;

[0040] The rating prediction formula for user u for item i is:

[0041]

[0042] where, is the average rating of item i;

[0043] I u The set of items rated by user u;

[0044] sim(i, j) represents the similarity between item i and item j.

[0045] Preferably, the establishment of real-time monitoring includes the following steps:

[0046] S4-1-1. The anomaly detection model identifies whether there are abnormal behaviors such as teenagers trying to frequently access inappropriate content. Based on the K-means and DBSCAN clustering algorithms, the content interaction behaviors of users are divided into different behavior clusters, and the behaviors that conform to the "normal" mode and the abnormal modes deviating from the normal are identified;

[0047] S4-1-2. Set multiple anomaly criteria, such as frequently clicking on content with high age restrictions and repeatedly trying to access high-risk content within a short period of time;

[0048] S4-1-3. Use the autoencoder algorithm to construct a real-time anomaly detection model to automatically identify and label abnormal behavior patterns; classify abnormal behaviors according to the risk level, such as minor anomalies, medium risks, and high risks, to facilitate subsequent adoption of different recommendation intervention measures.

[0049] Preferably, when the real-time anomaly detection model detects an abnormal behavior, the system adjusts the content recommendation algorithm in real time to control the exposure of inappropriate content, including the following steps:

[0050] S4-2-1. Use the deep Q-network (DQN) reinforcement learning algorithm to achieve dynamic weight adjustment of the recommendation algorithm; for users who frequently try to watch content for older ages, the system automatically reduces the weight of inappropriate content tags and increases the weight of low-risk and high-educational-value content;

[0051] S4-2-2. Introduce a feedback mechanism to adjust the recommended content in real time according to user behavior data; for content that does not match the age, a "cooling" mechanism can be set so that the age-inappropriate content does not appear in the recommendation list for a period of time; for users who repeatedly try to watch content for older ages, the system reduces the appearance frequency of the age-inappropriate content in the recommendation and temporarily restricts the recommendation of high-risk content;

[0052] S4-2-3. For the detected abnormal behaviors, the system triggers recommendation intervention measures to dynamically adjust the content recommendation frequency; for minor abnormal behaviors, reduce the recommendation of such content; for high-frequency abnormal behaviors, suspend the recommendation and replace the content.

[0053] S4-2-4. Dynamically adjust the recommendation rules based on the user's age and behavioral risk level; for low-risk users, recommend diverse content; while for high-risk users, the recommendation is limited to educational, safe, and educational content

[0054] Compared with the prior art, the advantages of the present invention are as follows:

[0055] (1) By monitoring user interactions in real time, the system can dynamically adjust content recommendations, avoiding the limitations of relying solely on static rules and improving adaptability and timeliness.

[0056] (2) The new method manages content classification based on the user's specific behavior, reducing blind intervention and providing corresponding recommendation solutions for different risk levels, thereby improving the accuracy of content display.

[0057] (3) When abnormal behavior is detected, the system can flexibly switch the filtering level or restrict the display of high-risk content, balancing content security and user experience.

[0058] (4) Through the reinforcement learning mechanism, the system continuously optimizes the recommendation effect based on user feedback, adapts to changes in user needs, and avoids the rigidity problem of traditional filtering.

[0059] (5) Parents can monitor the user's behavior in real time and adjust the content filtering intensity, enhancing the transparency and interactivity of content management.

[0060] (6) By using contrastive learning and multi-modal analysis of text, images, and videos, the content recognition accuracy is improved, misjudgments of single modalities are solved, and the judgment reliability of the system is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the overall process of a method for age adaptation of content displayed to teenagers according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0062] For the embodiments, please refer to Figure 1 , a method for age adaptation of content displayed to teenagers, a method for age adaptation of content displayed to teenagers includes the following steps:

[0063] S1. When the user registers and enters the platform, obtain the user's age information through birthday verification;

[0064] According to the nature of the content, pre-label the content on the platform with themes, sensitivities, and risk levels, mark the labels of "education", "entertainment", "violence", "adult", and subdivide the age levels of 7+, 13+, and 18+;

[0065] S2. Build a content filtering model using machine learning algorithms; the content filtering model automatically identifies content using natural language processing, image recognition, and video analysis technologies, determines the displayed content based on the matching of the user's age and content tags, and marks the violent and pornographic elements inappropriate for teenagers in images and videos with identification tags;

[0066] S3. Conduct multi-level content screening according to the user's age, classify the content into three types: suitable, suspicious, and inappropriate, and only display the "suitable" type of content for teenage users.

