Learning resource recommendation method, system and device, medium and product

Abstract

The invention relates to the technical field of intelligent education, in particular to a learning resource recommendation method, system and device, a medium and a product. According to the method, basic information and historical learning behavior data of a target object are obtained, an initial interest map is constructed in combination with a graph neural network algorithm, emotion calculation is performed according to multi-modal emotion data of the target object, a comprehensive emotion vector is generated, and therefore the target object is obtained according to a first interest vector of the initial interest map and the comprehensive emotion vector. And determining a learning resource recommendation result so as to improve the accuracy and adaptability of learning resource recommendation and further improve the learning efficiency of target objects such as children.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent education, and in particular to a learning resource recommendation method, system, device, medium and product. BACKGROUND

[0002] At present, the traditional learning resource recommendation system used in the field of intelligent education matches the preset fixed learning content (for example, animation, reading, and standardization forms such as text and pictures) for children by collecting basic information such as the age and subject of the children, so as to realize the automatic distribution of learning resources.

[0003] However, the traditional learning resource recommendation system only collects rough labels such as the age and subject of the children, which easily leads to the single dimension of the user portrait and cannot capture the differentiated needs of different users. Moreover, the traditional system usually adopts fixed animation, reading, and standardization learning paths such as text and pictures, which are difficult to meet the personalized needs of users in terms of learning content and form. For example, in the aspect of region, the preferences of children in the north for ice and snow themes and children in the south for marine science popularization cannot be accurately identified; in the aspect of interest, the game interaction learning form that is popular among young children is also often ignored by the traditional learning recommendation system which mainly uses static text and pictures. In addition, the traditional learning recommendation system cannot perceive and analyze the emotional state of children in real time, which easily leads to a vicious cycle of "rigidity-declining interest-loss of tolerance", seriously affecting the learning efficiency of children.

[0004] Therefore, there is an urgent need for a learning resource recommendation method to improve the accuracy and adaptability of learning resource recommendation, so as to improve the learning efficiency of target objects such as children. SUMMARY

[0005] The present application provides a learning resource recommendation method, system, device, medium and product to improve the accuracy and adaptability of learning resource recommendation, so as to improve the learning efficiency of target objects such as children.

[0006] In a first aspect, the present application provides a learning resource recommendation method, which comprises:

[0007] obtaining basic information and historical learning behavior data of a target object; the basic information includes the region, age and gender of the target object, and the historical learning behavior data includes course completion degree, answer accuracy rate and knowledge point residence time length;

[0008] constructing an initial interest graph based on the basic information and the historical learning behavior data in combination with a graph neural network algorithm; each graph node in the initial interest graph represents different subject categories, knowledge point labels and learning styles, and the edge weight of each graph node represents the preference degree of the target object for different learning resources;

[0009] perform emotion calculation based on multi-modal emotion data of the target object to generate a comprehensive emotion vector; the multi-modal emotion data comprises visual data, voice data and text data, and the comprehensive emotion vector represents a comprehensive emotion state of the target object in multiple dimensions;

[0010] determine a learning resource recommendation result based on a first interest vector of the initial interest graph and the comprehensive emotion vector; the first interest vector represents comprehensive interest preferences of the target object in multiple dimensions, and the learning resource recommendation result comprises recommended content and at least one presentation form corresponding to the recommended content.

[0011] In a first aspect, the present application provides a learning resource recommendation system, which comprises:

[0012] an image construction module configured to acquire basic information and historical learning behavior data of a target object; the basic information comprises a region, an age and a gender of the target object, and the historical learning behavior data comprises course completion degree, question answering accuracy and knowledge point staying time length;

[0013] construct an initial interest graph based on the basic information and the historical learning behavior data in combination with a graph neural network algorithm; each graph node in the initial interest graph represents different subject categories, knowledge point labels and learning styles, and edge weights of the graph nodes represent preference degrees of the target object for different learning resources;

[0014] a recommendation execution module configured to perform emotion calculation based on multi-modal emotion data of the target object to generate a comprehensive emotion vector; the multi-modal emotion data comprises visual data, voice data and text data, and the comprehensive emotion vector represents a comprehensive emotion state of the target object in multiple dimensions;

[0015] determine a learning resource recommendation result based on a first interest vector of the initial interest graph and the comprehensive emotion vector; the first interest vector represents comprehensive interest preferences of the target object in multiple dimensions, and the learning resource recommendation result comprises recommended content and at least one presentation form corresponding to the recommended content.

[0016] Optionally, the image construction module is specifically configured to:

[0017] map interaction records in the historical learning behavior data into a plurality of nodes, and calculate edge weights between each two nodes based on frequencies of co-occurrence in the interaction records;

[0018] initialize initial weights of each node based on the basic statistical information;

[0019] Based on the initial weight and the edge weight, node embedding generation and neighbor node aggregation and are performed to construct the initial interest graph.

[0020] Optionally, the recommendation execution module is specifically configured to:

[0021] The multi-modal emotional data are respectively subjected to feature extraction to obtain respective feature vectors corresponding to the visual data, the speech data and the text data;

[0022] Based on a preset attention mechanism, the feature vectors are subjected to weighted fusion to obtain the comprehensive emotional vector.

[0023] Optionally, the portrait framework module is further configured to:

[0024] Based on the historical learning behavior data, the weight of each node in the initial interest graph is determined;

[0025] Based on the weight of each node, the embedding vectors of the plurality of nodes are subjected to weighted summation to generate the first interest vector.

[0026] Optionally, the recommendation execution module is specifically configured to:

[0027] Based on the first interest vector and the label vector of each candidate content in the candidate learning resource set, a corresponding content fit degree is determined;

[0028] Based on the comprehensive emotional vector, an emotional gain factor is calculated;

[0029] Based on the emotional gain factor, the content fit degree is subjected to gain adjustment to determine a target interest score;

[0030] Based on the target interest score of each candidate content, a preset number of candidate contents are selected from high to low as the recommended contents;

[0031] Based on the comprehensive emotional vector, a preset emotional form mapping rule is combined to determine at least one presentation form matched with the recommended contents from a form combination strategy set.

