Learning task recommendation method and device based on content consumption sequence and skill gap identification, and medium

Abstract

The application is suitable for the field of artificial intelligence technology, and provides a learning task recommendation method and device based on content consumption sequence and skill gap identification and a medium. The method comprises the following steps: collecting heterogeneous learning behavior events of a target user in multiple business platforms, sorting the heterogeneous learning behavior events according to time sequence, and constructing a content consumption sequence; uniformly encoding and fusing features of each heterogeneous behavior event in the content consumption sequence to obtain a time sequence feature vector sequence; inputting the time sequence feature vector sequence into a skill gap sub-item prediction model for processing, and outputting a skill gap sub-item vector; calling fixed rule configuration data, performing hierarchical aggregation operation on the skill gap sub-item vector, and obtaining a skill gap aggregation vector and a comprehensive skill gap evaluation index; calling a preset directed acyclic graph, and generating a sequenced learning task sequence by using a deterministic topological sorting algorithm, thereby completing intelligent learning task recommendation, and improving the accuracy and scene adaptation flexibility of learning task recommendation.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a learning task recommendation method, device and medium based on content consumption sequence and skill gap identification. Background Technology

[0002] With the continuous advancement of digitalization in vocational education, various learning and training platforms are accumulating massive amounts of students' temporal learning behavior data. Leveraging artificial intelligence technology to conduct student ability assessments and learning task recommendations has become a mainstream research direction in the field of educational big data. Currently, various intelligent recommendation and learning analysis solutions generally adopt an end-to-end integrated modeling approach, integrating behavioral sequence feature extraction, ability level determination, and task ranking constraint logic into the neural network model for joint training. The model then relies on its own parameters to fit the inherent correlations in the data, directly outputting the final assessment results and recommended content.

[0003] However, the existing integrated modeling approach has the following technical drawbacks: First, business rules such as teaching and training standards, competency division dimensions, and course prerequisites are all implicit in the model training parameters, making it impossible to configure independently and manage versions. When the teaching and training program or job competency requirements are adjusted, samples need to be collected again and the entire model needs to be retrained, resulting in high iteration costs and poor flexibility in scenario adaptation. Second, the integrated model needs to undertake two types of tasks simultaneously: temporal feature learning and business rule judgment. The training objectives are complex, which significantly increases the difficulty of model training and makes it easy to have insufficient convergence stability. It is difficult to simultaneously take into account the accuracy of the underlying behavioral feature mining and the accuracy of the upper-level learning assessment, ultimately leading to low accuracy in learning task recommendations.

[0004] Therefore, improving the accuracy of learning task recommendations and the flexibility of scenario adaptation has become an urgent problem to be solved. Summary of the Invention

[0005] To address at least one of the aforementioned technical problems and improve the accuracy and flexibility of learning task recommendations in terms of scenario adaptation, this application proposes a learning task recommendation method, device, and medium based on content consumption sequence and skill gap identification.

[0006] Firstly, this application provides a learning task recommendation method based on content consumption sequence and skills gap identification, the method comprising: S1: Collect heterogeneous learning behavior events generated by target users across multiple business platforms, sort them chronologically, and construct a content consumption sequence. S2, perform unified feature encoding and feature fusion on each heterogeneous behavioral event in the content consumption sequence to obtain a temporal feature vector sequence; S3, input the time-series feature vector sequence into the skills gap item prediction model for processing, and output the skills gap item vector; S4, call the fixed rule configuration data, perform hierarchical aggregation operation on the skill gap sub-vectors, and obtain the skill gap aggregation vector and the comprehensive skill gap assessment index; S5, invoke a preset directed acyclic graph, combine the skill gap sub-vectors, the skill gap aggregation vectors, and the comprehensive skill gap assessment index, and use a deterministic topological sorting algorithm to generate a sequential learning task sequence for the target user, thereby completing the intelligent learning task recommendation.

[0007] In one possible implementation, step S2 includes: S21, the heterogeneous learning behavior events in the content consumption sequence are divided into three types of behavior: effective learning behavior, inefficient browsing behavior and invalid operation behavior, and different confidence weights are configured for different behavior types, with the confidence weights corresponding to effective learning behavior, inefficient browsing behavior and invalid operation behavior decreasing in sequence. S22, split the numerical features and discrete identifier features contained in each of the heterogeneous learning behavior events, normalize the numerical features, and retrieve the preset vector embedding table to perform vector embedding mapping for the discrete identifier features. S23, the various feature vectors after corresponding processing are concatenated and fused to obtain a unified feature vector corresponding to a single-step behavioral event; S24, the corresponding unified feature vector is weighted and corrected based on the confidence weight of each heterogeneous learning behavior event matching, and then integrated sequentially based on the arrangement order of the content consumption sequence to obtain the time-series feature vector sequence.

[0008] In one possible implementation, the skills gap prediction model includes a bidirectional long short-term memory network layer, a multi-head self-attention layer, a feedforward network layer, an attention pooling layer, and a network output layer connected in sequence. Accordingly, the processing procedure of the skills gap sub-item prediction model in step S3 includes: S31, the bidirectional long short-term memory network layer performs local temporal dependency feature extraction on the temporal feature vector sequence and outputs a temporal correlation feature sequence; S32, the multi-head self-attention layer mines cross-temporal behavioral correlation features within the temporal correlation feature sequence, adaptively adjusts the feature weight allocation, and outputs a global correlation feature sequence; S33, the feedforward network layer performs nonlinear transformation on the global correlation feature sequence, removes redundant features and strengthens effective features, and outputs an optimized feature sequence; S34, the attention pooling layer performs feature compression and dimension aggregation on the optimized feature sequence and outputs a global behavior representation vector; S35, the network output layer performs dimension adaptation mapping on the global behavior representation vector, and finally outputs a skill gap component vector with fixed dimensions.

[0009] In one possible implementation, the fixed rule configuration data includes a fixed aggregation matrix and fixed positive weight coefficients; Accordingly, step S4 includes: S41, call the fixed aggregation matrix, perform matrix operations and dimension-wise monotonic mapping on the skill gap component vectors, and transform them to obtain the skill gap aggregate vector; S42, invoke the fixed positive weight coefficient to perform a weighted summation operation on the data of each dimension in the skill gap aggregation vector; S43, determine the comprehensive skills gap assessment index based on the weighted summation result; wherein, the fixed rule configuration data are all pre-configured offline and remain unchanged throughout the process, and do not change with the training iteration process of the skills gap sub-item prediction model.

[0010] In one possible implementation, step S5 includes: S51, Load the preset directed acyclic graph and the association matrix between each learning task node and the skill dimension in the preset directed acyclic graph. The preset directed acyclic graph is used to represent the pre-dependency relationship between different learning tasks and to constrain the basic execution order of learning tasks. S52, Based on the correlation matrix, calculate the correlation strength between each learning task node and the skill gap component vector; S53, based on the skill gap aggregation vector, determine the overall skill deficiency level of each skill classification unit, and combine it with the comprehensive skill gap assessment index to obtain the global skill deficiency level of the target user; S54. Based on the execution order constraints of the preset directed acyclic graph, and combined with the correlation strength, the overall skill deficiency degree of each skill classification unit and the global skill deficiency level, a deterministic topological sorting algorithm is used to prioritize the learning task nodes and generate the sequential learning task sequence.