[0067] Provide content permission settings options for parents and guardians, allowing parents to customize the content viewing permissions of teenagers;

[0068] S4. Use real-time monitoring technology to dynamically analyze the user's content interaction situation, detect abnormal content browsing patterns including frequent attempts to view inappropriate content; dynamically adjust the content recommendation algorithm according to the user's age and usage situation;

[0069] Specifically, the system also introduces an adaptive learning and feedback mechanism to help the system adapt in a timely manner when the user's behavior changes, thereby dynamically optimizing the recommendation effect, mainly including the following content:

[0070] Online learning system: Enable the recommendation model to continuously learn and adapt to user changes through a reinforcement learning framework. For example, when the user's abnormal behavior decreases, resume the recommendation of diverse content suitable for the age group.

[0071] User feedback processing: Record the user's direct feedback on the recommended content (such as likes, reports), and use it as model training data to improve future recommendation accuracy.

[0072] A / B testing and algorithm optimization: Conduct A / B testing to evaluate the impact of dynamic adjustment strategies on user interaction effects, verify the effectiveness of model optimization, and avoid over-adjustment or insufficient intervention of the recommendation system.

[0073] S5. Introduce a user feedback and content review mechanism, allow users and parents to raise objections to content, and if the content is repeatedly reported, it will enter the manual review process; combine manual and algorithm adjustments to optimize the filtering mechanism;

[0074] S6. Set transparent content adaptation rules on the platform, explain the adaptation criteria for different content to users and parents, and regularly update the rules to reflect changes in new technology content; provide a prompt message of "Why can't I see this content" to show the reasons for content adaptation.

[0075] Specifically, real-time notification: For high-risk behaviors, push notifications through the platform to prompt parents or guardians. Parents can set customized notification frequencies to keep abreast of their children's browsing behaviors.

[0076] Parental control settings: Provide management options for content filtering and recommendation permissions on the parent side, allowing parents to regulate the type, frequency, and sensitivity of recommended content to better meet the growth needs of children.

[0077] Regular behavior reports: Generate content browsing and interaction reports for users for parents' reference, including detailed data such as content categories, time distribution, and abnormal behavior records.

[0078] In the content filtering model, the construction of the content classification model includes the following steps:

[0079] S2-1. Collect and organize the content data on the platform, and use the pre-trained BERT model to assign initial labels to each piece of collected content to form training data;

[0080] Specifically, the steps for the pre-trained BERT model to assign initial labels to the collected content include:

[0081] 1. Text input: First, convert the collected content into an input format that BERT can accept. This includes tokenizing the text, adding special tokens (such as [CLS] and [SEP]), and encoding words as IDs according to the BERT vocabulary.

[0082] Input vector generation: The ID vector of each word will be further transformed into an embedding vector. BERT combines the input word embedding vectors, position embeddings, and segment embeddings (such as distinguishing paragraphs or sentences) to form a complete input representation.

[0083] 2. Multi-layer bidirectional self-attention: The BERT model contains multi-layer bidirectional self-attention mechanisms, which allow the model to consider the semantic relationships of words in the context when encoding words. Each layer will generate a deeper semantic representation of the input.

[0084] Semantic representation of sentences / content: Usually, the [CLS] token vector of BERT is regarded as the semantic representation vector of the entire content for downstream classification tasks. This semantic vector represents the overall characteristics of the content.

[0085] 3. Adding a classification layer: Add one or more fully connected layers after the [CLS] vector as the classification layer, and output the probability distribution of the corresponding labels. The output dimension of the classification layer is equal to the number of labels (such as adolescent content labels, adult content labels, etc.), and the Softmax function is used to convert the scores into probabilities.

[0086] Fine-tuning of the pre-trained model: To adapt to specific content labeling tasks, BERT is usually fine-tuned on a specific label dataset. After fine-tuning, the model has the ability to assign initial labels to new content.

