Picture book live broadcast personalized book recommendation method and system based on artificial intelligence

Abstract

The invention discloses a picture book live broadcast personalized book recommendation method and system based on artificial intelligence, and relates to the technical field of live broadcast book recommendation, and the method comprises the following steps: obtaining multi-modal interaction data of a user in picture book live broadcast microphone connection, and generating a three-dimensional ability portrait vector and a psychological portrait vector representing the cognitive ability and psychological tendency of the user based on the multi-modal interaction data; performing time sequence analysis, and calculating an evolution direction vector and an evolution rate scalar; when the evolution direction is positive, according to a comparison result of an evolution rate scalar and an adaptive threshold range, updating a preset cooperation coefficient in the threshold range, if the evolution rate scalar exceeds the threshold range, adjusting an ability integrating degree weight and an emotion resonance degree weight, and dynamically updating parameters based on weight adjustment counting; and calculating a matching total score through a matching decision formula by using the dynamically adjusted parameters, and outputting picture book recommendation. According to the method and the device, the user state is adapted in real time in live broadcast interaction, the recommendation strategy is adaptively optimized, and the accuracy of the instant individuation degree of recommendation is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of live-streaming book recommendation technology, and more specifically, this application relates to a method and system for personalized live-streaming book recommendation of picture books based on artificial intelligence. Background Technology

[0002] In live-streaming interactive scenarios involving children's picture books, existing recommendation technologies struggle to achieve personalized book recommendations that are deeply integrated with real-time interaction. The core drawback lies in the disconnect between the recommendation process and the high-fidelity real-time interactive data stream. This prevents the technology from capturing and responding to changes in children's cognitive abilities and emotional states during instant conversations, resulting in recommendations that are lagging, static, and lack educational relevance.

[0003] Traditional recommendation methods largely rely on users' static historical data or preset tags. In the dynamic and spontaneous context of live-streaming interaction, their effectiveness is significantly reduced. Because they cannot access and process the multi-dimensional interactive data generated during the interaction, such as real-time voice, facial expressions, and text messages, these methods lose the best window to assess the child's current state. They often make recommendations after the interaction ends, based on historical behaviors with weak relevance to the current interaction. This makes it difficult for the recommendations to match the child's immediate focus of interest, language skills, or emerging emotional needs during the just-ended conversation. Essentially, it's a context-detached, reactive remedy rather than synchronous guidance integrated into the interactive process.

[0004] Furthermore, existing assessment models are typically one-dimensional and static. They may measure matching solely from a single perspective, such as language difficulty or content theme, lacking a comprehensive evaluation system that can simultaneously quantify children's cognitive expression level and psychological tendencies. For example, they cannot collaboratively analyze the emotional richness, identification with specific values, and complexity of children's linguistic logic when telling stories, and transform this multi-dimensional information into calculable recommendation criteria. Therefore, their recommendations often only partially satisfy one aspect of a need, perhaps meeting the cognitive difficulty but failing to address current emotions, or catering to superficial interests but failing to address their skill gaps, making it difficult to achieve the educational goal of synergistic development of abilities and emotional support. In addition, such recommendations usually lack interpretable educational development direction, merely presenting a list of books without clarifying the specific connection between the recommendation and the child's performance in this interaction, making the recommendation behavior more like an isolated product exposure than an evidence-based, logically coherent educational intervention.

[0005] To address this issue, there is an urgent need in this field for a technical solution that can be deeply embedded in the real-time live interactive process, fully utilize the data from the live interaction process to conduct real-time multi-dimensional status assessments of children, and generate personalized picture book recommendations with clear developmental guidance. In short, the key deficiency of existing technologies lies in their inability to simultaneously complete dynamic assessments of children's cognitive and psychological states through the analysis of real-time interactive data in the dynamic scenario of picture book live interaction, and to achieve personalized recommendations based on these assessments and immediate responses. Summary of the Invention

[0006] To address the technical challenges and provide a personalized book recommendation method for live picture book streaming based on artificial intelligence, this technical solution resolves the issues raised in the background section.

[0007] In a first aspect, embodiments of this application provide a method for personalized book recommendation in live-streaming picture book sharing based on artificial intelligence, comprising the following steps: acquiring interaction data of a target user, the interaction data including text data, voice feature data, and facial expression feature data; based on the interaction data, generating a three-dimensional ability profile vector sequence representing the target user's current cognitive expression ability and a three-dimensional psychological profile vector sequence representing the target user's current psychological tendency; performing time-series analysis on the three-dimensional ability profile vector sequence and the three-dimensional psychological profile vector sequence to calculate the evolution direction vector and the evolution rate scalar; when all components of the evolution direction vector are positive, determining whether the evolution rate scalar is within an adaptive threshold range, and if so, then calculating... Calculate the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold, and update the preset synergy coefficient by comparing the boundary value of the range limiting function. If not, and the evolution rate scalar is lower than the lower limit of the adaptive threshold range, no adjustment is made. If not, and the evolution rate scalar exceeds the upper limit of the adaptive threshold range, the ability fit weight and emotional resonance weight are adjusted. Record the adjustment counts of the ability fit weight and emotional resonance weight, and update the boundary value of the range limiting function accordingly. Based on the updated preset synergy coefficient, the adjusted ability fit weight, and the emotional resonance weight, calculate the total matching decision score using the matching decision formula. Select at least one target picture book based on the total matching decision score and output it.

[0008] Secondly, this application provides an AI-based personalized book recommendation system for live-streaming picture books, comprising: a data acquisition module for acquiring interaction data of the target user, the interaction data including text data, voice feature data, and facial expression feature data; a vector construction module for generating a three-dimensional ability profile vector sequence representing the target user's current cognitive expression ability and a three-dimensional psychological profile vector sequence representing the target user's current psychological tendency based on the interaction data; a time-series analysis module for performing time-series analysis on the three-dimensional ability profile vector sequence and the three-dimensional psychological profile vector sequence, and calculating the evolution direction vector and the evolution rate scalar; and a synergy coefficient acquisition module for determining whether the evolution rate scalar is within an adaptive threshold range when all components of the evolution direction vector are positive, and if so, calculating the evolution rate. The system updates the preset synergy coefficient by comparing the absolute difference between the scalar and the midpoint of the adaptive threshold with the boundary value of the range limiting function; the judgment module: if the judgment result is negative and the evolution rate scalar is lower than the lower limit of the adaptive threshold range, no adjustment is made; the weight adjustment module: if negative and the evolution rate scalar exceeds the upper limit of the adaptive threshold range, the ability fit weight and emotional resonance weight are adjusted; the statistics module: records the adjustment counts of the ability fit weight and emotional resonance weight, and updates the boundary value of the range limiting function accordingly; the total score calculation module: calculates the total matching decision score based on the updated preset synergy coefficient, the adjusted ability fit weight, and the emotional resonance weight using the matching decision formula; the output module: selects at least one target picture book based on the total matching decision score and outputs it.

[0009] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By processing text, voice, and facial expression data generated during live chat sessions in real time, and instantly generating 3D capability profile vector sequences and 3D psychological profile vector sequences, recommendation decisions are made entirely based on the user's immediate performance within the current live stream topic. Further analysis of the evolution direction vector and evolution rate scalar of the vector sequences allows for real-time perception of whether the user's state is consistently engaged or experiencing emotional fluctuations. Based on this, preset synergy coefficients or capability compatibility weights and emotional resonance weights are dynamically adjusted within seconds, enabling the recommendation strategy to evolve synchronously with the interaction process. This achieves dynamic recommendations deeply coupled with real-time live stream interaction, solving the problems of recommendation lag and contextual disconnect.

[0010] 2. This approach not only constructs a three-dimensional ability profile vector based on narrative structure features, emotional vocabulary features, and syntactic complexity features, but also builds a three-dimensional psychological profile vector through a multimodal emotion recognition model, value proposition lexicon matching, and role classification model, forming a complete quantitative description of the user's cognition and psychological state. In the matching stage, the compatibility coefficient and resonance coefficient are used to measure the matching degree with the picture book in terms of ability difficulty and psychological theme, respectively, and a preset synergy coefficient is introduced to enhance the recommendation gain that combines both. This multi-dimensional synergistic matching mechanism can simultaneously take into account the user's cognitive development needs and emotional resonance needs, resulting in more comprehensive and three-dimensional recommendation results. It constructs a multi-dimensional synergistic evaluation and matching model that integrates cognitive ability and psychological state, solving the problem of a single evaluation dimension.