[0032] Optionally, the system further comprises a feedback optimization module configured to:

[0033] Based on new learning behavior data and new multi-modal emotional data generated by the target object in response to the learning resource recommendation result, a new comprehensive emotional vector is generated;

[0034] Based on the new comprehensive emotional vector and a preset dynamic update model, the first interest vector is updated to obtain an updated second interest vector;

[0035] The second interest vector is taken as a first interest vector for next learning resource recommendation.

[0036] Optionally, the feedback optimization module is specifically configured to:

[0037] Based on the new learning behavior data and the new comprehensive emotion vector, a reward value is calculated in combination with a preset reinforcement learning algorithm, and the reward value represents effectiveness of the learning resource recommendation result.

[0038] Based on the reward value, weights in the first interest vector in multiple dimensions are adjusted to obtain the updated second interest vector.

[0039] In a second aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements any one of the learning resource recommendation methods in the first aspect when executing the computer program.

[0040] In a third aspect, the present application provides a computer storage medium, and the computer readable storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement any one of the learning resource recommendation methods in the first aspect.

[0041] In a fourth aspect, the present application provides a computer program product, comprising computer program instructions, and the computer program instructions are executed by a processor to implement any one of the learning resource recommendation methods in the first aspect.

[0042] The present application has the following advantages:

[0043] The present application provides a learning resource recommendation method, system, device, medium and product. The method obtains basic information and historical learning behavior data of a target object, constructs an initial interest graph in combination with a graph neural network algorithm, performs emotion calculation according to multi-modal emotion data of the target object to generate a comprehensive emotion vector, and determines a learning resource recommendation result according to a first interest vector of the initial interest graph and the comprehensive emotion vector, so as to improve the accuracy and adaptability of learning resource recommendation and further improve the learning efficiency of the target object such as a child. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the following will briefly introduce the drawings needed to be used in the embodiment or related art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0045] Figure 1A flowchart of a learning resource recommendation method provided by an embodiment of the present application is shown in FIG. 1.

[0046] Figure 2 A schematic diagram of a construction process of an initial interest graph provided by an embodiment of the present application is shown in FIG. 2.

[0047] Figure 3 A schematic diagram of a structure of a learning resource recommendation system provided by an embodiment of the present application is shown in FIG. 3.

[0048] Figure 4 A schematic diagram of a system model and flow provided by an embodiment of the present application is shown in FIG. 4.

[0049] Figure 5 A schematic diagram of a system analysis flow provided by an embodiment of the present application is shown in FIG. 5.

[0050] Figure 6 A schematic diagram of a structure of a computer device provided by an embodiment of the present application is shown in FIG. 6. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application. In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other at will. And, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be performed in an order different from that here.

[0052] The terms "first" and "second" in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the term "comprising" and any variations thereof are intended to cover non-exclusive protection. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device. "Multiple" in the present application can mean at least two, for example, can be two, three or more, and the embodiments of the present application are not limited.

[0053] The term "and / or" in the embodiments of the present application only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are in an "or" relationship.

[0054] It can be understood that in the following specific embodiments of the present application, data related to children's learning and the like are involved, and when the embodiments of the present application are applied to specific products or technologies, relevant permissions or consents need to be obtained, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of countries and regions. For example, the data of volunteers can be used by recruiting relevant volunteers and signing relevant agreements on the authorization of volunteer data; or the data management is implemented by using the data of internal members of an organization within the scope of the organization which has been authorized to allow; or the relevant data used in the specific implementation are all simulation data, for example, simulation data generated in a virtual scene.

[0055] The design idea of the embodiments of the present application will be briefly introduced below.

[0056] At present, the traditional learning recommendation system in the field of intelligent education often only collects coarse labels such as the age and subject of children, resulting in the single dimension of user portrait, and the traditional system usually adopts fixed animation, reading and other standardized learning paths, which are difficult to meet the personalized needs of children in learning content and form. For example, in the regional aspect, the preferences of children in the north for ice and snow theme and children in the south for marine popular science cannot be accurately identified; in the interest aspect, the game interaction learning form popular among young children is often ignored by the traditional learning recommendation system which mainly uses static graphics and text. In addition, the traditional learning recommendation system cannot realize real-time sensing and analysis of the emotional state of children, and is prone to a vicious cycle of "rigidity-declining interest-loss of tolerance", which seriously affects the learning effect of children.

[0057] In view of the above problems, the embodiments of the present application provide a learning resource recommendation method, system, device, medium and product. The method obtains the basic information and historical learning behavior data of a target object, constructs an initial interest graph combining a graph neural network algorithm, performs emotional calculation according to multi-modal emotional data of the target object, generates a comprehensive emotional vector, and determines a learning resource recommendation result according to a first interest vector of the initial interest graph and the comprehensive emotional vector, so as to improve the accuracy and adaptability of learning resource recommendation and further improve the learning efficiency of the target object such as children.

[0058] Further, the embodiment of the present application constructs a four-dimensional portrait model by collecting the child's region, age, gender, and historical learning data, and performs multi-modal sentiment calculation by fusing data such as cameras (expression recognition), microphones (speech emotion analysis), and text input (semantic understanding) to generate a comprehensive emotion vector containing interest preferences ("space exploration") and emotional states (such as "focus"). In this way, through the fusion of multi-modal data such as expressions, tones, and texts, the accuracy of emotion recognition is significantly improved. Next, the Transformer-LSTM hybrid model is used to predict the interest score of the combination of content theme and form, and the recommendation strategy is dynamically adjusted by combining reinforcement learning, thereby realizing the upgrade from "static tag recommendation" to "dynamic portrait-emotion linked recommendation", significantly improving the learning interest and patience of children, and being suitable for intelligent education in all scenarios. Moreover, unlike traditional methods that preset 3-5 fixed recommendation forms (such as animation or reading), the embodiment of the present application can support the dynamic combination of more than 10 recommendation forms (such as "experimental story + VR"), and adjust the proportion in real time according to the emotional state of the child (for example, 60% game proportion when positive, 70% animation proportion when negative).

