An education-oriented AI workbench application capability construction and resource matching system

Abstract

The application relates to the technical field of artificial intelligence and education information processing, and discloses an AI work station application ability construction and resource matching system for education, which comprises a behavior collection module, an ability modeling module, an ability reasoning module, a cross-domain fusion module, a path recommendation module and a strategy optimization module. The behavior collection module is used for acquiring multi-modal behavior data and constructing a behavior graph. The ability modeling module is used for generating an ability graph and identifying an ability gap. The ability reasoning module is used for modeling an ability dimension and generating an ability vector. The cross-domain fusion module realizes multi-graph atlas alignment and constructs a transfer graph. The path recommendation module generates a recommended path based on ability and resource matching. The strategy optimization module optimizes a strategy through reinforcement learning based on feedback. The application improves the ability modeling and potential relationship mining precision based on a graph neural network, realizes cross-domain ability graph fusion and personalized resource matching, enhances the dynamic adaptability and feedback optimization capability of a recommended strategy, and effectively improves the intelligent level of education.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and education information processing technology, in particular to an AI workbench application ability construction and resource matching system for education. BACKGROUND

[0002] Under the background of rapid development of artificial intelligence technology, the education field is gradually moving towards an intelligent and data-driven transformation stage. In particular, in the context of higher education and vocational education, students' learning behavior data is increasingly rich, covering click records, answer traces, video watching behavior, interactive communication and other multi-modal forms. These behavior data can provide important support for personalized teaching, but existing ability assessment and resource recommendation systems often fail to effectively utilize these deep and structured learning trajectory information.

[0003] Traditional education information systems mostly model abilities based on static labels, knowledge point trees or course scores. Such methods have problems such as single level of abstraction, weak dynamic updating ability, and are difficult to reflect the ability evolution path of learners in the actual learning process. In addition, current mainstream resource recommendation strategies mostly rely on collaborative filtering or shallow matching models, ignoring the deep coupling relationship between learner ability state and resource target, resulting in recommended results that often cannot accurately meet the current growth needs of learners.

[0004] At the resource construction and sharing level, due to inconsistent data standards between schools and platforms, scattered ability label systems, existing technologies are difficult to align and integrate multi-source ability graphs. This further limits the intelligent distribution and reuse of educational resources in a larger range, which is not conducive to promoting cross-institutional ability transfer and balanced development of regional educational resources.

[0005] On the other hand, although some intelligent education systems attempt to introduce machine learning for recommendation strategy optimization, most use offline modeling and static parameter tuning, lacking feedback mechanisms for real-time interaction with learners, and unable to dynamically adjust resource matching paths during continuous learning. This makes the system lack of response ability and adaptability when facing fluctuations in learner ability and interest shifts.

[0006] Therefore, the present application proposes an AI workbench application ability construction and resource matching system for education to solve the deficiencies of the prior art. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application provides an AI workbench application ability construction and resource matching system for education, which solves the problems of inaccurate learner ability modeling, low resource recommendation matching degree and lack of dynamic optimization mechanism in the prior art.

[0008] To achieve the above object, the application is implemented by the following technical solutions: An AI workbench application ability construction and resource matching system for education includes:

[0009] A behavior collection and graph construction module is configured to acquire multi-modal behavior data of learners and convert the multi-modal behavior data into a behavior graph.

[0010] An ability modeling module is configured to construct a learner ability graph based on the behavior graph and identify structural holes in the ability graph through topological compression analysis.

[0011] An ability reasoning module is configured to perform multi-granularity latent variable modeling on the ability graph, infer three ability dimensions of explicit ability, middle-layer ability and potential transfer ability, and generate an ability vector based on the ability dimensions.

[0012] A cross-domain fusion module is configured to perform structural alignment and fusion on the ability graphs of multiple learners to form a unified cross-domain ability transfer graph.

[0013] A path matching and recommendation module is configured to generate a structured resource label for candidate educational resources, generate a resource recommendation path according to the matching relationship between the ability vector and the resource label, and output a resource recommendation strategy.

[0014] A strategy optimization module is configured to perform reinforcement learning optimization on the resource recommendation strategy based on learner feedback.

[0015] Preferably, the behavior collection and graph construction module includes:

[0016] A data collection module is configured to receive multi-modal behavior data from learners, the multi-modal behavior data including at least action data, voice data, physiological data and environmental data of learners.

[0017] A data processing module is configured to pre-process, clean and extract features from the multi-modal behavior data.

[0018] A behavior graph construction module is configured to convert the pre-processed multi-modal behavior data into a behavior graph, the behavior graph including behavior nodes of learners and relationship edges between behaviors.

[0019] Preferably, the ability modeling module includes:

[0020] An ability graph construction module is configured to convert the behavior graph into an ability graph of learners, the ability graph including nodes representing abilities of learners and edges representing relationships between abilities.

[0021] A hole area identification module is configured to identify a hole area in the ability graph through a topological compression algorithm and identify potential ability gaps.

[0022] Preferably, the capability inference module comprises:

[0023] a latent variable modeling module configured to perform multi-granularity latent variable modeling on the capability graph to infer three capability dimensions of explicit capability, intermediate capability and latent transfer capability;

[0024] a vector generation module configured to generate a capability vector based on the capability dimensions to represent the capability level of the learner in different capability dimensions.