[0087] 4. Most Probable Label: For each content input, the output of the classification layer generates a probability distribution over multiple labels, and the label with the highest probability is selected as the initial label for that content.

[0088] Forming Training Data: These initial labels, together with the content itself, constitute the training dataset, which can be further used for supervised learning or as input to subsequent models (such as more complex classifiers).

[0089] S2-2. The content text generates semantic vectors through the embedding layer of the pre-trained BERT model, and is fed into the classification layer. The classified content is assigned to the corresponding age label to identify text content inappropriate for teenagers.

[0090] S2-3. Build a CNN-based image classification model ResNet to identify potential sensitive elements in images, such as violence and nudity. For real-time processing scenarios, use the YOLO model to detect objects in images and video frames.

[0091] Specifically, building a ResNet image classification model based on convolutional neural network (CNN) generally includes the following main steps:

[0092] Step 1: Use a deep learning framework such as TensorFlow or PyTorch.

[0093] Step 2: The core of the ResNet model lies in the "residual connection", which can effectively solve the vanishing gradient problem in deep networks. There are two common residual block designs:

[0094] BasicBlock: Used in shallow ResNet models (such as ResNet-18, ResNet-34), with a relatively simple structure.

[0095] BottleneckBlock: Used in deep ResNet models (such as ResNet-50, ResNet-101, ResNet-152), with a more complex structure.

[0096] Step 3: ResNet is constructed by stacking multiple residual blocks. The number of residual blocks in each layer depends on the specific model. The number of channels in each layer usually doubles layer by layer (such as 64, 128, 256, 512), and spatial downsampling is achieved through convolutional operations with a stride of 2.

[0097] Input Layer: The input layer is usually an image of 224x224x3.

[0098] Initial Convolutional Layer: Add a large convolutional layer (such as 7x7 convolution with a stride of 2) and a pooling layer to reduce the computational amount.

[0099] Residual layer stacking: After the initial convolutional layer, stack multiple layers composed of residual blocks according to the configuration. For example, the configuration of ResNet-18 is: [2, 2, 2, 2], that is, stack 2 BasicBlocks in each layer.

[0100] Global average pooling and output layer: Convert the feature map into a single vector through the global average pooling layer, and then connect a fully connected layer to output the final classification prediction.

[0101] Step 5: Configure an optimizer (such as Adam or SGD), a loss function (such as categorical_crossentropy), and evaluation metrics (such as accuracy) to compile the model.

[0102] Step 6: Use a dataset (such as CIFAR-10 or ImageNet) to train the model. Set training parameters such as batch size and number of training epochs, and preprocess the data to meet the input requirements of the model.

[0103] Step 7: Evaluate the model performance on the test data, calculate evaluation metrics such as classification accuracy, to measure the performance of the model in the actual scenario.

[0104] Use CNN combined with LSTM to process video sequences; use VisionTransformer to analyze risk elements and emotions in video frames; split the video sequence into frames, then extract features through CNN, and LSTM processes sequence information to detect sensitive information in the video;

[0105] The content classification model integrates mimetic data, combines text, image, and video features through a multi-modal neural network to generate a unified content understanding model; uses multi-modal Transformer and CLI to jointly analyze the text and visual features of the content.

[0106] Specifically, text feature extraction: Use a pre-trained BERT or RoBERTa model to convert text into fixed-length feature vectors. These feature vectors can well represent the semantic information in the text.

[0107] Image feature extraction: Use pre-trained ResNet, EfficientNet, or VisionTransformer (ViT) to encode image data and generate high-dimensional feature vectors of the image.

[0108] Video feature extraction: Use a pre-trained 3D convolutional network (such as C3D, I3D) or a Transformer-based video encoding model to extract the temporal features of video frames and form the temporal feature vectors of the video.

[0109] 1. Multimodal Feature Alignment: Since the feature dimensions of text, images, and videos are different, linear projection or fully connected layers can be used to map the features into a unified dimensional space for subsequent fusion and calculation.

[0110] 2. Multi-Head Attention Mechanism: Take the features of each modality (text, image, video) as input, and achieve feature interaction between different modalities through the multi-head self-attention mechanism of the multimodal Transformer. The Transformer will weigh the features of each modality through the attention allocation mechanism to ensure that the information of each modality is effectively utilized.