[0059] To facilitate understanding of the technical solutions provided by the embodiments of the present application, some key terms used by the embodiments of the present application are explained first:

[0060] Mel-Frequency Cepstral Coefficients (MFCC): a feature extraction technique widely used in speech processing, which converts speech signals into Mel-scale (Mel Scale) cepstral coefficients by simulating the perceptual characteristics of the human auditory system for different frequency sounds, capturing the spectral envelope features of speech

[0061] Long Short-Term Memory Network (LSTM): an improved Recurrent Neural Network (RNN) that solves the gradient vanishing problem of traditional RNN by introducing "memory cells" and "gating mechanisms" (input gate, forget gate, output gate), effectively handling long-term dependencies in sequence data.

[0062] The following is a brief introduction to the application scenarios that the technical solutions of the embodiments of the present application can be applied to. It should be noted that the following application scenarios are only used to illustrate the embodiments of the present application and are not limiting. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0063] The scheme provided by the embodiments of the present application can be applied to most learning education scenarios, and is used to improve the learning efficiency of users. For example, in the field of intelligent hardware, the system can be carried on an artificial intelligence (AI) desktop robot, so as to recommend a combination of "regional culture + interest theme + adaptive form" according to a child portrait. For example, when a child in Beijing learns "Palace Culture", an augmented reality (AR) panoramic tour + role-playing game is recommended. Or, an online education platform can recommend a typhoon simulation game + Cantonese dubbing animation for a 6-year-old boy in Guangdong (whose historical preference is "science experiment") when learning "water cycle". Or, in a special education scenario, for example, a self-abuse intervention agency adopts the system, and adjusts the content in real time by identification and calculation. For example, when an anxiety signal is detected, the system automatically switches to "stress relief music + repetitive shape pairing" to assist in improving the intervention efficiency. In this way, by using the method and system provided by the embodiments of the present application, an intelligent adaptive system of education equipment and content can be constructed, and learning equipment manufacturers (including AI desktop robots) can improve the user conversion rate, realize significant growth of the average revenue per user (ARPU) value, optimize the design of device functions, enhance user stickiness and product competitiveness, and build a demand prediction model for content providers to accurately locate user interest preferences, reduce the cost of invalid content production, realize high return rate and sustainable development of content output. On the other hand, the method and system provided by the embodiments of the present application can also narrow the gap between urban and rural education resources, and children in remote areas can obtain localized content (such as children in Yunnan learning "minority costumes"), improve the effect of cultural heritage, reduce the intervention cost of children with special needs, improve the learning participation, and promote education fairness.

[0064] Of course, the method provided by the embodiments of the present application is not limited to the above-mentioned application scenarios, but can also be used in other possible application scenarios, and the embodiments of the present application are not limited in this respect. The functions that can be achieved by each device in the above-mentioned application scenarios will be described in subsequent method embodiments, and will not be described in detail here.

[0065] In the following, the method and system provided by the exemplary embodiments of the present application will be described with reference to the above-described application scenarios and the accompanying drawings. It should be noted that the above-mentioned application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect.

[0066] Reference is made to Figure 1 A flowchart of a learning resource recommendation method provided by the embodiments of the present application is shown as follows:

[0067] Step 101: Obtain basic information and historical learning behavior data of the target object.

[0068] In the embodiment of the present application, before learning resource recommendation is performed for the target object such as a child, the basic information of the target object such as the child, such as the region where the child is located, the age of the child, and the gender of the child, and the historical behavior learning data of the child, such as the past course completion degree, the answer accuracy rate, and the knowledge point staying time of the child, are obtained. In this way, the embodiment of the present application can automatically associate the interest in "ice and snow sports" according to the "region-age-gender-historical behavior" hierarchical modeling framework, such as "northern winter boys", realize multi-dimensional feature cross through factorization machines (Factorization Machines, FM), and solve the problem of independent and unrelated traditional portrait labels.

[0069] In one possible implementation, the embodiment of the present application will import the basic statistical information of the target object such as the child, such as the region, age, and gender of the child, and the related historical learning behavior data of the child, such as the course completion degree, the answer accuracy rate, and the knowledge point staying time, through a data interface in batches.

[0070] Specifically, the present application can collect regional information through user application programs (Application, APP), global positioning systems (Global Positioning System, GPS), and Internet Protocol (Internet Protocol, IP) positioning, and construct a regional knowledge base through static data such as age and gender recorded by the APP, such as "ice and snow culture" in the northeast, "water town culture" in the south, and an age segmentation model, such as game priority for children aged 3-6, experiment priority for children aged 7-12, and an interest label library, such as "mechanical-boys" and "fairy-tale-girls".

[0071] Step 102: Based on the basic information and the historical learning behavior data, an initial interest graph is constructed by combining a graph neural network algorithm.

[0072] In the embodiment of the present application, the initial interest graph is constructed by combining the historical learning behavior data, such as click, staying time, and completion degree, and the graph neural network (Graph Neural Network, GNN) is used to extract the associated features, such as "animal-dinosaur-popular experiment", and generate a first interest vector containing interest preference weights. In this way, each graph node in the initial interest graph represents different subject categories, knowledge point labels, and learning styles, and the edge weight of the initial interest graph represents the preference degree of the target object such as the child for different learning resources, thereby laying a data foundation for subsequent learning resource recommendation.

[0073] Specifically, the historical learning behavior data in the embodiments of the present application is a quantitative record of the past learning track of a child, which can intuitively present the learning habits and knowledge mastery of the child. Each node in the initial interest graph constructed based on the historical learning behavior data covers dimensions such as subject classification, knowledge point label, and learning style, and the edge weight reflects the preference degree of the child for different content, thereby forming a multi-dimensional personalized portrait basis.

