Intelligent recommendation method and system based on children's digital picture book reading

Abstract

The embodiment of the application provides a smart recommendation method and system based on children's digital picture book reading, which comprises the following steps: constructing a recommendation model based on self-supervised learning and a convolutional neural network (CNN); in the self-supervised learning module, introducing a contrast mechanism combined with user cognition to optimize the feature representation of the model; constructing a dynamically updated user portrait, which contains the current cognitive level label of the user; in response to the reading request of the user, adaptively fusing the user portrait and the reading request, and introducing a picture book age-appropriateness evaluation factor in the fusion process, wherein the age-appropriateness evaluation factor is determined based on the matching degree of the text difficulty, picture complexity of the picture book and the user cognitive level, and taking the fused label vector as the input of the trained recommendation model to obtain a recommendation list. According to the embodiment of the application, the model feature representation is dynamically optimized by combining the user cognitive development stage, the reading ability difference of different age groups is accurately adapted, and the recommendation accuracy and adaptability are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to an intelligent recommendation method and system based on children's digital picture book reading. Background Technology

[0002] With the rapid development of digital technology, children's reading habits are also changing. Digital picture books, as an emerging reading medium, can stimulate children's reading potential and cultivate reading habits with their rich multimedia elements and interactive functions.

[0003] However, faced with a vast amount of digital picture book resources, children and parents often find themselves in a dilemma of choice, struggling to accurately identify high-quality reading materials that match their children's cognitive level, interests, and educational goals, resulting in poor reading outcomes.

[0004] Currently, most children's digital reading platforms use recommendation algorithms based on content similarity or collaborative filtering, generating recommendation lists by analyzing users' historical reading records or the preferences of similar users. With the application of deep learning technology, some platforms have introduced neural network models to process multimodal features such as text and images, attempting to improve recommendation accuracy.

[0005] However, existing recommendation technologies do not deeply integrate the characteristics of children's cognitive stages, making it difficult to dynamically reflect the differences in reading ability and comprehension needs at different developmental stages. Secondly, the feature processing of picture books is not optimized in stages according to children's cognitive characteristics, resulting in redundant features for younger children or insufficient features for older children. The recommendation results are insufficient in terms of cognitive adaptability and cross-stage interest continuity, and the accuracy and age-appropriateness need to be improved. Summary of the Invention

[0006] The purpose of this application is to provide an intelligent recommendation method and system based on children's digital picture book reading, so as to at least solve the technical problem that the recommendation results of traditional recommendation methods are insufficient in terms of cognitive adaptability and cross-stage interest continuity.

[0007] To achieve the above objectives, the embodiments of this application provide the following technical solutions.

[0008] In a first aspect, according to one embodiment of this application, an intelligent recommendation method based on children's digital picture book reading is provided;

[0009] Includes the following steps:

[0010] A recommendation model based on self-supervised learning and convolutional neural networks (CNN) is constructed. In the self-supervised learning, a comparison mechanism combining user cognition is introduced to optimize the feature representation of the model, and dynamic weights are introduced to calculate the similarity of positive and negative sample pairs to take into account the differences in users' reading abilities.

[0011] The system acquires the user's historical reading information, clusters the preprocessed historical reading information, and constructs a dynamically updated user profile based on the clustering results. The user profile includes the user's current cognitive tags.

[0012] In response to the user's reading request, the user profile and reading request are adaptively fused. During the fusion process, the age-appropriateness assessment factor of the picture book is introduced. The age-appropriateness assessment factor is determined based on the matching degree between the text difficulty and picture complexity of the picture book and the user's cognitive level. The fused tag vector is used as the input of the trained recommendation model to obtain the recommendation list.

[0013] When negative feedback behavior from a user towards the recommended list is detected, the user profile is enhanced, triggering a secondary screening of the picture book feature library to remove candidate picture books that do not match the user's current cognitive level. The weight vector of the adaptive fusion processing is adjusted, and the recommended list is updated based on the adjusted tag vector.

[0014] Preferably, the step of introducing a comparison mechanism that incorporates user cognition to optimize the feature representation of the model includes:

[0015] Construct positive and negative sample pairs of user features and picture book features;

[0016] For user characteristics, similar time-series behavioral data of users in the same age group and related behavioral data of the same user at different growth stages are used as positive sample pairs, while heterogeneous behavioral data of users in different age groups and unrelated cross-stage behavioral data of the same user are used as negative sample pairs.

[0017] For picture book features, picture books that fit the same cognitive stage and have a theme relevance higher than the threshold are considered as positive sample pairs, while picture books that fit different cognitive stages or have a theme relevance lower than the threshold are considered as negative sample pairs.

[0018] A dual-path momentum encoder and a query encoder are constructed. The momentum encoder includes a basic parameter layer and a cognitive stage adaptation layer. The parameters of the momentum encoder are dynamically obtained from the query encoder through a phased momentum update strategy. A negative sample queue is introduced, and the corresponding negative sample features are stored in partitions according to the cognitive stage. By comparing the correlation between positive and negative samples in the two dimensions of cognitive adaptation and feature similarity, the model learns a representation that has both stage adaptability and feature discriminability.

[0019] Based on the cognitive development weight matrix, the optimization weights for maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs are dynamically adjusted. An improved InfoNCE contrastive loss function with a fusion of cognitive stage penalty terms is used to optimize the feature representation of the model. The cognitive stage penalty terms are positively correlated with the degree of difference between the cognitive stages to which the sample pairs belong.

[0020] Preferably, the process of training the recommendation model includes the following steps:

[0021] Obtain authorized user age-based reading history data and picture book age classification information;

[0022] Initialize the parameters of the recommendation model, including the basic convolutional kernel parameters of the CNN feature extraction module, the initial weights of the dual-path momentum encoder and age-based query encoder in the self-supervised learning module, and the classification parameters of the fully connected layer;

[0023] The fusion vector of user profiles containing cognitive tags and reading requests is input into the model and forward propagated to obtain the predicted age-appropriate recommendation list. The behavioral loss value is calculated based on the matching degree between the prediction results and the user's actual reading behavior. At the same time, the improved InfoNCE contrastive loss function, which incorporates the cognitive stage penalty term in the self-supervised learning module, is combined with the dynamic allocation of the weight ratio of behavioral loss and contrastive loss according to the cognitive stage, and the total loss is calculated in a comprehensive manner.