[0025] Preferably, the step of performing multi-granularity latent variable modeling on the capability graph to infer three capability dimensions of explicit capability, intermediate capability and latent transfer capability comprises:

[0026] initializing a feature vector for each node in the capability graph initializing a feature vector wherein, represents the dimension of the capability feature;

[0027] performing iterative updating of node representation by a graph neural network model using the following formula:

[0028] ;

[0029] wherein, represents the feature representation of node in the -th layer, is a set of adjacent nodes of node , is an attention weight between nodes, is a trainable weight matrix, is an activation function, represents a layer index in the graph neural network, is an original feature vector of node ; represents the state of node when the -th layer is represented;

[0030] extracting a final representation of node in the output layer of the graph neural network , and inferring representations of three capability dimensions of explicit capability, intermediate capability and latent transfer capability by a multi-granularity decoder.

[0031] Preferably, the cross-domain fusion module comprises:

[0032] a heterogeneous feature alignment module configured to perform structural alignment on capability graphs of multiple learners;

[0033] The association relationship modeling module is configured to fuse the multiple ability graphs after structure alignment to form a unified cross-domain ability transfer graph.

[0034] Preferably, the step of fusing the multiple ability graphs after structure alignment to form a unified cross-domain ability transfer graph comprises:

[0035] The feature vector of each node is normalized, and the following formula is used:

[0036] ;

[0037] wherein, is the original feature vector of the node , is the normalized feature vector, is the Euclidean norm of the feature vector of the node ;

[0038] The multiple normalized ability graphs are fused at the node level using the following formula:

[0039] ;

[0040] wherein, is the fused node feature vector, is the feature vector of the node in the i-th graph, is the weighting coefficient of the i-th graph, is the number of fused graphs; A unified cross-domain ability transfer graph is constructed according to the fused node feature vector.

[0041] Preferably, the path matching and recommendation module comprises:

[0042] A resource label generation module is configured to perform structural processing on the candidate educational resources to generate resource labels containing knowledge points, skill dimensions, and use scenarios;

[0043] A matching relationship construction module is configured to calculate the matching degree between the learner ability vector and the resource label;

[0044] A path recommendation module is configured to construct a resource recommendation path based on the matching degree and generate a corresponding resource recommendation strategy.

[0045] Preferably, the strategy optimization module comprises:

[0046]

[0047] ​​​​a feedback modeling module configured to collect interaction feedback information of the learner on the recommended resource;

[0048] a state representation module configured to construct a state space for policy optimization based on the current capability state and the resource usage state;

[0049] a policy learning module configured to dynamically optimize the resource recommendation policy based on the state space using a reinforcement learning method, and generate an updated resource recommendation policy output.

[0050] The application provides an AI workbench application capability construction and resource matching system for education. The system has the following advantages:

[0051] 1. The application adopts a capability graph construction scheme based on a graph neural network (GNN) and multi-modal data fusion, which effectively improves the accuracy of depicting the real capability of the learner. Compared with the traditional method of relying only on questionnaires or linear behavior scores, the application solves the problem of fragmented and static evaluation of learner's capability, and achieves a more comprehensive and dynamic understanding of individual growth trajectory.

[0052] 2. The application can accurately find the breakpoints in the capability chain through the structure compression and capability hole identification technology of the behavior graph, thereby realizing targeted resource supplementation. Unlike the existing education recommendation system which only pushes content based on interest or course similarity, the application starts from the structure level and breaks through the limitation of resource recommendation being independent of real needs.

[0053] 3. The application constructs a cross-capability transfer graph by fusing multiple learner capability graphs, which realizes the sharing and collaboration of capability data between different colleges and platforms while ensuring the difference of individual characteristics. Traditional platforms are difficult to realize resource interconnection and matching optimization due to the data island effect, while the application breaks down the barriers and promotes the collaborative evolution of the intelligent education system.

[0054] 4. In terms of recommendation strategy optimization, the application introduces a reinforcement learning mechanism to adjust the strategy path in real time combined with learner feedback, making the recommendation result more sustainable and adaptive. Traditional systems are often static and fixed, with lagging adjustment, making it difficult to respond to changing needs during the learning process. The system effectively makes up for this shortcoming, making resource matching both real-time and long-term growth-oriented. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a system architecture diagram of the application;

[0056] Figure 2 is a method flowchart of the application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0058] Please refer to the drawings attached Figure 1 The embodiment of the present application provides an AI workbench application capability construction and resource matching system for education, comprising:

[0059] a behavior collection and graph construction module, configured to acquire multi-modal behavior data of a learner and convert the multi-modal behavior data into a behavior graph;

[0060] The behavior collection and graph construction module is responsible for sensing, processing and structure mapping of multi-modal behavior data of a learner in a learning process. The module plays a basic supporting role in the whole technical process, and its output result provides necessary data graph structure input for subsequent capability modeling. Therefore, the module and the capability graph construction module have a close logical dependency relationship, the input of the module is the original behavior data from the multi-modal sensing end in the learning environment, and the output of the module is a behavior graph with a graph structure, which is used to support accurate construction of the capability graph.

[0061] In the embodiment, the behavior collection and graph construction module comprises three sub-modules, namely a data collection module, a data processing module and a behavior graph construction module.

