Teaching resource personalized recommendation system based on user portraits

Abstract

The invention provides a teaching resource personalized recommendation system based on a user portrait. The system comprises a data acquisition module, a multi-source heterogeneous data processing module, a structured data module, a dynamic portrait generation module and a resource recommendation module. Infant behavior data are collected in multiple modes, a space-time diagram neural network is adopted for multi-modal feature fusion, and cross-modal data alignment is achieved in combination with a dynamic time warping algorithm and physical constraint optimization; a dynamic weight distribution mechanism is introduced, a dynamic weight matrix is generated through a time delay cross-correlation algorithm and real-time feature similarity calculation, and a multi-dimensional dynamic portrait is constructed in combination with a teaching outline knowledge graph and parent behavior data; according to the system, accurate quantification of unstructured behavior data is realized, the matching problem of rapid transition in an infant cognitive stage is solved, resource recommendation is performed by fusing teaching objectives of educational institutions, infant development characteristics and parent preferences, and the resource recommendation effect is improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically to a personalized recommendation system for teaching resources based on user profiles. Background Technology

[0002] Early childhood education is a crucial stage for children's character development and cognitive growth, with cognitive abilities of children aged 3-6 increasing by 18.7% monthly. Traditional kindergarten education generally adopts a standardized teaching model, which struggles to meet the diverse developmental needs of individuals. While personalized recommendation systems are increasingly being applied to education with the widespread use of AI technology, significant technical bottlenecks still exist in kindergarten settings.

[0003] Specifically, current systems rely heavily on structured behavioral data (such as click-through rates and task completion times) for recommendations, failing to capture unstructured behavioral characteristics of young children (such as the mechanical features of block construction and social tendencies in cooperative games) and physiological indicators (such as heart rate variability and micro-expression changes). Furthermore, traditional user profiles have long update cycles, making it difficult to adapt to the rapidly changing cognitive development stages of young children. For example, during the transition from the language-sensitive period (3-4 years old) to the logical thinking period (4-5 years old), existing systems cannot achieve real-time dynamic adjustments, resulting in a mismatch between recommended resources and the child's current developmental needs. Moreover, teacher syllabi, parental educational philosophies, and AI algorithms create data silos, leading to discrepancies between AI-recommended content and teaching objectives, and a lack of connection between home-based extension activities and kindergarten curriculum content.

[0004] Therefore, there is an urgent need to build a recommendation system that integrates multimodal perception, dynamic profile updates, and educational theory encoding. In particular, it is crucial to solve the key technical challenges of quantifying unstructured behaviors in young children and matching teaching resources using dynamic profiles, thereby achieving truly personalized education that aligns with the developmental patterns of young children. Summary of the Invention

[0005] To address the aforementioned constraints and limitations, this invention proposes a personalized teaching resource recommendation system based on user profiles. By introducing multi-source heterogeneous data analysis to understand children's cross-modal characteristics and dynamic profiles, and combining this with children's cognitive levels and parents' preferences, the system comprehensively recommends teaching resources, thereby improving the accuracy and adaptability of teaching resource recommendations.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The personalized teaching resource recommendation system based on user profiles consists of a data acquisition module, a multi-source heterogeneous data processing module, a structured data module, a structured data feature module, a dynamic profile generation module, and a resource recommendation module.

[0008] The data acquisition module is used to collect unstructured behavioral data generated by children during their activities in the kindergarten.

[0009] The multi-source heterogeneous data processing module is used to perform spatiotemporal feature fusion processing on the unstructured behavioral data to obtain unstructured behavioral feature data. The spatiotemporal feature fusion processing includes: sampling frequency difference alignment processing, cross-modal correlation feature extraction, and dynamic weight allocation on the unstructured behavioral data.

[0010] The structured data module is used to store structured behavioral data and provides a query interface for structured behavioral data.

[0011] The structured data feature module is used to extract target structured behavioral data from the structured data module and perform feature extraction to obtain structured behavioral feature data.

[0012] The dynamic profile generation module is used to generate dynamic profiles based on the unstructured behavioral feature data and the structured behavioral feature data.