[0024] The total loss is backpropagated to the model, and the model parameters are updated through a phased Adam optimization algorithm to minimize the total loss. The parameter update step size for different cognitive stages is dynamically adjusted according to the user's cognitive development rate. During the update process, the basic parameter layer of the momentum encoder is updated synchronously from the query encoder with a fixed momentum coefficient, and the parameters of the cognitive stage adaptation layer are dynamically adjusted based on the correlation between the current training sample and the cognitive stage, so as to realize phased momentum update.

[0025] The recommendation accuracy and age-appropriate matching of the model at each stage are evaluated on the age-specific validation set. Based on the evaluation results at each stage, the hyperparameters and model structure of the corresponding stage are adjusted in a differentiated manner to complete the staged training and optimization of the recommendation model.

[0026] Preferably, the step of constructing a user profile based on the clustering results includes:

[0027] We construct an age-based interaction matrix between users and picture books, extract temporal feature vectors of user behavior through a CNN feature extraction module, and encode the text descriptions, image elements, and difficulty labels of the picture books to obtain multimodal content feature vectors.

[0028] The text feature sequence is input and processed by combining a generative adversarial network (GAN) and a semantically constrained denoising autoencoder (DAE). Differentiated noise is added to the text features for different cognitive stages. The generator of the GAN is used to reconstruct the denoised feature sequence, and a discriminator with a semantically constrained loss function is used to distinguish the reconstructed features from the original features.

[0029] The temporal feature vector of user behavior and the multimodal content feature vector are merged to form an initial fusion vector;

[0030] A cognitive transfer weight matrix is ​​introduced and concatenated with the initial fusion vector in dimensional alignment to obtain a cognitive transfer-enhanced user content fusion vector that integrates cross-dimensional relationships.

[0031] By using an age-based attention mechanism to weight the content fusion vector of cognitive transfer-enhanced users, a user vector representing the stage-based reading preferences and cross-dimensional cognitive associations is obtained. The user vector is then input into a fully connected neural network of a cognitive stage classifier to output the user's current cognitive development stage label and corresponding confidence level, which is used to construct a user profile.

[0032] Preferably, the step of processing the text feature sequence input by combining a generative adversarial network (GAN) and a semantically constrained denoising autoencoder (DAE) includes:

[0033] The text feature sequences of picture books are preprocessed in stages, and the text complexity is divided into levels according to the cognitive stage. Visual element interference noise is added to the text at the low cognitive stage, and semantic ambiguity noise is added to the text at the high cognitive stage.

[0034] The noisy text feature sequence is input into the encoder of the DAE, and the noise reduction features are extracted through multi-layer convolution and pooling operations. The size and number of convolution kernels of the encoder are dynamically configured according to the text complexity level.

[0035] The formula for calculating the output feature map of the encoder's k-th layer is as follows:

[0036]

[0037] In the formula, This represents the output feature of the k-th layer. Represents the convolution weights. Indicates bias. This indicates a pooling operation. This represents the cognitive stage adjustment coefficient. Represents the staged noise vector;

[0038] The DAE decoder reconstructs the text sequence based on the noise reduction features output by the encoder. At the same time, it introduces the generator and discriminator of GAN for adversarial training. During training, the discriminator obtains a comprehensive score through a dual-path discrimination mechanism. The comprehensive score is used to guide the dynamic adjustment of the generator parameters and the gradient backpropagation of GAN adversarial training.

[0039] Preferably, the step of the discriminator obtaining a comprehensive score through a dual-path discrimination mechanism includes:

[0040] In the first path, the optimized features output by the generator and the original text features are input into the basic feature extraction layer of the discriminator. The cosine similarity of the feature vectors is calculated through a multilayer perceptron, and the authenticity probability score is output.

[0041] In the second approach, an age-appropriate cognitive assessment matrix is ​​constructed. , Let m represent the set of real numbers, k represent the number of cognitive stages, and k represent the feature dimension; matrix elements. The threshold for the adaptation of the s-th cognitive stage to the d-th dimension feature is represented; the optimized features are projected onto matrix M, the matching degree between each dimension feature and the corresponding stage threshold is calculated, and the cognitive matching degree score is obtained by weighted summation.

[0042] Discriminator output comprehensive score , represented as: ;in, Indicates the rating coefficient. Indicates the probability score of authenticity. Indicates cognitive fit score;

[0043] When the overall score is lower than the score threshold, the generator parameters are triggered to make reverse corrections, so that the generator optimizes the features iteratively.

[0044] Preferably, in the step where the DAE decoder reconstructs the text sequence based on the denoising features output by the encoder, and simultaneously introduces the generator and discriminator of the GAN for adversarial training:

[0045] The reconstruction loss is calculated by comparing the reconstructed features from the DAE decoder with the original features. , represented as:

[0046]

[0047] In the formula, This represents the i-th original text feature. Let i represent the i-th reconstructed text feature. This represents the cognitive weighting coefficient;

[0048] The adversarial loss is calculated by scoring the authenticity and cognitive fit of the generator's optimized features using a GAN discriminator. , represented as:

[0049]

[0050] In the formula, x represents the original feature. This represents the probability of the discriminator's output being true. The generator represents the reconstructed features. Optimized features generated with cognitive level s; The distribution representing the true characteristics, Indicates noise distribution;

[0051] The semantic constraint loss is calculated by using the semantic similarity between the generator's optimized features and the original features. , represented as:

[0052]

[0053] In the formula, This represents the semantic similarity of the i-th sample at cognitive stage s; This represents the i-th original text feature. Represents the i-th reconstructed text feature;

[0054] The total loss is expressed as: In the formula, and Indicates the weight of the loss in stages; Indicating resistance to loss, Indicates the reconstruction loss. This represents the semantic constraint loss.

[0055] Preferably, the step of weighting the fusion vector using an age-based attention mechanism includes:

[0056] The user-content fusion vector is weighted using a cognitive stage-based weight matrix. Perform mapping to obtain query vectors respectively. Key vector Sum value vector Weight matrix The parameters are dynamically updated based on the cognitive stage;

[0057] For user vectors at different cognitive stages, calculate attention weights. , represented as:

[0058]

[0059] In the formula, Represents the similarity function. The dimension of the key vector is represented by , where i represents the vector index and j represents an element in the vector set. This represents the query vector for vector i. This represents the key vector of the j-th element in the key vector set;

[0060] By weighting and summing the value vectors using attention weights, we obtain the age-specific attention-weighted user vector. , represented as:

[0061]

[0062] In the formula, Indicates attention weights, This represents the value vector of vector j;

[0063] User vectors weighted by age-based attention By fusing cross-stage interest-related vectors with adaptive fusion coefficients, a comprehensive user vector for constructing user profiles is generated.