[0062] Generally, the data collection module is configured to synchronously collect multi-modal behavior data of a learner in an interaction process from different sensing channels. The multi-modal data specifically include, but are not limited to, the following types:

[0063] action data: for example, mouse click trajectory, keyboard click frequency, eye movement trajectory, body posture, sitting posture change;

[0064] voice data: for example, learner's oral question voice, voice command response, reading audio;

[0065] physiological data: for example, heart rate, skin electricity, brain wave, facial expression index;

[0066] environmental data: for example, light intensity, noise level, temperature and humidity change, etc.

[0067] As an option, the data collection can be completed cooperatively by hardware collection devices such as SDK embedded in a teaching platform, a camera, a microphone, a wearable device, etc.

[0068] In one possible implementation, the system has independent sampling frequency and sampling window for each modality. For example, the action data sampling frequency is 30 frames per second, while the speech sampling is 16 kHz, and the environmental data is recorded in 5-second units.

[0069] Specifically, after data collection is completed, the data processing module is configured to preprocess, clean and extract features from the multi-modal data. The module aligns the data by timestamp, removes redundant or invalid data (such as static non-response behavior, background noise segment, non-human speech, etc.), and uses modality-specific feature extractors to encode the raw data into vector representation.

[0070] For example, for speech data, the system can use MFCC (Mel Frequency Cepstral Coefficients) features for extraction, converting each segment of speech into a 13-dimensional acoustic vector.

[0071] For action data, a sliding window-based statistical feature method can be used, such as calculating the moving speed, acceleration, and offset angle of the mouse trajectory.

[0072] In some embodiments, the data processing module can also introduce a lightweight deep neural network (such as MobileNet) to perform pose estimation on image data obtained by the camera, thereby inferring the learner's attention level.

[0073] After preprocessing is completed, the behavior graph construction module is configured to convert the preprocessed multi-modal behavior data into a behavior graph. The behavior graph is a structured graph model, with nodes representing atomic-level behavior events and edges representing the association between behaviors.

[0074] Specifically, the behavior graph can be defined as:

[0075] ;

[0076] wherein, represents a set of behavior nodes, each node represents a behavior unit with a timestamp and a feature vector; represents a set of edges, each edge represents a behavior and a behavior connection relationship in a certain semantic or time scale; each edge is attached with a weight , which is used to measure the correlation or sequential dependence between two behaviors, and the calculation method is as follows:

[0077] ;

[0078] wherein: representing the edge weight between behaviors in the behavior graph representing the edge weight between behaviors in the behavior graph representing the edge weight between behaviors in the behavior graph respectively representing the behavior feature vector of the node respectively representing the behavior feature vector of the node representing the cosine similarity of the behavior vectors representing the time difference between two behaviors is a hyperparameter weight coefficient, used to adjust the proportion of the influence of structural similarity and time proximity is a time decay factor, used to measure the time distance between two behaviors.

[0079] As a specific implementation, the system can divide the behavior events into time-ordered sliding windows, and construct the edge set based on the behavior pairs within the window.

[0080] In some other embodiments, the establishment of edges not only considers the time sequence, but also considers the similarity of the context environment, for example, the behavior pairs occurring in the same teaching task or scene will have a higher connection probability.

[0081] Generally, the behavior graph will be directly input into the ability modeling module after construction, to further generate the learner ability graph, and implement the subsequent ability identification and recommendation reasoning tasks.

[0082] Further, the behavior graph construction module supports dynamic updating and real-time graph structure maintenance, ensuring that the behavior graph is gradually improved as the learner's behavior evolves.

[0083] In summary, through the behavior collection and graph construction module in the present application, high-dimensional structured representation of learner behavior data can be effectively realized, providing accurate, timely, and expandable underlying support for subsequent ability modeling and resource matching.

[0084] The ability modeling module is used to construct the learner ability graph based on the behavior graph, and identify the structural voids in the ability graph through topological compression analysis.

[0085] After the collection, processing and behavior graph construction of multi-modal behavior data are completed, the behavior graph information is further converted into the learner ability graph by the ability modeling module, and the potential void areas in the structure are identified. As a core component that connects the previous and the next in the system, on the one hand, the module inherits the structural semantics of the behavior graph, and on the other hand, it provides a structural basis for subsequent ability recommendation, resource matching and other application modules.

[0086] Generally, the ability modeling module includes an ability graph construction module and a void area identification module.

[0087] The ability graph construction module is used to map the structured behavior graph to the ability graph where the nodes represent the ability items of the learners, and the edges represent the relationships between the ability items. The edges represent the intrinsic relationships between the abilities.

[0088] As an option, the graph supports dynamic updates according to the synergistic relationships between the abilities. In a possible implementation, the generation of the ability nodes depends on a pre-defined mapping function from the behavior nodes to the ability labels , and specifically, the mapping process satisfies the following expression:

[0089] ;

[0090] wherein: is a behavior node in the behavior graph; is the corresponding ability label; is a function for establishing the correspondence between the behavior and the ability, which is usually trained by artificial annotation samples or constructed based on a priori rule base.

[0091] In some embodiments, for constructing the edge relationships of the ability graph, the system converts the edge weights between the nodes in the original behavior graph . The edge weight is defined as follows:

[0092] ;

[0093] wherein, represents the edge weight between the nodes and in the behavior graph, i.e., the relationship strength between them; , is a weighting coefficient for balancing the influences of the behavior feature similarity and the time decay, satisfying ; represents the similarity between the feature vectors and of the nodes and ; is a time decay factor for representing the influence of the time difference between the behavior nodes and , and the edge weight decays as the time difference increases.

