Intelligent matching system for coach and student based on bidirectional user portrait

Abstract

This invention discloses an intelligent matching system for coaches and students based on bidirectional user profiles, relating to the field of intelligent education technology. The system includes a profile alignment module, a load monitoring module, a complementary calculation module, a strategy reordering module, a game-theoretic solution module, and a parameter update module. The profile alignment module acquires multi-source data from both the coach and student sides, mapping the data to a weighted bidirectional profile tensor based on preset cross-modal semantic anchor points. By constructing the weighted bidirectional profile tensor, accurate semantic alignment of the heterogeneous multi-source data is achieved. Cross-wavelet transform is used to extract phase coherence coefficients, and conditional information entropy is combined to dynamically penalize and correct the basic complementarity, effectively quantifying the real-time load matching status and uncertainty risk during the teaching interaction process. Simultaneously, an offline reinforcement learning strategy network is introduced, using the skill growth rate per unit time as a delayed reward signal to generate a reordering adjustment factor.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to an intelligent matching system for coaches and students based on two-way user profiles. Background Technology

[0002] Intelligent education technology refers to technologies that utilize artificial intelligence, multimodal signal processing, and data-driven decision-making to perceive, quantify, and adaptively optimize learner status, teacher-student interaction, and teaching effectiveness in real time during the teaching process.

[0003] Traditional methods for analyzing the temporal signals of student / coach behavior rely on human experience to set a fixed time window length. However, human behavior is highly non-stationary and has individual rhythm differences. A fixed window cannot simultaneously capture short-term sudden actions and preserve long-term periodic trends. Furthermore, when assessing the tacit understanding between teachers and students, current technologies often use linear indicators such as the Pearson correlation coefficient, which are easily affected by asynchronous factors such as individual baseline differences and common trend terms, leading to false synchronization misjudgments. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a coach-student intelligent matching system based on two-way user profiles. This solves the problem that traditional methods rely on human experience to set a fixed time window length when analyzing the behavioral time-series signals of students / coaches. However, human behavior is highly non-stationary and has individual rhythm differences, and a fixed window cannot simultaneously capture short-term sudden actions and preserve long-term periodic trends.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent matching system for coaches and students based on two-way user profiles, comprising:

[0008] Image alignment module, load monitoring module, complementary calculation module, strategy rearrangement module, game solving module, and parameter update module;

[0009] The portrait alignment module is used to acquire multi-source data from the coach's end and multi-source data from the student's end, and map the multi-source data from the coach's end and the multi-source data from the student's end into a bidirectional portrait tensor with confidence weights based on preset cross-modal semantic anchor points.

[0010] The load monitoring module is used to collect real-time behavior time-series signals from the coach's end and real-time response time-series signals from the student's end, perform cross wavelet transform on the real-time behavior time-series signals from the coach's end and real-time response time-series signals from the student's end to extract phase coherence coefficients, and construct a conditional probability distribution model of the phase coherence coefficients based on historical effective teaching data.

[0011] The complementary calculation module is used to receive bidirectional profile tensors and phase coherence coefficients, calculate the basic complementarity degree in the directional distortion potential space embedded with prior knowledge of teaching styles through an asymmetric metric function, calculate the conditional information entropy of the current candidate pair using a conditional probability distribution model, apply dynamic penalty correction to the basic complementarity degree based on the conditional information entropy to obtain the time-fit score, and generate an initial candidate matching pool based on the time-fit score.

[0012] The strategy reordering module is used to load an offline reinforcement learning strategy network, use the skill growth rate per unit time as a delayed reward signal to output a reordering adjustment factor to the candidate pairs in the initial candidate matching pool, and generate an effect-oriented adaptation score and a reordering candidate list based on the reordering adjustment factor and the time-effect adaptation score.

[0013] The game-solving module is used to obtain the coach's real-time remaining capacity and the student's priority weight, and to execute an improved bilateral matching algorithm with the reordered candidate list as the preference input. After an iterative proposal process with capacity constraints and priority awareness, a stable matching scheme is output.