[0064] Preferably, the step of adaptively fusing user profiles and reading requests includes:

[0065] The age-specific feature vectors and text features of reading requests in the user profile are encoded, and a staged pre-trained BERT model is used to extract semantic representations, resulting in user profile feature vectors adapted to different cognitive levels. and reading request feature vector ;

[0066] A dual-path self-attention mechanism is constructed, including: in the first path, calculating the feature association weights between user profiles and reading requests within the same stage; in the second path, calculating the cross-stage interest continuity weights; and finally, a dynamic weight matrix is ​​obtained. ;

[0067] User profile feature vector and reading request feature vector Input the age-specific weight adjustment network to obtain the fused weight vector. ;

[0068] The result of adaptive fusion is represented as follows:

[0069]

[0070] In the formula, This represents the fused label vector. The dimensions expand dynamically as the user's cognitive stage develops;

[0071] right A phased feature selection process is performed to remove features that are less than a threshold in terms of their fit with the user's current cognitive stage, ultimately yielding an optimized label vector for input into the recommendation model.

[0072] Secondly, according to another embodiment of this application, an intelligent recommendation system based on children's digital picture book reading is provided;

[0073] The recommendation system includes the following modules:

[0074] The model building module is used to build recommendation models based on self-supervised learning and convolutional neural networks (CNNs). In self-supervised learning, a comparison mechanism that combines user cognition is introduced to optimize the feature representation of the model, and dynamic weights are introduced to calculate the similarity of positive and negative sample pairs to take into account the differences in users' reading abilities.

[0075] The user profile module is used to obtain the user's historical reading information, cluster the preprocessed historical reading information, and construct a dynamically updated user profile based on the clustering results. The user profile includes the user's current cognitive tags.

[0076] The recommendation module is used to respond to users' reading requests by adaptively fusing user profiles and reading requests. During the fusion process, the age-appropriateness assessment factor of picture books is introduced. The age-appropriateness assessment factor is determined based on the matching degree between the text difficulty and picture complexity of the picture book and the user's cognitive level. The fused tag vector is used as the input of the trained recommendation model to obtain the recommendation list.

[0077] The update module is used to enhance the user profile when negative feedback behavior of the user towards the recommendation list is detected, trigger a secondary screening of the picture book feature library, remove candidate picture books that do not match the user's current cognitive level, adjust the weight vector of the adaptive fusion processing, and update the recommendation list based on the adjusted tag vector.

[0078] Compared with existing technologies, the beneficial effects of the intelligent recommendation method and system for children's digital picture book reading in this invention are:

[0079] The recommendation model in this application introduces a comparison mechanism that combines user cognition in the self-supervised learning module. By constructing positive and negative sample pairs that are adapted to children's cognitive stages, adopting a dual-path momentum encoder and a query encoder, and storing negative sample feature queues in partitions according to cognitive stages, the model optimizes feature representation in the process of comparing the cognitive fit and feature similarity of positive and negative samples, effectively improving the feature learning targeting and model generalization ability in children's reading scenarios.

[0080] This application preprocesses users' historical reading information and performs multi-dimensional clustering. Based on the clustering results, it constructs a dynamic user profile including cognitive stage labels. In profile construction, by combining Generative Adversarial Networks (GANs) and semantically constrained denoising autoencoders (DAEs), differentiated noise is added to and processed for text features at different cognitive stages. This significantly improves feature extraction accuracy and stage-specific adaptability of semantic understanding, enabling more precise segmentation of user groups and in-depth analysis of users' reading habits and preferences at different cognitive stages.

[0081] When responding to a user's reading request, this application encodes the user profile and reading request features through a phased pre-trained model, and adaptively fuses the user profile and reading request by combining a dual-path self-attention mechanism. This integrates the user's long-term preferences and real-time needs at different cognitive stages, effectively improving the fit between recommended content and the current cognitive level of child users. Attached Figure Description

[0082] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0083] Figure 1 This is a flowchart illustrating the implementation of the intelligent recommendation method for children's digital picture book reading in this application embodiment;

[0084] Figure 2 A sub-flowchart of the intelligent recommendation method provided in the embodiments of this application;

[0085] Figure 3 Another sub-flowchart of the intelligent recommendation method provided in the embodiments of this application;

[0086] Figure 4 This is a structural block diagram of an intelligent recommendation system based on children's digital picture book reading, as described in an embodiment of this application.

[0087] Figure 5 This is a hardware structure block diagram of an e-reader based on an intelligent recommendation method for children's digital picture book reading according to an embodiment of this application. Detailed Implementation

[0088] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0089] According to the embodiments of this application, a method embodiment of an intelligent recommendation method based on children's digital picture book reading is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0090] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0091] Please refer to Figure 1 In one embodiment of the present invention, an intelligent recommendation method based on children's digital picture book reading is provided.

[0092] Includes the following steps:

[0093] S101: Construct a recommendation model based on self-supervised learning and convolutional neural networks (CNN);

[0094] The recommendation model comprises a feature extraction layer, a self-supervised optimization layer, and a recommendation output layer. The feature extraction layer, the bottom layer of the recommendation model, is constructed using a CNN convolutional neural network to output multimodal features. The middle layer of the recommendation model is the self-supervised optimization layer, whose self-supervised learning module receives the multimodal features output by the CNN and optimizes them into feature vectors that fit the user's cognition through a comparison mechanism. The top-level recommendation output layer receives the fusion vector of the user profile and reading request, and combines it with the optimized feature vector to generate a recommendation list.

[0095] In the self-supervised learning module of this embodiment, a comparison mechanism that combines user cognition is introduced to optimize the feature representation of the model;

[0096] In the comparison mechanism, dynamic weights are introduced to calculate the similarity between positive and negative sample pairs, taking into account differences in users' reading abilities. The introduced dynamic weights include temporal correlation weights and semantic correlation weights. The temporal correlation weights are negatively correlated with users' cognitive levels, while the semantic correlation weights are positively correlated with users' cognitive levels. In this embodiment, as children's cognitive levels improve, the weights automatically shift from temporal-driven to semantic-driven, without the need for manual parameter adjustment. This allows the model to dynamically track the growth of children's reading abilities and avoid recommending content that lags behind or exceeds users' cognitive development.

[0097] In the recommendation model of this embodiment, the CNN feature extraction module consists of a text feature extraction layer and an image feature extraction layer. The text feature extraction layer uses a 1D convolution kernel to perform a convolution operation on the text description of the picture book and outputs a text feature vector. The image feature extraction layer uses a 3D convolution kernel to perform a convolution on the continuous frames of the picture book and outputs an image feature vector. The two types of feature vectors are concatenated into multimodal basic features after dimensionality reduction by a pooling layer.