[0111] Feature Fusion: The encoder structure of the Transformer will generate fused multimodal feature representations, which contain the correlation information between text, images, and videos, thereby improving the content understanding ability of the model.

[0112] Hierarchical Feature Integration: Features can be gradually fused in multiple Transformer layers, enabling the model to gradually deepen its understanding of multimodal data. In this way, the model can obtain basic information in the initial layer and capture complex cross-modal relationships in the higher layers.

[0113] 3. Feature Alignment and Joint Embedding: The key to CLI lies in the learning of the joint embedding space. Map the features of all modalities into the same embedding space for cross-modal alignment. Through sharing parameters or alternating training, the features of different modalities can be correlated with each other in this joint space.

[0114] Mutual Information Maximization: Adopt mutual information maximization technology to improve the mutual information between text features and visual features. This can better capture the correlation between text descriptions and image or video content.

[0115] Contrastive Learning: Introduce cross-modal contrastive learning to enhance the matching degree between different modality features. For example, pull closer the relevant text-image pairs or text-video pairs, while pull farther the non-relevant pairs. This can optimize the performance of the model in the joint embedding space.

[0116] 4. Unified Feature Representation: The feature representation fused by the Transformer and CLI will contain the overall information of text, images, and videos, generating a multimodal joint feature vector. This vector can be used as the input of the content understanding model, covering the text semantics and visual features of the content.

[0117] Semantic Reasoning: Using the generated multimodal joint features, more accurate semantic reasoning can be performed on the content. For example, the model can understand the theme of a video, the core content of an image, the emotional expression of text, etc., to achieve a comprehensive understanding of the content.

[0118] 5. Pre-training: Conduct pre-training on a large amount of labeled multi-modal data to enable the model to learn the basic cross-modal correlations.

[0119] Fine-tuning: Conduct fine-tuning on specific tasks (such as content classification, recommendation systems, etc.) to further adapt to the requirements of specific applications.

[0120] Loss function: Multi-modal contrast loss, cross-entropy loss, mutual information loss, etc. can be used to jointly optimize the model to improve the matching ability between text and visual modalities.

[0121] Attention visualization: By visualizing the attention weights of the Transformer, the relationships between the multi-modal features that the model focuses on can be understood, thereby helping to optimize the model structure and parameters.

[0122] In the content filtering model, the multi-label classification model random forest processes comprehensive content labels, and the output content is simultaneously assigned to multiple labels; the multi-label classification model matches content feature texts, images, videos, and adapted age labels;

[0123] The content labels output by the multi-label classification model are matched with the user age labels to filter out content suitable for the current user's age, and content display rules are set according to different labels and age levels.

[0124] In the content filtering model, the collaborative filtering recommendation system generates personalized content recommendations suitable for the user's age based on matrix factorization;

[0125] Based on the user's browsing history and behavior data, user-based collaborative filtering algorithms and item-based collaborative filtering algorithms are used to recommend content that matches the user's age; combined with the user's preference data, inappropriate high-age content is excluded;

[0126] Use the deep Q-network to optimize the content display strategy in real time; the reinforcement learning algorithm continuously learns based on the user's behavior feedback, dynamically adjusts the recommended content, and optimizes the content display suitable for users of different age groups; use the reinforcement learning strategy to continuously adjust the content display and recommendation strategies.

[0127] Specifically, the process of using the deep Q-network (DQN) to optimize the content display strategy in real time can be divided into the following main steps:

[0128] 1. State: In the scenario of content recommendation, the state usually includes relevant information of the user and the current content display situation. Specifically, it can include information such as the user's age, interest preferences, recent browsing history, viewing time, etc.

[0129] Content features: The labels, themes, popularity, etc. of the displayed content can be added to more comprehensively characterize the state.

[0130] 2. Action: An action is the content or strategy that the system can choose to present to the user. For example, presenting a set of content, presenting new recommended content, skipping the current content, etc.

[0131] The action space should include all categories and forms of recommendable content (such as videos, texts, pictures, etc.) so that DQN can make selections among different contents.