[0014] The parameter update module is used to collect real-time satisfaction data, process behavior data, and staged skill growth data corresponding to the stable matching scheme, and to adjust the topological constraint parameters and conditional probability distribution model in the complementary calculation module, the offline reinforcement learning policy network weights in the policy rearrangement module, and the capacity decay parameters in the game solving module.

[0015] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the profile alignment module maps multi-source data from the coach's end and multi-source data from the student's end into a bidirectional profile tensor with confidence weights. The specific steps are as follows:

[0016] Modal classification processing was performed on the multi-source data from the coach's side and the multi-source data from the student's side to obtain a structured label set, an unstructured text corpus, and a continuous behavior time series.

[0017] For structured tag sets, one-hot encoding is used to generate basic feature vectors; for unstructured text corpora, a pre-trained semantic encoder is used to extract semantic embedding vectors; and for continuous behavior time-series, a sliding window statistical method is used to generate behavior summary vectors.

[0018] The basic feature vector, semantic embedding vector, and behavior summary vector are concatenated into the original portrait matrix according to a preset dimension order;

[0019] Based on the preset cross-modal semantic anchor point system, the correlation strength between each dimension feature and the corresponding semantic anchor point in the original portrait matrix is ​​calculated, and the correlation strength is used as the confidence weight of the corresponding dimension.

[0020] The confidence weights are multiplied element-wise with the corresponding dimensional features in the original portrait matrix to obtain a bidirectional portrait tensor.

[0021] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the preset cross-modal semantic anchor points include three types of benchmark semantic nodes: teaching style anchor points, knowledge domain anchor points, and interaction preference anchor points.

[0022] Calculate the association strength between each dimension feature in the original portrait matrix and the corresponding semantic anchor point, including determining the modality category to which each dimension feature belongs, and selecting the subdivided semantic sub-nodes under the benchmark semantic node corresponding to the modality category as the target anchor point;

[0023] Calculate the cosine similarity between the dimensional feature vector and the target anchor vector in the normalized semantic space, and determine the cosine similarity value as the confidence weight.

[0024] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the load monitoring module performs cross-wavelet transform on the real-time behavior time-series signal from the coach's end and the real-time response time-series signal from the student's end to extract the phase coherence coefficient. The specific steps are as follows:

[0025] A continuous wavelet transform is performed on the real-time behavior time-series signal of the coach to obtain the wavelet power spectrum of the coach, and a continuous wavelet transform is performed on the real-time response time-series signal of the student to obtain the wavelet power spectrum of the student.

[0026] The cross wavelet spectrum is obtained by performing complex conjugate multiplication on the wavelet power spectrum of the coach and the wavelet power spectrum of the student.

[0027] A smoothed cross spectrum is obtained by performing Gaussian smoothing filtering on the cross wavelet spectrum in the time dimension;

[0028] The cross-wavelet coherence value is calculated based on the smooth cross spectrum, the wavelet power spectrum at the coach's end, and the wavelet power spectrum at the student's end.

[0029] The arithmetic mean of the cross-wavelet coherence values ​​within the preset effective teaching frequency band is determined as the phase coherence coefficient.

[0030] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the complementary calculation module calculates the conditional information entropy of the current candidate pair using a conditional probability distribution model, specifically through the following steps:

[0031] The coach profile features and student profile features of the current candidate pair are extracted from the bidirectional profile tensor. The coach profile features and student profile features are then input into the directed distorted latent space mapping network to obtain the joint representation vector.

[0032] Using the joint representation vector as the query condition, the matching phase coherence coefficient probability density curve is retrieved in the conditional probability distribution model.

[0033] The cumulative probability value is obtained by performing a definite integral operation on the probability density curve of the phase coherence coefficient within a preset effective teaching interval;

[0034] The negative natural logarithm of the cumulative probability value is defined as the conditional information entropy.

[0035] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the complementary calculation module applies a dynamic penalty correction to the basic complementarity based on conditional information entropy to obtain a timeliness adaptation score. The specific steps are as follows:

[0036] Set an upper limit and a lower limit for the information entropy threshold;

[0037] When the conditional information entropy is less than the lower limit of the information entropy threshold, the basic complementarity is directly determined as the timeliness adaptation score.