[0013] The resource recommendation module generates teaching resource recommendation results based on dynamic profiles using a hybrid recommendation algorithm.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] (1) A multimodal data fusion mechanism is adopted to break through the dependence of traditional recommendation systems on structured data. By fusing heterogeneous data such as motion trajectory, micro-expression dynamics, and pressure gradient through spatiotemporal graph neural networks, the unstructured behavior of young children can be accurately quantified;

[0016] (2) Construct a dynamic weight adaptive system to solve the problem of rapid transitions in early childhood cognitive stages. Through the time-delay cross-correlation algorithm (TDCC) and physical constraint optimization, real-time profile updates are achieved, improving the accuracy of teaching resource matching;

[0017] (3) An abnormal data compensation mechanism is adopted to ensure the stability of complex educational scenarios.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a structural diagram of a personalized recommendation system for teaching resources based on user profiles, provided in an embodiment of the present invention.

[0020] Figure 2This is a structural diagram of a data acquisition module provided in an embodiment of the present invention.

[0021] Figure 3 This is a flowchart illustrating a cross-modal correlation feature extraction method provided in an embodiment of the present invention. Detailed Implementation

[0022] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. To further understand the present invention, the present invention will be further described in detail below with reference to the preferred embodiments.

[0023] The inventive point of this invention lies in a personalized recommendation system for teaching resources based on user profiles, referencing... Figure 1 The system consists of a data acquisition module, a multi-source heterogeneous data processing module, a structured data module, a structured data feature module, a dynamic profile generation module, and a resource recommendation module.

[0024] The data acquisition module is used to collect unstructured behavioral data generated by children during their activities in the kindergarten.

[0025] The multi-source heterogeneous data processing module is used to perform spatiotemporal feature fusion processing on the unstructured behavioral data to obtain unstructured behavioral feature data. The spatiotemporal feature fusion processing includes: sampling frequency difference alignment processing, cross-modal correlation feature extraction, and dynamic weight allocation on the unstructured behavioral data.

[0026] The structured data module is used to store structured behavioral data and provides a query interface for structured behavioral data.

[0027] The structured data feature module is used to extract target structured behavioral data from the structured data module and perform feature extraction to obtain structured behavioral feature data.

[0028] The dynamic profile generation module is used to generate dynamic profiles based on the unstructured behavioral feature data and the structured behavioral feature data.

[0029] The resource recommendation module generates teaching resource recommendation results based on dynamic profiles using a hybrid recommendation algorithm.

[0030] Further, refer to Figure 2 The data acquisition module consists of a motion acquisition submodule, an image acquisition submodule, and a pressure acquisition submodule.

[0031] The motion acquisition submodule uses a UWB ultra-wideband positioning system to collect activity trajectory data, and a heart rate sensor to collect children's heart rate data.

[0032] The image acquisition subsystem uses an edge computing camera that integrates micro-expression recognition algorithm and body language analysis algorithm to acquire micro-expression data and body movement data.

[0033] The pressure acquisition submodule is a thin-film pressure sensor embedded in the teaching aid, used to record pressure distribution data of operational force and mechanical characteristics.

[0034] The data acquisition module also includes a sensor self-test module, which is used to monitor the data quality and working status of each sub-module of the data acquisition module in real time and output data anomaly detection signals.

[0035] As one embodiment, the data acquisition module is equipped with a hierarchical clock synchronization system to ensure that the data acquisition devices connected via a wired network are synchronized in time.

[0036] For wearable devices that cannot access wired networks (such as heart rate sensors), wireless transmission delay is eliminated through a software compensation algorithm. The compensation formula is as follows:

[0037] t corrected =t received -(t txDelay +t processing )

[0038] Where t received t is the preset delay compensation value. processing This provides a dynamic calibration value for the timestamp-based data transmission packets periodically sent by the device.

[0039] As one embodiment, the micro-expression recognition algorithm integrated in the image acquisition subsystem can adopt a lightweight model based on ResNet-50, which can recognize basic expressions.

[0040] The body language analysis algorithm extracts 21 key skeletal points using the OpenPose algorithm to quantify social tendencies in cooperative games.

[0041] As one embodiment, the sampling frequency difference alignment processing is specifically implemented by including the following steps:

[0042] S11. Unify the time reference axis and establish a timestamp mapping:

[0043] A global time reference system is constructed using the sensor with the highest sampling rate as the reference axis; the timestamp mapping relationship of other sensor data streams is established using the Dynamic Time Warping (DTW) algorithm.