[0074] In a possible implementation, the embodiments of the present application can map the interaction records in the historical learning behavior data into a plurality of nodes, and calculate the edge weight between each two nodes through the co-occurrence frequency in the interaction records. Then, the initial weight of each node is initialized according to the basic statistical information, so as to construct the initial interest graph through the initial weight and the edge weight in combination with the graph neural network algorithm.

[0075] In a possible implementation, the embodiments of the present application can determine the weight of each node in the initial interest graph through the historical learning behavior data, and perform weighted summation on the embedding vectors of the plurality of nodes through the weights of the nodes to generate a first interest vector to represent the comprehensive interest tendency of the target object in subject preference, knowledge point preference, and learning style preference.

[0076] Specifically, the process of constructing the initial interest graph through implicit feedback in the embodiments of the present application can be as follows:

[0077] (1) Data collection and processing:

[0078] The embodiments of the present application collect historical learning behavior data such as click, dwell time, and completion degree of a target object such as a child. For click data, the embodiments of the present application can record information such as click course title and video button through click events to generate nodes. For example, clicking “dinosaur video” generates the node “dinosaur”.

[0079] For dwell time data, the embodiments of the present application can monitor the dwell time of a user on a certain content through a timer, and map the dwell time to node weight. For example, it is monitored that the target object dwells on the “dinosaur encyclopedia” page for 3 minutes, and the 3 minutes correspond to a weight of 0.8.

[0080] For completion degree data, the embodiments of the present application can count the percentage of courses completed by a user, and convert the completion degree to node weight. For example, the course “dinosaur evolution history” is completed by 80%, and the corresponding weight is 0.7.

[0081] (2) Construction of initial interest graph:

[0082] The embodiments of the present application map the implicit feedback data to graph nodes to generate nodes, for example, “clicking dinosaur video → dinosaur node”, and calculate the edge weight between nodes through co-occurrence frequency, for example, “dinosaur” and “fossil” co-occur 10 times, and the edge weight = 10 / total co-occurrence times.

[0083] (3) First interest vector generation:

[0084] In the embodiments of the present application, the GNN can capture multi-hop associations such as "animal -> dinosaur -> popular science experiment", break through the single-hop limitation of traditional collaborative filtering, and generate a first interest vector I = {i1, i2,..., in} representing the comprehensive interest tendency of the target object in subject preference, knowledge point preference and learning style preference. Wherein i1 represents a subject label (for example, "science"), and the weight is determined by the cumulative weight of the subject related node; i2 represents a knowledge point label (for example, "Jurassic period"), and the weight is calculated by the edge weight of the knowledge point node; i3 represents an interest entity (for example, "dinosaur"), and the weight is calculated by the combination of clicks, dwell time and completion degree; and in represents a learning style label (for example, "game preference"), and the weight is determined by the proportion of game content selected by the user.

[0085] Step 103: Perform emotion calculation based on the multi-modal emotion data of the target object, and generate a comprehensive emotion vector of the target object.

[0086] In the embodiments of the present application, the multi-modal emotion data includes visual data such as facial expressions and body movements, voice data such as tone, speed, and keywords, and text data such as learning notes and question content, and the comprehensive emotion vector represents the comprehensive emotional state of the target object in multiple dimensions of pleasure, concentration and confusion.

[0087] In one possible implementation, the embodiments of the present application can respectively extract features from the multi-modal emotion data to obtain feature vectors of the visual data, the voice data and the text data, and perform weighted fusion on the feature vectors according to a preset attention mechanism to obtain the comprehensive emotion vector.

[0088] Specifically, the embodiments of the present application can use multi-modal information fusion technology to synchronously collect multi-modal emotion data in the learning process of children, including visual data (facial expressions, body movements), voice data (tone, speed, keywords) and text data (learning notes, question content), and extract feature vectors from the visual, voice and text three modal data. The specific process can be as follows:

[0089] (1) Extract visual modal features through a convolutional neural network (CNN):

[0090] Input the facial expression image (such as concentration, confusion) collected by the camera, and use MobileNetV3 to extract facial key point features, output the expression category, and calculate e vis Visual emotion value / visual concentration score, for example, concentration score 0.8.

[0091] (2) Extracting speech modal features using LSTM:

[0092] The speech signal (e.g., "too difficult", "like") captured by the input microphone is used to extract speech spectral features through MFCC, and combined with LSTM to classify speech sentiment polarity (e.g., positive / negative probability e). aud ∈[0,1]).

[0093] (3) Extracting text modal features using Transformer:

[0094] The BERT-wwm model extracts features from user-input text (e.g., "I like robots", "princess") and calculates the text sentiment polarity e based on the extracted semantic vectors. txt ∈[-1,1], for example, a positive sentiment score of 0.9.

[0095] Furthermore, after extracting the multimodal feature vectors, multi-dimensional sentiment vectors such as pleasure, focus, and confusion can be calculated to obtain a comprehensive sentiment vector, as follows:

[0096] (1) Enjoyment level:

[0097] via e txt (Textual emotional polarity) and e aud (Probability of positive emotional response in speech) calculation, such as e txt =0.9, e aud A score of 0.8 indicates high levels of enjoyment.

[0098] (2) Focus:

[0099] via e vis (Visual Emotion Score / Visual Attention Score) Calculation, e vis =0.7 indicates focus.

[0100] (3) Confusion level:

[0101] via e aud (Speech confusion keyword detection) and e txt (Text interrogative word frequency) calculation, such as e aud =0.3, e txt A value of -0.5 indicates a high level of confusion.

[0102] In summary, weights are dynamically allocated based on the attention mechanism, and a comprehensive sentiment vector E is obtained by weighted fusion of three-modal features to quantify multi-dimensional sentiment states. The fusion formula is shown below:

[0103] E = λ vis e vis +λ aud e aud +λ txt e txt

[0104] Here, λvis represents the visual modality weight, which is adjusted according to different age groups (e.g., λvis = 0.4 for younger children and λvis = 0.3 for older children).