[0038] When the conditional information entropy is greater than the upper limit of the information entropy threshold, the basic complementarity is multiplied by the preset minimum retention coefficient and then determined as the timeliness adaptation score.

[0039] When the conditional information entropy is between the lower limit and the upper limit of the information entropy threshold, the timeliness adaptation score is calculated according to the linear decay formula, which is:

[0040] ;

[0041] in, Indicates timeliness-appropriate score. Indicates the basic complementarity. Represents conditional information entropy. This represents the lower limit of the information entropy threshold. This represents the upper limit of the information entropy threshold. This indicates the preset minimum retention factor.

[0042] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the strategy reordering module uses the skill growth rate per unit time as a delayed reward signal to output a reordering adjustment factor for candidate pairs in the initial candidate matching pool. The specific steps are as follows:

[0043] Extract a set of historical samples that have the same combination of coach profile features and student profile features as the current candidate pair from historical matching records;

[0044] The changes in skill test scores and corresponding durations of all samples in the historical sample set within the preset observation period after matching are statistically analyzed, and the ratio of the change in skill test scores to the duration is calculated as the skill growth rate per unit time.

[0045] Input the skill growth rate per unit time into the value function header of the offline reinforcement learning policy network, and output the state value estimate;

[0046] The state value estimate is normalized, and the normalized value is used as the reordering adjustment factor.

[0047] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the load monitoring module constructs a conditional probability distribution model of the phase coherence coefficient based on historical effective teaching data, specifically through the following steps:

[0048] Valid teaching samples with post-lesson satisfaction scores higher than the preset passing score were selected from historical teaching records;

[0049] Extract the phase coherence coefficient time series corresponding to the valid teaching samples;

[0050] Adaptive bandwidth kernel density estimation is performed on the phase coherence coefficient time series to obtain the original probability density function;

[0051] The formula for calculating the adaptive bandwidth kernel density estimation is as follows:

[0052] ;

[0053] in, This indicates that the original probability density function is in Density value at that location, This represents the total number of valid teaching samples. Indicates the first The phase coherence coefficient of a valid teaching sample. Represents the Gaussian kernel function. Indicates the first The adaptive bandwidth of each sample; the original probability density function is stored as a conditional probability distribution model.

[0054] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the training process of the directional distortion latent space mapping network includes:

[0055] Collect positive and negative sample pairs with labeled matching results;

[0056] The coach profile features and student profile features of positive sample pairs and negative sample pairs are respectively input into the mapping network to be trained to obtain the joint representation vector of positive samples and the joint representation vector of negative samples.

[0057] A contrastive loss function is constructed with the objective of maximizing the difference between the joint representation vectors of positive samples and the joint representation vectors of negative samples.

[0058] The contrastive loss function is used to perform gradient descent updates on the mapping network to be trained until the loss function converges. The converged network parameters are then determined as the final parameters of the directional warped latent space mapping network.

[0059] As a preferred embodiment of the intelligent matching system for coaches and students based on bidirectional user profiles described in this invention, the parameter update module adjusts the topological constraint parameters and conditional probability distribution model in the complementary calculation module, specifically through the following steps:

[0060] The overall effectiveness score for this match is calculated based on real-time satisfaction data and phased skill growth data.

[0061] When the overall performance score is lower than the preset update trigger threshold, the conditional information entropy and basic complementarity corresponding to this match are extracted as abnormal samples.

[0062] Incremental correction of the adaptive bandwidth parameter of the conditional probability distribution model is performed using outlier samples;

[0063] The upper and lower limits of the information entropy threshold are adjusted based on the degree of deviation between the conditional information entropy and the basic complementarity in abnormal samples.

[0064] Write the corrected adaptive bandwidth parameters and the adjusted upper and lower limits of the information entropy threshold back to the configuration storage area of ​​the complementary computing module.