[0011] 3. It records adjustments to the weights of ability compatibility and emotional resonance, and dynamically calculates the upper limit of the range limiting function based on the upward adjustment count using a saturation growth function. When users frequently require emotional support, the upper limit of the weight adjustment is appropriately relaxed, giving the strategy greater flexibility; when the user's state stabilizes, this upper limit gradually returns to normal over time with a decay factor. This design enables it to learn and adapt to users' long-term interaction patterns, preventing excessive oscillations in parameter adjustments, and ensuring the robustness and personalization depth of the long-term recommendation strategy while maintaining sensitivity to immediate states. Attached Figure Description

[0012] Figure 1 A schematic diagram illustrating the steps of the AI-based personalized book recommendation method for live-streaming picture books, provided in an embodiment of this application. Figure 2 A schematic diagram of the logic flow of the AI-based personalized book recommendation method for live picture book streaming provided in this application embodiment; Figure 3 A logical diagram illustrating the updating of the boundary of the range limiting function provided in this application embodiment; Figure 4 A schematic diagram of the structure of the AI-based live-streamed personalized book recommendation system for picture books provided in this application embodiment. Detailed Implementation

[0013] This application's embodiments address the technical problem in the prior art of insufficient accuracy in real-time adaptation to user status and adaptive optimization of recommendation strategies during live-streaming picture book recommendations, through an AI-based personalized book recommendation method and system.

[0014] To address the problem that existing picture book live streaming recommendation technologies struggle to respond in real time to children's dynamic interactive states, the fundamental solution of this approach lies in constructing a closed-loop processing flow that can tightly couple with the live streaming data stream, instantly quantify children's multi-dimensional states, and adaptively adjust the recommendation strategy based on state changes.

[0015] This approach begins at the data source, requiring direct processing of high-fidelity data generated during live chat that reflects children's immediate reactions. Therefore, it necessitates real-time acquisition and synchronous processing of text data from the audio stream, speech feature data extracted from the audio, and facial expression feature data extracted from the video.

[0016] After obtaining the raw data, the core task is to transform the unstructured interactive information into standardized state indicators that are machine-computable and comparable. This requires designing two parallel quantization systems.

[0017] On the one hand, from the perspective of cognitive development, we analyze the language organization logic and emotional vocabulary usage in text data and the sentence complexity in speech data. Through a pre-set feature dimension mapping model, we quantify them into scores on three orthogonal dimensions: language ability, emotional expression and cultural cognition, thus forming a three-dimensional ability profile vector that represents cognitive expression ability.

[0018] On the other hand, from a psychological and emotional perspective, the analysis integrates the intonation and speed features of speech with facial expression features to identify the dominant emotional category. Simultaneously, it analyzes value-oriented keywords embedded in the text and, combined with the live stream theme, determines the child's role identification during the interaction. Finally, it maps emotional category, value orientation, and role type to three dimensions: emotional state intensity, value orientation intensity, and role projection intensity, forming a three-dimensional psychological profile vector representing psychological tendencies. These two vectors together constitute a digital twin of the child's current comprehensive state.

[0019] However, static snapshots are insufficient to support dynamic recommendations. To capture the evolutionary trends of children's states, time-series analysis of continuously generated ability and mental profile vector sequences is required. By calculating the direction and rate of change of the vector sequences, an evolutionary direction vector representing the overall developmental trend and an evolutionary rate scalar representing the degree of change are obtained. These two dynamic features become key criteria for determining whether a child is in a stable engagement or experiencing emotional fluctuations, among other interaction patterns.

[0020] Based on the results of dynamic perception, the corresponding picture book recommendation strategy needs to be adapted in real time and accurately. If the child shows a positive and stable trend of progress, the system enters the state adaptation adjustment mode. At this time, the optimization goal is to consolidate the learning state. Therefore, the synergy coefficient in the matching decision formula will be calculated and appropriately increased to enhance the importance of the synergistic effect of ability matching and interest resonance. If the child shows positive but overly intense emotional fluctuations, possibly accompanied by distraction, the system enters the emotion adaptation adjustment mode. At this time, emotional support needs to be prioritized. Therefore, the system will initiate a linkage adjustment of the weight of ability fit and emotional resonance, increasing the decision weight of emotional factors, so that the recommendation will lean towards content that can better evoke current emotional resonance.

[0021] To ensure long-term adaptability and stability and avoid strategy oscillations, a self-optimization mechanism at the strategy level is further introduced. It records the historical frequency of emotional weight adjustments and dynamically adjusts the allowable range of these adjustments based on this frequency. When a child frequently requires emotional attention, the adjustment limit is appropriately widened, providing greater flexibility to the strategy; as the child's state stabilizes, this limit gradually shrinks back to normal over time. This dynamic boundary mechanism based on historical behavioral feedback enables it to learn and adapt to the user's long-term interaction patterns.

[0022] Finally, using dynamically adjusted and boundary-constrained parameters, combined with the child's current abilities and psychological profile vector, a comprehensive matching score is calculated for each candidate book in the picture book database using a matching decision formula. Based on this, the most suitable picture book recommendation is output, tailored to the child's current state and developmental needs. The entire process achieves a complete closed loop from real-time data perception to state quantification, and then to dynamic strategy adjustment and self-optimization, enabling the recommendation to be deeply integrated into the interactive process and possessing a clear educational development orientation.

[0023] To better understand the technical solution, the following will provide a detailed explanation of the technical solution in conjunction with the accompanying drawings and specific implementation methods.

[0024] Figure 1 The schematic diagram of the steps of the personalized book recommendation method for live picture book streaming based on artificial intelligence provided in this application embodiment includes the following steps: acquiring the interaction data of the target user, the interaction data including text data, voice feature data, and facial expression feature data; based on the interaction data, generating a three-dimensional ability profile vector sequence representing the target user's current cognitive expression ability and a three-dimensional psychological profile vector sequence representing the target user's current psychological tendency; performing time-series analysis on the three-dimensional ability profile vector sequence and the three-dimensional psychological profile vector sequence to calculate the evolution direction vector and the evolution rate scalar; when all components of the evolution direction vector are positive, determining whether the evolution rate scalar is within the adaptive threshold range, if so, then... Calculate the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold, and update the preset synergy coefficient by comparing the boundary value of the range limiting function. If not, and the evolution rate scalar is below the lower limit of the adaptive threshold range, no adjustment is made. If not, and the evolution rate scalar exceeds the upper limit of the adaptive threshold range, the ability fit weight and emotional resonance weight are adjusted. Record the adjustment counts of the ability fit weight and emotional resonance weight, and update the boundary value of the range limiting function accordingly. Based on the updated preset synergy coefficient, the adjusted ability fit weight, and the emotional resonance weight, calculate the total matching decision score using the matching decision formula. Select at least one target picture book based on the total matching decision score and output it.

[0025] Figure 2A schematic diagram of the logic flow of the AI-based personalized book recommendation method for live picture book streaming provided in this application embodiment.

[0026] By acquiring multimodal interaction data of target users in real time, and dynamically generating a sequence of three-dimensional profile vectors representing their cognitive expression abilities and psychological tendencies, the evolution direction and rate of user states can be captured through time-series analysis. Based on this, the synergy coefficients and weights in the recommendation strategy can be adaptively adjusted.

[0027] This enables a deep integration of assessment and immediate response to children's cognitive and psychological states during live-streamed picture book interactions, overcoming the limitations of traditional recommendations that are lagging, static, and lack educational focus. It provides more dynamic, collaborative, and development-oriented personalized picture book recommendations.

[0028] Furthermore, the specific generation process of the three-dimensional ability profile vector is as follows: extract narrative structure features and emotional vocabulary features from text data; extract syntactic complexity features from speech feature data; input the narrative structure features, emotional vocabulary features, and syntactic complexity features into a preset feature dimension mapping model, and the feature dimension mapping model outputs the target user's score on three preset orthogonal ability dimensions, including language ability dimension, emotional expression dimension, and cultural cognition dimension; normalize the score on the three orthogonal ability dimensions, and construct a three-dimensional vector in the preset orthogonal three-dimensional space as the three-dimensional ability profile vector.

[0029] In this embodiment, narrative structure features refer to the patterns in logical organization, plot development, and information presentation of text content. Their function is to reflect the target user's logical thinking ability, information organization ability, and understanding of complex concepts.

[0030] In practical implementation, natural language processing techniques can be used, such as syntactic analysis to identify dependencies between sentences, discourse analysis to identify connections and transitions between paragraphs, topic chains to identify the evolution of the core issues in the text, or referential resolution techniques to track the consistency of entities in the text. These techniques can quantify the coherence, integrity, and complexity of the text, thereby characterizing the user's narrative ability.

[0031] Emotional lexical features refer to words or phrases in a text that express emotional tendencies. Their function is to reflect the target user's emotional understanding, the richness of their emotional expression, and their ability to empathize with the emotions of others.

[0032] In practical implementation, a pre-defined sentiment dictionary can be constructed or utilized to match words in the text and statistically analyze the frequency and intensity of positive, negative, or neutral sentiment words. Alternatively, a deep learning-based sentiment classification model can be used to determine the overall sentiment tendency of the text and extract key sentiment expression words as features.

