A personalized learning pushing method and system of a financial training platform

Abstract

The application provides a personalized learning pushing method and system of a financial training platform, including the following contents: obtaining a multi-dimensional cognitive state vector of a learner, inputting the multi-dimensional cognitive state vector into a deep reinforcement learning network, and deciding a target knowledge point set; constructing a pushing queue group according to a knowledge graph correlation relationship; quantifying a current cognitive load grade by collecting interactive behavior data in real time; adjusting a scheduling priority in combination with a cognitive consumption coefficient of each queue; selecting a non-empty queue head pushing unit according to the priority, and pushing the non-empty queue head pushing unit to a terminal; obtaining a current mastery value of each knowledge point based on learning feedback, determining a pushing interval according to a difference between the current mastery value and a preset target value, and reloading the knowledge point to a queue head for pushing when the interval ends; and when the difference of all knowledge points is less than a preset value, updating the cognitive state vector and re-planning a next stage learning task. The application realizes dynamic matching of pushing content and a learner cognitive state, and improves learning efficiency and experience.

Description

Technical Field

[0001] This invention relates to the field of Internet technology, specifically to a personalized learning push method and system for a financial training platform. Background Technology

[0002] Financial training platforms offer professional teaching services to practitioners, covering areas such as market fundamentals, investment analysis, risk management, and compliance. The knowledge system in these areas has a distinct hierarchical structure and a rigorous logical connection. Learners must follow specific prerequisite paths to gradually delve deeper. For example, understanding derivative pricing requires mastering the characteristics of basic financial instruments, while building an investment portfolio requires a solid foundation in asset pricing models.

[0003] However, in real-world applications, learners often face the dual challenges of the complexity of knowledge and individual differences. On the one hand, the vast network of financial knowledge and the close interrelationships of concepts can easily lead to excessive cognitive load for learners, making it difficult for them to systematically build a knowledge framework. On the other hand, different learners have significant differences in their basic level, learning preferences, cognitive abilities, and time investment, making it difficult for standardized and uniform teaching content to meet personalized learning needs.

[0004] The current push notification method suffers from significant gaps. It focuses solely on the relevance of knowledge points, disregarding whether learners are currently fatigued. Complex cases are often forcibly pushed to learners when their attention is wandering, ignoring their real-time comprehension abilities. There is a lack of effective tracking of learners' mastery of knowledge points, making it impossible to distinguish between knowledge on the verge of forgetting and knowledge that has been firmly grasped. This results in familiar content being repeatedly pushed, wasting learning time, weakening learners' interest in participation, and creating a continuous discrepancy between the content and learners' actual learning status. It is impossible to accurately determine when to introduce new knowledge and when to consolidate old knowledge. The speed of knowledge progression does not match the individual's cognitive pace, reducing learning efficiency and negatively impacting learners' confidence.

[0005] Therefore, existing technologies struggle to achieve an effective balance between the continuity of knowledge logic, the learner's immediate cognitive state, and the personalized learning pace. They cannot be flexibly adjusted according to the learner's actual mastery, resulting in a disconnect between the speed of learning and the individual's cognitive pace. This not only reduces learning efficiency but also easily leads to learners feeling fatigued and frustrated. Summary of the Invention

[0006] Therefore, the purpose of this invention is to provide a personalized learning push method and system for a financial training platform, which effectively solves the problem that existing push strategies are out of touch with learners' actual cognitive state.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a personalized learning recommendation method for a financial training platform, comprising, A learner's multidimensional cognitive state vector is obtained, and the multidimensional cognitive state vector is input into a deep reinforcement learning network. The deep reinforcement learning network then decides to output a target knowledge point set from the candidate learning object set based on a multi-objective reward function. Based on the target knowledge point set, push queue groups are constructed according to the relationships between knowledge points in the knowledge graph; Real-time acquisition of learners' interactive behavior data streams, quantification of the current cognitive load level; based on the current cognitive load level and combined with the cognitive consumption coefficient pre-configured for each queue, adjustment of the scheduling priority of each queue; According to the current scheduling priority from high to low, the head push unit in the non-empty queue is selected as the candidate push unit; and the candidate push unit is pushed to the learner terminal. Based on learning feedback data, the current mastery value of each knowledge point in the push queue group is obtained; the push interval is determined according to the difference between the current mastery value and the preset mastery target value; when the push interval timer ends, the knowledge point is loaded to the first position of the push queue and pushed to the learner's terminal. When the difference between each knowledge point and the preset mastery target value is less than the first preset value, the evaluation learning feedback data updates the multidimensional cognitive state vector, and based on the updated data, the learning task for the next stage is replanned.

[0008] The personalized learning push method for the financial training platform according to claim 1 is characterized in that: the multi-dimensional cognitive state vector is constructed using a neural network architecture that combines a multi-head self-attention mechanism with gated recurrent units; specifically including: The learner's basic attributes and learning preferences are encoded into a first feature vector; The knowledge concepts in the knowledge graph are mapped to graph structure features, and then encoded into a second feature vector through graph embedding. After concatenating and fusing the first feature vector and the second feature vector, the implicit interaction relationship between different knowledge dimensions is first captured by a multi-head self-attention layer, and then input into a gated recurrent unit layer to capture the change pattern of cognitive state over time, and output the multi-dimensional cognitive state vector.

[0009] Furthermore, the multidimensional cognitive state vector includes a knowledge mastery vector, a forgetting state vector, a cognitive load estimation vector, a preference state vector, and a learning objective embedding vector; wherein: The knowledge acquisition vector is iteratively updated by a gated cyclic unit combined with the time decay factor of the Ebbinghaus forgetting curve. The forgetting state vector is used to predict the forgetting risk value of each knowledge concept through the MemGNN network. The MemGNN network uses the knowledge graph as a graph structure, aggregates the state information of adjacent knowledge nodes through the graph attention network layer, and inputs the aggregation result into the gated recurrent unit to output the forgetting probability of each knowledge concept in the next time step. The cognitive load estimation vector is obtained by weighted fusion estimation based on learners’ response time, accuracy, single learning duration and interaction frequency. The preference state vector is decoupled from the learning behavior sequence through a continuous-time heterogeneous information network to obtain dynamic preference features.

[0010] Furthermore, the learner's learning objectives are obtained, and the learning objectives are classified in a fine-grained manner to obtain a classification result set; each knowledge concept in the classification result set is mapped to a first node set in the knowledge graph, and each learning content in the learning library is mapped to a second node set in the knowledge graph; the semantic embedding vector cosine similarity between the first node and the second node is calculated, and the learning content with similarity exceeding a preset dynamic threshold is included in the candidate learning object set.