[0094] Subsequently, the mapped ability node set and the edge relationships therebetween are constructed into the ability graph, and the edge weight thereof is calculated by the following formula:

[0095] ;

[0096] wherein, is a capability node; represents all the behavior nodes mapped to the capability node ; is the edge weight between the behavior nodes in the original behavior graph.

[0097] As an optimization strategy, the capability edge weight can be further normalized to eliminate the scale bias in the graph structure, and the normalization formula is as follows:

[0098] ;

[0099] wherein, is the normalized edge weight, so that the sum of the edge weights of the node pointing to all neighbors is 1;

[0100] After the construction of the capability graph, the system further detects the structural holes in the capability graph through the hole area identification module to assist in judging the potential missing areas in the learner's capability system. Specifically, one possible implementation way is to realize hole identification through topological compression analysis.

[0101] In some embodiments, the analysis method is based on the sparse adjacency matrix of the capability graph, and dimensionality reduction transformation and hole block compression are performed to identify topologically low-connected or non-connected sub-regions. The identification of the hole area can be based on the following indicators:

[0102] ;

[0103] wherein, represents the hole coefficient of the node ; represents the set of nodes adjacent to in the capability graph; is the normalized edge weight.

[0104] Generally, when is greater than a preset threshold , the node can be identified as a structural hole node, and its corresponding capability is considered to possibly exist knowledge omission or unachieved state.

[0105] As an extended implementation, the hole identification result can be compared with the learning resource library to further generate learning suggestions or resource recommendation paths for personalized capability improvement.

[0106] The capability modeling module, as the core hub of the system, realizes the conversion from the behavior graph to the capability graph, and identifies the structural holes through topological compression analysis to provide structural support for subsequent capability recommendation and resource matching.

[0107] a capability reasoning module configured to perform multi-granularity latent variable modeling on the capability graph, infer three capability dimensions of explicit capability, intermediate capability, and potential transfer capability, and generate a capability vector based on the capability dimensions;

[0108] The capability reasoning module is based on the capability graph described above, and performs in-depth mining and hierarchical representation of the learner's capability system through multi-granularity latent variable modeling technology. As a further processing unit of the system, the capability reasoning module is configured to perform deep modeling and dimension decomposition on the feature representation of the nodes in the capability graph, to infer three capability dimensions of explicit capability, intermediate capability, and potential transfer capability, and finally form a capability vector reflecting the comprehensive capability features of the learner, to provide a representation basis and structural support for subsequent resource matching and personalized capability intervention.

[0109] In this embodiment, the capability reasoning module includes a latent variable modeling module and a vector generation module.

[0110] Generally, the latent variable modeling module is configured to perform multi-granularity hierarchical analysis on the input capability graph, and mine different capability dimensions implied therein. In one possible implementation, the module adopts a representation learning method based on a graph neural network (GNN), to aggregate the structure and attribute information of the nodes in the capability graph layer by layer, to capture the complex semantic relationships between the capabilities.

[0111] Specifically, at the beginning of graph neural network modeling, the system initializes the feature vector of each capability node in the capability graph to wherein is the dimension size of the initial capability feature, which is usually determined by the aggregation representation from the behavior node to the capability label.

[0112] In some embodiments, to enhance the semantic distinguishability of node feature propagation, the system further introduces an attention mechanism between the nodes, to perform weighted processing on the adjacent node information, and realize differential modeling of heterogeneous information. The update iteration process of the graph neural network follows the following formula:

[0113] ;

[0114] wherein, represents the feature representation of the node in the th layer; is the set of adjacent nodes of the node ; is the attention weight of the node to the node The attention weight represents the importance of the neighbor node information to the target node's representation update. This is a trainable weight matrix used for linear transformations; For non-linear activation functions, ReLU, ELU, or Sigmoid are typically chosen. It serves as an index for network layers, controlling the receptive field range of aggregation; To indicate the first Layer Time Node The state.

[0115] After multi-layer propagation, the system extracts the final representation of each node at the output layer of the graph neural network. The node is deconstructed using a pre-defined multi-granularity capability decoder to obtain its vector representation across three capability dimensions:

[0116] Dominant ability vector ;

[0117] Mid-level capability vector ;

[0118] Potential transferability vector .

[0119] As an alternative, decoding of the three capability dimensions can be implemented based on linear mapping functions, attention-based decoding networks, or variational decoders. Each method corresponds to a different modeling focus; for example, linear mapping is suitable for capability structures with strong local consistency, while attention mechanisms are more suitable for scenarios where there are heterogeneous dependencies between capabilities.

[0120] In some embodiments, the system further introduces prior structural constraints on the relationships between capabilities to enhance the interpretability and separability of multi-granularity latent variables in the semantic space. For example, a capability transfer path graph is introduced, using mid-level capabilities as intermediary nodes connecting explicit and potential capabilities to construct a "mastered-not yet mastered-transferable" capability chain.

[0121] Finally, the vector generation module, based on the representations of the three capability dimensions mentioned above, merges and generates a comprehensive vector representation of the capability nodes. This indicates that the learner is at the node. The overall ability level for the corresponding ability item. Generally, the ability vector can be calculated as follows:

[0122] ;

[0123] in: , , For the weighting coefficients, satisfying , for regulating the weight of each dimension in the final vector; each component capability vector can be adjusted according to training data or empirical priori.