[0098] In addition, the self-supervised learning module of the recommendation model provided in this embodiment incorporates a comparison mechanism that combines user cognition. It receives multimodal basic features output by CNN, optimizes feature representation through similarity calculation of positive and negative sample pairs, and outputs an optimized feature vector that adapts to the user's cognitive level.

[0099] For details, please refer to Figure 2 In this embodiment, the step of optimizing the feature representation of the model by incorporating a comparison mechanism based on user cognition includes the following steps:

[0100] S201: Construct positive and negative sample pairs of user features and picture book features;

[0101] In step S201 of this application embodiment, the positive sample pair can be similar temporal behavior data of users in the same age group; for example: for children with an age difference of less than 12 months, their reading sequence within 7 consecutive days is extracted. When the cosine similarity of the frequency of the same picture book in the sequence is greater than 60%, it can be determined as similar temporal behavior to form a positive sample pair; the positive sample pair can also be related behavior data of the same user at different growth stages. For example, for the same user, when the overlap of the theme words of the picture book read in the later stage with those read in the previous stage is not less than 50%, it is determined as related behavior to form a positive sample pair;

[0102] In step S201 of this application embodiment, the negative sample pair can be heterogeneous behavioral data of users of different age groups. For example, for children with an age difference of not less than 36 months, when the difference in the proportion of text-picture ratio in picture books in their reading sequence is not less than 70%, a negative sample pair is formed. The negative sample pair can also be unrelated cross-stage behavioral data of the same user. For example, when the overlap of the theme words of picture books read by the same user at different stages is less than 20%, a negative sample pair is formed.

[0103] For user characteristics, similar temporal behavioral data of users in the same age group and related behavioral data of the same user at different growth stages are used as positive sample pairs, while heterogeneous behavioral data of users in different age groups and unrelated cross-stage behavioral data of the same user are used as negative sample pairs.

[0104] In addition, for picture book features, picture books that are adapted to the same cognitive stage and have a theme relevance higher than the threshold are used as positive sample pairs, while picture books that are adapted to different cognitive stages or have a theme relevance lower than the threshold are used as negative sample pairs.

[0105] S202: Construct a dual-path momentum encoder and a query encoder. The momentum encoder includes a basic parameter layer and a cognitive stage adaptation layer. The parameters of the momentum encoder are dynamically obtained from the query encoder through a phased momentum update strategy.

[0106] In this embodiment, the phased momentum update strategy includes: dividing the momentum coefficient range according to the user's cognitive stage, wherein the momentum coefficient increases with the increase of cognitive complexity to match the stability requirements of higher-order cognitive features; in addition, the momentum encoder parameter update frequency is adapted to the training batch, such as querying the encoder parameters once every 50 training batches, and calling the momentum coefficient of the corresponding range according to the cognitive stage to which the current training sample belongs during synchronization.

[0107] S203: Introduce a negative sample queue, store the corresponding negative sample features according to the cognitive stage, and compare the correlation between positive and negative samples in the two dimensions of cognitive adaptability and feature similarity so that the model learns a representation that has both stage adaptability and feature discrimination.

[0108] S204: Based on the cognitive development weight matrix, the optimization weights for maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs are dynamically adjusted, and the feature representation of the model is optimized by using an improved InfoNCE contrastive loss function that incorporates cognitive stage penalty terms.

[0109] In this embodiment, the cognitive stage penalty term is positively correlated with the degree of difference between the sample pairs in their respective cognitive stages.

[0110] Please continue to refer to Figure 1 The intelligent recommendation method based on children's digital picture book reading provided in this embodiment also includes the following steps:

[0111] S102: Obtain the user's historical reading information, perform multi-dimensional clustering on the preprocessed historical reading information, and construct a dynamically updated user profile based on the clustering results; the user profile includes the user's current cognitive tags.

[0112] In this embodiment, step S102, the step of constructing a user profile based on the clustering results, includes:

[0113] Construct an age-appropriate interaction matrix between users and picture books. The matrix elements of this interaction matrix include the user's reading time, interaction frequency, number of rereads, and age-appropriateness feedback rating for picture books at different cognitive stages.

[0114] The age-based interaction matrix is ​​normalized, and the temporal feature vector of user behavior is extracted through the CNN feature extraction module. At the same time, the text description, picture elements and difficulty labels of the picture book are encoded to obtain multimodal content feature vectors.

[0115] The text feature sequence is input and processed by combining a generative adversarial network (GAN) and a semantically constrained denoising autoencoder (DAE). Differentiated noise is added to the text features for different cognitive stages. The generator of the GAN is used to reconstruct the denoised feature sequence, and the discriminator distinguishes the cognitive adaptability of the reconstructed features and the original features. At the same time, a semantically constrained loss function is introduced so that the denoised features can adapt to the understanding ability of the corresponding cognitive stage while retaining the core semantics.

[0116] Furthermore, in this embodiment, the step of processing the text feature sequence input by combining a Generative Adversarial Network (GAN) and a semantically constrained Denoising Autoencoder (DAE) includes:

[0117] The text feature sequences of picture books are preprocessed in stages, and the text complexity is divided into levels according to the cognitive stage. Visual element interference noise is added to the text at the low cognitive stage, and semantic ambiguity noise is added to the text at the high cognitive stage.

[0118] The noisy text feature sequence is input into the encoder of the DAE, and the noise reduction features are extracted through multi-layer convolution and pooling operations. The size and number of convolution kernels of the encoder are dynamically configured according to the text complexity level.

[0119] The formula for calculating the output feature map of the encoder's k-th layer is as follows:

[0120]

[0121] In the formula, This represents the output feature of the k-th layer. Represents the convolution weights. Indicates bias. This indicates a pooling operation. This represents the cognitive stage adjustment coefficient. Represents the staged noise vector;

[0122] The DAE decoder reconstructs the text sequence based on the noise reduction features output by the encoder. At the same time, it introduces the generator and discriminator of GAN for adversarial training. During training, the discriminator obtains a comprehensive score through a dual-path discrimination mechanism. The comprehensive score is used to guide the dynamic adjustment of the generator parameters and the gradient backpropagation of GAN adversarial training.