[0011] Furthermore, the push queue group includes a text and image summary queue, an interactive exercise queue, and a case video queue. The push units in each queue are associated with the same target knowledge point identifier and carry a form type label. The text and image summary queue, the interactive exercise queue, and the case video queue are pre-configured with sequentially increasing cognitive consumption coefficients.

[0012] Furthermore, adjust the scheduling priority of each queue, including: Obtain the learner's cognitive load level quantification value and the cognitive consumption coefficient of each queue at the current moment; The cognitive consumption coefficients are normalized to obtain the normalized consumption intensity value of each queue. The remaining proportion of the quantified cognitive load level relative to the maximum value is taken as the learner's current remaining cognitive capacity; The normalized consumption intensity value of each queue is compared with the current remaining cognitive capacity. The smaller the deviation between the two, the higher the load fit score of the queue. The load fit score is used to characterize the degree of matching between the cognitive consumption intensity of the queue and the learner's current cognitive burden. The scheduling priority of each queue is adjusted based on its load adaptability score.

[0013] Furthermore, the deep reinforcement learning network adopts a deep reinforcement learning model based on the Actor-Critic architecture, whose policy evaluation module includes a state encoding layer, an action masking layer, an Actor branch, and a Critic branch; wherein: The state coding layer is used to perform feature fusion coding on the multidimensional cognitive state vector at the current moment and output a compact implicit state representation. The action masking layer is used to generate an action mask based on the set of candidate learning objects, and to mask non-candidate learning objects, so that the policy evaluation module can only select actions within the current set of candidate learning objects. The Actor branch is used to output the selection probability distribution of each learning object in the candidate learning object set; the Critic branch is used to output the value estimate of the current multidimensional cognitive state.

[0014] Furthermore, the deep reinforcement learning network also includes an experience replay pool for storing a quadruple of state, action, reward, and next state for a historical time step, and updates the network parameters during the training phase through random batch sampling.

[0015] Furthermore, based on the negative correlation between the difference and the push interval, the push interval includes multiple non-zero time intervals.

[0016] A personalized learning recommendation system for a financial training platform, applied to the aforementioned personalized learning recommendation method for the financial training platform, includes: The state-aware module is used to acquire the learner's multidimensional cognitive state vector and determine the set of candidate learning objects; The decision module is used to input the multidimensional cognitive state vector into the deep reinforcement learning network and, based on the multi-objective reward function, decide and output the target knowledge point set from the candidate learning object set. The queue module is used to construct a push queue group based on the target knowledge point set and according to the knowledge graph; The first push module is used to quantify the current cognitive load level based on the real-time collected interactive behavior data flow, and adjust the scheduling priority of each queue in combination with the pre-configured cognitive consumption coefficient of each queue; according to the current scheduling priority from high to low, the first push unit of the non-empty queue is selected and pushed to the learner terminal. The second push module is used to obtain the current mastery value of each knowledge point based on learning feedback data, and to determine the push interval based on the difference between the current mastery value and the preset mastery target value. When the push interval timer ends, the push unit corresponding to the knowledge point is reloaded to the first position of the push queue. The replanning module is used to update the multidimensional cognitive state vector and trigger the state perception module to replan the learning task for the next stage when the difference between each knowledge point and the preset mastery target value is less than the first preset value.

[0017] The beneficial effects of the above technical solution are as follows: In response to the problem that existing push strategies are out of touch with learners' actual cognitive states, this invention effectively solves the contradiction between the strong logic of the knowledge system and the fluctuating cognitive rhythm of learners in financial training by using multi-dimensional cognitive state modeling and deep reinforcement learning strategy decision-making.

[0018] At the cognitive perception level, this invention utilizes a cascaded neural network architecture constructed with a multi-head self-attention mechanism and gated recurrent units to uniformly represent multi-dimensional features such as knowledge mastery, forgetting trends, cognitive load, and learning preferences. The multi-head self-attention layer captures the implicit connections between knowledge points and the interaction relationships between different cognitive dimensions. The gated recurrent unit layer takes historical cognitive state sequences as input and extracts the temporal patterns of cognitive state evolution over time, forming a cognitive profile that can map knowledge blind spots and load states in real time. This overcomes the perceptual limitations of traditional methods that rely on static labels or single indicators for learner modeling.

[0019] This invention relies on the collaborative operation of deep reinforcement learning networks and financial knowledge graphs. An action shielding layer constrains the decision space within the current learning objective, while topological sorting ensures that the pushed sequence meets the pre-requirement dependencies of knowledge points, maintaining the inherent logical integrity of the financial knowledge system. A multi-objective reward function integrates five indicators: knowledge mastery gain, learning efficiency, content adaptability, path compliance, and behavioral negative feedback. This enables the decision model to globally optimize the learning path in terms of long-term cumulative rewards, balancing immediate learning gains with the coherence of the knowledge structure, and overcoming the suboptimal path problem caused by greedy strategies.

[0020] By constructing three progressively challenging queues for pushing content—text summaries, interactive exercises, and case study videos—and dynamically adjusting scheduling priorities based on the matching degree between real-time cognitive load levels and the cognitive load coefficients of each queue, an adaptive match between the content format and the learner's current cognitive capacity is achieved. When cognitive resources are sufficient, high-consumption content is prioritized to promote deep understanding; when cognitive resources are insufficient, low-consumption content is prioritized to avoid overload, ensuring the learning process remains within an appropriate cognitive load range.

[0021] The review push interval is determined based on the difference between the current level of mastery and the preset mastery target. The larger the difference, the higher the risk of forgetting, and the shorter the interval. When the timer ends, the corresponding knowledge point is reloaded to the top of the push queue. This mechanism accurately identifies knowledge points on the verge of forgetting and provides targeted review and recall, reducing ineffective repetition of solidified knowledge and optimizing the allocation efficiency of cognitive resources.