[0105] e vis Represents visual emotion value, the attention score output by the CNN model, and its range is [0,1].

[0106] λ aud λ represents the modal weights of speech, which are determined by the stability of speech emotion (e.g., when the speech rate changes significantly). aud =0.3).

[0107] e txt Represents the sentiment polarity of the text, output by the BERT-wwm model, and its range is [-1, 1].

[0108] λ txt The modal weight of the text is determined by the text length and keyword density (e.g., λtxt = 0.5 for long text and λtxt = 0.3 for short text).

[0109] Thus, this application can extract features from each modality of data through convolutional neural networks, recurrent neural networks, and Transformer models, and then use attention mechanisms to weightedly fuse them to finally generate a comprehensive emotional vector E containing dimensions such as pleasure, focus, and confusion, thereby quantifying the child's real-time emotional state.

[0110] In one possible implementation, refer to Figure 2 This is a schematic diagram illustrating the construction process of an initial interest graph provided in an embodiment of this application. This application collects historical learning behavior data generated by users, such as clicks, dwell time, and completion rate. When constructing the initial interest graph, this data will be mapped to resource nodes. Each resource node is associated with its corresponding knowledge point or interest tag. Subsequently, by aggregating neighbor information through a GNN, graph computation is performed on these knowledge point or interest tag nodes to generate node embedding vectors. The embedding vectors of these nodes are then weighted and summed to obtain an interest vector I = {i1, i2, ..., in} that represents the user's comprehensive interest tendency.

[0111] Specifically, taking "Animals → Dinosaurs → Science Experiments" as an example, the nodes of the initial interest graph include interest tags ("Animals", "Dinosaurs", and "Science Experiments"), knowledge points ("Dinosaur Classification" and "Fossil Formation"), and learning resources ("Dinosaur Videos", "Fossil Excavation Experiments"). The edges of the initial interest graph include: a strong association between "Animals" and "Dinosaurs" (e.g., a user clicks on dinosaur-related resources multiple times, weight 0.8), an edge between "Dinosaurs" and "Science Experiments" (e.g., a user completes a "Dinosaur Fossil Excavation Experiment", weight 0.7), and an association between "Dinosaur Videos" and the knowledge point "Dinosaur Classification" (weight 0.9). Next, neighbor node aggregation and node embedding generation are performed using GNN. The specific processing flow includes input node features, such as static labels (e.g., "dinosaur" belongs to "animal") and dynamic weights (e.g., click frequency); adjacency matrix: the graph structure is formed by the above edge weights; execution of GNN operations, including neighbor aggregation: information aggregation (e.g., mean, max pooling) is performed on the neighbor nodes of each node. For example, for the "dinosaur" node, the features of its neighbors "animal" and "science experiment" are aggregated; node embedding generation: after passing through multiple layers of GNN, each node generates a low-dimensional embedding vector (e.g., 64-dimensional). For example, the embedding vector of the "dinosaur" node captures its association with labels such as "animal", "science experiment", and "paleontology"; interest vector generation: the embedding vectors of all nodes that the user has interacted with are weighted and summed to generate an interest vector I = {i1, i2, ..., in}, for example, i1 = 0.85 (animal), i2 = 0.7 (science experiment), and i3 = 0.6 (space exploration).

[0112] Step 104: Determine the learning resource recommendation results based on the first interest vector and the comprehensive sentiment vector of the initial interest graph.

[0113] In this embodiment, the recommended content is scored based on dimensions such as knowledge fit, difficulty matching, and interaction form adaptability for candidate learning resources. The format score is also calculated based on historical preferences for different presentation formats such as animation, games, and text and images. The combined score is generated by weighted fusion and finally outputs the learning content and corresponding presentation format that meet the preset conditions, thereby achieving personalized and accurate recommendations.

[0114] Specifically, this application embodiment achieves joint modeling of content and form through a preset candidate learning resource set and a form combination strategy set. The candidate learning resource set includes preset content library tags, such as themes (ancient / modern / science, etc., 20 categories), difficulty levels (L1-L5), and regional adaptation (e.g., "Northwest folk customs"). The form combination strategy set includes preset form library tags, such as animation (proportion fa), games (proportion fg), and experiments (proportion fe), and supports dynamic combination from 0-100%. Thus, through the aforementioned emotion-form linkage mechanism and the designed mapping rules between emotional states and form proportions (e.g., "confusion → animation proportion +20%, game proportion -10%)", this application can automatically optimize rule parameters through reinforcement learning, significantly improving the learning efficiency of target users such as children.

[0115] In one possible implementation, this application embodiment determines the corresponding content fit by using the inner product between the first interest vector and the tag vectors of each candidate content in the candidate learning resource set. It then calculates an emotional gain factor by integrating the emotional vector, and adjusts the content fit based on this gain factor to determine the target interest score. A preset number of candidate contents are selected as recommended content based on their target interest scores, from highest to lowest. Combined with preset emotional form mapping rules, at least one presentation form matching the current emotional state is determined from a set of form combination strategies to determine the final learning resource recommendation result.

[0116] Specifically, the target interest score S in the embodiments of this application ij The calculation is as follows:

[0117] S ij =(I·C j )×(1+αE)

[0118] α=1+β×(EE avg )

[0119] Where I represents the first interest vector, C j Represents the vector of content tags;

[0120] α is the emotional gain factor; for example, α = 1.2 for positive emotions.

[0121] β represents the adjustment coefficient (e.g., β = 0.2).

[0122] E represents the comprehensive sentiment vector, i.e., the current sentiment score, ranging from [0,1]. avg The historical average sentiment score.

[0123] Furthermore, the embodiment of this application can adjust the form ratio according to the emotional state, such as fa=70% and fg=20% when negative, to achieve animation soothing + light game awakening.