[0124] As an extensible implementation, the capability inference module can also integrate a multi-head attention mechanism or a graph contrastive learning structure to further enhance the discriminability and generalizability of the capability vector.

[0125] The capability inference module, based on the capability graph constructed in the previous step, uses a graph neural network to model the multi-level features of the nodes and infer the three dimensions of explicit capability, intermediate capability, and potential transfer capability through latent variable modeling. Subsequently, through multi-granularity decoding and vector fusion, the capability vector is generated, comprehensively representing the learner's capability level in different capability dimensions, providing structured support for personalized recommendation and capability intervention.

[0126] The cross-domain fusion module is used to structure align and fuse the capability graphs of multiple learners to form a unified cross-domain capability transfer graph.

[0127] Through the cross-domain fusion module, the capability structure differences between individuals are bridged, and knowledge transfer relationship modeling is realized across individuals and in multiple scenarios. The multi-granularity capability vector output by the capability inference module serves as the input basis for the cross-domain fusion module, providing a unified feature representation form for its structure alignment and fusion operations. The design of the cross-domain fusion module not only takes in the capability graph and capability vector information output by the previous modules, but also provides a structured transfer graph support for the subsequent transfer path generation and generalization resource recommendation.

[0128] In this embodiment, the cross-domain fusion module mainly includes a heterogeneous feature alignment module and an association relationship modeling module, which are used to structure align and semantically fuse the capability graphs of multiple learners at the feature level, thereby constructing a unified cross-domain capability transfer graph.

[0129] Generally, the heterogeneous feature alignment module first performs structure mapping and feature normalization on the capability nodes in the capability graphs of different learners to ensure the comparability and consistency of the subsequent graph structure.

[0130] In one possible implementation, the system performs the following normalization operation on the feature representation vector of each capability node

[0131] ;

[0132] wherein, represents the original capability vector; represents the Euclidean norm (i.e., L2 norm) of the vector; represents the normalized unit vector.

[0133] ​​This normalization operation can eliminate structural mismatch between different capability graphs due to different representation scales, which is beneficial for unifying node similarity measurement and structural alignment in a unified semantic space.

[0134] In some embodiments, to achieve structure-level node alignment, the system further introduces a node semantic embedding matching mechanism to construct a cross-graph node matching function. ,in Indicates the first The first learner map in the first learner map For each node, the matching function calculates a matching score based on multiple factors such as vector cosine similarity, domain label consistency, and historical behavior context, in order to obtain the node equivalence relationship after structural alignment.

[0135] Specifically, the system can construct an alignment matrix. ,in Indicates the first The nodes in the graph are related to the first The nodes in the graph are semantically aligned. , These represent the number of graph nodes participating in the fusion. After node alignment, the association modeling module fuses the capabilities of multiple graphs according to their alignment relationships. In one possible implementation, the system uses a weighted feature fusion strategy to integrate node vectors, with the fusion formula as follows:

[0136] ;

[0137] in, The fused node feature vectors For the first Nodes in the graph eigenvectors, Indicates the first The weights of each graph in the fusion process typically satisfy the following: This weight can be determined by a combination of indicators such as sample contribution, behavioral coverage, or learning time weight. The number of graphs participating in the fusion;

[0138] In some embodiments, the fusion operation may also consider consistency verification of edge weights, and use a collaborative filtering mechanism to retain or adjust edges with structural conflicts in order to maintain the structural rationality of the fused graph.

[0139] Subsequently, the system based on the aforementioned fusion node vector Based on the corresponding alignment edge relationships, a unified cross-domain capability transfer graph is constructed as follows:

[0140] ;

[0141] wherein: is the fused capability node set; represents the cross-maps edge set, and the edge weight can be obtained by weighted combination of the edge relationship in the original map.

[0142] As an extended technical solution, the system can also construct a capability migration path model based on the cross-domain capability migration map, depict the reachability and path optimization degree of the learner from the existing capability state to the target capability state, and provide a graph structure basis for personalized learning path planning.

[0143] In some embodiments, the capability migration map can further learn the cross-map migration potential through a graph convolution network or a graph attention network to support efficient recommendation decisions.

[0144] In summary, the cross-domain fusion module fuses multiple capability maps into a unified structural migration map through feature normalization, node matching, and vector weighted fusion, effectively realizing the structural depiction of cross-individual capability migration potential. This module is closely connected to the capability modeling and reasoning module in structure, and provides unified migration network support for the resource recommendation module.

[0145] The path matching and recommendation module generates a structured resource label for the candidate educational resource, generates a resource recommendation path according to the matching relationship between the capability vector and the resource label, and outputs a resource recommendation strategy;

[0146] The path matching and recommendation module receives the fused capability vector and migration graph structure, and based on the candidate educational resource, constructs a matching mechanism and a recommendation path, and outputs a resource recommendation strategy for capability improvement. This module is not only closely integrated with the upstream capability representation model, but also provides structured and target-oriented resource pushing logic support for terminal learning services.

[0147] In this embodiment, the path matching and recommendation module mainly includes a resource label generation module, a matching relationship construction module, and a path recommendation module, which are used to complete key processes such as educational resource structured processing, capability resource matching calculation, and recommendation path generation.