[0033] Syntactic complexity features refer to the complexity of sentence structure, vocabulary usage, and grammatical rules in spoken expression. Their function is to reflect the target user's language organization ability, cognitive load, and the fluency and accuracy of spoken expression.

[0034] In practical implementation, speech feature data can first be converted into text data using speech recognition technology. Then, syntactic analysis is performed on the converted text, such as calculating average sentence length, number of clauses, lexical diversity, and frequency of use of complex sentence structures. In addition, prosodic features can be directly extracted from speech signals as indirect indicators of syntactic complexity, because these prosodic features are often associated with the syntactic structure and cognitive load of spoken expression.

[0035] The predefined feature dimension mapping model is a trained machine learning model, such as a multilayer perceptron, support vector machine, or a more complex neural network model. Its role is to effectively integrate and map the low-level, diverse features extracted from raw interaction data to higher-level, more interpretable capability dimensions.

[0036] In its implementation, the model establishes a non-linear relationship between features and ability dimension scores by learning from a large amount of labeled data. The language ability dimension refers to the target user's overall level in vocabulary mastery, grammar usage, fluency, and language organization. This dimension's score quantifies the user's ability to communicate effectively using language. The emotional expression dimension refers to the target user's ability to understand others' emotions, express their own emotions, and demonstrate emotional empathy in communication.

[0037] This dimension quantifies the user's level of emotional intelligence. The cultural cognition dimension refers to the target user's ability to understand and apply specific cultural backgrounds, social customs, allusions, and values. This dimension quantifies the breadth and depth of the user's cross-cultural communication and understanding. These three dimensions are designed to be orthogonal, meaning they are conceptually independent and can comprehensively assess the user's cognitive expression abilities from different perspectives.

[0038] This solution can also accurately capture the cognitive and expressive abilities of target users from multimodal interaction data. Specifically, by extracting narrative structure features and emotional vocabulary features from text data, and syntactic complexity features from speech feature data, it ensures a comprehensive consideration of users' language organization ability, emotional understanding and expression ability, and cognitive complexity. These features, processed by a pre-defined feature dimension mapping model, can output quantitative scores on three orthogonal dimensions: language ability, emotional expression, and cultural cognition, avoiding the limitations of single-dimensional evaluation. The scores are normalized and constructed into a three-dimensional vector, giving the user ability profile a unified dimension and an intuitive geometric representation. This greatly improves the accuracy and interpretability of the three-dimensional ability profile vector, providing a solid and refined foundation for subsequent picture book matching decisions, thereby significantly enhancing the personalization and accuracy of picture book recommendations.

[0039] Furthermore, the specific generation process of the 3D psychological profile vector is as follows: Multimodal feature fusion is performed on the tone and speed parameter sequences in the speech feature data and the facial action unit parameter sequences in the expression feature data, and the result is input into a preset multimodal emotion recognition model to output the target user's current emotional state category; semantic analysis and keyword extraction are performed on the text data, and it is matched with a preset value proposition lexicon to identify the core value orientation category contained in the target user's expression; the emotional state category, core value orientation category, and the topic tag data associated with the live chat session are input into a preset role classification model to output the target user's role projection type in the current interaction; the emotional state category, core value orientation category, and role projection type are mapped to initial values ​​for the three dimensions of emotional state intensity, value orientation intensity, and role projection intensity, respectively, according to a preset quantization mapping table; the initial values ​​of the three dimensions are normalized to construct a 3D vector in a preset orthogonal 3D space, which serves as the 3D psychological profile vector.

[0040] In this embodiment, the preset multimodal emotion recognition model is a trained machine learning or deep learning model used to receive fused multimodal features and identify the user's current emotional state category. This model is typically trained on a large multimodal emotion annotation dataset and can map the fused features to preset emotion categories, such as joy, sadness, surprise, and calmness. Through this model, the user's real-time emotional state during live streaming interactions can be objectively determined.

[0041] Semantic analysis and keyword extraction are processes for gaining a deeper understanding of textual data. Semantic analysis aims to reveal the deeper meaning and contextual relationships of text, while keyword extraction focuses on identifying the most important words or phrases in the text. These techniques can be implemented using methods from natural language processing, such as word embeddings, topic models, and dependency parsing. They are used to extract key information from user-input text that represents its core ideas and concerns.

[0042] A pre-built value proposition thesaurus is a collection of words or phrases related to different core values. For example, it might include value tags such as courage, exploration, friendliness, and responsibility, along with their corresponding synonyms, near-synonyms, or related expressions. By matching keywords extracted from user text with this thesaurus, it's possible to identify the core value categories implicit in the user's expressions or those they prioritize.

[0043] The topic tag data associated with the live-stream interaction segment refers to pre-set tag information related to the content of the picture book live-stream interactive segment. These tags can describe the theme, content type, and educational goals of the live stream, such as science enlightenment, emotional intelligence development, historical stories, and art appreciation. These tags provide important contextual information for understanding the user's psychological activities in a specific situation.

[0044] A pre-defined role classification model is used to determine the type of role a user plays or projects in a current interaction based on multi-dimensional information (emotional state category, core value orientation category, and topic tag data related to the live chat session). For example, a user might be classified as an active participant, a curious explorer, an emotional empathizer, or a critical thinker. This model is typically trained using supervised learning methods, utilizing interaction data with role tags to learn classification rules.

[0045] The pre-defined quantitative mapping table is a lookup table or function that converts qualitative emotional state categories, core value orientation categories, and role projection types into quantitative values. For example, different emotional categories can be mapped to different emotional intensity values, different value orientations can be mapped to intensity values ​​from 0 to 1, and different role types can also be mapped to corresponding intensity values. This mapping table ensures that the psychological characteristics of different categories can be uniformly quantified for subsequent mathematical calculations.

[0046] These multi-dimensional and refined psychological characteristics are quantified and normalized to construct a more representative and distinctive three-dimensional psychological profile vector. This vector can more accurately reflect the user's true psychological state and potential needs, providing a solid foundation for subsequent picture book matching decisions and significantly improving the accuracy of personalized recommendations and user satisfaction.

[0047] Furthermore, the specific acquisition process of the target picture book is as follows: For candidate picture books in the picture book database, the pre-stored difficulty feature vector and theme feature vector corresponding to the candidate picture book are obtained; the difficulty feature vector is composed of the preset difficulty level values ​​of the candidate picture book in three dimensions: language ability, emotional expression, and cultural cognition; the theme feature vector is composed of the preset theme intensity values ​​of the candidate picture book in three dimensions: emotional state, value orientation, and role projection; the absolute values ​​of the cosine similarity and modulus difference between the three-dimensional ability profile vector and the difficulty feature vector are calculated, and the absolute values ​​of the cosine similarity and modulus difference are weighted and averaged according to the preset first weight to generate the fit coefficient; the dot product between the three-dimensional psychological profile vector and the theme feature vector is calculated as the psychological profile vector. To determine the theme fit, the vector components representing emotional states are extracted from the three-dimensional psychological profile vector and the theme feature vector, respectively. The standardized absolute difference between these components is calculated and recorded as the emotional matching degree. The theme fit and emotional matching degree are weighted and averaged according to a preset second weight to generate a resonance coefficient. The first weighted component is obtained by multiplying the fit coefficient by the ability fit weight, the second weighted component is obtained by multiplying the resonance coefficient by the emotional resonance weight, and the synergy gain component is obtained by multiplying the product of the fit coefficient and the resonance coefficient by a preset synergy coefficient. The first weighted component, the second weighted component, and the synergy gain component are all added together to obtain the total matching decision score. Candidate picture books whose total matching decision score exceeds the recommendation threshold are selected as target picture books.

[0048] In this embodiment, the process of acquiring the target picture book first requires obtaining the pre-stored difficulty feature vector and theme feature vector of the candidate picture book from the picture book database. The difficulty feature vector is used to quantify the cognitive complexity of the picture book. It can be manually annotated by experts based on the picture book content, or automatically extracted by analyzing text complexity through a natural language processing model or a machine learning model.

[0049] In terms of language proficiency, the vocabulary difficulty and syntactic complexity of picture books can be assessed; in terms of emotional expression, the richness of emotional vocabulary and the fluctuations of emotional curves in picture books can be assessed; and in terms of cultural cognition, the depth of background knowledge and cultural metaphors required by picture books can be assessed.

[0050] These values ​​are typically standardized to facilitate subsequent calculations and comparisons. Thematic feature vectors, on the other hand, describe the psychological and emotional content tendencies of picture books. They can be obtained based on the picture book content through manual annotation, keyword extraction, thematic model analysis, or sentiment analysis models. For example, the emotional state dimension can assess the main emotions conveyed by the picture book; the value orientation dimension can assess the values ​​advocated by the picture book; and the character projection dimension can assess the types of characters in the picture book and their appeal to readers. These values ​​are also typically standardized.