[0011] In one possible implementation, step S51 includes: Load the preset directed acyclic graph; While maintaining the topological order constraint structure of the preset directed acyclic graph unchanged throughout the process, the internal weight values ​​of each association matrix are adjusted according to the target user's phased learning completion status and skill mastery feedback information, in order to continuously optimize the matching association effect between learning task nodes and various skill dimensions.

[0012] In one possible implementation, step S54 includes: S541, Based on the learning task execution order limited by the preset directed acyclic graph, lock the basic execution order of each learning task so that it cannot be changed. S542, without changing the basic execution order, increase the priority of the corresponding learning tasks based on the correlation strength; S543, based on the overall skill deficiency of each skill category, prioritize the arrangement of learning tasks corresponding to skill category units with higher deficiency. S544, combining the global skill deficiency level corresponding to the target user, adaptively adjust the density and difficulty of the learning tasks to generate the final sequential learning task sequence.

[0013] In one possible implementation, the training process of the skills gap component prediction model includes: Construct a joint loss function that includes a skill gap component error loss term, a hierarchical aggregation consistency auxiliary loss term, and a network parameter regularization loss term; Based on the labeled sample dataset, the joint loss function is used to iteratively train the skill gap prediction model. During the training process, the fixed rule configuration data and the structural parameters of the preset directed acyclic graph remain unchanged.

[0014] In a second aspect, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in the first aspect or any of the implementations thereof.

[0015] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in the first aspect or any of the implementations thereof.

[0016] Fourthly, this application provides a computer program product comprising a computer program that, when executed by a processor, implements the steps of the method as described in the first aspect or any of the implementations thereof.

[0017] Compared with existing technologies, the advantages of this application are as follows: It collects heterogeneous learning behavior events of target users to construct content consumption sequences, which are then fused through feature encoding and input into a skill gap prediction model for processing to obtain skill gap vectors. These vectors are then processed using offline fixed-rule calculations to obtain a skill gap aggregation vector and a comprehensive skill gap assessment index. Finally, a deterministic topological sorting algorithm is used to generate a learning task recommendation sequence based on a pre-defined directed acyclic graph. By externalizing and solidifying teaching-related business rules and decoupling them from the neural network model, this approach overcomes the shortcomings of traditional end-to-end integrated modeling methods, such as rule embedding, the need for retraining for scenario adjustments, high training difficulty, and insufficient recommendation accuracy. This reduces the cost of solution iteration, improves the flexibility of scenario adaptation, and can stably mine behavioral temporal features, effectively improving the accuracy of learning task recommendations.

[0018] It is understood that the electronic devices, computer-readable storage media, and computer program products provided in this application have the same beneficial effects as the learning task recommendation method based on content consumption sequence and skill gap identification described above, and will not be repeated here. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a learning task recommendation method based on content consumption sequence and skill gap identification, provided as an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a skills gap sub-item prediction model provided in an embodiment of this application. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0022] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0023] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0025] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] This application provides a learning task recommendation method based on content consumption sequence and skills gap identification, preferably applied to the learning progress recommendation scenario for higher vocational college students, but not limited to higher vocational students, and adaptable to various training participants. For example, in the learning progress recommendation scenario for higher vocational college students, it assesses the continuous behavioral data generated by higher vocational college students throughout their entire lifecycle of "course learning - skills training - internship - job preparation," evaluating the ability structure shortcomings reflected by a single user's content consumption event sequence within a set time window. The output consists of an M-dimensional skills gap component vector, an L-dimensional skills gap aggregation vector, a comprehensive skills gap assessment index S, and a sequential learning task sequence P on a pre-defined directed acyclic graph. P corresponds to executable tasks such as course remediation, remedial training tasks, job skills enhancement, and job preparation. Business specificity is reflected in: continuous cross-stage trajectory of the same account, clear sequential dependencies in the resource system, direct output results for scheduling adjustments, and consumption by the training task distribution and employment guidance subsystems. Stage tags can be optional metadata; if missing, inference is still based solely on the event sequence and resource identifiers.

[0028] For ease of understanding, the technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0029] Figure 1 This is a flowchart illustrating a learning task recommendation method based on content consumption sequence and skill gap identification, provided in one embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The method provided in this embodiment includes the following steps: S1 collects heterogeneous learning behavior events generated by target users across multiple business platforms, sorts them chronologically, and constructs a content consumption sequence.

[0030] In practice, target users generate various heterogeneous behavioral data on different learning platforms during their daily learning process, such as browsing courses, watching training videos, participating in exercises and answering questions, and submitting practical assignments. By uniformly collecting behavioral data from multiple platforms, the user's complete learning trajectory can be fully restored, effectively avoiding the problem of one-sided learning analysis caused by data collection from a single platform.

[0031] In one possible implementation, step S1 may include: S11 retrieves raw log data from multiple business platforms for target users.

[0032] Optionally, raw log data of target users can be uniformly captured from various business platforms such as course learning platforms, skills training platforms, and job internship guidance platforms. The captured raw log data includes multi-dimensional behavioral information such as user operation timestamps, learning behavior types, learning duration, task completion status, and learning content identifiers, realizing the comprehensive collection of diversified learning behavior data.

[0033] S12 performs data cleaning, outlier removal, and missing value completion on the original log data to filter out heterogeneous learning behavior events.

[0034] Optionally, the collected raw log data can be cleaned, outlier removed, and missing value filled in sequentially. Interference data with no actual learning value can be removed, such as invalid operation data that has timed out, duplicate and redundant operation data, and data with incorrect format. For missing data, the common mean of the same scenario can be used to fill in the missing data, and heterogeneous learning behavior events that can truly represent the user's learning status can be selected.

[0035] S13. First, sort the heterogeneous learning behavior events in ascending order based on the timestamps corresponding to each event. When multiple learning behaviors occur simultaneously under the same timestamp, sort the multiple heterogeneous learning behavior events under the same timestamp in a fixed order based on the preset secondary business sorting field to obtain the content consumption sequence corresponding to the target user.

[0036] Optionally, the learning behavior events are first sorted in ascending order based on their timestamps to achieve a monotonous temporal arrangement. For multiple learning behavior events generated under the same timestamp, they are then arranged in a fixed order according to preset secondary business sorting fields to ensure a consistent and stable event arrangement order at any given moment. Finally, the content consumption sequence corresponding to the target user is obtained. For example, business-specific fields such as behavior type code, learning content level, platform type number, and task priority number can be used as preset secondary business sorting fields, and the priority of each preset secondary business sorting field can be pre-set.

[0037] In addition, when the sequence length exceeds the set range, the end retention or sliding window method can be used to uniformly regulate the sequence length, and it is necessary to ensure that the sequence processing rules in the subsequent model training stage and the inference stage are consistent.

[0038] S2 performs unified feature encoding and feature fusion on each heterogeneous behavioral event in the content consumption sequence to obtain a temporal feature vector sequence.