[0148] Generally, the resource label generation module first uniformly structures the candidate educational resource, and extracts the knowledge points, skill dimensions, and applicable scene information associated therewith. Specifically, the system represents each educational resource as a structured label vector The vector is composed of the following three sub-vectors:

[0149] Knowledge point representation vector ;

[0150] Skill dimension representation vector ; ​

[0151] Using scene representation vectors .

[0152] Each sub-vector can be obtained by encoding resource metadata using a text semantic embedding model (such as BERT, RoBERTa), or generated by one-hot encoding of expert-labeled discrete labels.

[0153] As an option, the system can concatenate the above label vectors to form a complete label vector:

[0154] ;

[0155] Wherein, the symbol “ ” represents the vector concatenation operation.

[0156] Subsequently, the matching relationship construction module calculates the matching degree between the learner ability vector and the resource label vector. The ability vector is derived from the aforementioned ability inference and fusion process, and contains the weighted combination results of the explicit ability vector , the intermediate ability vector , and the latent transfer ability vector .

[0157] Specifically, the matching degree calculation can use cosine similarity:

[0158] ;

[0159] Wherein, represents the inner product of the ability vector and the label vector; , are the respective Euclidean norms; represents the matching degree between the learner and the resource .

[0160] In some embodiments, in order to improve the discriminability of the matching result, the system can also introduce a triple matching loss structure to train and optimize the matching function:

[0161] ;

[0162] Wherein: is the matching degree of the positive sample resource; is the matching degree of the negative sample resource; is a preset interval hyperparameter.

[0163] After the matching relationship is constructed, the path recommendation module calculates the matching degree based on the matching degree calculation result, and combines the aforementioned cross-domain ability transfer graph The node connection relationship in the graph is used to generate an optimal resource recommendation path. Generally, the path recommendation takes the current ability node of the learner as the starting point and the target ability node as the end point, and the resources in the path are selected according to the recommendation score of the corresponding ability node in the graph and the path accessibility.

[0164] In a possible implementation, the system builds a path scoring function , evaluates all reachable paths , and performs the following evaluations:

[0165] ;

[0166] wherein, is the position weight of the i th resource in the path; is the aforementioned ability-resource matching score; The greater the value is, the higher the value of the path recommendation is.

[0167] As an extensible way, the path recommendation module can also introduce a reinforcement learning strategy model, define the state as the current ability state, the action as the candidate resource set in the space, and the reward function based on the ability improvement potential and the learning load for comprehensive design, to improve the dynamic adaptability of the recommended path.

[0168] Finally, the system converts the scoring optimal recommended path into a structured resource recommendation strategy output form, which includes a recommended resource sequence, a recommended priority, an ability gain estimation, and a path prompt for supplementing, for guiding the learner to select resources for subsequent learning tasks.

[0169] In summary, the path matching and recommendation module constructs a resource recommendation path system for ability improvement based on the matching relationship between the ability vector and the structured resource label, in combination with the ability transfer path structure, to realize a complete closed loop from ability construction to learning resource pushing. The module maintains high coordination with the core modules such as ability reasoning and graph fusion in structure, and provides support for the intelligent adaptability and learning effect of the overall system.

[0170] The strategy optimization module is configured to perform reinforcement learning optimization on the resource recommendation strategy based on learner feedback.

[0171] The strategy optimization module continuously optimizes the recommended path and strategy parameters based on the interactive feedback information of the learner in the resource use process. The module closely connects the foregoing resource recommendation process, builds a feedback closed loop mechanism based on the ability modeling and resource label matching results, and realizes reinforcement learning optimization and dynamic updating of the resource strategy.

[0172] ​In this embodiment, the policy optimization module includes a feedback modeling module, a state representation module, and a policy learning module, which work together in the reinforcement iterative updating process of the resource recommendation path.

[0173] Generally, the feedback modeling module is used to collect the behavior feedback information generated by the learner after receiving the recommended resources. The feedback can include click rate, completion rate, learning duration, accuracy, learner self-evaluation, and other multi-dimensional data. The system standardizes these behavior data to form a unified feedback representation vector , which is the basic signal input for the subsequent optimization process.

[0174] In one possible implementation, the system sets a weighting parameter for different types of feedback signals , and constructs a feedback score function :

[0175] ;

[0176] wherein, represents the total feedback score of the learner for a recommended resource; represents the quantized value of the th feedback dimension; is the weight coefficient of the feedback dimension, satisfying ; represents the total number of feedback dimensions, such as click behavior, completion, learning duration, accuracy, and subjective score

[0177] Specifically, the system can adjust the feedback weight configuration according to task requirements, for example, emphasizing accuracy in knowledge mastery tasks and focusing on completion duration and task efficiency in skill transfer tasks.

[0178] The state representation module is used to construct a state space representation for reinforcement learning modeling by combining feedback information and current ability state. Generally, the system concatenates the ability state vector , the resource label vector , and the feedback representation vector into the current state vector :

[0179] ;

[0180] wherein, is the current learner ability vector , the label representation of the current resource , and the feedback representation vector; is the dimension of the state vector.

[0181] As an option, the system can further introduce time encoding or historical state features as extended state information to enhance the modeling capability of the policy on long-term learning trajectories.