[0051] After obtaining the feature vector of the picture book, the absolute value of the difference between the cosine similarity and the modulus between the 3D ability profile vector and the difficulty feature vector is calculated.

[0052] Cosine similarity is used to measure the similarity between two vectors, that is, the degree of relative matching between user ability profile and picture book difficulty features in each dimension. A higher cosine similarity indicates that user ability and picture book difficulty are highly consistent in structure.

[0053] The absolute value of the difference in modulus length measures the absolute strength difference between the two vectors across each dimension. A smaller difference in modulus length indicates that the user's ability and the difficulty of the picture book are similar at an overall level. Subsequently, the absolute values ​​of cosine similarity and the difference in modulus length are weighted and averaged according to a preset first weight to generate a fit coefficient. The preset first weight is used to balance the contributions of cosine similarity and the difference in modulus length in the calculation of the fit coefficient; it can be set empirically or obtained through training a machine learning model to optimize the recommendation effect. The fit coefficient comprehensively reflects the degree of matching between the user's cognitive expression ability and the difficulty of the picture book; a higher value indicates a better match.

[0054] It also calculates the dot product between the 3D psychological profile vector and the theme feature vector as the psychological theme fit. The dot product is used to measure the overall matching degree of two vectors in terms of direction and magnitude. When the two vectors are in the same direction and have a large magnitude, the dot product value is large, indicating that the user's psychological tendency is highly consistent with the theme of the picture book.

[0055] In addition, the vector components representing emotional states are extracted from the 3D psychological profile vector and the topic feature vector, respectively. Their standardized absolute differences are calculated and recorded as the emotional matching degree. The vector components representing emotional states refer to the components in the 3D psychological profile vector and the topic feature vector specifically used to describe the emotional dimension. The standardized absolute difference measures the absolute difference between the user's current emotional state and the emotional state expressed by the picture book, and is standardized to a value between 0 and 1, where 0 represents a perfect match and 1 represents a complete mismatch.

[0056] The psychological theme fit reflects the overall match between the user's psychological inclination and the picture book's theme, while the emotional fit specifically reflects the match between the user's emotional state and the picture book's emotional expression. Subsequently, the psychological theme fit and emotional fit are weighted and averaged according to a preset second weight to generate a resonance coefficient. This preset second weight is used to balance the contributions of psychological theme fit and emotional fit in the resonance coefficient calculation; it can be adjusted based on user feedback or expert experience to emphasize the user's identification with the overall theme or resonance with specific emotions. The resonance coefficient comprehensively reflects the degree of emotional resonance between the user's psychological inclination and the picture book's theme; a higher value indicates a stronger resonance.

[0057] The first weighted component is obtained by multiplying the fit coefficient by the ability fit weight, and the second weighted component is obtained by multiplying the resonance coefficient by the emotional resonance weight. The ability fit weight and emotional resonance weight are used to adjust the relative importance of the match between the user's cognitive ability and the picture book's difficulty, and the degree of resonance between the user's psychological inclination and the picture book's theme, in the final recommendation decision. These are parameters that are dynamically adjusted based on user interaction data and behavioral evolution to adapt to users' different needs for cognitive challenge and emotional satisfaction at different stages.

[0058] For example, when a user is judged to be more inclined towards cognitive development, the weight of ability fit may be increased; when a user values ​​emotional experience more, the weight of emotional resonance may be increased. Furthermore, the product of the fit coefficient and the resonance coefficient is multiplied by a preset synergy coefficient to obtain a synergy gain component. The preset synergy coefficient is used to quantify and amplify the positive synergistic effect when cognitive fit and emotional resonance coexist. When a user cognitively matches a picture book and also resonates emotionally, this coefficient can further increase the recommendation priority, thereby identifying picture books that can provide a comprehensive and high-quality experience. This coefficient is also dynamically adjusted according to the evolutionary trend of the user's state.

[0059] Finally, the first weighted component, the second weighted component, and the synergistic gain component are all summed to obtain the total matching decision score. This score comprehensively considers the match between the user's ability and the picture book's difficulty, the resonance between the user's psychology and the picture book's theme, and the synergistic effect between the two, serving as the basis for the final recommendation decision. Ultimately, candidate picture books whose total matching decision score exceeds the recommendation threshold are selected as target picture books. The recommendation threshold is a preset value used to filter out picture books that sufficiently match the user's needs. This threshold can be dynamically adjusted or optimized based on historical recommendation performance, user satisfaction, and other indicators.

[0060] This solution also provides a refined and dynamic picture book recommendation decision-making mechanism. By introducing difficulty feature vectors and theme feature vectors, the intrinsic attributes of picture books are quantitatively matched with user profiles, ensuring an objective basis for recommendations. By calculating the fit coefficient and resonance coefficient, the degree of matching between users and picture books is evaluated from two core dimensions: cognitive ability and psychological emotion, respectively, avoiding the limitations of single-dimensional recommendations.

[0061] By incorporating dynamically adjusted competency fit weights, emotional resonance weights, and preset synergy coefficients into the calculation of the total matching decision score, the recommendation strategy can be flexibly adjusted according to the user's real-time changing state. For example, when a user's cognitive ability improves, the competency fit weight can be appropriately increased to recommend more challenging picture books; when a user's emotional needs are strong, the emotional resonance weight can be increased to recommend picture books that are more likely to touch their heart.

[0062] In particular, the introduction of the synergistic gain component can capture the chemical reaction when both cognitive and emotional are matched, effectively identifying high-quality picture books that can satisfy users' cognitive development and evoke strong emotional resonance. This significantly improves the accuracy of recommendations and user satisfaction, solving the problem that static matching based solely on user profiles and picture book features may lead to recommendations that are not personalized enough and cannot adapt to dynamic changes in users.

[0063] Furthermore, the specific update process of the preset synergy coefficient is as follows: Within the evolution time window, three-dimensional capability profile vectors and three-dimensional psychological profile vectors are acquired at preset time intervals to obtain a temporal three-dimensional capability profile vector sequence and a temporal three-dimensional psychological profile vector sequence; the temporal three-dimensional capability profile vector sequence and the temporal three-dimensional psychological profile vector sequence are subjected to first-order difference operations to obtain capability change direction sub-vectors and psychological change direction sub-vectors; the weighted average vector of capability change direction sub-vectors and psychological change direction sub-vectors on the evolution time window is calculated to obtain the evolution direction vector; the Euclidean distance between three-dimensional capability profile vectors in adjacent preset time intervals and the Euclidean distance between three-dimensional psychological profile vectors in adjacent preset time intervals are calculated in the temporal three-dimensional capability profile vector sequence and the temporal three-dimensional psychological profile vector sequence to obtain two rate sub-sequences; the values ​​of the two rate sub-sequences are normalized and then added to obtain the evolution rate scalar; when the vector directions of the evolution direction vectors are all positive and the evolution rate scalar is within the adaptive threshold range, it is judged to be a state adaptation adjustment mode, and the preset synergy coefficient is updated accordingly.

[0064] In this embodiment, within the evolution time window, the three-dimensional capability profile vector and three-dimensional psychological profile vector of the target user at different times will be continuously acquired at preset time intervals.

[0065] The evolution time window can be a preset fixed duration, such as 10 minutes or the duration of a live session, used to capture short-term dynamics of user status.

[0066] The preset time interval determines the frequency of data sampling, such as acquiring data every 5 seconds or 1 minute, to balance real-time performance and computational overhead. In this way, sequential data reflecting the changes in user capabilities and psychological states over time can be constructed, namely, time-series 3D capability profile vector sequences and time-series 3D psychological profile vector sequences, laying the foundation for subsequent dynamic analysis.

[0067] Based on this, in order to capture the instantaneous changes in user status, the temporal three-dimensional capability profile vector sequence and the temporal three-dimensional psychological profile vector sequence are subjected to first-order difference operations respectively.

[0068] First-order difference operations can obtain a series of sub-vectors of ability change direction and psychological change direction by calculating the difference between adjacent vectors in the sequence.

[0069] Each sub-vector representing the direction of change in ability indicates the direction and magnitude of changes in a user's language ability, emotional expression, and cultural cognition at adjacent points in time.

[0070] Similarly, each sub-vector representing the direction of psychological change reflects the instantaneous changes in the user's emotional state, value orientation, and role projection. These sub-vectors provide detailed information on the micro-dynamics of the user's state.

[0071] To obtain the overall evolutionary trend of user status, the sub-vectors of ability change direction and psychological change direction are weighted and averaged over the evolutionary time window to obtain the evolutionary direction vector. The weighted average can assign different weights to different sub-vectors based on time proximity or other strategies to more accurately reflect the current trend. This evolutionary direction vector integrates the micro-changes in user ability and psychological state, representing the macro-development trend of the user's cognitive expression ability and psychological tendencies throughout the entire evolutionary time window. For example, when all components of the evolutionary direction vector are positive, it usually means that the user is showing positive growth or increased interest in multiple dimensions.