[0132] 3. Reward: The reward function is the key to measuring the effect of recommended actions and can be designed based on the user's interaction behavior. For example:

[0133] If the user clicks on or views the recommended content, a positive reward can be given.

[0134] If the user skips or unfollows, a negative reward is given.

[0135] For teenage users, the system can also adjust the reward according to the suitability of the content. For example, content with a high suitability has a higher positive reward.

[0136] The design of the reward should encourage the system to recommend content with high suitability and user interest, and prevent the recommendation of inappropriate content as much as possible.

[0137] 4. DQN is a deep neural network that takes the current state as input and outputs the Q-values of each action in that state. A multi-layer fully connected neural network can be used, or when dealing with large-scale content features, a convolutional neural network (CNN) can be combined to extract content features.

[0138] Network output: The output of the network is a set of Q-values of actions, representing the long-term reward expectations of performing each action in the current state.

[0139] 5. Training the DQN model

[0140] Experience replay: To improve the training stability of DQN, experience replay is used to store the user's interaction records (i.e., state, action, reward, next state). The system randomly samples from the replay pool during each training.

[0141] Q-learning objective: By calculating the TD error (temporal difference error), the Q-value update is optimized.

[0142] The calculation formula of the user-based collaborative filtering algorithm is as follows:

[0143] Assume that the ratings of user u and v for multiple items i are r ui and r vi , then the cosine similarity formula is:

[0144]

[0145] Where: Iuv Denote the set of users who have rated both user u and user v;

[0146] The prediction formula for the rating of user u and item i is:

[0147]

[0148] where, is the average rating of user u; U i is the set of users who have rated item i;

[0149] sim(u, v) represents the similarity between user u and v.

[0150] The calculation formula of the above-mentioned item-based rating prediction formula is as follows:

[0151] Assume that the ratings of items i and j are r ui and r vj , then the similarity formula of items i and j is:

[0152]

[0153] where, U ij represents the set of users who have rated both items i and j;

[0154] The prediction formula for user u's rating of item i is:

[0155]

[0156] where, is the average rating of item i;

[0157] I u is the set of items rated by user u;

[0158] sim(i, j) represents the similarity between items i and j.

[0159] The establishment of real-time monitoring includes the following steps:

[0160] S4-1-1. The anomaly detection model identifies whether there are abnormal behaviors such as teenagers trying to frequently access inappropriate content. Based on the K-means and DBSCAN clustering algorithms, the content interaction behaviors of users are divided into different behavior clusters, and the behaviors that conform to the "normal" mode and the abnormal modes that deviate from the normal are identified;

[0161] S4-1-2. Set multiple anomaly criteria, such as frequently clicking on content with high age restrictions and repeatedly trying to access high-risk content within a short period of time;

[0162] S4-1-3. Build a real-time anomaly detection model using the autoencoder algorithm to automatically identify and label abnormal behavior patterns; classify abnormal behaviors according to risk levels, such as minor anomalies, medium risks, and high risks, to facilitate subsequent adoption of different recommended intervention measures.

[0163] Specifically, the process of building a real-time anomaly detection model using an autoencoder and classifying abnormal behaviors according to risk levels is as follows:

[0164] 1. Feature selection: Select important features for real-time detection, which should be able to reflect the risks of abnormal behaviors. For example, the user's access frequency, stay time, click-through rate, content type preference, etc.

[0165] Sliding window: To capture the features of time series, the sliding window technique is commonly used, that is, each window contains data of multiple time steps, so as to model the short-term and long-term features in the real-time data stream.

[0166] 2. Model architecture: The autoencoder consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional latent space representation, and the decoder restores it to the original dimension.

[0167] Training objective: Train the autoencoder to make the input and the reconstructed output as close as possible. In this way, the model can learn normal behavior patterns. When new data is input into the model, the model is difficult to reconstruct abnormal data and generates a large reconstruction error.

[0168] Reconstruction error: Calculate the difference between the input data and the reconstructed output (such as mean squared error) as the basis for judging anomalies. The reconstruction error of normal behavior should be less than the set threshold, and abnormal behavior will have a higher reconstruction error.

[0169] 3. Training data: Use historical data of normal behaviors to train the autoencoder model to ensure that the features learned by the model mainly come from normal data.