[0022] This invention matches the knowledge transmission process with the learner's cognitive patterns, coordinating the speed of knowledge advancement with the individual's cognitive rhythm while ensuring the integrity of the knowledge structure. This shortens the conversion cycle from knowledge acquisition to skill internalization, and improves the overall effectiveness and user experience of financial professional training. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the implementation of the push method of the present invention; Figure 2 Flowchart for constructing multidimensional cognitive state vectors; Figure 3 Flowchart for filtering the set of candidate learning objects; Figure 4 Flowchart for review push implementation; Figure 5 A line graph comparing learning efficiency; Figure 6 A graph showing the change in mastery of different knowledge points after they were pushed to users. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 This embodiment aims to provide a personalized learning push method for a financial training platform. Addressing the technical challenges of significant differences in learners' knowledge base, dynamic changes in cognitive states, and the difficulty of existing push methods in balancing the continuity of knowledge logic with individual cognitive pace in financial training scenarios, this embodiment achieves effective matching of learning content push with learners' actual cognitive states by constructing a multi-dimensional cognitive state vector, making knowledge point decisions based on deep reinforcement learning, adaptively adjusting push priorities based on cognitive load, and driving review pushes based on mastery gaps.

[0025] The engineering scenario applied in this embodiment is an internal training platform of a financial institution, providing pre-job training for newly hired employees on modules such as financial market fundamentals, investment analysis, risk management, and compliance supervision. The platform has approximately 500 registered learners, and its learning content library includes teaching resources in three formats: text summaries, interactive exercises, and case study videos, totaling approximately 1200 push units covering over 350 knowledge concepts in the knowledge graph. In actual operation, the platform faces two prominent problems. First, learners have diverse professional backgrounds before joining the company. Some learners have a financial education background and are familiar with basic concepts, while others come from STEM or humanities backgrounds and lack knowledge of financial terminology and basic logic. A uniform pace strategy results in the former being limited in their progress and the latter falling behind. Second, learners utilize fragmented time outside of their daily work for training, and their attention and cognitive resources fluctuate significantly at different times. In the afternoon, learners have a higher cognitive load after continuous business processing; pushing highly complex content at this time significantly reduces learning effectiveness.

[0026] In response to the above problems, such as Figure 1 As shown, this embodiment achieves personalized learning recommendations through the following steps.

[0027] S1. Construction of Multidimensional Cognitive State Vectors like Figure 2 As shown, this embodiment employs a neural network architecture combining a multi-head self-attention mechanism and a gated recurrent unit to construct the multidimensional cognitive state vector. In specific implementation, the learner's basic attributes (including years of employment, prior education, and job category) and learning preferences (including preferred content formats and average daily study time) are first encoded as the first feature vector. The education background field in the basic attributes is encoded using one-hot encoding according to five categories: finance, economics, management, science and engineering, and others. The job category is categorized and encoded according to front-end, middle-end, and back-end. Learning preferences are statistically analyzed based on the learner's learning behavior logs over the past 30 days, extracting the percentage of clicks and dwell time for different content formats as preference features.

[0028] Simultaneously, knowledge concepts in the knowledge graph are mapped to graph structure features and encoded into a second feature vector through graph embedding. The knowledge graph is constructed with knowledge concepts as nodes and semantic relationships such as prerequisites, associations, and complements as edges. Graph embedding adopts a representation learning method based on graph convolutional networks, mapping each knowledge concept to an embedding vector. This vector encodes the semantic position of the concept in the knowledge graph and its relationship features with adjacent concepts.

[0029] After concatenating and fusing the first and second feature vectors, the result is input into a multi-head self-attention layer. This layer has eight attention heads, each independently calculating attention weights between different feature dimensions to capture implicit interactions between dimensions such as knowledge mastery, cognitive load, and learning preferences. For example, when a learner's cognitive load is high, the attention mechanism should reduce the feature response intensity for more difficult knowledge points. The output of the self-attention layer is then input into a gated recurrent unit layer. This layer takes the learner's historical cognitive state sequence as input and, through its update and reset gate mechanisms, captures the changing patterns of cognitive states over time, outputting a multi-dimensional cognitive state vector for the current moment.

[0030] The multidimensional cognitive state vector includes the knowledge mastery vector, the forgetting state vector, the cognitive load estimation vector, the preference state vector, and the learning objective embedding vector.

[0031] Among them, the knowledge mastery vector records the learner's estimated mastery of each knowledge concept, with a value range from 0 to 1, and is iteratively updated by gated cyclic units combined with the time decay factor of the Ebbinghaus forgetting curve.

[0032] The forgetting state vector is used to predict the forgetting risk value of each knowledge concept through the MemGNN network. This network uses a knowledge graph as its graph structure, aggregates the state information of adjacent knowledge nodes through a graph attention network layer, and inputs the aggregation result into a gated recurrent unit to output the forgetting probability of each knowledge concept in the next time step. In this embodiment, the MemGNN network uses two graph attention layers, each with four attention heads, and the hidden layer has a dimension of 128. The training loss function is binary classification cross-entropy loss.

[0033] The cognitive load estimation vector is obtained by weighted fusion estimation based on learners' response time, accuracy, single learning duration, and interaction frequency.

[0034] The preference state vector is derived from the learning behavior sequence by decoupling dynamic preference features based on the temporal heterogeneous graph neural network; the learning target embedding vector encodes the learner's phased learning targets.

[0035] S2. Determining the set of candidate learning objects like Figure 3As shown, after obtaining the learner's learning objectives, these objectives are categorized in a finer granular manner according to the hierarchical structure of the financial knowledge system. For example, the learning objectives of a newly hired fixed income trading assistant can be subdivided into several sub-objectives, such as understanding the basic elements and pricing principles of bonds, mastering the business process of repurchase transactions, and becoming familiar with the main trading instruments in the bond market. The knowledge concepts in the categorized result set are mapped to the first set of nodes in the knowledge graph, and the learning content in the learning base is mapped to the second set of nodes in the knowledge graph.

[0036] Calculate the cosine similarity of the semantic embedding vectors between the first node and the second node, and classify learning content such as image and text summaries, interactive exercises and case videos with similarity exceeding a preset dynamic threshold into the candidate learning object set.

[0037] The preset dynamic threshold is dynamically adjusted according to the size of the current candidate set. When there are fewer than 10 candidates, the threshold is lowered to expand the candidate range, and when there are more than 50 candidates, the threshold is raised to narrow the candidate range.

[0038] S3, Knowledge Point Set for Decision-Making Objectives Based on Deep Reinforcement Learning Networks The multidimensional cognitive state vector is input into a deep reinforcement learning network, which then decides to output a set of target knowledge points from the set of candidate learning objects based on a multi-objective reward function.

[0039] In this embodiment, the deep reinforcement learning network adopts a model based on the Actor-Critic architecture, and its policy evaluation module includes a state encoding layer, an action masking layer, an Actor branch, and a Critic branch.