[0065] The beneficial effects of this invention are as follows: By constructing a bidirectional profile tensor with confidence weights, accurate semantic alignment of multi-source heterogeneous data is achieved. The phase coherence coefficient is extracted using cross wavelet transform, and the basic complementarity is dynamically penalized and corrected by combining conditional information entropy. This effectively quantifies the real-time load matching status and uncertainty risk in the teaching interaction process. At the same time, an offline reinforcement learning strategy network is introduced to generate a reordering adjustment factor using the skill growth rate per unit time as a delayed reward signal. Long-term teaching effect feedback is integrated into the immediate matching decision. Then, an improved bilateral matching algorithm with capacity constraints and priority awareness is used to solve the stable solution. Based on multi-dimensional feedback data, the topological constraint parameters, probability distribution model, and network weights are continuously iterated. Under the premise of ensuring the stability of the matching results, the timeliness, effect orientation, and adaptive evolution capability of the system in bidirectional adaptation between coaches and students are improved. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the intelligent matching system for coaches and students based on two-way user profiles in Example 1. Detailed Implementation

[0068] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0069] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0070] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0071] Example 1, referring to Figure 1 As one embodiment of the present invention, this embodiment provides a coach-student intelligent matching system based on two-way user profiles, including:

[0072] The module includes a profile alignment module, a load monitoring module, a complementary calculation module, a strategy rearrangement module, a game theory solution module, and a parameter update module.

[0073] The profile alignment module is used to acquire multi-source data from both the coach and student sides, and map the multi-source data from both sides into bidirectional profile tensors with confidence weights based on preset cross-modal semantic anchor points.

[0074] Furthermore, modal classification processing is performed on the multi-source data from the coach's side and the multi-source data from the student's side to obtain a structured label set, an unstructured text corpus, and a continuous behavior time series. For the structured label set, a one-hot encoding method is used to generate basic feature vectors. For the unstructured text corpus, a pre-trained semantic encoder is used to extract semantic embedding vectors. For the continuous behavior time series, a sliding window statistical method is used to generate behavior summary vectors.

[0075] The basic feature vector, semantic embedding vector, and behavior summary vector are concatenated into the original portrait matrix in a preset dimension order. The correlation strength between each dimension feature and the corresponding semantic anchor in the original portrait matrix is ​​calculated according to the preset cross-modal semantic anchor system, and the correlation strength is used as the confidence weight of the corresponding dimension. The confidence weight is multiplied element-wise with the corresponding dimension feature in the original portrait matrix to obtain the bidirectional portrait tensor.

[0076] The preset cross-modal semantic anchors include three types of benchmark semantic nodes: teaching style anchors, knowledge domain anchors, and interaction preference anchors. The association strength between each dimension feature in the original profile matrix and the corresponding semantic anchor is calculated, including determining the modality category to which each dimension feature belongs and selecting the subdivided semantic sub-nodes under the benchmark semantic nodes corresponding to the modality category as target anchors. The cosine similarity between the dimension feature vector and the target anchor vector in the normalized semantic space is calculated, and the cosine similarity value is determined as the confidence weight.

[0077] The calculation process for the confidence weight is as follows:

[0078] Calculate the dot product and Euclidean norm of the semantic embedding vectors of the two modalities respectively, and divide the dot product by the product of the two norms to obtain the original cosine similarity; calculate the maximum and minimum values ​​of the similarity across all dimensions in the current modality, perform min-max linear normalization on the original similarity, and map it to the [0,1] interval as the final weight, expressed as:

[0079] ;

[0080] In the formula, To normalize the confidence weights, , It is a two-modal semantic embedding vector. This represents the dot product operation. Describes the Euclidean norm. , These are the maximum and minimum similarity values ​​across all dimensions of the current modality, respectively.

[0081] It should be noted that by performing modal classification and targeted encoding on multi-source heterogeneous data respectively, and combining the cross-modal semantic anchor system to calculate the feature association strength as the confidence weight, the semantic gap between different data types can be effectively eliminated. This ensures that the generated bidirectional profile tensor retains the integrity of the original information and has a comparable unified semantic benchmark, thereby improving the accuracy of profile representation and the reliability of subsequent matching calculations.

[0082] Specifically, the profile alignment uses a combination of one-hot encoding and domain-fine-tuned semantic models to eliminate the misleading numerical representation of discrete features and the semantic shift of general models in teaching scenarios; the behavior summary window is dynamically adjusted according to the autocorrelation period of the signal to take into account both short-term events and long-term trends; semantic anchors are automatically generated through clustering of highly matched samples to adapt to data evolution; the confidence weight uses cosine similarity and is normalized to strengthen the consistency of high-dimensional space orientation and eliminate dimensional interference, so that each dimension of the profile tensor carries verified reliability information, providing a noise-resistant unified semantic benchmark for matching.