[0044] S12. Multi-band signal reconstruction and resampling:

[0045] For high-frequency signals, downsampling is performed using a 64th-order FIR low-pass filter;

[0046] For missing frames in the video signal, interpolation is performed based on the Lagrange interpolation formula, and a Lanczos resampling filter (window width 5, frequency domain stopband attenuation -50dB) is introduced to suppress Gibbs oscillation.

[0047] For low-frequency signals, upsampling is performed, and downsampling is achieved through cubic spline interpolation;

[0048] S13. Data alignment optimization and verification:

[0049] Cyclic redundancy check (CRC) is employed, with a CRC-32 checksum appended to the end of the data packet. If the check fails, a forward linear prediction model is activated to fill in the missing data. Simultaneously, a Hankel matrix H∈R is constructed for large-scale missing data. m×n After optimization and verification, output multimodal data.

[0050] As one embodiment, the cross-modal correlation feature extraction is achieved through spatiotemporal graph modeling and deep feature fusion techniques, referencing... Figure 3 The specific implementation method includes the following steps:

[0051] S21, Submodule Feature Analysis;

[0052] S22. Construct a spatiotemporal heterogeneous graph model, wherein the spatiotemporal heterogeneous graph model is a graph neural network model;

[0053] S23. Perform joint optimization on the spatiotemporal heterogeneous graph model to obtain the optimized fused feature vector;

[0054] S24. Based on the real-time data collected by the data acquisition module, perform dynamic weight allocation to obtain a dynamic weight matrix.

[0055] As one embodiment, the sub-module feature parsing includes motion feature extraction, visual feature extraction, and pressure feature extraction.

[0056] The specific methods for motion feature extraction include:

[0057] S2111. Calculate the trajectory features of the activity trajectory data, including movement speed features, acceleration features, and motion smoothness features;

[0058] S2112. Perform Lomb-Scargle spectral analysis on children's heart rate data to extract the ratio of low-frequency power to high-frequency power (LF / HF).

[0059] S2113. Combine the results of S2111 and S2112 to form a motion feature vector.

[0060] The specific methods for visual feature extraction include:

[0061] S2121. The MobileNetV3-Small model is used to extract micro-expression codes from video frames based on the micro-expression data, and the dynamic feature differences between adjacent frames are calculated. The calculation method is as follows:

[0062]

[0063] Where, Δf t f represents the difference in dynamic features between adjacent frames. t For the micro-expression encoding at time t, f t-1 Encoding of micro-expressions at time t-1;

[0064] S2122. Synchronously analyze OpenPose key points of limb movement data and calculate limb opening and closing degree;

[0065] S2123. Combining the results of S2121 and S2122, output the visual feature vector.

[0066] The calculation method for the limb opening and closing degree is as follows:

[0067]

[0068] Among them, L Height For children's height, k leftwrist k rightwrist These are the two-dimensional coordinates of key points on the left and right wrists, k. torso These are the coordinates of the center point of the torso.

[0069] The visual feature vector can be represented as F v =[Δf t ,O]∈R 256 .

[0070] The specific methods for extracting pressure features include:

[0071] S2131. Perform a two-dimensional discrete cosine transform on the pressure distribution data, retaining the first 16 low-frequency coefficients to construct the pressure compression feature C. DCT ;

[0072] S2132. Calculate the pressure gradient and its mean, and obtain the pressure gradient characteristic G. avg ;

[0073] S2133. Combining the results of S2131 and S2132, we obtain the pressure feature vector.

[0074] The pressure feature vector can be represented as F p =[C DCT G avg ]∈R 64

[0075] As one embodiment, the spatiotemporal heterogeneous graph model is constructed through the following steps:

[0076] S221. Map the motion feature vector, visual feature vector, and pressure feature vector to graph nodes of the spatiotemporal heterogeneous graph model to obtain the node feature matrix.

[0077] The node feature matrix can be represented as V = [F m ;F v ;F p ]∈R 3×448 ;where F m F v F p These are motion feature vectors, visual feature vectors, and pressure feature vectors, respectively.

[0078] S222. Calculate the cross-modal causal strength, retain the edges that meet the preset strength conditions as significant associated edges, and construct a directed edge weight matrix;

[0079] S223. Perform graph initialization to obtain spatiotemporal heterogeneous graph structure data, which consists of node feature matrix, edge set and directed edge weight matrix.