[0039] In one possible implementation, step S2 may include: S21, the heterogeneous learning behavior events in the content consumption sequence are divided into three types of behavior: effective learning behavior, inefficient browsing behavior, and invalid operation behavior. Differentiated confidence weights are assigned to different behavior types, with the confidence weights corresponding to effective learning behavior, inefficient browsing behavior, and invalid operation behavior decreasing sequentially.

[0040] Optionally, three types of behaviors are pre-defined: effective learning behaviors, inefficient browsing behaviors, and ineffective operational behaviors. Effective learning behaviors may include in-depth learning behaviors such as completing a full course, practical training, and answering formal questions; inefficient browsing behaviors may include shallow learning behaviors such as quickly flipping through courses and briefly previewing content; and ineffective operational behaviors may include meaningless operations such as accidental clicks and rapid page jumps. Further, different confidence weights are assigned to different types of learning behaviors: high weights for effective learning behaviors, medium weights for inefficient browsing behaviors, and low weights for ineffective operational behaviors. Each behavioral event within the content consumption sequence is matched with its corresponding confidence weight, and the behavioral features are weighted and corrected to weaken the data interference caused by ineffective behaviors, strengthen the proportion of data from truly effective learning behaviors, accurately reflect the user's true learning state, and then proceed with subsequent feature encoding and fusion processing after weighted correction.

[0041] It should be noted that typical examples of effective learning behavior, inefficient browsing behavior, and invalid operation behavior are complete course learning, quick page browsing, and accidental invalid clicks, respectively. The boundaries of the three categories are completely consistent with the classification rules defined in claim 2.

[0042] S22, split the numerical features and discrete label features contained in each heterogeneous learning behavior event, normalize the numerical features, and retrieve the preset vector embedding table to perform vector embedding mapping for the discrete label features.

[0043] Optionally, numerical features include quantitative data such as learning duration, answer scores, and number of practice sessions, which are mapped to a fixed numerical range using linear normalization; discrete identifier features include non-quantitative textual data such as course type, training category, and learning stage, which are digitized using vector embedding.

[0044] As an example, the original processing materials for each heterogeneous learning behavior event only contain structured data and text fields, without introducing multimedia file entities such as images, audio, and video. A single learning behavior event can be decomposed into a combination of multiple feature vectors. Among them, numerical features are preferably truncated by quantiles or logarithmic transformation before linear normalization to reduce the impact of extreme values. Text content is first segmented into words or sub-words to obtain standardized discrete identifiers, and then mapped into low-dimensional feature vectors according to a preset vector embedding table. When performing batch data processing, placeholders are padded at the end of sequences that have not reached the maximum time sequence length of the batch. Preset dedicated placeholders are used to complete the filling operation. At the same time, the number of rows and columns of the preset vector embedding table is reasonably set, and dedicated mapping positions for placeholders are reserved to meet the batch operation format requirements. If there are missing text features, zero vectors or dedicated missing embedding vectors are used to fill them in uniformly to ensure the integrity and uniformity of feature dimensions.

[0045] For example, to facilitate reproducible experiments, single-step events are further specified as three-dimensional numerical features (stay, completion, click depth, etc.), discrete text features, and discrete identifier features; during batch processing, the time dimension that is less than the maximum length in the batch is padded at the end, and the padded position uses the placeholder identifier agreed with the vocabulary. The number of embedded table rows is the resource base plus one and the vocabulary size plus one to accommodate the placeholder rows.

[0046] It should be noted that the feature fields and vector dimensions can be flexibly adjusted during actual deployment, as long as the overall vectorization mapping logic remains consistent.

[0047] S23, the various feature vectors after corresponding processing are concatenated and fused to obtain a unified feature vector corresponding to a single-step behavioral event.

[0048] Optionally, the normalized numerical feature vectors and the various discrete feature vectors that have completed the embedding mapping are dimensionally concatenated and fused to form a unified feature vector that can fully represent the information of single-step learning behavior.

[0049] For example, the various feature vectors after corresponding processing are linearly fused and mapped to... The dimension is used to obtain a unified feature vector, as shown in equation (1), where the semicolon indicates vector concatenation; if discrete embedding is involved, the embedding is performed first and then the concatenation is performed. and The input dimension is consistent with the splicing dimension.

[0050] (1); in, This is the weight matrix. For bias vectors, For text features, For numerical features, For identification features.

[0051] S24. The corresponding unified feature vector is weighted and corrected based on the confidence weight of each heterogeneous learning behavior event matching, and then integrated sequentially based on the arrangement order of the content consumption sequence to obtain the time-series feature vector sequence.

[0052] Optionally, based on the pre-matched confidence weights of each learning behavior event, the corresponding single-step unified feature vector is weighted and corrected to weaken the proportion of low-value behavior features and strengthen effective learning behavior features. Then, strictly following the predetermined order of content consumption sequence, all corrected single-step feature vectors are sequentially and orderly integrated to finally construct a complete temporal feature vector sequence.

[0053] In addition, a reasonable upper limit should be set for the overall length of the sequence. For ultra-long time series sequences, the end should be retained or a sliding window should be used for uniform truncation. It should also be strictly ensured that the sequence truncation rules used in the model training phase and the inference phase are completely consistent to ensure that the data processing standards are consistent before and after.

[0054] The technical solution provided in this step achieves feature differentiation weighting by dividing the learning behavior into weights. Combined with numerical feature optimization preprocessing and discrete feature standardization embedding mapping, it completes unified feature encoding and feature fusion. This not only effectively removes invalid data interference and fills in missing feature information, but also accurately highlights the features of effective learning behavior, greatly improving the data representation capability of time-series feature sequences.

[0055] S3 inputs the time-series feature vector sequence into the skills gap sub-item prediction model for processing and outputs skills gap sub-item vectors.

[0056] Specifically, the skills gap prediction model in this embodiment only undertakes the work of learning behavior feature extraction and temporal pattern mining, without embedding any teaching rules or skills evaluation standards. This achieves the separation of feature learning and business evaluation rules, reducing the model modification cost caused by subsequent adjustments to teaching rules.

[0057] In one possible implementation, such as Figure 2 As shown, the skills gap prediction model consists of a bidirectional long short-term memory network layer, a multi-head self-attention layer, a feedforward network layer, an attention pooling layer, and a network output layer connected in sequence. Correspondingly, the processing steps for the skills gap sub-item prediction model in step S3 may optionally include: S31, the bidirectional long short-term memory network layer performs local temporal dependency feature extraction on the temporal feature vector sequence and outputs a temporal correlation feature sequence.

[0058] Optionally, the bidirectional long short-term memory network layer performs local temporal dependency feature extraction on the temporal feature vector sequence, fully explores the correlation between adjacent learning behaviors, captures the user's short-term learning habits and continuous learning state, and outputs a temporal correlation feature sequence.

[0059] S32 uses a multi-head self-attention layer to mine cross-temporal behavioral correlation features within temporal correlation feature sequences, and adaptively adjusts feature weight allocation to output a global correlation feature sequence.

[0060] Optionally, a multi-head self-attention layer can be used to mine long-distance behavioral relationships within temporal correlation feature sequences, accurately correlate learning behavior data with long time intervals, and adaptively adjust the weight ratio of different features according to the importance of the behavior to highlight the feature role of key learning behaviors and output a globally correlated feature sequence.