[0170] Training process: Adjust the model parameters (such as the number of layers and neurons of the encoder and decoder), and perform multiple rounds of training on the training data until the model can reconstruct normal data well.

[0171] Select threshold: Based on the reconstruction error distribution of the training data, set an appropriate anomaly threshold to distinguish normal behaviors from abnormal behaviors.

[0172] 4. Error range division: Divide the range of reconstruction errors according to risk levels, for example:

[0173] Minor anomaly: The reconstruction error is slightly higher than the threshold but does not reach the level of obvious anomaly.

[0174] Medium risk: The reconstruction error is relatively large, indicating that the data deviates from normal behavior and there are potential risks.

[0175] High risk: The reconstruction error is extremely large, indicating a very high probability of abnormal behavior and requires priority attention.

[0176] Dynamic adjustment: Dynamically adjust the threshold range according to the distribution or feedback of real-time data to adapt to changing behavior patterns.

[0177] 5. Data preprocessing: Perform necessary cleaning and transformation on the real-time data stream to ensure it conforms to the input format of the autoencoder.

[0178] Real-time reconstruction error calculation: Input new data into the autoencoder, calculate the reconstruction error and compare it with the set threshold to determine in real-time whether it is abnormal.

[0179] Risk level annotation: Automatically add mild, medium or high risk labels to the detected abnormal behaviors according to the risk interval to which the reconstruction error belongs.

[0180] 6. Multi-level alarm mechanism: Different risk levels correspond to different levels of alarm measures. For example:

[0181] Mild anomaly: Record and store it as a log without immediately triggering an alarm.

[0182] Medium risk: Send an alarm to the monitoring personnel to prompt potential risks.

[0183] High risk: Immediately trigger a high-level alarm mechanism, such as locking the account, restricting access or sending an emergency notice.

[0184] Automated response: Automatic response rules can be set to take corresponding measures according to different risk levels.

[0185] When the real-time anomaly detection model detects abnormal behavior, the system adjusts the content recommendation algorithm in real-time to control the exposure of inappropriate content, including the following steps:

[0186] S4-2-1. Use the deep Q-network (DQN) reinforcement learning algorithm to achieve dynamic weight adjustment of the recommendation algorithm; for users who frequently attempt to view high-age content, the system automatically reduces the weight of inappropriate content tags and increases the weight of low-risk and high-educational-value content;

[0187] Specifically, implementing the DQN reinforcement learning algorithm to dynamically adjust the content weights in the recommendation system can help the system adapt to changes in user behavior, especially when there are frequent attempts to view high-age content. The specific implementation process is as follows:

[0188] 1. State: Define the user's state to include the basic attributes of user behavior and content, such as user age, viewing frequency, recent viewing preferences, content tags (such as education, entertainment), and content suitability (such as for low age, high age).

[0189] Action: In this scenario, the action is to dynamically adjust the weights of recommended content tags. The adjustment range of weights for different tags can be set, and the action set can include:

[0190] Increase the weight of low-risk content;

[0191] Increase the weight of high-education-value content;

[0192] Decrease the weight of unsuitable high-age content.

[0193] Reward: The design of the reward function should reflect the recommendation effect and adaptability:

[0194] If the user selects the recommended low-risk, high-education content, give a relatively high positive reward;

[0195] If the user attempts to view unsuitable high-age content, give a negative reward to encourage the system to reduce the recommendation of such content;

[0196] At the same time, according to the user's long-term adaptability and viewing behavior changes, accumulate rewards to ensure the optimality of the long-term strategy.

[0197] 2. Model Structure: The DQN network takes the user's state as input and outputs the Q value for each action. A multi-layer neural network (such as a fully connected layer or an LSTM layer) can be used to capture the complex relationship between user behavior patterns and content preferences.

[0198] Q-value Calculation: Estimate the Q value for each action through the DQN network, and select the action with the largest Q value to guide the adjustment of the weights of recommended content.

[0199] 3. Experience Replay: Store the user's interaction data (including state, action, reward, and next state) in the experience replay pool, and randomly sample from it to train the model to reduce the correlation between data.

[0200] 4. State Update: Update the current state according to the user's behavior and recent recommendation results.