[0123] Preferably, in the step where the DAE decoder reconstructs the text sequence based on the denoising features output by the encoder, and simultaneously introduces the generator and discriminator of the GAN for adversarial training:

[0124] The reconstruction loss is calculated by comparing the reconstructed features from the DAE decoder with the original features. , represented as:

[0125]

[0126] In the formula, This represents the i-th original text feature. Let i represent the i-th reconstructed text feature. This represents the cognitive weighting coefficient;

[0127] The adversarial loss is calculated by scoring the authenticity and cognitive fit of the generator's optimized features using a GAN discriminator. , represented as:

[0128]

[0129] In the formula, x represents the original feature. This represents the probability of the discriminator's output being true. The generator represents the reconstructed features. Optimized features generated with cognitive level s; The distribution representing the true characteristics, Indicates noise distribution;

[0130] The semantic constraint loss is calculated by using the semantic similarity between the generator's optimized features and the original features. , represented as:

[0131]

[0132] In the formula, This represents the semantic similarity of the i-th sample at cognitive stage s; This represents the i-th original text feature. Represents the i-th reconstructed text feature;

[0133] The total loss is expressed as: In the formula, and Indicates the weight of the loss in stages; Indicating resistance to loss, Indicates the reconstruction loss. This represents the semantic constraint loss.

[0134] The discriminator obtains a comprehensive score through a dual-path discrimination mechanism, including:

[0135] In the first path, the optimized features output by the generator and the original text features are input into the basic feature extraction layer of the discriminator. The cosine similarity of the feature vectors is calculated through a multilayer perceptron, and the authenticity probability score is output.

[0136] In the second approach, an age-appropriate cognitive assessment matrix is ​​constructed. , Let m represent the set of real numbers, k represent the number of cognitive stages, and k represent the feature dimension; matrix elements. The threshold for the adaptation of the s-th cognitive stage to the d-th dimension feature is represented; the optimized features are projected onto matrix M, the matching degree between each dimension feature and the corresponding stage threshold is calculated, and the cognitive matching degree score is obtained by weighted summation.

[0137] Discriminator output comprehensive score , represented as: ;in, Indicates the rating coefficient. Indicates the probability score of authenticity. Indicates cognitive fit score;

[0138] When the overall score is lower than the score threshold, the generator parameters are triggered to make reverse corrections, so that the generator optimizes the features iteratively.

[0139] Furthermore, the step of constructing a user profile based on the clustering results in this embodiment of the application also includes:

[0140] First, based on users' historical reading behavior data, such as text comprehension interaction feedback, frequency of image detail recognition, and accuracy of plot logic reasoning, calculate users' scores in multiple cognitive dimensions, including language cognition score, logical cognition score, spatial cognition score, and social cognition score.

[0141] Then, a cognitive dimension association network is constructed through a graph neural network (GNN), with multiple cognitive dimensions as network nodes. Based on the transfer influence coefficient, a dynamically updated cognitive transfer weight matrix is ​​formed. For the transfer influence coefficient, for example, for every 10 points increase in language cognition score, the logic cognition score increases by an average of 6 points, and the corresponding language-logic transfer coefficient is 0.6.

[0142] The initial fusion vector is formed by fusing the temporal feature vector of user behavior and the denoised multimodal content feature vector of picture books.

[0143] Based on the cognitive transfer weight matrix, the transfer contribution value of each cognitive dimension to other dimensions is calculated. For example, if a user's language cognition score is 80 points, its transfer contribution value to logical cognition is the product of the user's language cognition score and the transfer coefficient.

[0144] Further, all migration contribution values ​​are used to construct a cognitive migration feature vector, which is then concatenated with the initial fusion vector in dimensional alignment to obtain a cognitive migration-enhanced user content fusion vector that integrates cross-dimensional relationships.

[0145] We use an age-based attention mechanism to weight the content fusion vector of cognitive transfer-enhanced users. The attention weight matrix is ​​dynamically adjusted according to the cognitive stage, and the influence of the cognitive transfer coefficient is introduced. Through attention weighting, we obtain a user vector that can simultaneously represent stage-based reading preferences and cross-dimensional cognitive associations.

[0146] The user vector is input into a fully connected neural network of a cognitive stage classifier, which outputs the user's current cognitive development stage label and corresponding confidence score. Meanwhile, if the confidence score output by the classifier is less than the threshold, there is ambiguity in the stage determination. In this case, the dimensional correlation data in the cognitive transfer weight matrix is ​​called for correction. Finally, based on the corrected label and confidence score, a user profile containing basic attributes, dynamic preferences, cognitive stage and cross-dimensional correlation features is constructed.

[0147] In the confidence adjustment process, if a user's language cognition has reached the standard of middle age, but their logical cognition is still at the lower age level, the user is judged to be in the transition stage from lower to middle age based on the language-logic transfer coefficient of 0.6, and the confidence level is increased.

[0148] When a user's cumulative reading behavior reaches a preset threshold or the change in the confidence level of the cognitive stage label exceeds the threshold, the profile is automatically updated. During the update process, cross-stage interest-related features are retained, and the transfer influence coefficient of the cognitive dimension association network is updated simultaneously. For example, if a user adds new reading behaviors related to logical cognition and the logical cognition score improves, the corresponding language-logic transfer coefficient is also improved simultaneously, ensuring that the cognitive association features of the user profile correspond to the user's actual cognitive development.

[0149] In the step of preserving cross-stage interest association features, based on the user's historical cross-stage reading records, the cosine similarity of reading topics in different stages is calculated, topic pairs with similarity greater than a threshold are selected to construct an association matrix, and then the temporal association features of the matrix are extracted through convolution to obtain the cross-stage interest association vector.

[0150] Furthermore, in this embodiment, the step of weighting the fusion vector using an age-based attention mechanism includes:

[0151] The user content fusion vector is processed using a weight matrix based on cognitive stages. Perform mapping to obtain query vectors respectively. Key vector Sum value vector Weight matrix The parameters are dynamically updated based on the cognitive stage;

[0152] For user vectors at different cognitive stages, calculate attention weights. , represented as:

[0153]

[0154] In the formula, Represents the similarity function. The dimension of the key vector is represented by , where i represents the vector index and j represents an element in the vector set. This represents the query vector for vector i. This represents the key vector of the j-th element in the key vector set;

[0155] By weighting and summing the value vectors using attention weights, we obtain the age-specific attention-weighted user vector. , represented as:

[0156]

[0157] In the formula, Indicates attention weights, This represents the value vector of vector j;

[0158] User vectors weighted by age-based attention By fusing cross-stage interest-related vectors with adaptive fusion coefficients, a comprehensive user vector for constructing user profiles is generated.