[0083] The calculation process for the sliding window length of the behavior summary vector is as follows:

[0084] Calculate the autocorrelation function for the behavioral time series signal, starting from lag order 0 and iterating through the integer lag orders where the autocorrelation value first drops below the peak proportion threshold (default 0.5). ,Will Add 1 to the sliding window length to ensure the window covers at least one complete behavior cycle. The expression is:

[0085] ;

[0086] In the formula, The length of the dynamically sliding window. Lag order The autocorrelation function value at that location, The peak value of the autocorrelation is zero hysteresis. The peak ratio threshold, The smallest integer lag order that satisfies the conditions. It means "to be bound by".

[0087] The load monitoring module is used to collect real-time behavior time-series signals from the coach's end and real-time response time-series signals from the student's end. It performs cross-wavelet transform on the real-time behavior time-series signals from the coach's end and real-time response time-series signals from the student's end to extract the phase coherence coefficients, and constructs a conditional probability distribution model of the phase coherence coefficients based on historical effective teaching data.

[0088] Furthermore, a continuous wavelet transform is performed on the real-time behavior time-series signal of the coach to obtain the wavelet power spectrum of the coach, and a continuous wavelet transform is performed on the real-time response time-series signal of the student to obtain the wavelet power spectrum of the student. Complex conjugate multiplication is performed on the wavelet power spectra of the coach and the student to obtain a cross-wavelet spectrum. Gaussian smoothing filtering is performed on the cross-wavelet spectrum in the time dimension to obtain a smoothed cross-spectrum. The cross-wavelet coherence value is calculated based on the smoothed cross-spectrum, the wavelet power spectrum of the coach, and the wavelet power spectrum of the student. The arithmetic mean of the cross-wavelet coherence values ​​within a preset effective teaching frequency band is determined as the phase coherence coefficient.

[0089] Valid teaching samples with post-class satisfaction scores higher than the preset passing line were selected from historical teaching records; the phase coherence coefficient time series corresponding to the valid teaching samples were extracted; adaptive bandwidth kernel density estimation was performed on the phase coherence coefficient time series to obtain the original probability density function.

[0090] The formula for calculating adaptive bandwidth kernel density estimation is as follows:

[0091] ;

[0092] in, This indicates that the original probability density function is in Density value at that location, This represents the total number of valid teaching samples. Indicates the first The phase coherence coefficient of a valid teaching sample. Represents the Gaussian kernel function. Indicates the first Adaptive bandwidth for each sample; store the original probability density function as a conditional probability distribution model.

[0093] It should be noted that by extracting the phase coherence coefficient between the real-time behavioral signals of the coach and the students through cross wavelet transform, and constructing an adaptive bandwidth kernel density estimation model based on historical effective teaching samples, the dynamic synchronization characteristics and statistical distribution patterns of the behavioral rhythm of both parties during the teaching interaction can be accurately captured. This provides a probabilistic metric basis with both time-frequency resolution and data-driven adaptability for quantitatively evaluating the real-time teaching load matching status.

[0094] Specifically, load monitoring uses complex Morlet wavelets to capture transient synchronization features in non-stationary interactions, and uses complex conjugate multiplication to remove individual rhythm differences and retain only relative phase information; the smoothing bandwidth is adaptively set according to the cross-spectral energy concentration to avoid erasing the real coupling structure; and pseudo-coherent regions are removed by significance testing before the coherence value is averaged.

[0095] The calculation process for the cross-wavelet coherence coefficient is as follows: Perform continuous complex Morlet wavelet transforms on both signals, calculate the cross-wavelet spectrum, and apply Gaussian smoothing. Calculate the normalized coherence value point-by-point within the target frequency band. Generate a null hypothesis distribution through Monte Carlo simulation, retain only the frequency points that pass the significance test, and take the arithmetic mean of their normalized coherence values ​​to obtain the final coherence coefficient, expressed as:

[0096] ;

[0097] In the formula, The phase coherence coefficient, , The result is the continuous wavelet transform of the two signals. For cross-wavelet spectrum, This represents the Gaussian smoothing operator. Let be the set of frequency points that pass the significance test. The number of elements in the set.