[0040] The state encoding layer is responsible for fusing and encoding the features of the multidimensional cognitive state vector input at the current moment, outputting a compact implicit state representation. In practice, the multidimensional cognitive state vector includes multiple sub-vectors such as knowledge mastery vector, forgetting state vector, cognitive load estimation vector, preference state vector, and learning target embedding vector, with differences in data distribution and scale among the sub-vectors. The state encoding layer uses two fully connected network layers to remap and fuse the features of these sub-vectors, compressing the high-dimensional heterogeneous input space into a unified-dimensional implicit representation for subsequent policy evaluation and decision-making.

[0041] The action masking layer generates an action mask based on the candidate learning object set, blocking non-candidate learning objects. In the push scenario of the financial training platform, the candidate learning object set consists of knowledge content within the current learning objective range; knowledge points outside this range should not be pushed. By forcibly setting the policy probability of non-candidate objects to zero, the action masking layer ensures that the policy evaluation module only selects actions within the current candidate learning object set, guaranteeing that the decision output always remains within the constraints of the learning objective.

[0042] The Actor branch outputs the selection probability distribution of each learning object in the candidate learning object set. It employs a two-layer fully connected network structure. The first layer maps the implicit state representation to intermediate features and activates them using ReLU. The second layer maps these intermediate features to logits values ​​(equivalent to the number of candidate actions), which are then normalized using the Softmax function to obtain the probability distribution. This probability distribution reflects the relative advantage of selecting each learning object in the candidate set under the current cognitive state. During training, the Actor branch updates the policy gradient based on the value evaluation signal provided by the Critic branch, gradually shifting the selection probability towards actions that yield higher cumulative rewards.

[0043] The Critic branch outputs a value estimate of the current multidimensional cognitive state. This value estimate is a prediction of the expected long-term cumulative reward in the current state, reflecting the total reward the learner can obtain by executing subsequent learning processes according to the current policy in the current cognitive state. The Critic branch shares the implicit state representation output by the state encoding layer with the Actor branch, and value mapping is performed on this basis through an independent fully connected layer.

[0044] The Actor-Critic architecture updates based on temporal difference learning. At each decision step, the system executes the action sampled by the Actor branch, receives an immediate reward, and proceeds to the next state. The Critic branch calculates the temporal difference error, which is the difference between the immediate reward plus the next state value estimate and the current state value estimate. This error reflects the deviation between the current policy and the optimal policy. The Critic branch updates the network parameters with the goal of minimizing this error, gradually bringing the value estimate closer to the true state value function. The Actor branch then uses this error as an estimate of the advantage function, adjusting the policy parameters in the direction that increases the advantage function, thus assigning a higher probability to actions that bring higher rewards.

[0045] In this embodiment, the deep reinforcement learning network also includes an experience replay pool, which stores a quadruple of state, action, reward, and next state for historical time steps. During the training phase, network parameters are updated through random batch sampling. The experience replay pool breaks the temporal correlation between data from adjacent time steps, making the training samples approximately satisfy the independent and identically distributed assumption, thereby improving the stability and convergence efficiency of parameter updates. The experience replay pool capacity is set to 10,000 records, the batch size is 64, the Adam optimizer is used for network parameter updates, the learning rate is set to 0.001, and the target network uses a soft update method with an update coefficient of 0.01.

[0046] Deep reinforcement learning networks make decisions based on a multi-objective reward function. At each decision step, the multi-objective reward function calculates the comprehensive reward value R using the following formula: The multi-objective reward function calculates the comprehensive reward value at each decision step using the following formula. :

[0047] in: To reward knowledge acquisition gains, a linear increment is assigned based on the positive change in the learner's knowledge acquisition vector after learning the optimal learning object; the larger the positive change, the higher the reward. For example, if a learner's mastery of a corresponding concept in their knowledge acquisition vector increases from 0.35 to 0.55 after completing a text-image summary of bond pricing, the positive increment is 0.20. The value is assigned proportionally as 0.20 × the base reward.

[0048] To reward learning efficiency, a reward is calculated based on the ratio of the learner's actual time spent completing a learning task to a preset standard time. The preset standard time is determined in advance based on the difficulty and format of the content. For example, the preset standard time for a medium-difficulty case video is 8 minutes. If the learner completes the task in 6 minutes with the required accuracy, the ratio is less than 1, and a positive reward is given. If the actual time is 12 minutes, the ratio is greater than 1, and a negative reward is given. The purpose of this reward system is to guide learners to improve their learning efficiency while ensuring comprehension.

[0049] For content-adaptive rewards, a negative correlation is assigned between the difficulty coefficient of the learning object and the Euclidean distance of the learner's current cognitive load estimation vector; the smaller the distance, the higher the reward. The difficulty coefficient is calibrated by experts based on the level of financial knowledge and the complexity of the content when it is added to the database, with a value ranging from 0 to 1. The load level in the cognitive load estimation vector is also normalized to the interval of 0 to 1. The smaller the Euclidean distance between the two, the better the content difficulty matches the learner's current capacity, and the higher the reward.

[0050] As a compliance reward, a fixed positive reward is given when the learner meets the prerequisite relationships defined in the knowledge graph and the current learning path plan; otherwise, no reward is given. For example, if a learner has not yet mastered the prerequisite knowledge point of "characteristics of basic financial instruments," and the decision model chooses to push content related to "derivative pricing," then... The item is set to zero, thereby suppressing the behavior of skipping levels for push notifications.

[0051] As a negative feedback penalty for behavior, when it is detected that the learner drags the progress bar for more than a preset time threshold, frequently pauses operations, or the error rate of the answer suddenly increases by more than a preset percentage threshold, a penalty value proportional to the abnormal magnitude is applied.

[0052] These are the preset weight coefficients for each item, and the values ​​of each weight coefficient can be adjusted based on platform operating data through grid search or Bayesian optimization.

[0053] This embodiment presents a personalized learning delivery framework based on multi-dimensional cognitive state perception, deep reinforcement learning decision-making, and adaptive scheduling of cognitive load. This framework comprehensively characterizes learners' knowledge mastery, forgetting trends, cognitive load, and learning preferences. It utilizes a reinforcement learning network driven by a multi-objective reward function to dynamically determine the optimal sequence of knowledge points. During the delivery process, it adaptively adjusts the scheduling priority of multi-form queues based on real-time cognitive load levels. Furthermore, it incorporates a review delivery mechanism driven by mastery gaps to ensure timely knowledge consolidation, thereby achieving effective coordination between the logical continuity of knowledge, learners' cognitive capacity, and personalized learning pace.