[0159] Please continue to refer to Figure 1 The intelligent recommendation method based on children's digital picture book reading provided in this embodiment also includes the following steps:

[0160] S103: In response to the user's reading request, the user profile and reading request are adaptively fused. During the fusion process, the age-appropriateness assessment factor of the picture book is introduced. The age-appropriateness assessment factor is determined based on the matching degree between the text difficulty and picture complexity of the picture book and the user's cognitive level. The fused tag vector is used as the input of the trained recommendation model to obtain the recommendation list.

[0161] In step S103 of this embodiment, the step of adaptively fusing user profiles and reading requests includes:

[0162] The age-specific feature vectors and text features of reading requests in the user profile are encoded, and a staged pre-trained BERT model is used to extract semantic representations, resulting in user profile feature vectors adapted to different cognitive levels. and reading request feature vector ;

[0163] A dual-path self-attention mechanism is constructed, including: in the first path, calculating the feature association weights between user profiles and reading requests within the same stage; in the second path, calculating the cross-stage interest continuity weights; and finally, a dynamic weight matrix is ​​obtained. ;

[0164] User profile feature vector and reading request feature vector Input the age-specific weight adjustment network to obtain the fused weight vector. ;

[0165] The result of adaptive fusion is represented as follows:

[0166]

[0167] In the formula, This represents the fused label vector. The dimensions expand dynamically as the user's cognitive stage develops;

[0168] right A phased feature selection process is performed to remove features that are less than a threshold in terms of their fit with the user's current cognitive stage, ultimately yielding an optimized label vector for input into the recommendation model.

[0169] Please refer to Figure 3In one implementation of this application, the process of training the recommendation model includes the following steps:

[0170] S301: Obtain authorized user age-based historical reading behavior data and picture book age classification information;

[0171] S302: Initialize the parameters of the recommendation model, including the basic convolutional kernel parameters of the CNN feature extraction module, the initial weights of the dual-path momentum encoder and age-based query encoder in the self-supervised learning module, and the classification parameters of the fully connected layer;

[0172] S303: Input the fusion vector of user profiles containing cognitive tags and reading requests into the model, perform forward propagation to obtain the predicted age-appropriate recommendation list; calculate the behavioral loss value based on the matching degree between the prediction results and the user's actual reading behavior; at the same time, combine the improved InfoNCE contrastive loss function that integrates the cognitive stage penalty term in the self-supervised learning module, dynamically allocate the weight ratio of behavioral loss and contrastive loss according to the cognitive stage, and comprehensively calculate the total loss.

[0173] S304: The total loss is backpropagated to the model, and the model parameters are updated through a phased Adam optimization algorithm to minimize the total loss. The parameter update step size for different cognitive stages is dynamically adjusted according to the user's cognitive development rate. During the update process, the basic parameter layer of the momentum encoder is updated synchronously from the query encoder with a fixed momentum coefficient, and the parameters of the cognitive stage adaptation layer are dynamically adjusted based on the correlation between the current training sample and the cognitive stage, so as to realize phased momentum update.

[0174] S305: Evaluate the recommendation accuracy and age-appropriate matching of the model at each stage on the age-specific validation set. Adjust the hyperparameters and model structure of the corresponding stage according to the stage evaluation results to complete the stage-specific training and optimization of the recommendation model.

[0175] Please continue to refer to Figure 1 The intelligent recommendation method based on children's digital picture book reading provided in this embodiment also includes the following steps:

[0176] S104: Establish a real-time update mechanism for the recommendation list. When negative feedback behavior of the user towards the recommendation list is detected, the user profile is enhanced, triggering a secondary screening of the picture book feature library. Candidate picture books that do not match the user's current cognitive level are removed, and the weight vector of the adaptive fusion processing is adjusted. The recommendation list is updated based on the adjusted tag vector as the input of the recommendation model.

[0177] In step S104 of this embodiment, the negative feedback behavior is divided into:

[0178] Mild negative feedback: If the recommended book is read for a short time or is skipped, reduce the weight of the book's theme in the user profile.

[0179] Moderate negative feedback: If the same topic of recommended books is skipped multiple times in a row, the user profile interest features are reconstructed and the related features of that topic are removed.

[0180] Severe negative feedback: If the parent clicks the "not interested" button or reports that the content is not suitable, then the profile recognition label verification and feature library secondary screening will be performed.

[0181] This embodiment, through dynamic updates to the recommendation list, can promptly correct recommendation results that do not conform to the user's cognitive stage or interests, significantly improving the user experience and the intelligent iteration capability of the recommendation system.

[0182] like Figure 4 As shown, in another embodiment of this application, an intelligent recommendation system based on children's digital picture book reading is provided;

[0183] The recommendation system includes the following modules:

[0184] Model building module 401 is used to build recommendation models based on self-supervised learning and convolutional neural networks (CNNs).

[0185] User profile module 402 is used to obtain the user's historical reading information, perform multi-dimensional clustering processing on the preprocessed historical reading information, and construct a dynamically updated user profile based on the clustering results. The user profile includes the user's current cognitive tags.

[0186] The recommendation module 403 is used to respond to the user's reading request, adaptively fuse the user profile and the reading request, and introduce the age-appropriate evaluation factor of the picture book in the fusion process. The age-appropriate evaluation factor is determined based on the matching degree between the text difficulty and picture complexity of the picture book and the user's cognitive level. The fused tag vector is used as the input of the trained recommendation model to obtain the recommendation list.

[0187] The update module 404 is used to establish a real-time update mechanism for the recommendation list. When negative feedback behavior of the user towards the recommendation list is detected, the user profile is enhanced, triggering a secondary screening of the picture book feature library to remove candidate picture books that do not match the user's current cognitive level. The weight vector of the adaptive fusion processing is adjusted, and the recommendation list is updated based on the adjusted tag vector as the input of the recommendation model.

[0188] Figure 5 An internal structural diagram of an e-reader in one embodiment is shown.

[0189] like Figure 5As shown, the e-reader includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an intelligent recommendation method based on children's digital picture book reading.

[0190] The internal memory may also store computer programs, which, when executed by the processor, enable the processor to perform intelligent recommendation methods based on children's digital picture book reading. The e-reader's display screen can be an LCD screen or an e-ink screen. The input device of the e-reader can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad set on the e-reader's casing, or an external keyboard, touchpad, or mouse, etc.

[0191] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the e-reader to which the present application is applied. A specific e-reader may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0192] In one embodiment, the intelligent recommendation system based on children's digital picture book reading provided in this application can be implemented as a computer program, which can be implemented in the form of, for example... Figure 5 The system runs on the e-reader shown. The e-reader's memory can store the various program modules that make up this intelligent recommendation system.