[0061] S33, the feedforward network layer performs nonlinear transformation on the globally correlated feature sequence, removes redundant features and strengthens effective features, and outputs the optimized feature sequence.

[0062] Optionally, the feedforward network layer incorporates a Gaussian error linear unit activation function. This activation function enables nonlinear transformation of features, deep purification and feature optimization of the globally correlated feature sequence, elimination of redundant feature information, enhancement of the core learning information feature expression capability, and output of the optimized feature sequence.

[0063] S34, the attention pooling layer performs feature compression and dimension aggregation on the optimized feature sequence and outputs a global behavior representation vector.

[0064] Optionally, the attention pooling layer performs feature compression and aggregation on the optimized feature sequence, simplifies redundant sequence data, and integrates time series of varying lengths into a unified overall representation data of fixed dimensions, outputting a global behavior representation vector.

[0065] S35, the network output layer performs dimension adaptation mapping on the global behavior representation vector, and finally outputs a skill gap component vector with fixed dimensions.

[0066] Optionally, the network output layer performs dimensional mapping on the global behavior representation vector, adjusts the vector dimension according to the preset skill subdivision dimensions, and finally outputs a fixed-dimensional skill gap sub-vector that can accurately represent the user's weakness in various subdivision skills.

[0067] As an example, the integrated temporal feature sequence is input into the skill gap prediction model. The bidirectional long short-term memory network layer completes the local temporal dependency capture and outputs the spliced ​​hidden state. After layer normalization, it is then fed into the multi-head self-attention layer to complete long-range correlation modeling. The attention operation result is input into a two-layer linear feedforward network layer equipped with Gaussian error linear units. At the same time, a residual connection structure is added to the feedforward operation path. Finally, the output is a context feature sequence that can be used for subsequent pooling operations. In this embodiment, no additional independent residual branch is added between the attention output end and the hidden state of the long short-term memory network.

[0068] Furthermore, the attention pooling layer compresses the context feature sequence using temporal attention pooling to generate a fixed-dimensional global behavior representation vector. During pooling, the original scores corresponding to ineffective padding time steps within the sequence are assigned extremely small values, and after normalization activation, the attention weights corresponding to the padding positions approach zero. Simultaneously, the self-attention sub-layer configures padding masks matching the effective temporal step size for the key sequence, completely blocking invalid padding positions from participating in the attention weight distribution calculation. To ensure network operation compliance and stability, the total dimension of the hidden states obtained by bidirectional concatenation is limited to be divisible by the number of attention heads. A random deactivation mechanism can also be introduced into the attention operation branch and the feedforward network branch to optimize the model's generalization ability. After the model enters the online inference stage, the random deactivation-related configuration is turned off. The network output layer only sets a single-layer linear transformation structure with a Sigmoid activation function to construct the second output branch, without adding any extra output paths.

[0069] For example, a sequence of time-series feature vectors The input is a bidirectional long short-term memory network layer to capture local temporal dependencies, as shown in Equation (2); after the hidden states are concatenated, they are first normalized by layers, then passed through a multi-head self-attention layer, and the attention output is then passed through a two-layer linear feedforward network layer containing Gaussian error linear units, and the feedforward path contains residual connections to the feedforward input, to obtain the context sequence used for subsequent pooling, as shown in the attention calculation form summarized in Equation (3): (2); (3); Attention pooling layers perform temporal attention pooling on the context sequence to obtain a global behavior representation vector. See equations (4) and (5).

[0070] (4); (5); in, Let be the context feature vector at time step t; Here is the weight matrix of the attention pooling layer; This is a learnable attention query vector. Let be the attention weight coefficient at time step t.

[0071] The network output layer contains only one linear transformation and a sigmoid function, as shown in equation (6): (6); in, This is the weight matrix of the network output layer; σ is the bias vector of the network output layer; σ is the sigmoid activation function. This is the skill gap component vector output by the network.

[0072] In the above structure, the bidirectional long short-term memory network layer processes the sequential dependencies and local patterns of the temporal feature vector sequence; the multi-head self-attention layer reweights long-range dependencies and key events; and the attention pooling layer compresses variable-length sequences into a fixed-dimensional global representation, corresponding to the skill gap regression objectives. The network output layer maps resource identifiers and textual information to a space that can be jointly optimized with numerical features, corresponding to the fragmented and multi-source event input characteristics.

[0073] As an example, the training process of a skills gap component prediction model may optionally include: Construct a joint loss function that includes a skill gap component error loss term, a hierarchical aggregation consistency auxiliary loss term, and a network parameter regularization loss term; Based on the labeled sample dataset, the joint loss function is used to iteratively train the skill gap prediction model. During the training process, the fixed rule configuration data and the structural parameters of the preset directed acyclic graph remain unchanged.

[0074] Optionally, the skill gap sub-item error loss term is used to reduce the deviation between the model output vector and the actual learning situation labeled data; the hierarchical aggregation consistency auxiliary loss term is used to ensure that the subsequent vector aggregation operation results are logically consistent with the subdivided skill gap data; and the network parameter regularization loss term is used to suppress model overfitting. Based on the labeled sample dataset, the joint loss function is used to iteratively train the skill gap sub-item prediction model. Throughout the entire process of overall model iterative optimization, the offline fixed rule configuration data and the structural parameters of the preset directed acyclic graph (i.e., the directed acyclic graph of learning resource dependencies) remain constant. Only the model's own weight parameters are adjusted, which avoids the business rules from shifting with model training and ensures the long-term uniformity and stability of skill evaluation standards.

[0075] As an example, the training objective simultaneously constrains the error of the skill gap component vector and the error of the skill gap component vector through the same fixed aggregation matrix on mini-batch samples. The predicted skills gap aggregation vector is used to assist the consistency error and includes parameter L2 regularization, as shown in equations (7) to (9). Fixed weights for each dimension of the skills gap component vector; , For hyperparameters; Standard values ​​of the skills gap component vector The reference value of the aggregated vector of skill gaps is obtained by the same rule as in Equation (10).

[0076] set up This is the set of sample indices for the mini-batch samples during training iterations. , Samples The corresponding skill gap component vectors and standard values, , The first term obtained by applying the corresponding skill gap component vector from equation (10) The aggregated vector of skill gaps and the reference value are then used to calculate the skill gap component error loss term. Auxiliary loss term with hierarchical aggregation consistency The calculation formulas are shown in equations (7) and (8) respectively, and equation (9) is the expression for the joint loss function: (7); (8); in, Depend on Obtained through the same calculation path as equation (10); Standard values ​​of the skills gap component vector The result is obtained from equation (10). In batch training, equations (7) and (8) can be implemented by calling the mean square error on the mini-batch sample tensor. When they are equal, the result is the same as the above formula.

[0077] (9); in, This is the summation of the squares of the learnable parameters of the network, i.e., the network parameter regularization loss term. The standard values ​​of the skill gap component vectors can be obtained through expert annotation, rule generation weakly aligned with academic results, or simulation annotation; the missing dimension can be masked using equation (7). Additionally, if... Equal; the optimizer can be AdamW, and the parameter norm penalty is only applied through equation (9). Including the total loss, weight decay is not added to the optimizer side to avoid [affecting] the total loss. Item repeated.