[0098] Kernel density estimation dynamically adjusts the bandwidth based on local sample density, balancing the stability of sparse regions with the details of dense regions; the conditional probability model is constructed only based on high-satisfaction samples, transforming the implicit experience of effective teaching into statistical priors, so that the phase coherence coefficient anchors the true teaching effect.

[0099] The complementary calculation module receives bidirectional profile tensors and phase coherence coefficients, calculates the basic complementarity using an asymmetric metric function within a directional distortion latent space embedded with prior knowledge of teaching styles, calculates the conditional information entropy of the current candidate pair using a conditional probability distribution model, applies a dynamic penalty correction to the basic complementarity based on the conditional information entropy to obtain the time-fit score, and generates an initial candidate matching pool based on the time-fit score.

[0100] Furthermore, the coach profile features and student profile features of the current candidate pair are extracted from the bidirectional profile tensor. The coach profile features and student profile features are input into the directed distortion latent space mapping network to obtain a joint representation vector. Using the joint representation vector as the query condition, the matching phase coherence coefficient probability density curve is retrieved in the conditional probability distribution model. The phase coherence coefficient probability density curve is subjected to definite integral operation within a preset effective teaching interval to obtain the cumulative probability value. The negative natural logarithm of the cumulative probability value is determined as the conditional information entropy.

[0101] The calculation process for conditional information entropy is as follows:

[0102] After concatenating student and coach profiles, a joint representation is obtained via a latent spatial mapping network. Based on post-class satisfaction data corresponding to identical or adjacent joint representations in historical matching samples, an adaptive bandwidth kernel density estimation method is used to obtain the probability density function of the interaction result. Taking the natural logarithm of this probability density function, the integral expectation is approximated on its support set using Monte Carlo sampling. Taking the negative value yields the conditional information entropy, expressed as:

[0103] ;

[0104] In the formula, For conditional information entropy, For joint representation vector, For the interaction result variable, Let be the conditional probability density function. The number of sampling points in Monte Carlo. For the first One sampling point.

[0105] The training process of the Directional Twisted Latent Space Mapping Network includes collecting positive and negative sample pairs with labeled matching results; inputting the coach profile features and student profile features of the positive and negative sample pairs into the mapping network to be trained to obtain the joint representation vector of positive samples and the joint representation vector of negative samples; constructing a contrastive loss function with the objective of maximizing the difference between the distance between the joint representation vectors of positive samples and the distance between the joint representation vectors of negative samples; using the contrastive loss function to perform gradient descent updates on the mapping network to be trained until the loss function converges, and determining the converged network parameters as the final parameters of the Directional Twisted Latent Space Mapping Network.

[0106] Set an upper limit and a lower limit for the information entropy threshold; when the conditional information entropy is less than the lower limit, the basic complementarity is directly determined as the timeliness fit score; when the conditional information entropy is greater than the upper limit, the basic complementarity is multiplied by the preset minimum retention coefficient and then determined as the timeliness fit score; when the conditional information entropy is between the lower and upper limits of the information entropy threshold, the timeliness fit score is calculated according to the linear decay formula.

[0107] The linear decay formula is:

[0108] ;

[0109] in, Indicates timeliness-appropriate score. Indicates the basic complementarity. Represents conditional information entropy. This represents the lower limit of the information entropy threshold. This represents the upper limit of the information entropy threshold. This indicates the preset minimum retention factor.

[0110] It should be noted that calculating the basic complementarity in the directional distortion potential space embedded with prior knowledge of teaching styles, and applying dynamic penalty correction to the complementarity in combination with conditional information entropy, can effectively suppress the risk of mismatch caused by the uncertainty of real-time interaction while preserving the potential for personalized matching. This makes the generated time-appropriate matching score reflect both the static profile fit and the dynamic interaction stability, thereby improving the quality and robustness of the initial candidate matching pool.