[0182] The policy learning module is configured to optimize and update the recommendation policy using reinforcement learning algorithms on the basis of the state representation. In some embodiments, the system trains the policy network using policy gradient-based reinforcement learning algorithms (e.g., REINFORCE, PPO, or DDPG) to maximize the long-term learning rewards of the learner. In one possible implementation, the policy objective function is as follows:

[0183] ;

[0184] wherein, represents the current policy parameters , the expected cumulative reward (objective function value) to be obtained represents the mathematical expectation under the sampling distribution of the policy ; represents a parameterized policy function, whose input is the current state , and whose output is the probability distribution of the resource recommendation action; is the learnable parameter vector of the policy function ; represents the interaction round (or time step); represents the maximum length of the recommended path or the termination time of the policy execution; represents the immediate feedback score generated by the learner according to the recommended result at the th time step; is a discount factor, which is used to control the degree of influence of future rewards on the current policy update, and is usually set as a fixed constant.

[0185] The system updates the policy parameters by optimizing so that the recommended result can obtain a higher cumulative feedback score in future interactions.

[0186] In some embodiments, the system can use an Actor-Critic framework, in which the policy network acts as an actor to select the recommended action, and the value network acts as a critic to evaluate the performance of the current policy and guide the adjustment direction of the policy network.

[0187] In addition, the policy optimization module can introduce a mechanism combining offline learning and online fine-tuning, train an initial policy through historical data at the initial stage, and dynamically fine-tune according to real-time feedback after deployment, to achieve a balance between performance and efficiency.

[0188] In summary, the policy optimization module continuously optimizes the resource recommendation strategy through the three-stage process of feedback modeling, state construction and reinforcement learning, forms a learner behavior feedback driven closed-loop enhancement system, and effectively supports the personalized learning service capability of the system in a dynamic teaching environment.

[0189] Please refer to the attached Figure 2 The application also provides an AI workbench application capability construction and resource matching method for education, comprising the following steps:

[0190] S1, acquiring multi-modal behavior data of learners and converting the data into a behavior graph;

[0191] S2, constructing a learner ability graph based on the behavior graph and identifying structural holes in the ability graph through topological compression analysis;

[0192] S3, performing multi-granularity latent variable modeling on the ability graph, inferring three ability dimensions of explicit ability, intermediate ability and potential transfer ability, and generating an ability vector based on the ability dimensions;

[0193] S4, structurally aligning and fusing the ability graphs of multiple learners to construct a unified cross-domain ability transfer graph;

[0194] S5, structurally processing candidate educational resources to generate resource tags, and generating a resource recommendation path based on the matching relationship between the ability vector and the resource tags, and outputting a resource recommendation strategy;

[0195] S6, based on the learner feedback information, performing reinforcement learning optimization on the resource recommendation strategy.

[0196] For S1, the multi-modal behavior data of learners includes but is not limited to click records, answer conditions, learning duration, interaction frequency and behavior order, etc. Through preprocessing and feature extraction, the original behavior data is represented as a structured behavior graph to depict behavior characteristics such as learning path and learning preference.

[0197] For S2, the behavior graph is used to construct the corresponding ability graph of each learner, which is used to represent the mastery degree and ability association structure of the learner in different knowledge units or skill modules. Through topological compression and graph analysis method, the structural holes in the graph are further identified, so as to find the ability missing area or weak link in the learning path.

[0198] For S3, multi-granularity latent variable modeling is performed on the constructed ability graph, and the ability dimensions of explicit ability (already mastered), intermediate ability (forming) and potential transfer ability (not yet manifested but has the possibility of transformation) are inferred from the graph structure and historical behavior. Based on the ability division, a unified ability vector representation is generated for downstream matching and recommendation tasks.

[0199] For S4, the ability maps of multiple learners are structurally aligned, the mapping relationship between heterogeneous ability nodes is integrated through graph fusion method, and then a unified cross-domain ability transfer graph is constructed to realize the ability transfer path discovery and shared representation between different fields or different learning stages.

[0200] For S5, the candidate educational resources in the system are structurally processed, the target ability labels, knowledge point associations, difficulty levels, interaction modes and other meta-information are extracted, and the standardized resource labels are formed. Then, based on the matching relationship between the ability vector and the resource label, a personalized resource recommendation path is generated, and an adaptive resource recommendation strategy is output to guide the learners to improve their ability coverage and level.

[0201] For S6, based on the behavior feedback of the learners in the resource use process, such as click preference, completion, learning efficiency, etc., the resource recommendation strategy is continuously optimized. The system introduces a reinforcement learning mechanism to dynamically adjust the recommendation order and strategy parameters according to the historical recommendation performance and feedback effect, realizing the closed-loop optimization and individualization enhancement of the recommendation system.