[0078] As an example, in scenarios where the number of effective samples of newly registered users and learning behaviors is small, the skill gap sub-item prediction model can still complete the inference calculation and output the corresponding skill gap sub-item vectors normally. At the same time, the external interface pushes low confidence identification markers and preset business rollback strategy markers, and can flexibly choose to output only skill gap sub-item vectors or not generate a complete learning layout path for the time being.

[0079] As an example, during batch training of the model, the prediction error of skill gap sub-items is simultaneously constrained within the dimension of small batch samples. At the same time, hierarchical vector consistency constraints are completed according to the established mapping rules, and overall parameter optimization is completed with parameter regularization constraints. The standard value data used for training can be obtained through various methods such as manual annotation by industry experts, generation of rules linked to academic assessment results, and simulation annotation. For samples with missing data dimensions, a loss mask can be used to shield the error calculation corresponding to the missing dimensions. The model optimizer preferably adopts AdamW, which completes the numerical constraints of network parameters only through regularization loss terms, and does not repeatedly set weight decay mechanisms at the optimizer end to avoid the problem of double parameter penalty.

[0080] The technical solution provided in this step is based on a skills gap item prediction model to extract multi-scale learning behavior features, explore local temporal patterns and long-term behavioral correlations, avoid interference from invalid padding data on feature modeling by using padding masks and minimum value assignment, and train the model using a joint loss function that includes a skills gap item error loss term, a hierarchical aggregation consistency auxiliary loss term, and a network parameter regularization loss term. This effectively improves the accuracy of skills gap item identification and the consistency of hierarchical evaluation, while also being suitable for cold start scenarios with sparse samples and being able to stably output standardized skills gap item vectors.

[0081] S4 calls the fixed rule configuration data to perform hierarchical aggregation operation on the skill gap sub-vectors to obtain the skill gap aggregate vector and the comprehensive skill gap assessment index.

[0082] Specifically, all the operation rules, operation parameters, and rule configuration data in this application are pre-stored offline according to the teaching evaluation standards uniformly formulated in the industry. There is no need to rely on model training and learning, and rule adjustments only require modification of the offline configuration.

[0083] In one possible implementation, the fixed rule configuration data includes a fixed aggregation matrix and fixed positive weight coefficients; correspondingly, step S4 may optionally include: S41, call the fixed aggregation matrix, perform matrix operations and dimension-wise monotonic mapping on the skill gap component vectors, and transform them to obtain the skill gap aggregate vector.

[0084] Optionally, the fixed aggregation matrix maintains constant parameters throughout the model training and online inference application phases, without participating in iterative updates. The internal data normalization process is also fixed through offline configuration. During the aggregation process, the skill gap component vectors output by the network are first subjected to a dimension-wise monotonic mapping process to unify the values ​​within a standardized range. In practical applications, a dimension-wise Sigmoid function or an extremum calibration method (min-max calibration) can be used to achieve the range constraint. After the mapping is completed, the linear dimension aggregation operation is performed in conjunction with the fixed aggregation matrix. In addition, row normalization constraints can be applied to the aggregation matrix during the offline parameter calibration phase to ensure that each row element is non-negative and the sum of the row elements is 1. After calibration, the matrix can be directly archived for use. Alternatively, column normalization or other constraint forms can be selected according to actual business needs.

[0085] As an example, a fixed aggregation matrix No updates are made during training and inference; the normalization method is fixed by the configuration. Skill gap aggregation vector. Skill gap component vector The result obtained by monotonically normalizing and linearly aggregating is shown in equation (10): (10); in, For a dimension-wise monotonic mapping, each component is constrained to... For example, dimensional sigmoid or min-max calibration. Additionally, and Matrix multiplication can be combined with [other methods] in engineering implementation. (Dimensional compatibility under batch processing) is equivalent and does not change the meaning of equation (10).

[0086] S42 calls a fixed positive weight coefficient to perform a weighted summation operation on the data of each dimension within the skills gap aggregation vector.

[0087] Optionally, the fixed positive weighting coefficient can be determined by combining offline business statistical analysis with constraint optimization algorithm. After the assignment is completed, it is archived and managed in a unified manner in a versioned manner. The larger the weight value, the higher the proportion of the impact of the corresponding dimension's skill deficiency on the overall evaluation result. The final summation value can intuitively reflect the severity of the overall skill gap.

[0088] S43. The comprehensive skills gap assessment index is determined based on the weighted summation result. Among them, the fixed rule configuration data are all pre-configured offline and remain unchanged throughout the process, without changing with the training iteration process of the skills gap sub-item prediction model.

[0089] Optionally, the result of the weighted summation and fusion of dimensions can be determined as the final comprehensive skills gap assessment index. All fixed rule configuration data used in the overall hierarchical aggregation process are pre-configured offline and parameters are solidified. The parameters remain stable throughout the iterative training of the skills gap sub-item prediction model and the entire process of daily business reasoning, and will not change in value or be adjusted in sync with the network learnable parameters.

[0090] As an example, The calculation formula is shown in equation (11): (11); in, The positive weight coefficients are fixed and determined jointly by offline statistics and constraint optimization, and are saved in a versioned manner. The larger the value, the more significant the overall gap.

[0091] For example, the skills gap aggregation vector total Dimension represents the aggregated state quantity at the business module level, and its value is mapped to a fixed sequence of rules. Interval. The total number of skill gap component vectors. Dimension, representing the fine-grained sub-items of the state, is consistent with the dimension directly output by the skills gap sub-item prediction model, and The comprehensive skills gap assessment index S is a scalar, obtained by linearly weighting each dimension of the skills gap aggregation vector using fixed positive weight coefficients, with a weight sum of 1. Optional, based on... Discrete level numbers are obtained by comparing with a preset threshold sequence, which are used for graded display; the threshold and grade name are fixed by the configuration and do not participate in network training.

[0092] In this embodiment, the fixed rule configuration data and calculation rules are not manually specified temporarily during the online inference phase, but rather are versioned configurations that have been calibrated and solidified offline based on historical sample statistics, co-occurrence relationships, and sequence constraints. During deployment, only the corresponding version configuration is loaded to perform deterministic calculations.

[0093] For example, Table 1 shows an example structure of the meaning of the aggregated vector dimension of the skills gap and the index of the sub-vectors of the skills gap. This example is... Can be For divisibility cases, the specific semantics are bound by the business mapping configuration. Table 2 shows the meanings of the skill gap symbols: Table 1. Meaning of the aggregated vector dimension of the skills gap and the index of the sub-vectors representing the skills gap:

[0094] For example, when and At that time, the above division is equivalent to four consecutive skill gap sub-items in each module. Cannot be When dividing by integers, the above table's equal blocks are not used. Instead, the business mapping matrix separately defines the set of skill gap item indices covered by each skill gap aggregation vector dimension. In actual deployment, the business mapping matrix takes precedence; the index range in the table is only an illustrative example.