[0124] Specifically, the emotional form mapping rules can be set as follows:

[0125] Positive emotion (E≥0.7): The proportion of gamification form is 60% (such as "Dinosaur Battle Game"), and the proportion of animation is 20%.

[0126] Neutral emotion (0.3<E<0.7): The proportion of experimental form is 50% (such as "Fossil Digging Simulation"), and the proportion of animation is 30%.

[0127] Negative emotion (E≤0.3): The proportion of animation soothing is 70%, and the proportion of mild game arousal is 20%.

[0128] Next, the emotional state is monitored in real time through a sliding window (such as every 10 minutes), and the form ratio is updated every minute.

[0129] In summary, the embodiments of this application can calculate the recommended content scores for the candidate learning resource set from dimensions such as knowledge fit, difficulty matching degree, and interaction form adaptation degree, and calculate the form scores from the historical preferences of different presentation forms such as animation, games, and graphics and texts. In this way, the combined scores are generated through weighted fusion, and finally the top three learning contents and corresponding presentation forms with comprehensive scores are output to achieve personalized and accurate recommendation.

[0130] In a possible implementation manner, the embodiments of this application can generate a new comprehensive emotion vector according to the new learning behavior data and new multi-modal emotion data generated by the target object for the learning resource recommendation result, and update the first interest vector through the new comprehensive emotion vector and the new learning behavior data, combined with a preset dynamic update model, and use the updated second interest vector as the first interest vector for the next learning resource recommendation to achieve closed-loop feedback optimization.

[0131] Specifically, the embodiments of this application can input the comprehensive emotion vector and the newly generated learning behavior data, such as the latest course learning records, test results, etc. into the dynamic update model. Among them, the newly generated learning behavior data can continuously track the current learning performance of children, combine it with the emotion vector, and through the adaptive weight adjustment algorithm, the weights of each dimension in the first interest vector I can be corrected in real time, and the emotion preference weights can be redistributed. In this way, the content tags associated with high focus and high positive emotion are strengthened, and the learning elements that cause negative emotions are weakened, realizing the dynamic optimization of the portrait. <​​The recommendation form (fa, fg, fe) is combined with the reward function to calculate the reward value. This reward value represents the comprehensive effectiveness of the learning resource recommendation results in improving learning effectiveness, maintaining learning duration, and promoting positive emotions. In this way, the weights of subject preference, knowledge point preference, and learning style preference in the first interest vector can be adjusted through the reward value to generate an updated second interest vector.

[0133] Specifically, in this embodiment, the profile model can be updated every 10 minutes, and the recommendation strategy is optimized using the Proximal Policy Optimization (PPO) reinforcement learning algorithm. The reward function is as follows:

[0134] R=γ1S r +γ2T s +γ3A e

[0135] Where Sr represents the learning completion rate = actual completed learning quantity / planned learning quantity, for example, completing 80% of the course earns 1 point;

[0136] Ts represents the study duration; for example, studying continuously for 30 minutes earns 1 point.

[0137] Ae represents the percentage of positive emotions; for example, if the percentage of positive emotions E > 0.7, 1 point is awarded.

[0138] R represents the reward value;

[0139] γ1 represents the learning completion rate weight (e.g., γ1 = 0.4), which is determined by education experts based on learning effect experiments.

[0140] γ2 represents the weight of learning time (e.g., γ2 = 0.3), which is related to user stickiness. The longer the time, the higher the stickiness, but the weight cannot be higher than the weight of learning completion rate.

[0141] γ3 represents the weight of positive emotion (e.g., γ3 = 0.3), which is determined by research on the correlation between emotion and learning efficiency.

[0142] In summary, the embodiments of this application can randomly recommend content based on an initial strategy, calculate the reward value R based on user feedback, and update the policy network through the aforementioned PPO algorithm to maximize the accumulated reward value.

[0143] In summary, this application embodiment can record learning behavior data such as children's learning time and emotional feedback (such as actively submitted emotion tags and voice emotion recognition results) in real time during the learning interaction process. Emotional feedback supplements the learning experience information from a subjective emotional perspective, serving as optimization samples for the next round of model training, continuously improving the accuracy and adaptability of the recommendation system. Thus, this application embodiment can construct a time-series-based user behavior prediction model, combined with real-time learning data, updating the weight of recommended content every 15 minutes. This allows for the automatic reduction of the recommendation frequency of related content when a child repeatedly skips a certain type of knowledge point, and the discovery of alternative learning resources from similar knowledge graph nodes.

[0144] On the other hand, reference Figure 3 The diagram shown is a structural schematic of a learning resource recommendation system provided in an embodiment of this application. The system includes the following unit modules:

[0145] The profile construction module 301 is used to acquire basic information and historical learning behavior data of the target object, and to construct an initial interest map by combining graph neural network algorithm; and to perform sentiment calculation based on the multimodal sentiment data of the target object to generate a comprehensive sentiment vector.

[0146] The recommendation execution module 302 is used to determine the learning resource recommendation results based on the first interest vector and the comprehensive sentiment vector of the initial interest graph.

[0147] The feedback optimization module 303 is used to generate a new comprehensive sentiment vector based on new learning behavior data and new multimodal sentiment data generated by the target object in response to the learning resource recommendation results. Based on the new comprehensive sentiment vector and a preset dynamic update model, the first interest vector is updated, and the resulting second interest vector is used as the first interest vector for the next learning resource recommendation.

[0148] In one possible implementation, the profile construction module in this application embodiment further includes a basic information submodule, a history learning submodule, a multimodal emotion submodule, and a fusion module.

[0149] Specifically, the basic information submodule is used for profile initialization. It collects geographic information through the user's app and uses static data such as age and gender recorded by the app to build a geographic knowledge base, an age segmentation model, and a gender interest tag library. The historical learning submodule constructs an initial interest graph through implicit feedback, extracts related features using a graph neural network, and generates the first interest vector. The multimodal emotion submodule includes a voice emotion submodule, a text emotion submodule, and a visual submodule. The voice emotion submodule can extract features through MFCC and classify voice emotion using LSTM, while extracting voice emotion keywords such as "like" and "too difficult". The text emotion submodule analyzes the text emotion polarity and extracts interest entities using the BERT-wwm model. The visual submodule extracts visual modal features and calculates visual emotion value / visual attention score using CNN. The fusion module dynamically allocates weights through an attention mechanism to generate a comprehensive emotion vector.