[0111] In particular, complementary computation embeds prior constraints on teaching styles into the latent space to prevent the model from learning invalid statistical shortcuts, and the asymmetric metric function explicitly characterizes the role inequality between the coach and the student; contrastive loss uses hard example mining to force the network to focus on subtle complementary signals, and conditional information entropy maps static profiles to dynamic interaction probability distributions, quantifying the correlation between matching potential and uncertainty; the minimum retention coefficient of linear decay is derived from the statistical analysis of historical mismatch deviations, ensuring that potentially valuable candidates are still retained in uncertain scenarios, so that dynamic penalty has both risk avoidance and opportunity retention functions.

[0112] The strategy reordering module is used to load the offline reinforcement learning policy network, use the skill growth rate per unit time as a delayed reward signal to output a reordering adjustment factor for candidate pairs in the initial candidate matching pool, and generate effect-oriented adaptation scores and a reordering candidate list based on the reordering adjustment factor and the time-effect adaptation score.

[0113] Furthermore, a historical sample set with the same combination of coach profile features and student profile features as the current candidate pair is extracted from historical matching records; the change in skill test scores and corresponding duration of all samples in the historical sample set within the preset observation period after matching are statistically analyzed, and the ratio of the change in skill test scores to the duration is calculated as the skill growth rate per unit time; the skill growth rate per unit time is input into the value function header of the offline reinforcement learning policy network, and the state value estimate is output; the state value estimate is normalized, and the normalized value is determined as the reordering adjustment factor.

[0114] The calculation process for the reordering adjustment factor is as follows:

[0115] The conservative Q-learning network predicts the value estimate of a given state-action pair. The maximum and minimum value estimates of all candidate actions in the current state are calculated. After performing min-max normalization on the current action value, it is multiplied by a preset upper limit coefficient to obtain a reordering adjustment factor limited to a reasonable range. The expression is:

[0116] ;

[0117] In the formula, This is the reordering adjustment factor. To conserve the state of the Q-learning network output Next action The value estimate, , They are respectively states The maximum and minimum values ​​of the estimated value of all candidate actions. This is a preset upper limit coefficient used to prevent long-term signals from overshadowing the importance of real-time adaptation.

[0118] It should be noted that using the skill growth rate per unit time as a delayed reward signal to drive the reordering adjustment factor of the offline reinforcement learning strategy network output can effectively integrate long-term teaching effect feedback into the immediate matching decision-making process. This overcomes the short-sighted bias caused by traditional matching methods relying solely on short-term indicators, and enables the reordered candidate list to have an optimization orientation towards long-term learning outcomes while taking into account the current fit.

[0119] In particular, the value function output reflects long-term skill growth expectations rather than single, immediate gains, fundamentally correcting short-sighted biases; the historical sample set limits the time period to control for external factor drift; and the reordering adjustment factor is restricted to a reasonable range to prevent long-term signals from overshadowing the importance of real-time load matching, making effect-oriented adaptation a balance between short-term adaptation and long-term effectiveness.

[0120] The game-solving module is used to obtain the coach's real-time remaining capacity and the student's priority weight. It uses the reordered candidate list as preference input to execute an improved bilateral matching algorithm and outputs a stable matching solution through an iterative proposal process with capacity constraints and priority awareness.

[0121] Furthermore, initialize the remaining capacity vector on the coach side and the priority weight vector on the student side, sort each candidate pair in the reordering candidate list from high to low according to the effect-oriented adaptation score to generate an ordered preference sequence; perform descending sorting on all students to be matched according to the student priority weight vector to obtain a proposal order queue, and select the current highest priority student as the active proposer from the proposal order queue.

[0122] The current initiator sends a matching request to the target coach ranked first according to the ordered preference sequence. After receiving the matching request, the target coach checks whether its remaining capacity is greater than zero.