[0095] Table 2, Meaning of Skills Gap Symbols:

[0096] The technical solution provided in this step relies on offline fixed parameters that are static and do not update throughout the process to complete hierarchical aggregation operations. The operation rules are unified and the results are stable and reproducible. Numerical standardization and control are achieved through dimension-wise monotonic mapping. Multi-dimensional gap information fusion is completed by combining dedicated fixed weights. At the same time, the rule parameters and network learning parameters are strictly decoupled, which effectively avoids the evaluation rule deviation caused by the model training process and greatly improves the objectivity, uniformity and business adaptability of the quantitative assessment of skill gaps.

[0097] S5 calls a preset directed acyclic graph, combines the skill gap component vectors, skill gap aggregation vectors, and comprehensive skill gap assessment index, and uses a deterministic topological sorting algorithm to generate a sequential learning task sequence for the target user, thus completing the intelligent learning task recommendation.

[0098] In one possible implementation, step S5 may include: S51, load the preset directed acyclic graph and the association matrix between each learning task node and skill dimension in the preset directed acyclic graph. The preset directed acyclic graph is used to represent the pre-dependency relationship between different learning tasks and to constrain the basic execution order of learning tasks.

[0099] As an example, step S51 includes: Load a pre-defined directed acyclic graph; While maintaining the topological constraints of the pre-defined directed acyclic graph throughout the process, the internal weight values ​​of each association matrix are adjusted based on the target user's phased learning progress and skill mastery feedback, in accordance with the pre-defined method of unified updates after the end of each phase, in order to continuously optimize the matching and association effects between learning task nodes and various skill dimensions.

[0100] For example, a pre-defined directed acyclic graph is loaded. This graph is built based on the professional talent training program, which clarifies the irreversible learning order among all learning courses and practical training tasks. At the same time, the association matrix between each learning task node and skill dimension in the pre-defined directed acyclic graph is loaded to establish the corresponding relationship between learning tasks and subdivided skills and ability modules.

[0101] Preferably, after loading the overall structure of the directed acyclic graph, the topological order constraint structure of the graph is kept locked throughout the process without any structural changes; based on the target user's phased learning completion status and actual skill mastery feedback data, and following the preset operation mode of unified update at the end of each phase, the internal weight values ​​of the association matrix are fine-tuned. Under the premise of not disrupting the established learning order, the matching association accuracy between learning task nodes and various skill dimensions is continuously iterated and optimized, so that task matching fits the user's dynamic learning progress.

[0102] S52, based on the correlation matrix, calculates the correlation strength between each learning task node and the skill gap component vector.

[0103] Optionally, a quantitative calculation relationship can be established based on the loaded association matrix to accurately measure the association strength between each learning task node and the skill gap sub-vector. This association strength can be calculated according to a predetermined formula. The association matrix between task nodes and skill dimensions can be jointly constructed by combining the historical learning behavior transfer probability, the co-occurrence frequency of learning resources, and the inherent dependency constraints of industry business. After the construction is completed, it can be frozen and fixed after offline compliance verification and will not be changed arbitrarily.

[0104] As an example, Nodes in a directed graph The correlation strength with the component vector of the skills gap is shown in equation (12). The correlation matrix between learning task nodes and skill gap sub-items along the vector dimension: (12); in, Learning task nodes The correlation strength with the current skills gap component vector. F represents the preset value. E represents the set of all learning task nodes in the graph, with each node corresponding to a specific learning task; E represents the directed dependency relationship between nodes in the predefined directed acyclic graph, that is, the directed edge from one node to another.

[0105] S53 determines the overall skill deficiency level of each skill category unit based on the skill gap aggregation vector, and obtains the global skill deficiency level of the target user by combining the comprehensive skill gap assessment index.

[0106] As an example, step S53 may optionally include: Based on the preset skill classification rules, the values ​​of each dimension in the skill gap aggregation vector are assigned to the corresponding skill classification unit; The overall skill deficiency level of the corresponding skill category unit is quantified by integrating and calculating all the dimension gap values ​​within the same skill category unit. Multiple skill value judgment ranges are predefined, and the comprehensive skill gap assessment index is matched to the corresponding judgment range to obtain the global skill deficiency level of the target user.

[0107] S54. Based on the execution order constraints of the pre-defined directed acyclic graph, and combined with the correlation strength, the overall skill deficiency degree of each skill category unit and the global skill deficiency level, a deterministic topological sorting algorithm is used to prioritize the learning task nodes and generate a sequential learning task sequence.

[0108] As an example, step S54 may optionally include: S541, based on the execution order of learning tasks constrained by a pre-defined directed acyclic graph, locks the unchangeable basic execution order of each learning task; Optionally, the learning sequence defined by the directed acyclic graph can be used as the basic framework for the overall arrangement, locking the fixed execution order of each learning task before and after it that cannot be arbitrarily changed, thus obtaining the basic execution order.

[0109] S542 prioritizes the arrangement of corresponding learning tasks based on the strength of association without changing the basic execution order.

[0110] Optionally, without changing the established basic execution order, based on the correlation strength of each learning task, the priority of subdivided learning tasks with higher correlation can be reasonably increased, and learning content that fits the user's weak skills can be pushed first.

[0111] As an example, under the premise of satisfying the directed acyclic graph structure specification, the calculated association strength is used as the core sorting criterion. The set of candidate task nodes with an in-degree of zero is selected and sorted in descending order of association strength. Then, the sequence is reasonably truncated by combining the learning path length and the learning stage prefix rule. The entire sorting process strictly follows the deterministic operation logic and does not participate in the backpropagation parameter update of the model.

[0112] S543, based on the overall skill deficiency of each skill category, prioritize the arrangement of learning tasks corresponding to skill category units with higher deficiency levels.

[0113] Optionally, based on the overall skill deficiency level corresponding to different skill categories, priority can be given to arranging various learning tasks corresponding to skill categories with higher deficiency levels, so as to specifically fill in the user's weak skill areas.

[0114] S544, based on the target user's corresponding global skill deficiency level, adaptively adjusts the density and difficulty of learning tasks to generate the final sequential learning task sequence.

[0115] Optionally, based on the user's actual global skill deficiency level, the overall arrangement density and rhythm of learning tasks and the difficulty gradient of individual tasks can be adaptively and flexibly adjusted to balance the rationality of the learning pace and the adaptability of the learning content, ultimately generating a standardized sequential learning task sequence that fits the user's actual learning situation.

[0116] In addition, during actual deployment and operation, if abnormal configuration problems such as loop structure or fragmented node regions are detected in the preset directed acyclic graph, and a valid topology learning sequence cannot be generated, the system will automatically refuse to output the learning path and pop up an error message. The preset directed acyclic graph used in actual deployment is preferably set to a single weakly connected overall structure. If multiple independent learning and training systems are involved, the graph can be split and partitioned in advance according to the connected components to ensure stable task sorting.

[0117] In one embodiment, the technical solution provided in this application specifically includes the following process: Collect time-ordered content consumption events of a single user, and form a content consumption sequence after preprocessing. The features at each step are encoded and input into the skills gap component prediction model to obtain the skills gap component vector. ; Calculate the skills gap aggregation vector according to formulas (10) and (11). Comprehensive Skills Gap Assessment Index According to equation (12) and the graph algorithm, in a pre-defined directed acyclic graph... Generate a sequence of sequential learning tasks. ;Will , , and Return to the business system or interface. The network's learnable parameters do not include a fixed aggregation matrix. Fixed positive weighting coefficients Preset directed acyclic graph Correlation matrix The semantic content is learned by simply mapping content consumption sequences to skill gap component vectors. Each of these steps can be implemented by executing program instructions from storage media using a general-purpose or special-purpose processor.