[0202] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An education-oriented AI workbench application capability construction and resource matching system, characterized in that, The method comprises the following steps: a behavior collection and graph construction module is used to acquire multi-modal behavior data of a learner and convert the multi-modal behavior data into a behavior graph; an ability modeling module is used to construct an ability graph of the learner based on the behavior graph and identify structural holes in the ability graph through topological compression analysis; an ability reasoning module is used to perform multi-granularity latent variable modeling on the ability graph, infer three ability dimensions of explicit ability, intermediate ability and potential transfer ability, and generate an ability vector based on the ability dimensions; a cross-domain fusion module is used to perform structural alignment and fusion on the ability graphs of multiple learners to form a unified cross-domain ability transfer graph; a path matching and recommendation module is used to generate a structured resource label for a candidate educational resource, generate a resource recommendation path according to a matching relationship between the ability vector and the resource label, and output a resource recommendation strategy; a strategy optimization module is used to perform reinforcement learning optimization on the resource recommendation strategy based on learner feedback. The behavior collection and graph construction module comprises: a data collection module is used to receive multi-modal behavior data from a learner, wherein the multi-modal behavior data at least comprises action data, voice data, physiological data and environmental data of the learner; a data processing module is used to perform preprocessing, cleaning and feature extraction on the multi-modal behavior data; a behavior graph construction module is used to convert the preprocessed multi-modal behavior data into a behavior graph, wherein the behavior graph comprises behavior nodes of the learner and relationship edges between behaviors; The ability modeling module comprises: an ability graph construction module is used to convert the behavior graph into an ability graph of the learner, which maps the behavior nodes in the behavior graph into corresponding ability labels based on a predefined mapping function from behavior nodes to ability labels, and the ability graph comprises nodes representing the ability of the learner and edges representing the relationship between the abilities; a hole region identification module is used to identify a hole region in the ability graph and identify a potential ability gap through a topological compression algorithm, wherein the identification of the hole region in the ability graph through the topological compression algorithm comprises: analyzing a sparse adjacency matrix of the ability graph through the topological compression algorithm to identify a sub-region with low connectivity or no connectivity in topology as a structural hole; The ability reasoning module comprises: a latent variable modeling module is used to perform multi-granularity latent variable modeling on the ability graph to infer three ability dimensions of explicit ability, intermediate ability and potential transfer ability; a vector generation module is used to generate an ability vector based on the ability dimensions to represent the ability level of the learner in different ability dimensions; The step of performing multi-granularity latent variable modeling on the ability graph to infer three ability dimensions of explicit ability, intermediate ability and potential transfer ability comprises: For each node v in the capability graph i Initialize the feature vector where d denotes the dimension of the capability feature. iterative updating of node representation through a graph neural network model using the following formula: wherein, denotes the feature vector of node v i in the l-th layer, is the set of neighboring nodes of node v i , α ij is the attention weight between nodes, W is a trainable weight matrix, σ is an activation function, l denotes the layer index in the graph neural network, h i is the original feature vector of node v i , denotes the feature vector of node v j in the l-th layer; extracting node final representations at the output layer of the graph neural network and obtaining representations of the three capability dimensions of explicit capability, middle-layer capability, and potential migration capability through multi-granularity decoder inference. 2.The education-oriented AI workbench application capability construction and resource matching system according to claim 1, characterized in that, The cross-domain fusion module comprises: a heterogeneous feature alignment module is used to perform structural alignment on the ability graphs of multiple learners; an association relationship modeling module is used to fuse the multiple ability graphs after structural alignment to form a unified cross-domain ability transfer graph. 3.The education-oriented AI workbench application capability construction and resource matching system according to claim 2, characterized in that, The step of fusing the plurality of capability graphs after structure alignment to form a unified cross-domain capability transfer graph comprises: The characteristic vector h i of each node v i is normalized using the following formula: where h i is the original feature vector of node v i , h i ′ is the normalized feature vector, and ||h i || is the Euclidean norm of the feature vector of node v i . The node-level weighted fusion of the plurality of normalized capability graphs is performed using the following formula: wherein, is the feature vector of the fused node, is the feature vector of node v i in the kth graph, w k is the weighting coefficient of the kth graph, and m is the number of fused graphs. A unified cross-domain capability transfer graph is constructed according to the fused node feature vectors. 4.The education-oriented AI workbench application capability construction and resource matching system according to claim 1, wherein, The path matching and recommendation module comprises: A resource label generation module is configured to perform structural processing on the candidate educational resources to generate resource labels containing knowledge points, skill dimensions and use scenarios; A matching relationship construction module is configured to calculate the matching degree between the learner capability vector and the resource label; A path recommendation module is configured to construct a resource recommendation path based on the matching degree and generate a corresponding resource recommendation strategy. 5.The education-oriented AI workbench application capability construction and resource matching system according to claim 1, wherein, The strategy optimization module comprises: A feedback modeling module is configured to collect and represent the interactive feedback information of the learner on the recommended resources; A state representation module is configured to construct a state space for strategy optimization based on the current capability state and the resource use state; A strategy learning module is configured to dynamically optimize the resource recommendation strategy based on the state space using a reinforcement learning method to generate an updated resource recommendation strategy output.

6. An education-oriented AI workbench application capability construction and resource matching method applied to an education-oriented AI workbench application capability construction and resource matching system according to any one of claims 1-5, characterized in that, The method comprises the following steps: Obtain multi-modal behavior data of the learner and convert the data into a behavior graph; Construct a learner capability graph based on the behavior graph and identify structural voids in the capability graph through topological compression analysis; Perform multi-granularity latent variable modeling on the capability graph to infer three capability dimensions of explicit capability, intermediate capability and potential transfer capability, and generate a capability vector based on the capability dimensions; Align and fuse the capability graphs of multiple learners to construct a unified cross-domain capability transfer graph; Perform structural processing on the candidate educational resources to generate resource labels, and generate a resource recommendation path based on the matching relationship between the capability vector and the resource label, and output a resource recommendation strategy; Optimize the resource recommendation strategy based on the feedback information of the learner.

Images