[0054] S4. Construction and Scheduling of Push Queue Groups Based on the target knowledge point set, push queue groups are constructed according to the relationships between knowledge points in the knowledge graph. The push queue group includes a text and image summary queue, an interactive exercise queue, and a case video queue. The push units in each queue are associated with the same target knowledge point identifier and carry a form type label.

[0055] The target knowledge point set consists of multiple knowledge points selected by the deep reinforcement learning network in the current decision round. Each knowledge point is accompanied by its node identifier and associated edge information in the knowledge graph. Because the financial knowledge system has a strict prerequisite dependency structure, there is a clear logical order among the knowledge points in the target knowledge point set. For example, the Capital Asset Pricing Model (CAPM) depends on risk and return measurement, which in turn depends on probability theory. If the target knowledge point set is directly pushed out sequentially according to the decision order, it is possible that higher-level content will be pushed out before the prerequisite content has been learned.

[0056] To address the aforementioned issues, this embodiment first performs topological sorting on the knowledge points in the target knowledge point set when constructing the push queue group. Using the predefined prerequisite relationships in the knowledge graph as partial order constraints, a directed acyclic graph topological sort is performed on all nodes in the target knowledge point set to ensure that all prerequisite knowledge points for any knowledge point in the push sequence are arranged before it.

[0057] After completing the topological sorting, the system constructs a corresponding push unit group for each knowledge point according to the sorting results. Each push unit group contains three types of push units: text and image summary units, interactive exercise units, and case video units. These three types of units are associated with the same target knowledge point identifier and carry a type tag. The text and image summary unit presents the core content of the knowledge point using structured text and illustrations, with a length controlled between 800 and 1200 words. It has the lowest cognitive load and is designed to help learners build a basic cognitive framework in a short time. The interactive exercise unit designs 3 to 5 multiple-choice or true / false questions around key concepts, requiring learners to answer them one by one and receive immediate feedback. It has a moderate cognitive load and is designed to strengthen memory encoding through active recall and immediate error correction. The case video unit uses real financial cases as backgrounds, embedding knowledge points into specific business scenarios for explanation and demonstration. Its duration is controlled between 5 and 8 minutes, with the highest cognitive load. It is designed to promote contextualized knowledge transfer and comprehensive understanding.

[0058] The three types of push notification units belong to the image and text summary queue, the interactive exercise queue, and the case video queue, respectively. Each queue is organized in topological order, and units at the same sequence number in the three queues correspond to the same knowledge point. The three queues are physically independent of each other, but logically connected horizontally through knowledge point identifiers.

[0059] After the push queue groups are constructed, the scheduling priority of each queue is dynamically adjusted by the subsequent cognitive load adaptive scheduling module based on the learner's real-time cognitive load level. When the cognitive load level is low, the case video queue is given a higher scheduling priority, allowing learners to complete the understanding of complex content when they are energetic; when the cognitive load level is high, the text and image summary queue is given a higher priority to avoid increasing the cognitive burden when cognitive resources are insufficient. As the push is executed, when the first unit of a queue is pushed, that unit is removed from the queue, and the subsequent units move forward in sequence; when all three queues are empty at the same time, it indicates that all knowledge points in the current target knowledge point set have been pushed for the first time, and the system enters the review push and stage convergence judgment stage.

[0060] As an alternative implementation of this embodiment, the morphology of the push queue group can be adjusted according to the characteristics of the specific training content. For knowledge points that emphasize calculation and formula derivation, a calculation training queue can be added on top of the text and image summaries, interactive exercises, and case study videos. For knowledge points that emphasize compliance processes and regulations, the case study video queue can be replaced with a scenario simulation queue. All of the above alternative solutions are within the protection scope of this invention.

[0061] In this embodiment, the text and image summary queue is used to store knowledge point explanations presented in text and chart form, which has the lowest cognitive consumption; the interactive exercise queue is used to store practice content in the form of multiple choice questions, true or false questions, which requires learners to actively participate in answering, and has a moderate cognitive consumption; the case video queue is used to store teaching videos containing actual case analysis, which requires learners to maintain their attention and make comprehensive understanding, and has the highest cognitive consumption.

[0062] The system collects learners' interactive behavior data streams in real time and quantifies their current cognitive load level. The quantification process is based on a weighted fusion of four indicators: learner's response time, accuracy rate, duration of a single learning session, and interaction frequency. Specifically, a longer response time and a lower accuracy rate indicate a higher cognitive load; a longer duration of a single learning session indicates a higher level of fatigue accumulation; and a higher interaction frequency (such as the frequency of pause and rewind operations) indicates a higher cognitive load. Each of the four indicators is normalized and then weighted according to preset weights to obtain a quantified cognitive load level value, ranging from 0 to 1.

[0063] Based on the current cognitive load level and the pre-configured cognitive consumption coefficients for each queue, the scheduling priority of each queue is adjusted. The adjustment process is as follows: Obtain the quantified value of the learner's cognitive load level at the current moment. Cognitive consumption coefficients of the three queues ( =1, 2, and 3 correspond to text and image summaries, interactive exercises, and case study videos, respectively. The cognitive consumption coefficients are normalized to obtain the normalized consumption intensity value for each queue. The specific calculation formula is as follows:

[0064] The remaining proportion of the cognitive load level quantified relative to the maximum value This represents the learner's current remaining cognitive capacity. The normalized consumption intensity value for each queue is then calculated. With current remaining cognitive capacity Compare the two and the degree of deviation between them. The smaller the value, the higher the load fit score for the cohort. The load fit score characterizes the degree of match between the cognitive exertion intensity of the cohort and the learner's current cognitive load capacity. For example, the cognitive load level after a learner works continuously in the afternoon... When =0.75, remaining cognitive capacity =0.25, at this time the image and text summary queue 0.3, and deviation The highest compatibility; the case study video queue ,and deviation The adaptation is the lowest, so the scheduling priority of the text and image summary queue is set to the highest to avoid pushing high-consumption case videos when students are fatigued.

[0065] Based on the current scheduling priority from highest to lowest, the head unit of a non-empty queue is selected as a candidate push unit and pushed to the learner terminal. In implementation, the scheduling priority is quantified based on load adaptability scores. After the scheduling priorities are sorted, each queue is checked from highest to lowest, and the head unit of the first non-empty queue is selected as a candidate push unit and pushed to the learner terminal. If the highest priority queue is empty, the next highest priority queue is used, and so on.