[0150] In one possible implementation, the recommendation execution module can construct content library tags and form library tags through content-form joint modeling, calculate the target interest score of each candidate content through the recommendation algorithm, and adjust the presentation format ratio of the candidate content according to the sentiment state.

[0151] In one possible implementation, the feedback optimization module can update the profile model at preset time intervals and optimize the recommendation strategy using the PPO reinforcement learning algorithm to continuously improve the accuracy and adaptability of the learning resource recommendation system.

[0152] It is worth mentioning that the actual application process of the learning resource recommendation system in this application embodiment is the same as the method flow in the aforementioned learning resource recommendation method. Therefore, this process can be referred to the detailed introduction of the aforementioned training process, and will not be elaborated further here.

[0153] In one possible implementation, refer to Figure 4 The diagram shown is a system model and flowchart provided in an embodiment of this application. Figure 4In this architecture, the input layer collects basic user information (region / age / gender), historical learning data (explicit / implicit feedback), real-time video streams with features extracted using MobileNet V3, and real-time audio streams with text and acoustic features extracted using Automatic Speech Recognition (ASR). The basic information is fed into the child profile construction module, where features are fused using region encoding (8-dimensional one-hot encoding), age embedding (4-dimensional segmented vector), and gender distribution (2-dimensional probability) to generate a GNN interest graph. This graph is then processed with dynamic updates and a time decay function before being stored as a profile. Meanwhile, the real-time streaming data is fed into the multimodal sentiment analysis module, where extracted visual features (128D facial expression vector), acoustic features (39D speech vector), and text features (768D semantic vector) are quantified for sentiment through cross-modal fusion and an attention mechanism, outputting a 12D comprehensive vector to the recommendation decision module. This recommendation decision-making module utilizes Transformer for global feature interaction and LSTM for temporal pattern mining, driving the policy network to generate action probability distributions. The PPO algorithm is then used to update the reinforcement learning decision tree to optimize the recommendation strategy. The final output layer pushes the recommendation results to the terminal in a topic-plus format, while simultaneously collecting sentiment feedback and clickstream data generated through explicit evaluations and behavioral tracking as a weakened signal (1D composite vector) fed back to the input layer. This forms a closed-loop, personalized learning resource recommendation system driven by child profiling, multimodal sentiment computing, and reinforcement learning.

[0154] In one possible implementation, refer to Figure 5The diagram illustrates a system analysis process according to an embodiment of this application. This embodiment utilizes children's basic information (including region, age, and gender) and historical learning behavior data (including course completion rate, answer accuracy, and knowledge point dwell time) as input for profile initialization. This involves constructing an initial interest graph and generating a first interest vector using a preset graph neural network algorithm. Simultaneously, multimodal emotional data (i.e., visual, speech, and text data) is collected in real-time for emotion calculation. A comprehensive emotion vector is generated through cross-modal fusion and weighted fusion. This comprehensive emotion vector and the initial interest vector are then input into a profile update module. An adaptive weight adjustment algorithm dynamically corrects the weights of each dimension in the interest vector. The updated data is then input into a joint recommendation model to calculate content relevance scores. Finally, through emotion-form mapping rules, the learning resource recommendation result is determined, i.e., a combination of specific learning resource content and presentation formats is pushed to the user. Next, real-time user feedback data is collected, along with user behavior and emotional response data to the recommendation results. This data, as new learning behavior data, is reintegrated into the system loop along with multimodal emotional data, driving continuous iterative optimization of the user profile. This results in a data-driven, multimodal, and self-evolving personalized recommendation closed-loop system.

[0155] Please see Figure 6 As shown, based on the same technical concept, this application also provides a computer device 60. In one embodiment, the computer device can be a device specifically for recommending learning resources, or it can be an auxiliary device for learning and education. Figure 6 As shown, it includes a memory 601, a communication module 603, and one or more processors 602.

[0156] The memory 601 is used to store computer programs executed by the processor 602. The memory 601 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0157] Memory 601 may be volatile memory, such as random-access memory (RAM); memory 601 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 601 may be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 601 may be a combination of the above-described memories.

[0158] Processor 602 may include one or more central processing units (CPUs) or digital processing units, etc. Processor 602 is used to implement the above-mentioned learning resource recommendation method when calling computer programs stored in memory 601.

[0159] The communication module 603 is used to communicate with learning and education systems or other platforms.

[0160] This application embodiment does not limit the specific connection medium between the memory 601, communication module 603, and processor 602 described above. This application embodiment... Figure 6 The memory 601 and the processor 602 are connected via a bus 604, and the bus 604 is in Figure 6 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered limiting. The 604 bus can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 6 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0161] The memory 601 stores a computer storage medium, which stores computer-executable instructions. The computer-executable instructions are used to implement the learning resource recommendation method of the embodiments of this application, and the processor 602 is used to execute the learning resource recommendation methods of the above embodiments.

[0162] Based on the same inventive concept, embodiments of this application also provide a storage medium storing a computer program that, when run on a computer, causes the computer to perform the steps in the learning resource recommendation method according to various exemplary embodiments of this application described above.

[0163] In some possible implementations, various aspects of the learning resource recommendation method provided in this application can also be implemented in the form of a computer program product, which includes a computer program that, when run on a computer device, causes the computer device to perform the steps in the learning resource recommendation method according to various exemplary embodiments of this application described above. For example, the computer device can perform the steps of the various embodiments.

[0164] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0165] The program product of the embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and may run on a computer device. However, the program product of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium that contains or stores a program, and the computer program included therein may be used by or in conjunction with a command execution system, apparatus, or device.