[0201] Action Selection and Weight Adjustment: Select an action according to the Q value output by the DQN, that is, adjust the tag weights. If the user frequently attempts to view high-age content, the DQN will select to reduce the weights of the tags of unsuitable content and increase the weights of low-risk and high-education content.

[0202] Feedback and Reinforcement Learning: Record the user's feedback on the recommended content (such as click, skip, etc.) in real time, and update the experience replay pool according to the feedback for subsequent training.

[0203] S4-2-2. Introduce a feedback mechanism to adjust the recommended content in real time according to user behavior data; for content that does not meet the age requirements, a "cooling" mechanism can be set so that content that is not suitable for the age will not appear in the recommended list for a period of time; for users who repeatedly attempt to view content for the elderly, the system reduces the frequency of appearance of content that is not suitable for their age in the recommendations and temporarily restricts the recommendation of high-risk content.

[0204] S4-2-3. In response to detected abnormal behaviors, the system triggers recommendation intervention measures to dynamically adjust the content recommendation frequency; for minor abnormal behaviors, the recommendation of such content is reduced; for high-frequency abnormal behaviors, the recommendation is suspended and the content is replaced.

[0205] S4-2-4. Dynamically adjust the recommendation rules based on the user's age and behavioral risk level; for low-risk users, diverse content is recommended; while for high-risk users, the recommendation is limited to educational, safe, and educational content.

[0206] The above shows and describes the basic principles, main features, and advantages of the present invention; those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed; the scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for displaying age suitability of content for teenagers, characterized in that: The method for displaying age suitability of content for teenagers comprises the following steps: S1. When a user registers and enters the platform, the user's age information is obtained through birthday verification; Based on the nature of the content, the content on the platform is pre-labeled with themes, sensitivity, and risk levels, and labeled as "educational", "entertainment", "violence", and "adult", and subdivided into age levels of 7+, 13+, and 18+; S2. Use machine learning algorithms to build a content filtering model; the content filtering model uses natural language processing, image recognition and video analysis technology to automatically identify content, determine the content to be displayed based on the user's age and content label matching, and identify violent and pornographic elements in images and videos that are not suitable for teenagers; S3. Perform multi-level content screening based on user age, classifying content into three categories: suitable, suspicious, and inappropriate. For young users, only "suitable" content is displayed; Provide parents and guardians with content permission settings, allowing parents to customize their teenagers' content viewing permissions; S4. Use real-time monitoring technology to dynamically analyze users’ content interactions and detect abnormal content viewing patterns, including frequent attempts to view inappropriate content; dynamically adjust content recommendation algorithms based on user age and usage; S5. Introduce user feedback and content review mechanisms, allowing users and parents to raise objections to content. If content is repeatedly reported, it will enter the manual review process; combine manual and algorithmic adjustments to optimize the filtering mechanism; S6. Set up transparent content adaptation rules on the platform, explain the adaptation standards of different content to users and parents, and regularly update the rules to reflect changes in new technical content; provide prompt information such as "Why can't I see this content" to show the reasons for content adaptation.

2. According to claim 1, a method for displaying age suitability of content for teenagers, characterized in that: In the content filtering model, the content classification model construction includes the following steps: S2-1. Collect and organize content data on the platform, and use the pre-trained BERT model to assign initial labels to each collected content to form training data; S2-2. The content text generates a semantic vector through the pre-trained BERT model embedding layer, adds a classification layer, assigns the classified content to the corresponding age label, and identifies text content that is not suitable for teenagers; S2-3. Build a CNN-based image classification model ResNet to identify potential sensitive elements such as violence and nudity in images; for real-time processing scenarios, use the YOLO model to detect objects in images and video frames; S2-4. Use CNN combined with LSTM to process video sequences; use VisionTransformer to analyze risk elements and emotions in video frames; split the video sequence into frames, and then use the CNN to extract features, and the LSTM to process sequence information to detect sensitive information in the video.

3. According to claim 2, a method for displaying age suitability of content for teenagers, characterized in that: The content classification model integrates mimetic data, combines text, image and video features through a multimodal neural network, and generates a unified content understanding model; it uses a multimodal Transformer and CLI to jointly analyze the text and visual features of the content.