[0080] The spatiotemporal heterogeneous graph structure data can be represented as G = (V, E, W), where V is the node feature matrix, E is the edge set, and W is the directed edge weight matrix.

[0081] Where, the edge set E = {e m→v ,e v→m ,e m→p ,e p→m ,e v→p ,e p→v}

[0082] Specifically, the cross-modal causality strength is calculated using the Time Delay Cross-Correlation (TDCC) algorithm, and the calculation method is as follows:

[0083]

[0084] Among them, W i→j (τ) refers to the cross-causal strength from mode i to mode j (with a time delay of τ); modes include m (motion features), v (visual features), and p (stress features); τ is the time delay parameter, with a value range of [-500ms, +500ms]; cross_cov is the cross-covariance function; denoted as the standard deviations of the features of modes i and j, respectively.

[0085] As one example, the joint optimization of spatiotemporal heterogeneous graph models achieves deep fusion of cross-modal features through joint optimization guided by multi-head graph attention mechanism and physical constraints;

[0086] Specifically, it includes:

[0087] S231. Using a multi-head attention mechanism, perform four-head attention calculation on the spatiotemporal heterogeneous structure data; after concatenating the output results of the four-head attention calculation, compress them with the dimension of the gated linear unit to obtain the cross-modal attention feature matrix;

[0088] S232. The acceleration features obtained in S2111 and the pressure gradient features obtained in S2132 are used as physical constraints to optimize the cross-modal attention feature matrix, resulting in a cross-modal feature correction matrix.

[0089] S233. Based on the historical data collected by the data acquisition module, construct positive sample pairs and negative sample pairs; use the AdamW optimizer to iteratively update the cross-modal feature correction matrix to obtain the fused feature vector.

[0090] As one embodiment, during multi-head attention calculation in S231, each attention head generates a query matrix, and the calculation method for multi-head attention is as follows:

[0091]

[0092] H h =A′ h V h

[0093] Among them, A h ' represents the attention weight of the h-th attention head, H h This is the output of the h-th attention head; The time dimension mean of the directed edge weight matrix; the query matrix, key matrix, and value matrix of the h-th attention head are respectively:

[0094]

[0095] Where V is the node feature matrix, W h Q W h K W h V These are the trainable query matrix parameters, key matrix parameters, and value matrix parameters, respectively.

[0096] The cross-modal attention feature matrix is ​​obtained as follows:

[0097] V fusion =GLU(H cat )∈R 3×128

[0098] H cat = [H1||H2||H3||H4]∈R 3×256

[0099] Where GLU is a gated linear unit, V fusion H is the cross-modal attention feature matrix. cat The output result of the cascaded calculation of the four-head attention calculation results.

[0100] It is understandable that the cross-modal attention feature matrix can be represented as V fusion =[F m ′,F v ′,F p ′],F p '、F m '、F v 'These represent the stress node features, motion node features, and visual node features of the cross-modal attention feature matrix, respectively.

[0101] As one embodiment, the specific method for obtaining the cross-modal feature correction matrix in S232 includes:

[0102] S2321. Calculate the pressure-motion correlation coefficient ρ, as follows:

[0103]

[0104] Among them, a t G t These are acceleration characteristics and pressure gradient characteristics, respectively; σ a σ G These are the standard deviations of the acceleration characteristics and the pressure gradient characteristics, respectively.

[0105] S2322. Determine whether the correlation coefficient is less than the preset threshold. If it is less, proceed to step S2323; otherwise, directly output the cross-modal attention feature matrix.

[0106] S2323. Perform feature compensation correction on the cross-modal attention feature matrix, specifically as follows:

[0107] By correcting the pressure feature vector using the motion feature vector, we obtain the corrected pressure feature vector, which can be expressed as:

[0108] F″ p =F′ p +0.3·MLP(F′ m )

[0109] Among them, F p "F represents the corrected pressure node characteristics." p '、F m 'These represent the stress node features and motion node features of the cross-modal attention feature matrix, respectively; MLP refers to the multilayer perceptron neural network structure, and MLP uses the GELU function as the activation function.

[0110] It is understandable that the cross-modal feature correction matrix obtained after feature compensation correction can be expressed as V′ fusion =[L′ m ,F′ v ,F″ p ].