[0066] like Figure 5 As shown, the trend verification results obtained in the simulation environment in this embodiment are as follows: the horizontal axis is the quantitative value of the learner's real-time cognitive load level (the value ranges from 0.1 to 0.9, and the higher the value, the more strained the learner's current cognitive resources are and the higher the fatigue level), and the vertical axis is the normalized value of the knowledge mastery gain per unit time (the value ranges from 0 to 1, and the higher the value, the higher the learning efficiency).

[0067] The fixed-schedule strategy represents a traditional approach that pushes text summaries, interactive exercises, and case study videos in a fixed ratio, without considering the learner's real-time cognitive load. Experimental data shows that as the cognitive load level increases from 0.1 to 0.9, the learning efficiency under the fixed-schedule strategy continuously decreases from 0.57 to 0.24, a drop of 57.9%. The decline in efficiency is particularly pronounced after the load level exceeds 0.5, reflecting a significant deterioration in learning effectiveness when learners are forced to receive cognitively demanding content under conditions of high fatigue.

[0068] The scheduling strategy in this invention represents the adoption of the cognitive load adaptive scheduling method described in this invention, based on the current remaining cognitive capacity. Normalized consumption intensity of each queue Deviation between Dynamically adjust queue priorities. Experimental data shows that, within the same cognitive load variation range, the learning efficiency of the adaptive scheduling strategy of this invention only decreased from 0.71 to 0.52, a reduction of 26.8%. In the high load range (L≥0.7), the average learning efficiency is about 95.2% higher than that of the fixed scheduling strategy.

[0069] The above comparison results fully demonstrate that this invention, by matching the cognitive consumption intensity of the pushed content with the learner's current remaining cognitive capacity in real time, effectively suppresses the precipitous drop in learning efficiency when cognitive resources are exhausted, maintaining a relatively stable and high efficiency level throughout the learning process across the entire load range. This verifies the technical effectiveness of the cognitive load adaptive scheduling mechanism. S5. Review push notifications based on mastery feedback like Figure 4 As shown, based on learning feedback data, the system obtains the current mastery value of each knowledge point in the push queue group. The learning feedback data includes learners' completion status of the pushed content, exercise answer records, and self-test results after learning. For a knowledge point that has been pushed, the system updates the mastery estimate of that knowledge point by analyzing the learner's correct answer rate and answer time on the related exercises.

[0070] The push interval is determined based on the difference between the current mastery level and the preset mastery target value. The larger the difference, the further the mastery level is from the target, and the more urgently review is needed; therefore, the push interval is shorter. In this embodiment, the push interval includes multiple non-zero time intervals. When the difference is greater than a second preset value, the push interval is negatively correlated with the difference; the larger the difference, the smaller the non-zero time interval. When the difference is less than the second preset value, the upper limit is taken, indicating that no push is needed. When the push interval timer ends, the push unit corresponding to the knowledge point is loaded to the first position of the push queue and re-pushed to the learner's terminal, ensuring that knowledge points at the threshold of forgetting can receive timely review opportunities, while knowledge points that have been firmly mastered will not frequently disturb the learner.

[0071] S6. Phased Convergence and Reprogramming When the difference between each knowledge point and the preset mastery target value is less than the first preset value, it indicates that the target knowledge points for the current stage have been basically mastered. The learning feedback data is used to update the multidimensional cognitive state vector. Based on the updated data, the learning tasks for the next stage are replanned. During the replanning, the current knowledge weaknesses are identified based on the updated knowledge mastery vector. Combined with the learner's learning progress, a new set of target knowledge points is generated, and the above-mentioned push process is repeated.

[0072] like Figure 6 As shown, the horizontal axis represents the number of consecutive push interactions for the same knowledge point (from the 1st to the 6th time), and the vertical axis represents the current mastery value of the knowledge point obtained by the system based on learning feedback data (ranging from 0 to 1). The three curves in the figure represent simple knowledge points (basic concepts), medium knowledge points (application calculations), and difficult knowledge points (comprehensive analysis), respectively. The horizontal dashed line represents the preset mastery target value of 0.92. From the overall trend, the mastery of the three types of knowledge points shows a monotonically increasing trend during the continuous push process, indicating that the target knowledge point set selection strategy based on deep reinforcement learning decision-making in this invention can effectively drive the continuous improvement of learners' knowledge level.

[0073] This embodiment integrates multi-dimensional cognitive state perception, deep reinforcement learning decision-making, adaptive cognitive load scheduling, and forgetting curve-driven review push into a system. This allows the push system to simultaneously consider the logical integrity of the knowledge structure, the learner's real-time cognitive capacity, and the continuous tracking of learning outcomes. Compared to traditional solutions that only focus on knowledge point matching, this embodiment introduces cognitive load perception into the push decision-making process, avoiding the push of highly difficult content when learners lack sufficient cognitive resources. Compared to solutions that only focus on the logical coherence of knowledge, this embodiment achieves dynamic optimization of the push strategy through the sequential decision-making capabilities of deep reinforcement learning. Compared to solutions that only optimize path planning, this embodiment ensures the consolidation of knowledge through a mastery gap-driven review push mechanism, effectively coordinating the speed of knowledge acquisition with the individual's cognitive pace. This completes the systematic knowledge transfer without increasing cognitive burden, providing a complete technical solution for personalized learning services on financial training platforms.

[0074] Example 2 This embodiment provides a personalized learning recommendation system for a financial training platform. This system is applied to the personalized learning recommendation method described in Embodiment 1 and is used to execute the process in Embodiment 1. This embodiment aims to illustrate the specific composition of each functional module in the system, the data interaction relationships between modules, and the deployment and collaboration methods in the platform's operating environment.

[0075] The system deployment environment used in this embodiment is the backend service cluster of an online training platform for a financial institution. This platform adopts a microservice architecture, is deployed in a private cloud environment, and includes infrastructure such as application servers, database servers, cache servers, and file storage servers. Learners access the training service through a web browser or mobile application. The system server and client interact via a RESTful API, and real-time interactive data is communicated bidirectionally via a WebSocket long-lived connection.

[0076] The state awareness module is responsible for acquiring the learner's multidimensional cognitive state vector and determining the set of candidate learning objects. This module consists of three parts: a data acquisition submodule, a state inference submodule, and a candidate selection submodule.

[0077] The data acquisition submodule monitors various interactive events of learners, including login / logout, page browsing, content clicking, answering exercises, video playback control, etc., and organizes these raw event data into a learning behavior sequence according to time order and stores it in a time-series database.

[0078] The state inference submodule reads the current learner's behavior sequence data, historical mastery records, and knowledge graph data at preset time intervals (set to be after each push unit is completed in this embodiment), and constructs a multidimensional cognitive state vector according to the method described in Embodiment 1.