4. According to claim 1, a method for displaying age suitability of content for teenagers, characterized in that: In the content filtering model, comprehensive content labels are processed by a multi-label classification model random forest, and the output content is simultaneously assigned to multiple labels; the multi-label classification model matches content feature text, images, videos and adapted age labels; The content labels output by the multi-label classification model are matched with the user age labels to filter out content suitable for the current user age, and content display rules are set according to different labels and age levels.

5. According to claim 1, a method for displaying age suitability of content for teenagers, characterized in that: In the content filtering model, the collaborative filtering recommendation system generates personalized content recommendations suitable for the user's age based on matrix decomposition; Based on the user's browsing history and behavior data, the collaborative filtering algorithm based on users and the collaborative filtering algorithm based on items are used to recommend content that matches the user's age. Combine user preference data to exclude inappropriate content for older people; Use deep Q-network to optimize content display strategy in real time; The reinforcement learning algorithm continuously learns based on user behavior feedback, dynamically adjusts recommended content, and optimizes content display for users of different age groups; Use reinforcement learning strategies to continuously adjust content display and recommendation strategies.

6. A method for displaying age suitability of content for teenagers according to claim 5, characterized in that: The calculation formula of the user-based collaborative filtering algorithm is as follows: Assume that users u and v give r scores to multiple items i ui and r vi , then the cosine similarity formula is: Where: I uv represents the set of users rated by both user u and user v; The rating prediction formula for user u and item i is: in, The average rating of user u; U i is the set of users who have rated item i; sim(u, v) represents the similarity between users u and v.

7. A method for displaying age suitability of content for teenagers according to claim 5, characterized in that: The calculation formula of the item-based rating prediction formula is as follows: Assume that the ratings of items i and j are r ui and r vj , then the similarity formula between items i and j is: Among them, U ij represents the set of users who have rated both items i and j; The formula for predicting the rating of user u for item i is: in, is the average rating of item i; I u The set of items rated by user u; sim(i, j) represents the similarity between items i and j.

8. A method for displaying age suitability of content for teenagers according to claim 1, characterized in that: The establishment of real-time monitoring includes the following steps: S4-1-1. The anomaly detection model identifies whether adolescent users have abnormal behaviors such as frequent attempts to access inappropriate content. Based on the K-means and DBSCAN clustering algorithms, the user's content interaction behaviors are divided into different behavior clusters, and behaviors that conform to the "normal" pattern and abnormal patterns that deviate from the normal pattern are identified; S4-1-2. Set multiple abnormal standards, frequently click on high-age restricted content, and try to access high-risk content multiple times in a short period of time; S4-1-3. Use the autoencoder algorithm to build a real-time anomaly detection model to automatically identify and mark abnormal behavior patterns; classify abnormal behaviors according to risk levels, including minor anomalies, medium risks, and high risks.

9. A method for displaying age suitability of content for teenagers according to claim 8, characterized in that: The real-time anomaly detection model detects abnormal behavior, and the system adjusts the content recommendation algorithm in real time to control the exposure of inappropriate content, including the following steps: S4-2-1. Use the reinforcement learning algorithm DQN to implement dynamic weight adjustment of the recommendation algorithm; for users who frequently try to watch old-age content, the system automatically reduces the weight of inappropriate content labels and increases the weight of low-risk and high-educational-value content; S4-2-2. Introduce a feedback mechanism to adjust recommended content in real time based on user behavior data; set a "cooling-off" mechanism for age-inappropriate content so that the age-inappropriate content no longer appears in the recommendation list for a period of time; for users who have repeatedly tried to watch age-inappropriate content, the system will reduce the frequency of age-inappropriate content in the recommendation and temporarily limit the recommendation of high-risk content; S4-2-3. For detected abnormal behaviors, the system triggers recommended intervention measures and dynamically adjusts the frequency of content recommendations; for minor abnormal behaviors, the recommendation of such content is reduced; for high-frequency abnormal behaviors, the recommendation is suspended and the content is replaced; S4-2-4. Dynamically adjust recommendation rules based on user age and behavior risk level; recommend diversified content to low-risk users; and for high-risk users, recommendations are limited to educational, safety, and content.

Images