[0111] As one embodiment, positive sample pairs and negative sample pairs can be represented as (V i V j ) and (V i V k The loss function used in iterative updates during S233 is as follows:

[0112] L total =L tri +0.3L phy

[0113] Among them, L total For the comprehensive loss function used, L tri For triplet loss, L phy The physical constraint loss is defined as follows:

[0114] L tri =max(||V i -V j ||2-‖V i -V k ||2+m,0)

[0115] L phy =ReLU(0.6-ρ)

[0116] In the above formula, ρ is the pressure-motion correlation coefficient calculated in S2321; V i The anchor sample feature matrix of positive sample pairs (or negative sample pairs), V j The positive sample feature matrix of positive sample pairs, V k is the negative sample feature matrix of the negative sample pair; m is the minimum distance difference between the positive and negative sample pairs, and the value of m ranges from 0.3 to 0.7.

[0117] It is understood that when the AdamW optimizer iteratively updates the cross-modal feature correction matrix, the update objects are the attention parameters (i.e., query matrix parameters, key matrix parameters, and value matrix parameters) and the MLP parameters. Iteratively updating the cross-modal feature correction matrix using the AdamW optimizer is a prior art technique, which can be readily implemented by those skilled in the art based on the foregoing embodiments, and will not be elaborated further here.

[0118] As one embodiment, a dynamic weight allocation step is further included after S23, wherein the dynamic weight allocation specifically includes:

[0119] S241. Obtain the data anomaly detection signal output by the sensor self-test module; the data anomaly detection signal includes environmental interference markers and modal confidence scores between each mode;

[0120] S242. The real-time data stream acquired by the input data acquisition module is processed in step S21 to extract real-time features and obtain a real-time feature vector. The real-time feature vector includes a real-time motion feature vector, a real-time visual feature vector, and a real-time pressure feature vector.

[0121] S243. Perform time offset compensation on the real-time feature vector, calculate the time delay cross-correlation value between each mode based on the real-time feature vector, and calculate the cross-modal time delay cross-correlation weight matrix.

[0122] S244. Calculate the similarity between pairs of features to obtain the real-time feature similarity matrix and generate the real-time weight matrix.

[0123] S245. Update the directed edge weight matrix;

[0124] S246. The updated directed edge weight matrix is ​​mixed with the real-time weight matrix in proportion to obtain the dynamic weight matrix.

[0125] Furthermore, in S243, the calculation method for the time delay cross-correlation value is as follows:

[0126]

[0127] Where, ρ i→j (τ) refers to the time-delay cross-correlation value from mode i to mode j (time delay is τ), F real_i Let F be the real-time feature vector of mode i. real_j Let be the real-time feature vector of mode j.

[0128] The cross-modal time-delay cross-correlation weight matrix is ​​calculated using the Time Delay Cross-Correlation (TDCC) algorithm, where the time delay used for calculating the weights from mode i to mode j is: the time delay cross-correlation value ρ from mode i to mode j. i→j The greatest time delay.

[0129] In S244, the similarity between pairwise features is calculated using cosine similarity. The method for calculating the real-time weight matrix is ​​as follows:

[0130] W real_time =softmax(0.8C+0.5S)

[0131] Among them, W real_time Let S be the real-time weight matrix, C be the confidence matrix, and S be the real-time feature similarity matrix.

[0132] The element C of the confidence matrix ij The modal confidence score between mode i and mode j is calculated as follows:

[0133]

[0134] Where ∈ represents a decimal value that is excluded from zero.

[0135] In S245, the method for updating the directed edge weight matrix is ​​as follows:

[0136] Obtain the historical directed edge weight matrix and perform an exponential moving average to obtain the current directed edge weight matrix; when the environmental interference flag is 1, set the current directed edge weight matrix to the preset default directed edge weight matrix.

[0137] The above exponential moving average process can be expressed as:

[0138]

[0139] Among them, W hist (t) Let W be the current directed edge weight matrix. hist (t-1) W is the directed edge weight matrix for history. realtime (t) The real-time weight matrix is ​​calculated by S244, and α is the preset matrix update coefficient.

[0140] The dynamic weight matrix in S246 is calculated as follows:

[0141] W dynamic =γW hist +(1-γ)W TDCC

[0142] Among them, W dynamic W is a dynamic weight matrix. hist W is the updated directed edge weight matrix obtained in S245. TDCC γ is the cross-modal time delay cross-correlation weight matrix obtained from S243, and γ is the preset mixing coefficient.