[0118] From a business data closed-loop perspective: the platform aggregates learning logs, training logs, internship records, and job preparation behaviors by account and forms feature sequences; during the offline training phase, network parameters are updated and the configuration version is periodically reassessed; during the online inference phase, only the frozen model and configuration output are loaded. , , , The execution side then issues supplementary learning resources, supplementary training tasks, and employment preparation tasks based on this information, and returns the completion results to the log again for the next round of offline calibration. This closed loop can directly serve the integrated business process of academic affairs, practical training, and employment guidance in higher vocational colleges.

[0119] The structured data returned by the external interface may include, but is not limited to: , , , It includes an optional global skill deficiency level and an optional low confidence flag; the transmission format can be JSON, HTTP API, or message queue.

[0120] Specifically, Table 3 lists the main formula numbers and brief descriptions involved in this application: Table 3. Main Formula Numbers and Brief Descriptions:

[0121] The technical solution provided in this application collects heterogeneous learning behavior events of target users to construct a content consumption sequence. After feature encoding and fusion, the sequence is input into a skill gap sub-item prediction model for processing to obtain a skill gap sub-item vector. Then, an offline fixed rule operation is used to obtain a skill gap aggregation vector and a comprehensive skill gap assessment index. Finally, a deterministic topological sorting algorithm is used to generate a learning task recommendation sequence based on a preset directed acyclic graph. By externalizing and solidifying teaching-related business rules and decoupling them from the neural network model, this approach overcomes the shortcomings of traditional end-to-end integrated modeling methods, such as rule embedding, the need for retraining for scenario adjustments, high training difficulty, and insufficient recommendation accuracy. This reduces the cost of solution iteration, improves the flexibility of scenario adaptation, and can stably mine behavioral temporal features, effectively improving the accuracy of learning task recommendations.

[0122] Corresponding to the above method steps, the apparatus provided in this application may include a memory for storing program instructions and configuration data (including...). , , , The system includes a processor for executing program instructions to perform: event sequence acquisition and preprocessing, deep network inference according to equations (1) to (6), continuation calculation according to equations (10) and (11), path generation according to equation (12) and graph algorithm, and writing the results to memory or outputting them via a communication interface. Each functional unit is logically consistent with the aforementioned modules and can run on a single machine, distributed node, or cloud-side service, without constituting a limitation on the hardware form.

[0123] The learning task recommendation method based on content consumption sequence and skill gap identification provided in this application can be divided into the following logical modules (without limiting specific software packages or process boundaries): Data Acquisition and Access Module: Reads raw events from logs or event tracking interfaces of learning platforms, training platforms, and employment guidance platforms, and sorts them by user and time; Preprocessing and serialization module: performs timestamp alignment, outlier handling, and sequence truncation or window splitting, outputting the content to be consumed in the sequence. ; Feature encoding module: maps each event in the content consumption sequence to the vector required by equation (1); Sequence deep network inference module: Execute equations (2) to (6), and output... ; Indicator continuation calculation module: After the network forward pass is completed, execute equations (10) and (11) according to the continuation fixed rules to obtain the following results. and ; Graph path generation module: Loads data bound to the vocational college training objectives and job competency maps. , Execute equation (12) and the topology algorithm to output ; Task orchestration and distribution module: The task list is mapped to supplementary learning, supplementary training, and employment preparation tasks and pushed to the corresponding business subsystems. Configuration and Version Management Module: Maintenance , , , , , The version identifiers for the model weights ensure that the same configuration set is used for training, inference, and auditing.

[0124] Furthermore, to verify the technical solution proposed in this application, the following experiment was conducted, with the following experimental setup: a total of 840 user sequence samples, of which 600 were used for training, 120 for validation, and 120 for testing; the dimension of the skills gap component vector... Skill gap aggregation vector dimension The length of a single sequence is random between 16 and 128 steps. Single-step input: numerical feature dimension is 3; resource identifier discrete cardinality is 48; word segmentation vocabulary size is 128; embedding table rows are 49 and 129 respectively (including a placeholder row); bidirectional LSTM unidirectional hidden layer dimension is 64 (concatenation dimension...). The number of multi-head self-attention heads is 4, and the feature width after single-step fusion is... Resource identifier embedding dimension and word segmentation embedding dimension are each set to 24; attention and feedforward network path dropout are set to 0.1 (enabled during training, disabled during inference). Random seed is set to 42. The optimizer uses AdamW, with a learning rate of [missing value]. No weight decay is set on the optimizer side; the coefficient of the auxiliary term in equation (9) Take 0.35, L2 term coefficient Pick The training run consisted of 25 epochs with a batch size of 32. During training, in each epoch, the mean squared error of the skill gap component vector was calculated on the validation set per batch, and the average was taken over the batch. If the average was lower than the historical best, the current network weights were preserved. The optimal values ​​for the validation set listed in Table 4 represent the minimum of this average over the entire training process. In Table 5 of the test set, MSE and MAE are global averages over all test samples and all dimensions. Standard values ​​were generated by adding noise to rules related to behavioral statistics and were used to verify the completeness of the process; they do not represent the true labeled distribution.

[0125] The experimental environment included: a CPU as the computing device (one embodiment did not use a GPU; if a GPU was used, the wall clock and video memory usage were recorded separately); PyTorch 2.x or later as the deep learning framework; Python 3 as the programming language; and the operating system based on the machine used on the day of the experiment. The precise framework minor version number, operating system, and processor model should be recorded separately for comparison with the wall clock time.

[0126] Under the same sample partitioning, random seed, and the aforementioned hyperparameters, the training and inference processes should yield losses and wall clock times of the same order of magnitude. When changing hardware or the main version of the framework, floating-point differences may cause slight fluctuations in values, which is normal. In addition to error metrics, it is recommended to simultaneously record the path topology validity rate, configuration version hit rate, and recalculation consistency rate with the same input as engineering acceptance indicators.

[0127] For example, Table 4 shows the actual test results of a complete training and validation (the computing device is a CPU, and the specific machine model and framework version are based on the records on the day of the experiment), and Table 5 shows the actual test results of accuracy and inference time on the test set (120 test samples, batch size 32, computing device is a CPU).

[0128] Table 4. Experimental results of a complete training and validation cycle:

[0129] Table 5 shows the actual test results of accuracy and inference time on the test set:

[0130] Connection rule link verification: Take any sample from the test set and output the network output. Substituting equation (12) with a fixed directed acyclic graph, and using nodes with an in-degree of zero, we can obtain the following: Kahn topological sort variant for keys (with maximum) After obtaining a topologically valid total order (with priority dequeuing), the first 8 nodes are extracted, with an example sequence of (0,2,1,3,4,5,6,7); the comprehensive skills gap assessment index is calculated according to formulas (10) and (11). It is 0.621109. The above sequence number is... This varies depending on the sample and weight; this example only illustrates the connectivity of the link.