[0193] In one embodiment, an e-reader is provided, the e-reader including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent recommendation method based on children's digital picture book reading provided in the above embodiment.

[0194] The above provides a detailed description of the real-time detection method for the operating status of electromechanical equipment disclosed in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An intelligent recommendation method based on children's digital picture book reading, characterized in that, Includes the following steps: Construct a recommendation model based on self-supervised learning and convolutional neural networks (CNNs); In self-supervised learning, a contrast mechanism combining user cognition is introduced to optimize the model's feature representation, including: constructing positive and negative sample pairs for user features and picture book features; for user features, similar temporal behavioral data of users of the same age group and related behavioral data of the same user at different developmental stages are used as positive sample pairs, while heterogeneous behavioral data of users of different age groups and unrelated cross-stage behavioral data of the same user are used as negative sample pairs; for picture book features, picture books that fit the same cognitive stage and have a theme relevance higher than a threshold are used as positive sample pairs, while picture books that fit different cognitive stages or have a theme relevance lower than a threshold are used as negative sample pairs; a dual-path momentum encoder and a query encoder are constructed, with the momentum encoder containing... The model includes a basic parameter layer and a cognitive stage adaptation layer. The momentum encoder's parameters are dynamically obtained from the query encoder through a phased momentum update strategy. A negative sample queue is introduced, storing corresponding negative sample features according to cognitive stage. By comparing the correlation between positive and negative samples in two dimensions—cognitive adaptation and feature similarity—the model learns feature representations with stage adaptability and feature discriminative power. Based on the cognitive development weight matrix, the optimization weights for maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs are dynamically adjusted. An improved InfoNCE contrastive loss function that incorporates cognitive stage penalty terms is used to optimize the model's feature representation. To address differences in users' reading abilities, dynamic weights are introduced to calculate the similarity of positive and negative sample pairs. The system acquires the user's historical reading information, clusters the preprocessed historical reading information, and constructs a dynamically updated user profile based on the clustering results. The user profile includes the user's current cognitive tags. In response to the user's reading request, the user profile and reading request are adaptively fused. During the fusion process, the age-appropriateness assessment factor of the picture book is introduced. The age-appropriateness assessment factor is determined based on the matching degree between the text difficulty and picture complexity of the picture book and the user's cognitive level. The fused tag vector is used as the input of the trained recommendation model to obtain the recommendation list. When negative feedback behavior from a user towards the recommended list is detected, the user profile is enhanced, triggering a secondary screening of the picture book feature library to remove candidate picture books that do not match the user's current cognitive level. The weight vector of the adaptive fusion processing is adjusted, and the recommended list is updated based on the adjusted tag vector.

2. The intelligent recommendation method based on children's digital picture book reading according to claim 1, characterized in that, The process of training a recommendation model includes the following steps: Obtain user age-based reading history data and picture book age classification information; The parameters of the recommendation model are initialized by inputting the fusion vector of user profiles containing cognitive tags and reading requests into the model and performing forward propagation to obtain the predicted age-appropriate recommendation list. The behavioral loss value is calculated based on the matching degree between the prediction results and the user's actual reading behavior. At the same time, the improved InfoNCE contrastive loss function, which incorporates the cognitive stage penalty term in the self-supervised learning module, is combined with the dynamic allocation of the weight ratio of behavioral loss and contrastive loss according to the cognitive stage, and the total loss is calculated comprehensively. The total loss is backpropagated into the model, and the model parameters are updated through a phased Adam optimization algorithm to minimize the total loss. During the update process, the basic parameter layer of the momentum encoder is updated synchronously from the query encoder with a fixed momentum coefficient, and the parameters of the cognitive stage adaptation layer are dynamically adjusted based on the correlation of the cognitive stage to which the current training sample belongs, so as to realize phased momentum update.

3. The intelligent recommendation method based on children's digital picture book reading according to claim 2, characterized in that, The steps for constructing user profiles based on clustering results include: We construct an age-based interaction matrix between users and picture books, extract temporal feature vectors of user behavior through a CNN feature extraction module, and encode the text descriptions, image elements, and difficulty labels of the picture books to obtain multimodal content feature vectors. The text feature sequence is input and processed by combining a generative adversarial network (GAN) and a semantically constrained denoising autoencoder (DAE). Differentiated noise is added to the text features for different cognitive stages. The generator of the GAN is used to reconstruct the denoised feature sequence, and a discriminator with a semantically constrained loss function is used to distinguish the reconstructed features from the original features. The temporal feature vector of user behavior and the multimodal content feature vector are merged to form an initial fusion vector; A cognitive transfer weight matrix is ​​introduced and concatenated with the initial fusion vector in dimensional alignment to obtain a cognitive transfer-enhanced user content fusion vector that integrates cross-dimensional relationships. By using an age-based attention mechanism to weight the content fusion vector of cognitive transfer-enhanced users, a user vector representing the stage-based reading preferences and cross-dimensional cognitive associations is obtained. The user vector is then input into a fully connected neural network of a cognitive stage classifier to output the user's current cognitive development stage label and corresponding confidence level, which is used to construct a user profile.

4. The intelligent recommendation method based on children's digital picture book reading according to claim 3, characterized in that, The steps of processing text feature sequences by combining them with generative adversarial networks (GANs) and semantically constrained denoising autoencoders (DAEs) include: The text feature sequences of picture books are preprocessed in stages, and the text complexity is divided into levels according to the cognitive stage. Visual element interference noise is added to the text at the low cognitive stage, and semantic ambiguity noise is added to the text at the high cognitive stage. The noisy text feature sequence is input into the encoder of the DAE, and denoising features are extracted through multiple convolution and pooling operations. The output feature mapping of the k-th layer of the encoder is calculated using the following formula: In the formula, This represents the output feature of the k-th layer. Represents the convolution weights. Indicates bias. This indicates a pooling operation. This represents the cognitive stage adjustment coefficient. Represents the staged noise vector; The DAE decoder reconstructs the text sequence based on the noise reduction features output by the encoder. At the same time, it introduces the generator and discriminator of GAN for adversarial training. During training, the discriminator obtains a comprehensive score through a dual-path discrimination mechanism. The comprehensive score is used to guide the dynamic adjustment of the generator parameters and the gradient backpropagation of GAN adversarial training.