[0143] It should be noted that dynamically adjusting the fusion weights of multimodal features (i.e., calculating the dynamic weight matrix) can optimize the model's adaptability to the current environment. When the environment changes, the overall system can adjust in a timely manner and output the most accurate results.

[0144] Understandably, once the dynamic weight matrix is ​​obtained, the fused feature vector (dynamic weight fused feature) can be calculated. The calculation method is as follows:

[0145] V dynamic =W dynamic ·V concat

[0146] Among them, W dynamic V is a dynamic weight matrix. dynamic For dynamic weight fusion features, V concat This is the fusion feature vector obtained through S21-S23.

[0147] The dynamic weighted fusion feature is the unstructured behavioral feature data output by the multi-source heterogeneous data processing module.

[0148] As one embodiment, S24 further includes:

[0149] S25. Based on the outlier classifier, classify and detect the dynamic weight fusion features. If the classification and detection results meet the preset outlier range, return to S24 to perform dynamic weight allocation.

[0150] The outlier classifier is trained based on any one of the random forest algorithm, 1D-CNN model, or logistic algorithm. The specific training method is the existing technology and can be successfully implemented by those skilled in the art based on the description of the foregoing embodiments, and will not be described in detail here.

[0151] As one example, structured behavioral data refers to parent user behavior data in the teaching management system, including user interaction data, task completion data, and cognitive test data.

[0152] The user interaction data includes: click-through rate of each module, number of times resources are collected, number of times teaching resources are viewed, and frequency of teaching video playback.

[0153] The task completion data includes: answer accuracy rate, task completion rate, average task duration, and frequency of incorrect questions.

[0154] The cognitive test data includes: Piaget's cognitive development stage test scores and vocabulary size assessment results during the language-sensitive period.

[0155] As one example, the structured data feature module performs feature extraction in the following specific way:

[0156] For the user interaction data, time-series statistical aggregation and PCA dimensionality reduction are used to obtain interaction features;

[0157] For the task completion data, sliding window mean filtering and XGBoost feature importance calculation are used to obtain task completion features;

[0158] The cognitive test data is processed using data standardization to obtain cognitive test features.

[0159] It is understood that the above feature extraction method is the current existing technology, and those skilled in the art can successfully implement it based on the description of the foregoing embodiments, so it will not be described again here.

[0160] By concatenating interaction features, task completion features, and cognitive test features, structured behavioral feature data can be obtained.

[0161] As one embodiment, the dynamic image generation module generates dynamic images in the following specific ways:

[0162] S31. Align the dimensions of structured and unstructured behavioral feature data.

[0163] S32. Merge the dimension-aligned structured behavioral feature data with the unstructured behavioral feature data to obtain a comprehensive feature vector;

[0164] S33. Generate dynamic portrait vectors based on comprehensive feature vectors.

[0165] As one embodiment, in S32, the fusion can be performed using a static weight-based method, which can be specifically expressed as follows:

[0166] H fusion =θ·F struct +(1-θ)·V compressed

[0167] Among them, H fusion For the comprehensive feature vector, F struct V is structured behavioral feature data after dimension alignment. compressed This represents the unstructured behavioral feature data after dimension alignment; θ represents the preset structured feature weights. The method for generating dynamic profile vectors in S33 can be implemented through a fully connected mapping, which can be represented as:

[0168] P profile =sigmoid(W map ·H fusion +b map )

[0169] In the formula, P profile For dynamic image vectors; H fusionFor the comprehensive feature vector, W map To preset the weights of the fully connected layer, b map This is the preset offset.

[0170] The dynamic image vector obtained after processing by S33 can be represented as:

[0171] P profile =[P cognitive ,P motor ,P social ,P parent ]

[0172] P cognitive P motor P social P parent These represent the vectors of children's cognitive level, children's motor ability, children's social tendency, and parents' preferences, respectively.

[0173] As one embodiment, the resource recommendation module generates teaching resource recommendation results through a hybrid recommendation algorithm, specifically including the following steps:

[0174] S41. Obtain dynamic portraits and a knowledge graph of the teaching syllabus;

[0175] S42. Filter teaching resources in the knowledge graph of the teaching syllabus based on the cognitive level vector of children in the dynamic portrait;

[0176] S43. Based on the vectors of children's motor ability, children's social tendency, and parents' preferences, the teaching resources are scored to obtain a recommended list of teaching resources.