[0131] The absolute values ​​of accuracy metrics are affected by the construction rules of standard values ​​and noise levels. For formal application, they must be recalibrated using actual annotations or a business-approved gold standard. Wall clock times vary depending on hardware, drivers, parallelism, and batch size; experimental environments and hardware / software versions should be recorded separately for verification.

[0132] The pre-production validation process is as follows: offline playback is performed on the anonymized real logs, and gray-scale evaluation is carried out by time slice; during the gray-scale period, only the suggested sequence is output and the manual confirmation entry is retained, and the path execution rate, rollback rate and delay index are recorded; after the evaluation is passed, it is switched to automatic execution mode.

[0133] On the other hand, this application also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned learning task recommendation methods based on content consumption sequence and skill gap identification.

[0134] On the other hand, this application also provides an electronic device, including a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute any of the above-mentioned learning task recommendation methods based on content consumption sequences and skill gap identification.

[0135] For example, the program code may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in an electronic device.

[0136] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the electronic device may also include input / output devices, network access devices, buses, etc.

[0137] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0138] The memory can be an internal storage unit of the electronic device, such as a hard drive or RAM. It can also be an external storage device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units. The memory is used to store the program code and other programs and data required by the electronic device. The memory can also be used to temporarily store data that has been output or will be output.

[0139] The computer storage medium and electronic device described above are created based on the above method. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A learning task recommendation method based on content consumption sequence and skills gap identification, characterized in that, The method includes: S1: Collect heterogeneous learning behavior events generated by target users across multiple business platforms, sort them chronologically, and construct a content consumption sequence. S2, perform unified feature encoding and feature fusion on each heterogeneous behavioral event in the content consumption sequence to obtain a temporal feature vector sequence; S3, input the time-series feature vector sequence into the skills gap item prediction model for processing, and output the skills gap item vector; S4, call the fixed rule configuration data, perform hierarchical aggregation operation on the skill gap sub-vectors, and obtain the skill gap aggregation vector and the comprehensive skill gap assessment index; S5, invoke a preset directed acyclic graph, combine the skill gap sub-vectors, the skill gap aggregation vectors, and the comprehensive skill gap assessment index, and use a deterministic topological sorting algorithm to generate a sequential learning task sequence for the target user, thereby completing the intelligent learning task recommendation.

2. The method according to claim 1, characterized in that, Step S2 includes: S21, the heterogeneous learning behavior events in the content consumption sequence are divided into three types of behavior: effective learning behavior, inefficient browsing behavior and invalid operation behavior, and different confidence weights are configured for different behavior types, with the confidence weights corresponding to effective learning behavior, inefficient browsing behavior and invalid operation behavior decreasing in sequence. S22, split the numerical features and discrete identifier features contained in each of the heterogeneous learning behavior events, normalize the numerical features, and retrieve the preset vector embedding table to perform vector embedding mapping for the discrete identifier features. S23, the various feature vectors after corresponding processing are concatenated and fused to obtain a unified feature vector corresponding to a single-step behavioral event; S24, the corresponding unified feature vector is weighted and corrected based on the confidence weight of each heterogeneous learning behavior event matching, and then integrated sequentially based on the arrangement order of the content consumption sequence to obtain the time-series feature vector sequence.

3. The method according to claim 1, characterized in that, The skills gap prediction model comprises a bidirectional long short-term memory network layer, a multi-head self-attention layer, a feedforward network layer, an attention pooling layer, and a network output layer connected in sequence. Accordingly, the processing procedure of the skills gap sub-item prediction model in step S3 includes: S31, the bidirectional long short-term memory network layer performs local temporal dependency feature extraction on the temporal feature vector sequence and outputs a temporal correlation feature sequence; S32, the multi-head self-attention layer mines cross-temporal behavioral correlation features within the temporal correlation feature sequence, adaptively adjusts the feature weight allocation, and outputs a global correlation feature sequence; S33, the feedforward network layer performs nonlinear transformation on the global correlation feature sequence, removes redundant features and strengthens effective features, and outputs an optimized feature sequence; S34, the attention pooling layer performs feature compression and dimension aggregation on the optimized feature sequence and outputs a global behavior representation vector; S35, the network output layer performs dimension adaptation mapping on the global behavior representation vector, and finally outputs a skill gap component vector with fixed dimensions.

4. The method according to claim 1, characterized in that, The fixed rule configuration data includes a fixed aggregation matrix and fixed positive weight coefficients; Accordingly, step S4 includes: S41, call the fixed aggregation matrix, perform matrix operations and dimension-wise monotonic mapping on the skill gap component vectors, and transform them to obtain the skill gap aggregate vector; S42, invoke the fixed positive weight coefficient to perform a weighted summation operation on the data of each dimension in the skill gap aggregation vector; S43, determine the comprehensive skills gap assessment index based on the weighted summation result; wherein, the fixed rule configuration data are all pre-configured offline and remain unchanged throughout the process, and do not change with the training iteration process of the skills gap sub-item prediction model.

5. The method according to claim 1, characterized in that, Step S5 includes: S51, Load the preset directed acyclic graph and the association matrix between each learning task node and the skill dimension in the preset directed acyclic graph. The preset directed acyclic graph is used to represent the pre-dependency relationship between different learning tasks and to constrain the basic execution order of learning tasks. S52, Based on the correlation matrix, calculate the correlation strength between each learning task node and the skill gap component vector; S53, based on the skill gap aggregation vector, determine the overall skill deficiency level of each skill classification unit, and combine it with the comprehensive skill gap assessment index to obtain the global skill deficiency level of the target user; S54. Based on the execution order constraints of the preset directed acyclic graph, and combined with the correlation strength, the overall skill deficiency degree of each skill classification unit and the global skill deficiency level, a deterministic topological sorting algorithm is used to prioritize the learning task nodes and generate the sequential learning task sequence.

6. The method according to claim 5, characterized in that, Step S51 includes: Load the preset directed acyclic graph; While maintaining the topological order constraint structure of the preset directed acyclic graph unchanged throughout the process, the internal weight values ​​of each association matrix are adjusted according to the target user's phased learning completion status and skill mastery feedback information, in order to continuously optimize the matching association effect between learning task nodes and various skill dimensions.

7. The method according to claim 5, characterized in that, Step S54 includes: S541, Based on the learning task execution order limited by the preset directed acyclic graph, lock the basic execution order of each learning task so that it cannot be changed. S542, without changing the basic execution order, increase the priority of the corresponding learning tasks based on the correlation strength; S543, based on the overall skill deficiency of each skill category, prioritize the arrangement of learning tasks corresponding to skill category units with higher deficiency. S544, combining the global skill deficiency level corresponding to the target user, adaptively adjust the density and difficulty of the learning tasks to generate the final sequential learning task sequence.

8. The method according to any one of claims 1 to 7, characterized in that, The training process of the skills gap prediction model includes: Construct a joint loss function that includes a skill gap component error loss term, a hierarchical aggregation consistency auxiliary loss term, and a network parameter regularization loss term; Based on the labeled sample dataset, the joint loss function is used to iteratively train the skill gap prediction model. During the training process, the fixed rule configuration data and the structural parameters of the preset directed acyclic graph remain unchanged.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images