[0079] The candidate selection submodule receives the learning target input, categorizes the learning target, maps it to a set of first nodes in the knowledge graph, calculates the semantic similarity between each piece of content in the learning base and the target nodes, and includes learning content with a similarity exceeding a dynamic threshold into the candidate learning object set. The output of this module is a multi-dimensional cognitive state vector and a set of candidate learning objects, both of which serve as input data for the decision-making module.

[0080] The decision-making module is responsible for inputting the multi-dimensional cognitive state vector into the deep reinforcement learning network and, based on a multi-objective reward function, deciding on the target knowledge point set from the candidate learning object set. The core of the decision-making module is a pre-trained and deployed deep reinforcement learning network model, which adopts the Actor-Critic architecture described in Example 1. The model parameters are stored in the model management service in PB format. Upon startup, the decision-making module loads the latest version of the model parameters from the model management service into memory. After the state awareness module completes a state update, the decision-making module receives the multi-dimensional cognitive state vector and the candidate learning object set, calls the model inference interface to perform forward computation, outputs the selection probability distribution of each candidate object, and determines the selected learning object for the current step after action sampling. The decision-making module repeatedly executes the above decision-making process according to a preset single-round push limit until the limit is reached or the candidate knowledge point coverage threshold is reached. All knowledge points selected in this round are then aggregated into a target knowledge point set and passed to the queue module.

[0081] The queue module is responsible for constructing push queue groups based on the target knowledge point set and according to the knowledge graph. After receiving the target knowledge point set output by the decision module, the queue module first reads the prerequisite relationship data between each knowledge point from the knowledge graph service, performs topological sorting, and generates an ordered sequence of knowledge points. The queue module then generates text and image summary push units, interactive exercise push units, and case video push units for each knowledge point in the sequence. Each push unit contains a content identifier, a knowledge point identifier, and a format type tag. The content data of the push unit is read from the content management service according to the identifier, and the unit itself is stored in an in-memory queue as a lightweight object structure. The queue module pushes the three types of push units into the corresponding text and image summary queue, interactive exercise queue, and case video queue according to the order of the knowledge point sequence, completing the construction of the push queue group. After construction, the queue module passes the reference of the queue group to the first push module.

[0082] The first push module is responsible for quantifying the current cognitive load level based on the real-time collected interactive behavior data flow, and adjusting the scheduling priority of each queue in combination with the pre-configured cognitive consumption coefficient of each queue; according to the current scheduling priority from high to low, it selects the first push unit in the non-empty queue and pushes it to the learner's terminal.

[0083] The first push module includes a cognitive load quantification submodule, a priority scheduling submodule, and a push execution submodule.

[0084] The cognitive load quantification submodule monitors the learner's interactive behavior data stream in real time and calculates the current cognitive load level quantification value every 30 seconds.

[0085] The priority scheduling submodule receives the cognitive load level quantification value and the cognitive consumption coefficient of each queue. According to the load adaptability score calculation method described in Example 1, it calculates the scheduling priority score of the three queues respectively, and sorts them from high to low to form the scheduling priority list at the current time.

[0086] The push execution submodule checks each queue sequentially according to the scheduling priority list, selects the first non-empty queue's head push unit, calls the content push interface to send the unit's content to the learner's terminal, and then removes the unit from the queue. The push execution submodule also maintains push status records, including push time, push unit identifier, and push result status.

[0087] The second push module is responsible for obtaining the current mastery value of each knowledge point based on the learning feedback data, and determining the push interval based on the difference between the current mastery value and the preset mastery target value. When the push interval timer ends, the push unit corresponding to the knowledge point is reloaded to the first position of the push queue.

[0088] The second push module includes a feedback collection submodule, a mastery assessment submodule, and a review scheduling submodule.

[0089] The feedback collection submodule listens to the interaction data generated by learners after they complete the pushed learning unit, including the results of answering exercises, learning time, learning completion rate, etc., and formats this data into standardized feedback events.

[0090] The mastery assessment submodule updates the mastery estimate of the corresponding knowledge point after each feedback event is received, and compares it with the preset mastery target value to calculate the difference.

[0091] The review scheduling submodule determines the push interval and starts a timer based on the difference according to a preset mapping rule (0 push interval when the difference is greater than 0.3, 1 day when the difference is between 0.15 and 0.30, 3 days when the difference is between 0.05 and 0.15, and 7 days when the difference is less than 0.05). When the timer expires, the review scheduling submodule reloads the three push units corresponding to the knowledge point—text summary, interactive exercises, and case video—to the beginning of their respective queues, waiting for the first push module to execute the push.

[0092] The replanning module is responsible for updating the multidimensional cognitive state vector and triggering the state perception module to replan the learning task for the next stage when the difference between each knowledge point and the preset mastery target value is less than the first preset value. After each knowledge point completes the review push and updates the mastery, the replanning module calculates the mastery difference of all target knowledge points in the current stage. When the mastery of all knowledge points reaches the expected level, the replanning module sends a replanning signal to the state perception module, triggering the state perception module to reassess the learner's current cognitive state and generate the target learning task for the next stage.

[0093] The modules interact with each other via message queues and shared caches to achieve decoupling and asynchronous processing. The multidimensional cognitive state vector output by the state awareness module is written to a Redis cache in JSON format with the key "user:userId:cognitivestate", and a state update notification is sent to the message queue of the decision module. After consuming the notification, the decision module reads the latest cognitive state vector from Redis, performs decision reasoning, writes the output target knowledge point set to the Redis cache, and notifies the queue module. After completing the construction of the push queue group, the queue module stores the queue group information in an in-memory data structure and notifies the first push module of the queue reference address.

[0094] The first push module and the second push module cooperate through a shared queue group data structure. The first push module is responsible for taking the unit from the head of the queue for the first push, and the second push module is responsible for reloading the review unit to the head of the queue when the review time ends.