[0177] As one embodiment, S42 includes the following steps:

[0178] S421. Calculate the similarity between each cognitive stage and each teaching resource path; the calculation method is as follows:

[0179]

[0180] Wherein, π(r) i →S stage ) refers to the process from resource node r i Cognitive stage node S stage The shortest teaching logic path; |π(r i )|For resource r i The number of jumps to its directly related knowledge points; |π(S stage )| for stage S stage The number of jumps to its root knowledge point;

[0181] S422. Based on the child's cognitive level vector, determine the current dominant cognitive stage S. current_stageThe determination method is: S current_stage =arg max(P cognitive );

[0182] S423. Based on the results of S421 and S422, select the teaching resource nodes that meet the preset resource similarity conditions under the current dominant cognitive stage; this can be represented as:

[0183]

[0184] Among them, R filtered PathSim0 is the set of candidate teaching resources, and PathSim0 is the preset resource similarity condition.

[0185] As one embodiment, S43 includes the following steps:

[0186] S431. Based on the vectors of children's motor ability, children's social tendency, parents' preferences, and the corresponding preset motor requirement vectors, preset social tag sets, and preset parents' preference requirement vectors, calculate the motor ability suitability score, social tendency matching score, and parents' preference fit score for each teaching resource in the candidate teaching resource set.

[0187] S432. The comprehensive resource score is calculated by weighting and summing the scores of athletic ability suitability, social tendency matching, and parental preference compatibility.

[0188] S433. Based on the comprehensive scores of each teaching resource, select recommended teaching resources and obtain a list of recommended teaching resources.

[0189] Specifically, the dynamic capability fit score is calculated as follows:

[0190]

[0191] Among them, P motor A(r) represents the vector of children's cognitive level. i ) is the preset motion requirement vector.

[0192] The social tendency matching score is calculated as follows:

[0193] Score2 = Jaccard(P social ,T(r i ))

[0194] Among them, P social T(r) represents the vector of a child's social tendency. i ) is a preset set of social tags.

[0195] The calculation method for the parent preference matching score is as follows:

[0196]

[0197] Among them, P parent Let I(r) represent the parent preference vector. i ) represents the preset parent preference requirement vector.

[0198] As one embodiment, the method described in this invention can be implemented in software and / or a combination of software and hardware, for example, using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device.

[0199] The method described in this invention can be implemented as a software program, which can be executed by a processor to achieve the steps or functions described above. Similarly, the software program (including associated data structures) can be stored in a computer-readable recording medium, such as RAM memory, magnetic or optical drives, floppy disks, and similar devices.

[0200] In addition, some steps or functions of the method described in this invention can be implemented in hardware, for example, as a circuit that works with a processor to perform the various steps or functions.

[0201] Furthermore, a portion of the methods described in this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions invoking the methods described in this invention can be stored in a fixed or removable recording medium, and / or transmitted via a data stream in a broadcast or other signal carrying medium, and / or stored in the working memory of a computer device operating according to the program instructions.

[0202] As one embodiment, the present invention also provides an apparatus comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the apparatus is triggered to run a method and / or technical solution based on the foregoing plurality of embodiments.

[0203] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0204] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0205] Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0206] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A personalized recommendation system for teaching resources based on user profiles, characterized in that, The system consists of the following modules: The data acquisition module is used to collect unstructured behavioral data generated by children during their activities in the kindergarten. A multi-source heterogeneous data processing module is used to perform spatiotemporal feature fusion processing on the unstructured behavioral data to obtain unstructured behavioral feature data; the spatiotemporal feature fusion processing includes sampling frequency difference alignment processing, cross-modal correlation feature extraction, and dynamic weight allocation; The structured data module is used to store structured behavioral data and provides a query interface for structured behavioral data. The structured data feature module is used to extract features from structured behavioral data to obtain structured behavioral feature data. The dynamic profile generation module is used to generate dynamic profiles based on the unstructured behavioral feature data and the structured behavioral feature data. The resource recommendation module generates teaching resource recommendations based on dynamic profiles using a hybrid recommendation algorithm.