[0072] To quantify the rate of change in user states, the Euclidean distances between adjacent 3D capability profile vectors and adjacent 3D psychological profile vectors in the time-series 3D capability profile vector sequence are calculated. Euclidean distance intuitively measures the difference between two points in a multidimensional space; here, it is used to represent the rate of change of user capabilities and psychological states within adjacent time intervals, resulting in two rate subsequences. Subsequently, the values ​​of these two rate subsequences are normalized to eliminate dimensional differences, and then summed to obtain a single evolution rate scalar. This scalar comprehensively reflects the overall rate of change of user capabilities and psychological states within the evolutionary time window.

[0073] When all components of the evolution direction vector are positive and the evolution rate scalar is within the adaptive threshold range, the user is determined to be in state adaptation adjustment mode. The fact that all components of the evolution direction vector are positive indicates that the user is exhibiting a positive and upward development trend across all key capabilities and psychological dimensions. The adaptive threshold range is a dynamically adjusted interval used to determine whether the evolution rate scalar is in a moderate, stable state—neither too fast nor too slow. Only when the user state meets this positive and stable condition is it considered the optimal time to update the preset coordination coefficient, and the coordination coefficient is updated accordingly.

[0074] This solution also provides a more refined and dynamic preset collaborative coefficient update mechanism. By continuously monitoring and generating a time-series profile vector sequence within the evolution time window, and performing first-order difference operations on it to capture instantaneous changes, the evolution direction vector and evolution rate scalar of user capabilities and psychological states are accurately quantified through weighted averaging and Euclidean distance calculations.

[0075] This meticulous dynamic analysis enables accurate identification of whether a user is in a positive and stable state, allowing for appropriate adjustment. The preset synergy coefficients are only updated when the user is in this state—meaning their abilities and psychological state are trending positively and changing at a moderate rate. This avoids inappropriate coefficient adjustments when the user's state is unstable or changing too quickly / too slowly, ensuring more accurate and effective updates to the synergy coefficients. Ultimately, this helps improve the accuracy of the overall matching decision score, enabling picture book recommendations to respond more promptly and accurately to the user's current learning and psychological needs, significantly enhancing the personalization of recommendations and the user experience.

[0076] Furthermore, the specific update process of the preset synergy coefficient also includes: taking the angle between the calculated evolution direction vector and the preset positive reference vector as input, processing it through the arctangent function to obtain the basic adjustment factor; calculating the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold, and processing this difference through the negative exponential function to obtain the rate stability decay factor; multiplying the basic adjustment factor and the rate stability decay factor, and using their product as the increment of the synergy coefficient in this calculation; after obtaining the increment, adding the preset synergy coefficient to the increment to obtain the preliminary update value; if the preliminary update value is lower than the lower limit of the range limiting function, the range limiting function outputs the lower limit value as the output result; if the preliminary update value is higher than the upper limit of the range limiting function, the range limiting function outputs the upper limit value as the output result; if the preliminary update value is between the upper and lower limits of the range limiting function, the preliminary update value is used as the output result; and updating the preset synergy coefficient with the output result.

[0077] In this embodiment, the angle between the calculated evolution direction vector and a preset positive reference vector is used as input, and processed through the arctangent function to obtain the basic adjustment factor. This is used to quantify the degree of alignment between the evolution direction of the three-dimensional ability profile vector sequence (for the target user's cognitive expression ability) and the desired positive growth direction of the three-dimensional psychological profile vector sequence (for psychological tendencies).

[0078] A positive baseline vector is predefined, representing the ideal trend of user ability and mental health development. For example, in three-dimensional space, this could be a vector with all components being positive. By calculating the angle between the current evolution direction vector and this positive baseline vector, the consistency between their directions can be intuitively reflected. The smaller the angle, the closer the evolution direction is to the ideal positive trend. This angle is used as input to the arctangent function, which maps the angle value to a continuous factor with specific gain characteristics—the base adjustment factor. For example, when the angle is 0, the arctangent function outputs its maximum value, representing a completely positive trend; as the angle increases, the arctangent function output gradually decreases. This ensures that the closer the user's evolution direction matches the ideal positive trend, the larger the base adjustment factor, providing a stronger positive driving force for subsequent adjustments to the synergy coefficient.

[0079] Simultaneously, the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold is calculated. This difference is then processed through a negative exponential function to obtain a rate stability decay factor, which is used to assess the stability of the target user's ability and psychological evolution rate, i.e., whether its rate of change is within an ideal and stable range. First, the absolute difference between the evolution rate scalar and the preset adaptive threshold midpoint is calculated. The adaptive threshold midpoint represents the most suitable user evolution rate. The smaller the absolute difference, the closer the user evolution rate is to the ideal stable state. Subsequently, this absolute difference is processed by a negative exponential function. The negative exponential function has the characteristic of rapidly decaying as the input value increases. Therefore, when the evolution rate scalar is closer to the adaptive threshold midpoint, the smaller the absolute difference, the larger the rate stability decay factor output by the negative exponential function, indicating a more stable evolution rate; conversely, when the evolution rate deviates more from the midpoint, the decay factor is smaller, thus introducing a suppressive effect in the subsequent adjustment of the coordination coefficient to avoid over-adjustment under unstable or excessively fast / slow evolution rates.

[0080] Based on this, the basic adjustment factor is multiplied by the rate stability decay factor, and the product is used as the increment of the synergy coefficient in this calculation. This step comprehensively considers both the positivity of the user's evolutionary direction and the stability of the evolution rate to determine the adjustment magnitude of the preset synergy coefficient. The basic adjustment factor reflects the degree of alignment between the user's evolutionary direction and the ideal positive trend, while the rate stability decay factor reflects the stability of the evolution rate. Multiplying these two factors ensures that the synergy coefficient can only achieve a significant increment when the user's evolutionary direction is positive and the evolution rate is stable.

[0081] For example, even if the evolutionary direction is very positive, if the evolutionary rate is extremely unstable, the increment will be correspondingly smaller, avoiding aggressive adjustments to the coordination coefficient when user states fluctuate significantly. Conversely, if the evolutionary direction is unclear, even if the rate is stable, the increment will be small. This multiplicative combination mechanism makes the adjustment of the coordination coefficient more prudent and reasonable.

[0082] After obtaining the increment, the preset coordination coefficient is added to the increment to obtain a preliminary update value. After calculating the increment of the coordination coefficient, it is directly added to the current preset coordination coefficient to obtain a preliminary update value. This step is the basis for the actual adjustment of the coordination coefficient, reflecting a strategy of dynamic adaptation based on the user's current evolutionary state.

[0083] If the initial update value is lower than the lower limit of the range limiting function, the range limiting function outputs the lower limit value as the output result. If the initial update value is higher than the upper limit of the range limiting function, the range limiting function outputs the upper limit value as the output result. If the initial update value is between the upper and lower limits of the range limiting function, the initial update value is used as the output result. The output result updates the preset coordination coefficient. This ensures that the update of the preset coordination coefficient is always kept within a reasonable and effective range, preventing it from being too high or too low, thereby avoiding unreasonable deviations in the picture book recommendation results.

[0084] The range-limiting function sets a minimum and maximum value for the cooperability coefficient. If the initial update value exceeds this preset range, it is forcibly limited to the corresponding boundary values. For example, if the initial update value is less than the lower limit, the lower limit value is used; if it is greater than the upper limit, the upper limit value is used. If the initial update value is between the upper and lower limits, the initial update value is directly adopted. Through this limiting process, the dynamic range of the cooperability coefficient can be effectively controlled, ensuring its stability and effectiveness in recommendations and avoiding imbalances in recommendation performance due to extreme values.

[0085] This solution also provides a more refined and intelligent pre-set synergy coefficient update mechanism. By introducing the angle between the evolution direction vector and a pre-set positive reference vector to quantify directional fit, and using the arctangent function to generate a basic adjustment factor, the adjustment of the synergy coefficient can more sensitively respond to the degree of positive growth trends in users. Simultaneously, by calculating the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold, and combining it with a negative exponential function to generate a rate stability decay factor, the stability of the user's evolution rate is effectively assessed, avoiding inappropriate adjustments when user states fluctuate significantly. Multiplying these two factors as the increment of the synergy coefficient ensures that the update of the synergy coefficient is based on a comprehensive consideration of the user's positive and stable evolution, thus making the adjustment more prudent and reasonable. Finally, a range limiting function constrains the initial update value, ensuring that the synergy coefficient always remains within an effective and reasonable range. This refined update strategy enables more accurate capture of subtle changes in the target user's cognitive expression ability and psychological tendencies, thereby achieving a more precise and dynamic synergy effect in the personalized book recommendation process during picture book live streaming, significantly improving the personalization of recommendations and user satisfaction.