[0095] The above embodiments are merely preferred embodiments of the present invention and do not constitute any limitation on the scope of protection of the present invention. The basic concept of the present invention has been fully embodied in the foregoing embodiments. Any modifications, equivalent substitutions, improvements, etc., made to these embodiments within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A personalized learning recommendation method for a financial training platform, characterized in that: include, A learner's multidimensional cognitive state vector is obtained, and the multidimensional cognitive state vector is input into a deep reinforcement learning network. The deep reinforcement learning network then decides to output a target knowledge point set from the candidate learning object set based on a multi-objective reward function. Based on the target knowledge point set, push queue groups are constructed according to the relationships between knowledge points in the knowledge graph; Real-time acquisition of learners' interactive behavior data streams, quantification of the current cognitive load level; based on the current cognitive load level and combined with the cognitive consumption coefficient pre-configured for each queue, adjustment of the scheduling priority of each queue; According to the current scheduling priority from high to low, the head push unit in the non-empty queue is selected as the candidate push unit; and the candidate push unit is pushed to the learner terminal. Based on learning feedback data, obtain the current mastery level value of each knowledge point in the push queue group; The push interval is determined based on the difference between the current mastery level value and the preset mastery level target value; When the push interval timer expires, the knowledge point is loaded to the top of the push queue and pushed to the learner's terminal; When the difference between each knowledge point and the preset mastery target value is less than the first preset value, the evaluation learning feedback data updates the multidimensional cognitive state vector, and based on the updated data, the learning task for the next stage is replanned.

2. The personalized learning push method for the financial training platform according to claim 1, characterized in that: The multidimensional cognitive state vector is constructed using a neural network architecture that combines a multi-head self-attention mechanism with gated recurrent units; specifically, it includes: The learner's basic attributes and learning preferences are encoded into a first feature vector; The knowledge concepts in the knowledge graph are mapped to graph structure features, and then encoded into a second feature vector through graph embedding. After concatenating and fusing the first feature vector and the second feature vector, the implicit interaction relationship between different knowledge dimensions is first captured by a multi-head self-attention layer, and then input into a gated recurrent unit layer to capture the change pattern of cognitive state over time, and output the multi-dimensional cognitive state vector.

3. The personalized learning push method for the financial training platform according to claim 2, characterized in that: The multidimensional cognitive state vector includes a knowledge mastery vector, a forgetting state vector, a cognitive load estimation vector, a preference state vector, and a learning objective embedding vector; wherein: The knowledge acquisition vector is iteratively updated by a gated cyclic unit combined with the time decay factor of the Ebbinghaus forgetting curve. The forgetting state vector is used to predict the forgetting risk value of each knowledge concept through the MemGNN network. The MemGNN network uses the knowledge graph as a graph structure, aggregates the state information of adjacent knowledge nodes through the graph attention network layer, and inputs the aggregation result into the gated recurrent unit to output the forgetting probability of each knowledge concept in the next time step. The cognitive load estimation vector is obtained by weighted fusion estimation based on learners’ response time, accuracy, single learning duration and interaction frequency. The preference state vector is decoupled from the learning behavior sequence through a continuous-time heterogeneous information network to obtain dynamic preference features.

4. The personalized learning push method for the financial training platform according to claim 1, characterized in that: Obtain learners' learning objectives, classify the learning objectives in a fine-grained manner, and obtain a classification result set; Each knowledge concept in the classification result set is mapped to a first node set in the knowledge graph, and each learning content in the learning library is mapped to a second node set in the knowledge graph. The cosine similarity of the semantic embedding vectors between the first node and the second node is calculated, and the learning content with a similarity exceeding a preset dynamic threshold is included in the candidate learning object set.

5. The personalized learning push method for the financial training platform according to claim 1, characterized in that: The push queue group includes a text and image summary queue, an interactive exercise queue, and a case video queue. The push units in each queue are associated with the same target knowledge point identifier and carry a form type label. The text and image summary queue, the interactive exercise queue, and the case video queue are pre-configured with cognitive consumption coefficients that increase sequentially.

6. The personalized learning push method for the financial training platform according to claim 5, characterized in that: Adjust the scheduling priority of each queue, including: Obtain the learner's cognitive load level quantification value and the cognitive consumption coefficient of each queue at the current moment; The cognitive consumption coefficients are normalized to obtain the normalized consumption intensity value of each queue. The remaining proportion of the quantified cognitive load level relative to the maximum value is taken as the learner's current remaining cognitive capacity; The normalized consumption intensity value of each queue is compared with the current remaining cognitive capacity. The smaller the deviation between the two, the higher the load fit score of the queue. The load fit score is used to characterize the degree of matching between the cognitive consumption intensity of the queue and the learner's current cognitive burden. The scheduling priority of each queue is adjusted based on its load adaptability score.

7. The personalized learning push method for the financial training platform according to claim 1, characterized in that: The deep reinforcement learning network adopts a deep reinforcement learning model based on the Actor-Critic architecture, and its policy evaluation module includes a state encoding layer, an action masking layer, an Actor branch, and a Critic branch; wherein: The state coding layer is used to perform feature fusion coding on the multidimensional cognitive state vector at the current moment and output a compact implicit state representation. The action masking layer is used to generate an action mask based on the set of candidate learning objects, and to mask non-candidate learning objects, so that the policy evaluation module can only select actions within the current set of candidate learning objects. The Actor branch is used to output the selection probability distribution of each learning object in the candidate learning object set; the Critic branch is used to output the value estimate of the current multidimensional cognitive state.

8. The personalized learning push method for the financial training platform according to claim 7, characterized in that: The deep reinforcement learning network also includes an experience replay pool for storing quadruples of state, action, reward and next state for historical time steps, and updates the network parameters during the training phase through random batch sampling.

9. The personalized learning push method for the financial training platform according to claim 1, characterized in that: Based on the fact that the difference is negatively correlated with the push interval, the push interval includes multiple non-zero time intervals.

10. A personalized learning recommendation system for a financial training platform, applied to the personalized learning recommendation method of the financial training platform according to any one of claims 1-9, characterized in that, include: The state-aware module is used to acquire the learner's multidimensional cognitive state vector and determine the set of candidate learning objects; The decision module is used to input the multidimensional cognitive state vector into the deep reinforcement learning network and, based on the multi-objective reward function, decide and output the target knowledge point set from the candidate learning object set. The queue module is used to construct a push queue group based on the target knowledge point set and according to the knowledge graph; The first push module is used to quantify the current cognitive load level based on the real-time collected interactive behavior data flow, and adjust the scheduling priority of each queue in combination with the pre-configured cognitive consumption coefficient of each queue; according to the current scheduling priority from high to low, the first push unit of the non-empty queue is selected and pushed to the learner terminal. The second push module is used to obtain the current mastery value of each knowledge point based on learning feedback data, and to determine the push interval based on the difference between the current mastery value and the preset mastery target value. When the push interval timer ends, the push unit corresponding to the knowledge point is reloaded to the first position of the push queue. The replanning module is used to update the multidimensional cognitive state vector and trigger the state perception module to replan the learning task for the next stage when the difference between each knowledge point and the preset mastery target value is less than the first preset value.

Images