[0166] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with a command execution system, apparatus, or device.

[0167] Computer programs contained on readable media may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0168] Computer programs for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages.

[0169] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0170] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0171] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0172] Although preferred embodiments of this application 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 the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0173] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A learning resource recommendation method, characterized in that, The method includes: Acquire basic information and historical learning behavior data of the target object; the basic information includes the target object's region, age, and gender, and the historical learning behavior data includes course completion rate, answer accuracy rate, and time spent on knowledge points; Based on the aforementioned basic information and historical learning behavior data, an initial interest graph is constructed using a graph neural network algorithm. Each node in the initial interest graph represents a different subject category, knowledge point label, and learning style, and the edge weight of each node represents the target object's preference for different learning resources. Sentiment calculation is performed based on the multimodal sentiment data of the target object to generate a comprehensive sentiment vector; the multimodal sentiment data includes visual data, speech data and text data, and the comprehensive sentiment vector represents the multidimensional comprehensive sentiment state of the target object; Based on the first interest vector of the initial interest graph and the comprehensive sentiment vector, a learning resource recommendation result is determined; the first interest vector represents the multi-dimensional comprehensive interest preferences of the target object, and the learning resource recommendation result includes recommended content and at least one presentation format corresponding to the recommended content.

2. The method as described in claim 1, characterized in that, The initial interest graph is constructed based on the aforementioned basic information and historical learning behavior data, combined with a graph neural network algorithm, including: The interaction records in the historical learning behavior data are mapped to multiple nodes, and the edge weight between each pair of nodes is calculated based on the frequency of co-occurrence of the interaction records. Based on the aforementioned basic information, initialize the initial weight of each node; Based on the initial weights and edge weights, node embedding and neighbor node aggregation are performed to construct the initial interest graph.

3. The method as described in claim 1, characterized in that, The process of performing sentiment calculations based on the multimodal sentiment data of the target object to generate a comprehensive sentiment vector includes: Feature extraction is performed on the multimodal emotion data to obtain feature vectors corresponding to the visual data, the speech data, and the text data respectively; Based on a preset attention mechanism, the feature vectors are weighted and fused to obtain the comprehensive sentiment vector.

4. The method as described in claim 1, characterized in that, Before determining the learning resource recommendation result based on the first interest vector of the initial interest graph and the comprehensive sentiment vector, the method further includes: Based on the historical learning behavior data, the weight of each node in the initial interest graph is determined; Based on the weights of each node, the embedding vectors of each node are weighted and summed to generate the first interest vector.

5. The method as described in claim 1, characterized in that, The process of determining the learning resource recommendation result based on the first interest vector of the initial interest graph and the comprehensive sentiment vector includes: Based on the first interest vector and the tag vectors of each candidate content in the candidate learning resource set, the corresponding content fit is determined; Based on the comprehensive sentiment vector, the sentiment gain factor is calculated; Based on the aforementioned emotional gain factor, the content relevance is adjusted to determine the target interest score. Based on the target interest scores of each candidate content, a preset number of candidate content are selected from high to low as the recommended content. Based on the comprehensive sentiment vector and combined with the preset sentiment form mapping rules, at least one presentation form that matches the recommended content is determined from the set of form combination strategies.

6. The method as described in claim 1, characterized in that, After determining the learning resource recommendation results, the method further includes: Based on the new learning behavior data and new multimodal sentiment data generated by the target object in response to the learning resource recommendation results, a new comprehensive sentiment vector is generated. Based on the new comprehensive sentiment vector and the preset dynamic update model, the first interest vector is updated to obtain the updated second interest vector; The second interest vector will be used as the first interest vector for the next learning resource recommendation.

7. The method as described in claim 6, characterized in that, The step of updating the first interest vector based on the new comprehensive sentiment vector and a preset dynamic update model to obtain the updated second interest vector includes: Based on the new learning behavior data and the new comprehensive sentiment vector, a reward value is calculated using a preset reinforcement learning algorithm; the reward value represents the effectiveness of the learning resource recommendation result. Based on the reward value, the weights of multiple dimensions in the first interest vector are adjusted to obtain the updated second interest vector.

8. A learning resource recommendation system, characterized in that, The system includes: The profile building module is used to obtain basic information and historical learning behavior data of the target object; the basic information includes the target object's region, age and gender, and the historical learning behavior data includes course completion rate, answer accuracy rate and knowledge point dwell time; Based on the aforementioned basic information and historical learning behavior data, an initial interest graph is constructed using a graph neural network algorithm. Each node in the initial interest graph represents a different subject category, knowledge point label, and learning style, and the edge weight of each node represents the target object's preference for different learning resources. The recommendation execution module is used to perform sentiment calculation based on the multimodal sentiment data of the target object and generate a comprehensive sentiment vector; the multimodal sentiment data includes visual data, speech data and text data, and the comprehensive sentiment vector represents the multidimensional comprehensive sentiment state of the target object; Based on the first interest vector of the initial interest graph and the comprehensive sentiment vector, a learning resource recommendation result is determined; the first interest vector represents the multi-dimensional comprehensive interest preferences of the target object, and the learning resource recommendation result includes recommended content and at least one presentation format corresponding to the recommended content.

9. The system as described in claim 8, characterized in that, The system also includes: The feedback optimization module is used to generate a new comprehensive sentiment vector based on the new learning behavior data and new multimodal sentiment data generated by the target object in response to the learning resource recommendation results; Based on the new comprehensive sentiment vector and the preset dynamic update model, the first interest vector is updated to obtain the updated second interest vector; The second interest vector will be used as the first interest vector for the next learning resource recommendation.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

11. A computer storage medium storing computer program instructions thereon, characterized in that, When executed by a processor, the computer program instructions implement the steps of the method according to any one of claims 1 to 7.

12. A computer program product comprising computer program instructions, characterized in that, When executed by a processor, the computer program instructions implement the steps of the method according to any one of claims 1 to 7.

Images