[0086] Furthermore, the specific adjustment process for the ability fit weight and emotional resonance weight is as follows: By performing cluster analysis on the temporal three-dimensional psychological profile vector sequence, the average Euclidean distance from all vector points to their cluster centroids is calculated. After normalizing the average Euclidean distance, it is used as a scalar of state discreteness. Two components corresponding to the language ability and emotional expression dimensions of the three-dimensional ability profile vector are extracted from the evolutionary direction vector. The root mean square of the sum of the squares of these two components is calculated to obtain the intensity of positive change in ability. Two components corresponding to the value orientation and role projection dimensions of the three-dimensional psychological profile vector are extracted from the evolutionary direction vector. The absolute values ​​of these two components are calculated. The sum of these factors yields the intensity of psychological orientation change. Multiplying the intensity of positive change in ability by the intensity of psychological orientation change, and then taking the square root of the product, yields the expression coordination factor, which represents the coordination between cognition and psychology. The ratio of the state dispersion scalar to the expression coordination factor is calculated to obtain the attention density coefficient. The emotional resonance weight, the state dispersion scalar, and the attention density coefficient are multiplied to obtain the adjustment amount. The adjustment amount is added to the emotional resonance weight to obtain a new emotional resonance weight. Based on the new emotional resonance weight and the ability fit weight, a normalization operation is performed again to obtain the adjusted ability fit weight and the adjusted emotional resonance weight.

[0087] In this embodiment, cluster analysis is performed on the time-series three-dimensional psychological profile vector sequence to calculate the average Euclidean distance from all vector points to their cluster centroids. After normalizing the average Euclidean distance, it is used as a state discreteness scalar, aiming to quantify the stability and volatility of the target user's psychological state over a period of time.

[0088] In implementation, clustering algorithms such as K-means, DBSCAN, or Gaussian mixture models can be used to cluster the temporal three-dimensional psychological profile vector sequence within a preset time window to identify typical patterns or central trends in user psychological states. Subsequently, the Euclidean distance from each vector point to the centroid of its cluster is calculated, and the average of these distances is taken to reflect the degree of concentration or dispersion of psychological states. Finally, this average Euclidean distance is mapped to a preset numerical range using methods such as min-max normalization or Z-score normalization to obtain a state dispersion scalar. The higher the scalar value, the greater the volatility of the user's psychological state.

[0089] Two components corresponding to the language ability and emotional expression dimensions of the three-dimensional ability profile vector are extracted from the evolution direction vector. The square root of the sum of the squares of these two components is calculated to obtain the intensity of positive change in ability, which is used to assess the intensity of the positive development trend of users in the two key cognitive dimensions of language ability and emotional expression.

[0090] The evolutionary direction vector reflects the overall direction of change in a user's abilities and psychological state. By identifying and extracting components related to the language ability and emotional expression dimensions from this vector, we can focus on specific aspects of a user's cognitive development. Subsequently, we perform square and root mean square operations on these two components, i.e., calculate the Euclidean norm of the sub-vectors they form, to obtain a comprehensive quantitative value that represents the intensity of the user's positive changes in these two ability dimensions.

[0091] This process extracts two components from the evolutionary direction vector, corresponding to the value orientation and role projection dimensions of the three-dimensional psychological profile vector. The sum of the absolute values ​​of these two components yields the intensity of psychological orientation change, aiming to measure the magnitude of change in the user's value orientation and role projection dimensions. Similar to the intensity of positive change in ability, this process extracts the components corresponding to the value orientation and role projection dimensions from the evolutionary direction vector.

[0092] Since psychological changes can involve directionality, the sum of the absolute values ​​of these two components can be used to comprehensively reflect the overall intensity of changes in these two psychological dimensions without distinguishing whether the evolution is positive or negative.

[0093] Multiplying the intensity of positive changes in ability with the intensity of changes in psychological orientation, and then taking the square root of the product, yields the expression coordination factor, which represents the coordinated changes in cognition and psychology. Its function is to comprehensively assess the coordination and consistency between the development of users' cognitive abilities and changes in their psychological tendencies.

[0094] By multiplying the intensity of positive changes in ability by the intensity of changes in psychological orientation and taking the square root, a balanced index can be obtained. This index reflects whether a user's development at both the cognitive and psychological levels is synchronized and mutually reinforcing. A higher expression coordination factor indicates that the user's cognitive and psychological development is more coordinated, which contributes to more stable and effective learning.

[0095] The ratio of the state dispersion scalar to the expression coordination factor is used to obtain the attention density coefficient, which comprehensively considers the stability of the user's mental state and cognitive psychological coordination to infer the user's current level of focus or learning readiness. This coefficient provides a comprehensive indicator by calculating the ratio of the state dispersion scalar to the expression coordination factor. For example, when the state dispersion is low and the expression coordination factor is high, the attention density coefficient may be high, indicating that the user is in a more focused and better-suited learning state to receive new information.

[0096] The adjustment amount is obtained by multiplying the emotional resonance weight, the state dispersion scalar, and the attention density coefficient. This adjustment aims to dynamically calculate the specific value for adjusting the emotional resonance weight based on the user's current complex psychological and cognitive state. This adjustment amount comprehensively considers the current emotional resonance weight, the fluctuation of the user's psychological state, and the coordination between cognition and psychological development, making the adjustment process more intelligent and refined.

[0097] The emotional resonance weight is adjusted by adding an adjustment amount to obtain a new emotional resonance weight. Then, a normalization operation is performed based on this new emotional resonance weight and the ability fit weight to obtain the adjusted ability fit weight and the adjusted emotional resonance weight, ensuring the effectiveness and internal balance of the weight adjustment. After obtaining the new emotional resonance weight, to maintain the relative importance of the ability fit weight and the emotional resonance weight in the recommendation decision and to ensure that their sum conforms to a preset proportional relationship, both need to be re-normalized. This ensures that the ability fit weight is adjusted accordingly while the emotional resonance weight is adjusted, thus maintaining the logical consistency within the recommendation model.

[0098] Through its technical solution, this application enables more precise adjustment of the weights for ability compatibility and emotional resonance, overcoming the limitations of coarse adjustments based solely on the evolution rate scalar. Specifically, by performing cluster analysis on the temporal three-dimensional psychological profile vector sequence and calculating the state dispersion scalar, the stability of the user's psychological state can be accurately assessed, avoiding inappropriate weight adjustments when the user's psychological state fluctuates significantly. Simultaneously, by extracting key components from the evolution direction vector, calculating the intensity of positive changes in ability and the intensity of changes in psychological orientation, and further generating an expression coordination factor, this application can quantify the degree of coordination between cognition and psychological changes.

[0099] By combining the state dispersion scalar with the expression coordination factor to generate the attention density coefficient, the weight adjustment not only considers the stability of the user's psychology but also incorporates the coordination between cognitive and psychological development, thus generating a more adaptive adjustment amount. Finally, by precisely adjusting the emotional resonance weight and renormalizing it with the ability fit weight, the recommendation can more sensitively and accurately respond to subtle dynamics in the user's cognitive development and emotional psychological needs, significantly improving the personalization and accuracy of picture book recommendations and enabling users to obtain picture book content highly matched to their current overall state.

[0100] Furthermore, the specific adjustment process for the ability fit weight and emotional resonance weight includes: within a statistical time window, counting the number of times the ability fit weight and emotional resonance weight are adjusted together to obtain an upward adjustment count; inputting this upward adjustment count into a preset saturation growth function for mapping, the function outputs a dynamic upper limit value between 1 and a preset maximum allowable value; using the dynamic upper limit value to update the upper bound of the range limiting function; recording the moment when the ability fit weight and emotional resonance weight were last adjusted and starting the timer, from that moment on, after each statistical time window, the dynamic upper limit value is multiplied by a preset decay factor until its value decays to 1.

[0101] In this embodiment, Figure 3 This is a logical diagram illustrating the updating of the boundary of the range limiting function provided in an embodiment of this application.

[0102] Within the statistical time window, the frequency with which the competency fit weight and emotional resonance weight are adjusted simultaneously is continuously monitored and recorded, resulting in an upward adjustment count. This statistical time window can be a preset period, such as the most recent hour, day, or week, and its length can be configured according to the actual application scenario and the frequency of user behavior changes. The upward adjustment count reflects the responsiveness to changes in user behavior, that is, the frequency with which adjustments to user preferences are deemed necessary within a specific period.

[0103] The upward adjustment count is input to a preset saturation growth function for mapping. The saturation growth function is a mathematical model whose output value increases with the input value, but the growth rate gradually slows down and eventually approaches an upper limit. Here, the upward adjustment count is used as input, and through this function mapping, a dynamic upper limit value can be obtained. This dynamic upper limit value is used to limit the magnitude of subsequent weight adjustments, ensuring that the adjustments are responsive to user changes without increasing indefinitely. It lies between 1 and the preset maximum allowable value, meaning that even if adjustments are infrequent, there is at least a basic upper limit (1), while with frequent adjustments, the upper limit can be moderately relaxed, but will not exceed a preset maximum value to prevent runaway.