5. The intelligent recommendation method based on children's digital picture book reading according to claim 4, characterized in that, The discriminator obtains a comprehensive score through a dual-path discrimination mechanism, including the following steps: In the first path, the optimized features output by the generator and the original text features are input into the basic feature extraction layer of the discriminator. The cosine similarity of the feature vectors is calculated through a multilayer perceptron, and the authenticity probability score is output. In the second approach, an age-appropriate cognitive assessment matrix is ​​constructed. , Let m represent the set of real numbers, k represent the number of cognitive stages, and k represent the feature dimension; matrix elements. The threshold for the adaptation of the s-th cognitive stage to the d-th dimension feature is represented; the optimized features are projected onto matrix M, the matching degree between each dimension feature and the corresponding stage threshold is calculated, and the cognitive matching degree score is obtained by weighted summation. Discriminator output comprehensive score When the overall score falls below the scoring threshold, the generator parameters are adjusted, and the overall score is adjusted accordingly. Represented as: in, Indicates the rating coefficient. Indicates the probability score of authenticity. This indicates the cognitive matching score.

6. The intelligent recommendation method based on children's digital picture book reading according to claim 5, characterized in that, In the DAE decoder, the text sequence is reconstructed based on the denoised features of the encoder output, while simultaneously introducing the generator and discriminator of the GAN for adversarial training: The reconstruction loss is calculated by comparing the reconstructed features from the DAE decoder with the original features. , represented as: In the formula, This represents the i-th original text feature. Let i represent the i-th reconstructed text feature. This represents the cognitive weighting coefficient; The adversarial loss is calculated by scoring the authenticity and cognitive fit of the generator's optimized features using a GAN discriminator. , represented as: In the formula, x represents the original feature. This represents the probability of the discriminator's output being true. The generator represents the reconstructed features. Optimized features generated with cognitive level s; The distribution representing the true characteristics, Indicates noise distribution; The semantic constraint loss is calculated by using the semantic similarity between the generator's optimized features and the original features. , represented as: In the formula, This represents the semantic similarity of the i-th sample at cognitive stage s; This represents the i-th original text feature. Represents the i-th reconstructed text feature; The total loss is expressed as: In the formula, and Indicates the weight of the loss in stages; Indicating resistance to loss, Indicates the reconstruction loss. This represents the semantic constraint loss.

7. The intelligent recommendation method based on children's digital picture book reading according to claim 6, characterized in that, The steps for weighting the fusion vector using an age-based attention mechanism include: The fusion vector is passed through a weight matrix based on cognitive stages. Perform mapping to obtain query vectors respectively. Key vector Sum value vector Weight matrix The parameters are dynamically updated based on the cognitive stage; For user vectors at different cognitive stages, calculate attention weights. , represented as: In the formula, Represents the similarity function. The dimension of the key vector is represented by , where i represents the vector index and j represents an element in the vector set. This represents the query vector for vector i. This represents the key vector of the j-th element in the key vector set; By weighting and summing the value vectors using attention weights, we obtain the age-specific attention-weighted user vector. , represented as: In the formula, Indicates attention weights, This represents the value vector of vector j; User vectors weighted by age-based attention By fusing cross-stage interest-related vectors with adaptive fusion coefficients, a comprehensive user vector for constructing user profiles is generated.

8. The intelligent recommendation method based on children's digital picture book reading according to claim 7, characterized in that, The steps for adaptively fusing user profiles and reading requests include: The age-specific feature vectors and text features of reading requests in the user profile are encoded, and a staged pre-trained BERT model is used to extract semantic representations, resulting in user profile feature vectors adapted to different cognitive levels. and reading request feature vector ; A dual-path self-attention mechanism is constructed, including: in the first path, calculating the feature association weights between user profiles and reading requests within the same stage; in the second path, calculating the cross-stage interest continuity weights; and finally, a dynamic weight matrix is ​​obtained. ; User profile feature vector and reading request feature vector Input the age-specific weight adjustment network to obtain the fused weight vector. ; The result of adaptive fusion is represented as follows: In the formula, This represents the fused label vector. The dimensions expand dynamically as the user's cognitive stage develops; right A phased feature selection process is performed to remove features that are less than a threshold in terms of their fit with the user's current cognitive stage, ultimately yielding an optimized label vector for input into the recommendation model.

9. A recommendation system for implementing the intelligent recommendation method based on children's digital picture book reading as described in any one of claims 1 to 8, characterized in that, The recommendation system includes the following modules: The model building module is used to build recommendation models based on self-supervised learning and convolutional neural networks (CNNs). In self-supervised learning, a contrast mechanism combining user cognition is introduced to optimize the model's feature representation, including: constructing positive and negative sample pairs for user features and picture book features; for user features, similar temporal behavioral data of users of the same age group and related behavioral data of the same user at different developmental stages are used as positive sample pairs, while heterogeneous behavioral data of users of different age groups and unrelated cross-stage behavioral data of the same user are used as negative sample pairs; for picture book features, picture books that fit the same cognitive stage and have a theme relevance higher than a threshold are used as positive sample pairs, while picture books that fit different cognitive stages or have a theme relevance lower than a threshold are used as negative sample pairs; a dual-path momentum encoder and a query encoder are constructed, with the momentum encoder containing... The model includes a basic parameter layer and a cognitive stage adaptation layer. The momentum encoder's parameters are dynamically obtained from the query encoder through a phased momentum update strategy. A negative sample queue is introduced, storing corresponding negative sample features according to cognitive stage. By comparing the correlation between positive and negative samples in two dimensions—cognitive adaptation and feature similarity—the model learns feature representations with stage adaptability and feature discriminative power. Based on the cognitive development weight matrix, the optimization weights for maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs are dynamically adjusted. An improved InfoNCE contrastive loss function that incorporates cognitive stage penalty terms is used to optimize the model's feature representation. To address differences in users' reading abilities, dynamic weights are introduced to calculate the similarity of positive and negative sample pairs. The user profile module is used to obtain the user's historical reading information, cluster the preprocessed historical reading information, and construct a dynamically updated user profile based on the clustering results. The user profile includes the user's current cognitive tags. The recommendation module is used to respond to users' reading requests by adaptively fusing user profiles and reading requests. During the fusion process, an age-appropriate assessment factor for picture books is introduced. The age-appropriate assessment factor is determined based on the matching degree between the text difficulty and picture complexity of the picture book and the user's cognitive level. The fused tag vector is used as the input of the trained recommendation model to obtain the recommendation list. The update module is used to enhance the user profile when negative feedback behavior of the user towards the recommendation list is detected, trigger a secondary screening of the picture book feature library, remove candidate picture books that do not match the user's current cognitive level, adjust the weight vector of the adaptive fusion processing, and update the recommendation list based on the adjusted tag vector.

Images