2. The system according to claim 1, characterized in that, The data acquisition module includes: The motion acquisition submodule employs an UWB ultra-wideband positioning system and a heart rate sensor; The image acquisition subsystem employs an edge computing camera that integrates micro-expression recognition and body language analysis algorithms; The pressure acquisition submodule employs an embedded thin-film pressure sensor. The sensor self-test module is used to monitor the data quality and working status of each sub-module in real time and output data anomaly detection signals.

3. The system according to claim 1, characterized in that, The sampling frequency difference alignment process includes: S11. A unified time reference axis is established, and a dynamic time warping algorithm is used to establish a timestamp mapping. S12, Multi-band signal reconstruction and resampling; S13. Cyclic redundancy check is used and a Hankel matrix is ​​constructed for large-scale missing data to achieve data alignment optimization and verification.

4. The system according to claim 1, characterized in that, The cross-modal association feature extraction is achieved through a spatiotemporal graph neural network, specifically including: S21, Sub-module feature parsing, including motion feature extraction, visual feature extraction and stress feature extraction; S22. Construct a spatiotemporal heterogeneous graph model based on graph neural networks; S23. Perform joint optimization on the spatiotemporal heterogeneous graph model; S24. Dynamic weight allocation is performed based on real-time data to obtain a dynamic weight matrix.

5. The system according to claim 3, characterized in that, Joint optimization of spatiotemporal heterogeneous graph models achieves deep fusion of cross-modal features through joint optimization guided by multi-head graph attention mechanisms and physical constraints, specifically including: - S231. Using a multi-head attention mechanism, perform four-head attention calculation on spatiotemporal heterogeneous structural data; after concatenating the output results of the four-head attention calculation, obtain the cross-modal attention feature matrix by compressing it with the dimension of the gated linear unit. S232. Optimize the cross-modal attention feature matrix by using acceleration features and pressure gradient features as physical constraints to obtain the cross-modal feature correction matrix; S233. The AdamW optimizer is used to iteratively update the cross-modal feature correction matrix to obtain the fused feature vector.

6. The system according to claim 3, characterized in that, S24 dynamic weight allocation includes: S241. Obtain the abnormal data detection signal output by the sensor self-test module; S242. Input the real-time data stream collected by the input data acquisition module, execute step S21 to perform real-time feature extraction, and obtain the real-time feature vector; S243. Perform time offset compensation on the real-time feature vector, calculate the time delay cross-correlation value between each mode based on the real-time feature vector, and calculate the cross-modal time delay cross-correlation weight matrix. S244. Calculate the similarity between pairs of features to obtain the real-time feature similarity matrix and generate the real-time weight matrix. S245. Update the directed edge weight matrix; S246. The updated directed edge weight matrix is ​​mixed with the real-time weight matrix in proportion to obtain the dynamic weight matrix.

7. The system according to claim 3, characterized in that, The cross-modal association feature extraction also includes: S25. Based on the outlier classifier, classify and detect the dynamic weight fusion features. If the classification and detection results meet the preset outlier range, return to S24 to perform dynamic weight allocation.

8. The system according to claim 1, characterized in that, The structured behavioral data refers to parent user behavior data from the teaching management system, including user interaction data, task completion data, and cognitive test data.

9. The system according to claim 1, characterized in that, The specific methods by which the dynamic image generation module generates dynamic images include: S31. Align the dimensions of structured and unstructured behavioral feature data. S32. Merge the dimension-aligned structured behavioral feature data with the unstructured behavioral feature data to obtain a comprehensive feature vector; S33. Generate dynamic portrait vectors based on comprehensive feature vectors.

10. The system according to claim 8, characterized in that, S32 uses a static weight-based fusion method; In S33, dynamic image vectors are generated through fully connected mapping; The dynamic image vector can be represented as: P profile =[P cognitive ,P motor ,P social ,P parent ] P cognitive P motor P social P parent These represent the vectors of children's cognitive level, children's motor ability, children's social tendency, and parents' preferences, respectively.

11. The system according to claim 1, characterized in that, The specific methods by which the resource recommendation module generates teaching resource recommendation results include: S41. Obtain dynamic portraits and a knowledge graph of the teaching syllabus; S42. Filter teaching resources in the knowledge graph of the teaching syllabus based on the cognitive level vector of children in the dynamic portrait; S43. Based on the vectors of children's motor ability, children's social tendency, and parents' preferences, the teaching resources are scored to obtain a recommended list of teaching resources.

Images