[0104] The upper bound of the range limiting function is updated using this dynamic upper bound. The range limiting function ensures that the competence fit weight and emotional resonance weight remain within a reasonable numerical range after adjustment. By using the calculated dynamic upper bound as the upper bound of this function, the maximum allowed weight value can be adaptively adjusted based on the activity level of user behavior changes. This means that when user behavior changes frequently and requires a larger adjustment of the weights, the upper bound can be appropriately increased; conversely, when changes are infrequent, the upper bound will be tightened, thereby enhancing stability and adaptability.

[0105] To prevent the dynamic upper limit from remaining too high after the active user behavior period ends, thus leading to over-adjustment during the subsequent stable period, this application also introduces a decay mechanism. Specifically, the last time the ability compatibility weight and emotional resonance weight were adjusted is recorded and a timer is started. From that moment on, after each statistical time window, the dynamic upper limit is multiplied by a preset decay factor until its value decays to 1. The decay factor is a positive number less than 1, and each decay brings the upper limit closer to 1 until it reaches 1. This ensures that as user behavior stabilizes, the upper limit of its weight adjustment gradually returns to a conservative state, avoiding unnecessary fluctuations and improving long-term stability.

[0106] The system can dynamically adjust the upper limits of the weights for ability compatibility and emotional resonance based on changes in user activity levels. This provides greater adjustment flexibility when user behavior is active, enhancing responsiveness and adaptability. Simultaneously, by introducing a decay mechanism, it ensures that the adjustment upper limits gradually revert to their previous values ​​as user behavior stabilizes, avoiding over-adjustment and instability. This dynamic and adaptive weight adjustment upper limit management mechanism allows the personalized book recommendation method for picture book live streaming to better balance the accuracy and stability of recommendations, thus providing more personalized picture book recommendations that align with the user's current cognitive and expressive abilities and psychological inclinations.

[0107] Furthermore, the saturated growth function is specifically as follows: ,in, The number of times the weights for ability compatibility and emotional resonance were adjusted. This is the preset maximum allowed value. The growth rate constant is preset. This is the natural exponential function, and its output value is... This is the calculated dynamic upper limit value.

[0108] In this embodiment, the saturated growth function model describes a growth process in which the growth rate gradually slows down as the independent variable increases, eventually approaching an upper limit value. In this application, it maps the number of weight adjustments, x, to a dynamic upper limit value f(x). This upper limit value grows smoothly and non-linearly with the number of adjustments, but it does not grow indefinitely; instead, it gradually approaches a preset maximum allowable value, MaxLimit. This functional form can simulate the diminishing marginal utility phenomenon in the learning or adaptation process, that is, as experience accumulates, the impact of each adjustment on the upper limit value gradually decreases.

[0109] x, as the independent variable, represents the cumulative frequency or experience of adjusting the weights of ability compatibility and emotional resonance. It serves as input to the saturation growth function, quantifying the degree of validation for adjustments made in a specific direction.

[0110] The more adjustments are made, the more the effectiveness of the current adjustment strategy is validated, thus allowing for greater growth potential in the dynamic upper limit. MaxLimit is the upper limit of the saturation growth function, representing the maximum value that the dynamic upper limit can reach.

[0111] It is a preset constant used to limit the unlimited growth of the dynamic upper limit value, ensuring that a certain stability boundary is maintained during adaptive adjustment, preventing the upper limit value of weight adjustment from being too high, thereby avoiding over-adaptation or introducing instability.

[0112] k is a positive parameter in the saturated growth function, determining the steepness or speed of the function's growth. A larger k value results in a faster function reaching MaxLimit, meaning the dynamic upper limit is more sensitive to the number of adjustments; a smaller k value results in a smoother function growth and a slower response. This is used to finely adjust the rate at which the dynamic upper limit increases with the number of adjustments, adapting to different application scenarios and response speed requirements.

[0113] exp is the natural exponential function in mathematics, which is e raised to the power of x. In this function, the term exp(−k∗x) ensures that as x increases, the value of this term rapidly decays from close to 1 and approaches 0, thus causing the entire function f(x) to gradually increase from close to 1 and approach MaxLimit.

[0114] It is the core mathematical component for achieving saturated growth characteristics. f(x) is the output of the saturated growth function, representing the dynamic upper limit of the range limit function calculated based on the current number of weight adjustments x. This value changes dynamically, reflecting the accumulation of confidence in the weight adjustment strategy and providing a wider or more restricted range for subsequent weight adjustments.

[0115] By calculating the dynamic upper limit value using a saturated growth function, this application ensures that the upper limit value of the range limiting function grows non-linearly in a smooth and controlled manner as the ability fit weight and emotional resonance weight are adjusted more times. This growth pattern avoids the upper limit value from expanding rapidly in a linear or unlimited manner, thus effectively preventing problems such as insufficient adjustment in the early stage or excessive adjustment in the later stage.

[0116] Specifically, when the number of adjustments is small, the upper limit increases rapidly, enabling quick adaptation to new adjustment strategies. However, when the number of adjustments increases, the growth rate of the upper limit gradually slows down and approaches the preset maximum allowable value, MaxLimit. This reflects the accumulation of confidence in the current adjustment strategy and the need for stability, avoiding instability caused by over-reliance on adjustments.

[0117] Meanwhile, by using a preset growth rate constant k, the sensitivity of the upper limit growth can be flexibly adjusted, enabling the best balance between response speed and stability to be achieved according to the needs of actual application scenarios, thereby improving the robustness and adaptability of the entire picture book recommendation in long-term operation.

[0118] Figure 4 This is a schematic diagram of the structure of an AI-based personalized book recommendation system for live-streaming picture books provided in this application embodiment. The system includes: a data acquisition module for acquiring interaction data of the target user, including text data, voice feature data, and facial expression feature data; a vector construction module for generating a three-dimensional ability profile vector sequence representing the target user's current cognitive expression ability and a three-dimensional psychological profile vector sequence representing the target user's current psychological tendency based on the interaction data; a time-series analysis module for performing time-series analysis on the three-dimensional ability profile vector sequence and the three-dimensional psychological profile vector sequence to calculate the evolution direction vector and the evolution rate scalar; and a coordination coefficient acquisition module for determining whether the evolution rate scalar is within an adaptive threshold when all components of the evolution direction vector are positive. Within the specified range, if yes, the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold is calculated, and the preset synergy coefficient is updated by comparing the boundary value of the range limiting function. The judgment module is used to ensure no adjustment is made if the judgment result is no and the evolution rate scalar is below the lower limit of the adaptive threshold range. The weight adjustment module is used to adjust the ability fit weight and emotional resonance weight if no, and the evolution rate scalar exceeds the upper limit of the adaptive threshold range. The statistics module records the adjustment counts of the ability fit weight and emotional resonance weight, and updates the boundary value of the range limiting function accordingly. The total score calculation module calculates the total matching decision score using the matching decision formula based on the updated preset synergy coefficient, the adjusted ability fit weight, and the emotional resonance weight. The output module selects at least one target picture book based on the total matching decision score and outputs it.

[0119] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0124] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based personalized book recommendation method for picture book live streaming, characterized by, Includes the following steps: Acquire interaction data from the target user, which includes text data, voice feature data, and facial expression feature data; Based on interactive data, a three-dimensional ability profile vector sequence representing the target user's current cognitive expression ability and a three-dimensional psychological profile vector sequence representing the target user's current psychological tendency are generated. Temporal analysis was performed on the three-dimensional ability profile vector sequence and the three-dimensional psychological profile vector sequence to calculate the evolution direction vector and the evolution rate scalar. When all components of the evolution direction vector are positive, determine whether the evolution rate scalar is within the adaptive threshold range. If so, calculate the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold and update the preset synergy coefficient by combining the comparison result of the boundary value of the range limiting function. If not, and the evolution rate scalar is below the lower limit of the adaptive threshold range, then no adjustment is made; If not, and the evolution rate scalar exceeds the upper limit of the adaptive threshold range, then adjust the ability fit weight and emotional resonance weight. Record the adjustment counts of the ability compatibility weight and emotional resonance weight, and update the boundary value of the range limiting function accordingly; Based on the updated preset synergy coefficient, the adjusted ability compatibility weight, and the emotional resonance weight, the total matching decision score is calculated using the matching decision formula. Select at least one target picture book based on the total matching decision score and output it.

2. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 1, characterized in that, The specific generation process of the three-dimensional capability profile vector is as follows: Extract narrative structure features and sentiment lexical features from text data; Extracting syntactic complexity features from speech feature data; Narrative structure features, emotional vocabulary features, and syntactic complexity features are input into a preset feature dimension mapping model. The feature dimension mapping model outputs a score representing the target user on three preset orthogonal ability dimensions, including language ability dimension, emotional expression dimension, and cultural cognition dimension. The scores on the three orthogonal ability dimensions are normalized, and a three-dimensional vector is constructed in the preset orthogonal three-dimensional space to serve as the three-dimensional ability profile vector.

3. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 1, characterized in that, The specific generation process of the three-dimensional psychological profile vector is as follows: The tone and speed parameter sequences in the speech feature data are fused with the facial action unit parameter sequences in the expression feature data, and then input into a preset multimodal emotion recognition model to output the current emotion state category of the target user. Semantic analysis and keyword extraction are performed on the text data, and it is matched with a preset value proposition thesaurus to identify the core value orientation category contained in the target user's expression; The emotional state category, the core value orientation category, and the topic tag data associated with the live broadcast interaction are input into a preset role classification model to output the role projection type of the target user in the current interaction. The emotional state category, the core value orientation category, and the role projection type are mapped to initial values ​​for the three dimensions of emotional state intensity, value orientation intensity, and role projection intensity, respectively, according to a preset quantitative mapping table. The initial values ​​of the three dimensions are normalized, and a three-dimensional vector is constructed in a preset orthogonal three-dimensional space, which serves as the three-dimensional psychological profile vector.

4. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 1, characterized in that, The specific process for obtaining the target picture book is as follows: For candidate picture books in the picture book database, obtain the pre-stored difficulty feature vector and theme feature vector corresponding to the candidate picture books; The difficulty feature vector is composed of preset difficulty level values ​​for candidate picture books in three dimensions: language ability, emotional expression, and cultural cognition; The theme feature vector is composed of preset theme intensity values ​​of candidate picture books in three dimensions: emotional state, value orientation, and role projection; Calculate the absolute value of the difference between the cosine similarity and the modulus length between the 3D ability profile vector and the difficulty feature vector, and then perform a weighted average of the absolute values ​​of the difference between the cosine similarity and the modulus length according to the preset first weight to generate the fit coefficient. The dot product between the 3D psychological profile vector and the topic feature vector is calculated as the psychological topic fit. The vector components representing emotional state in the 3D psychological profile vector and the topic feature vector are extracted respectively, and their standardized absolute difference is calculated and recorded as the emotional matching degree. The psychological theme fit and emotional match are weighted and averaged according to a preset second weight to generate a resonance coefficient. The first weighted component is obtained by multiplying the compatibility coefficient by the ability compatibility weight, the second weighted component is obtained by multiplying the resonance coefficient by the emotional resonance weight, and the synergy gain component is obtained by multiplying the product of the compatibility coefficient and the resonance coefficient by the preset synergy coefficient. The first weighted component, the second weighted component and the synergy gain component are all added together to obtain the total matching decision score. Candidate picture books whose total matching decision score exceeds the recommendation threshold are selected as target picture books.

5. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 4, characterized in that, The specific update process for the preset coordination coefficient is as follows: Within the evolution time window, three-dimensional capability profile vectors and three-dimensional psychological profile vectors are acquired at preset time intervals to obtain temporal three-dimensional capability profile vector sequences and temporal three-dimensional psychological profile vector sequences. Perform first-order difference operations on the temporal three-dimensional ability profile vector sequence and the temporal three-dimensional psychological profile vector sequence to obtain the ability change direction sub-vector and the psychological change direction sub-vector respectively. The weighted average vector of the sub-vectors of ability change direction and psychological change direction over the evolution time window is calculated to obtain the evolution direction vector; Calculate the Euclidean distance between the three-dimensional ability profile vectors at adjacent preset time intervals and the Euclidean distance between the three-dimensional mental profile vectors at adjacent preset time intervals in the temporal three-dimensional ability profile vector sequence and the temporal three-dimensional mental profile vector sequence respectively to obtain two rate subsequences; normalize the values ​​of the two rate subsequences and add them together to obtain the evolution rate scalar. When all the evolution direction vectors are positive and the evolution rate scalar is within the adaptive threshold range, it is determined to be a state adaptation adjustment mode, and the preset cooperative coefficient is updated accordingly.

6. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 5, characterized in that, The specific update process of the preset coordination coefficient also includes: The basic adjustment factor is obtained by taking the angle between the calculated evolution direction vector and the preset positive reference vector as input and processing it through the arctangent function. The absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold is calculated, and this difference is processed through a negative exponential function to obtain the rate stability decay factor. The product of the basic adjustment factor and the rate stability decay factor is used as the increment of the synergy coefficient in this calculation. After obtaining the increment, the preset synergy coefficient is added to the increment to obtain the preliminary update value; If the initial update value is lower than the lower limit of the range limiting function, the range limiting function outputs the lower limit value as the output result; if the initial update value is higher than the upper limit of the range limiting function, the range limiting function outputs the upper limit value as the output result. If the initial update value is between the upper and lower limits of the range limiting function, then the initial update value will be used as the output result. Update the preset synergy coefficient with the output results.

7. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 4, characterized in that, The specific adjustment process for the competence compatibility weight and emotional resonance weight is as follows: By performing cluster analysis on the temporal three-dimensional psychological profile vector sequence, the average Euclidean distance from all vector points to their cluster centroids is calculated. After normalizing the average Euclidean distance, it is used as a scalar of state discreteness. Extract the two components of the evolution direction vector corresponding to the language ability and emotional expression dimensions of the three-dimensional ability profile vector, calculate the root mean square of the sum of the squares of the two components, and obtain the intensity of positive change in ability. Extract the two components of the evolution direction vector corresponding to the value orientation and role projection dimensions of the three-dimensional psychological profile vector, calculate the sum of the absolute values ​​of the two components, and obtain the intensity of the psychological orientation change. Multiply the intensity of positive change in ability by the intensity of change in psychological orientation, and take the square root of the product to obtain the expression coordination factor that represents the coordinating change between cognition and psychology. The ratio of the state discreteness scalar to the expression coordination factor is calculated to obtain the attention density coefficient; The adjustment amount is obtained by multiplying the emotional resonance weight, the state dispersion scalar, and the attention density coefficient; Add the adjustment amount to the emotional resonance weight to obtain the new emotional resonance weight; The normalization operation is performed again based on the new emotional resonance weight and ability compatibility weight to obtain the adjusted ability compatibility weight and the adjusted emotional resonance weight.

8. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 7, characterized in that, The specific adjustment process for the competency compatibility weight and emotional resonance weight also includes: Within the statistical time window, the number of times the statistical ability compatibility weight and the emotional resonance weight are adjusted together is the upward adjustment count. The upward count is mapped to a preset saturation growth function, which outputs a dynamic upper limit value between 1 and the preset maximum allowable value. Update the upper bound of the range limiting function using a dynamic upper limit value; Record the moment when the ability compatibility weight and emotional resonance weight were last adjusted and start timing. From that moment on, for each statistical time window, the dynamic upper limit value is multiplied by a preset decay factor until its value decays to 1.

9. The method for personalized book recommendation in live picture book streaming based on artificial intelligence according to claim 8, characterized in that, The saturated growth function is specifically: ,in, The number of times the weights for ability compatibility and emotional resonance were adjusted. This is the preset maximum allowed value. The growth rate constant is preset. This is the natural exponential function, and its output value is... This is the calculated dynamic upper limit value.

10. A personalized book recommendation system for live-streamed picture book reading based on artificial intelligence, characterized in that: include: Data acquisition module: used to acquire the interaction data of the target user, which includes text data, voice feature data and facial expression feature data; Vector Construction Module: Used to generate a three-dimensional ability profile vector sequence representing the target user's current cognitive expression ability and a three-dimensional psychological profile vector sequence representing the target user's current psychological tendency based on interactive data; The time series analysis module is used to perform time series analysis on the three-dimensional ability profile vector sequence and the three-dimensional psychological profile vector sequence, and calculate the evolution direction vector and the evolution rate scalar. Synergy coefficient acquisition module: When all components of the evolution direction vector are positive, it determines whether the evolution rate scalar is within the adaptive threshold range. If so, it calculates the absolute difference between the evolution rate scalar and the midpoint of the adaptive threshold and updates the preset synergy coefficient by combining the comparison result of the boundary value of the range limiting function. Judgment module: If the judgment result is negative and the evolution rate scalar is lower than the lower limit of the adaptive threshold range, then no adjustment will be made; Weight adjustment module: If not, and the evolution rate scalar exceeds the upper limit of the adaptive threshold range, then adjust the ability fit weight and emotional resonance weight. Statistics module: Used to record the adjustment counts of ability compatibility weight and emotional resonance weight, and update the boundary values ​​of the range limiting function accordingly; Total score calculation module: It is used to calculate the total score of matching decision based on the updated preset synergy coefficient, the adjusted ability compatibility weight and the emotional resonance weight, through the matching decision formula; Output module: Used to select at least one target picture book based on the total score of the matching decision and output it.

Images