[0123] If the target coach has a remaining capacity greater than zero and does not currently hold any pending requests with higher fit scores, then the current match request is temporarily accepted and its remaining capacity is reduced by one. At the same time, the performance-oriented fit score corresponding to the current match request is recorded as a temporary retention threshold. If the target coach has a remaining capacity equal to zero or already holds any pending requests with higher fit scores, then the performance-oriented fit score of the current match request is compared with the temporary retention threshold. When the performance-oriented fit score of the current match request is higher than the temporary retention threshold, the original temporary match is released and the corresponding capacity is restored. Instead, the current match request is accepted and the temporary retention threshold is updated. Otherwise, the current match request is directly rejected.

[0124] The rejected proposer marks the current target coach as unavailable in the ordered preference sequence and resends the matching request to the next target coach until all trainees have been matched or are marked as having no available candidates.

[0125] After traversing all trainees in the proposal order queue, collect all matching pairs in the temporary acceptance state to form a preliminary matching set. Perform a stability check on the preliminary matching set to check if there exists any pair of unmatched coaches and trainees whose bidirectional effect-oriented adaptation scores are higher than the adaptation scores of their respective currently matched objects. If the stability check finds a blocking pair, add the trainees in the blocking pair to the secondary proposal queue and increase their priority weight. Re-execute the above proposal process until no new blocking pairs are generated. Output the final matching set without blocking pairs as the stable matching scheme.

[0126] It should be noted that by introducing student priority weights and real-time remaining capacity constraints for coaches, and combining effect-oriented adaptation with temporary retention and blocking elimination mechanisms during the iterative proposal process, resource limitations and user differences can be fully respected while ensuring the stability of matching results. This avoids high-priority students being inefficiently occupied or coaches being overloaded, thereby achieving an organic unity of fairness, efficiency, and system feasibility.

[0127] Specifically, in the game theory solution, the priority of trainees is exponentially decayed with the historical waiting time to prevent priority accumulation from causing system oscillation. A temporary threshold is reserved to break the rigid lock and allow for better matching and replacement. The blocking check simultaneously verifies that the fit score of both parties after replacement is not lower than the minimum threshold to avoid sacrificing basic service quality. The priority increase of the secondary proposal is capped to prevent individual trainees from repeatedly triggering rematching. The iteration process is terminated early when the global fit score does not improve significantly. The confidence score attached to the final matching scheme integrates multiple indicators to provide fine-grained feedback for parameter updates.

[0128] The parameter update module is used to collect real-time satisfaction data, process behavior data, and staged skill growth data corresponding to stable matching schemes, and to adjust the topological constraint parameters and conditional probability distribution model in the complementary calculation module, the offline reinforcement learning policy network weights in the policy rearrangement module, and the capacity decay parameters in the game solving module.

[0129] Furthermore, the comprehensive effect score of this match is calculated based on real-time satisfaction data and phased skill growth data. When the comprehensive effect score is lower than the preset update trigger threshold, the conditional information entropy and basic complementarity corresponding to this match are extracted as abnormal samples. The adaptive bandwidth parameters of the conditional probability distribution model are incrementally corrected using abnormal samples. The upper limit and lower limit of the information entropy threshold are adjusted according to the degree of deviation between the conditional information entropy and the basic complementarity in the abnormal samples. The corrected adaptive bandwidth parameters and the adjusted upper and lower limits of the information entropy threshold are written back to the configuration storage area of ​​the complementarity calculation module.

[0130] It should be noted that by identifying abnormal matching samples based on multidimensional feedback data and incrementally correcting the probability distribution model parameters and dynamically adjusting the information entropy threshold boundary, the system can continuously calibrate its internal evaluation criteria based on actual operating results, forming a closed-loop adaptive mechanism from matching execution to effect verification to parameter optimization. This enhances the system's long-term evolution capability and anti-drift performance in non-stationary teaching environments.

[0131] Specifically, the comprehensive effect score in parameter updates adopts time-adaptive weighting and applies a credibility discount to abnormal feedback. Abnormal sample identification eliminates external interference to ensure that the correction targets the model's inherent biases. Adaptive bandwidth increment correction increases the probability density estimation of abnormal sample regions, enhancing future attention to similar situations. The information entropy threshold is adjusted according to teaching style groups to adapt to differences in sensitivity. Before writing back the corrected parameters, a safety boundary check is performed to prevent a single abnormal update from causing system instability, forming a protected online learning loop that allows the system to gradually converge to a matching logic that conforms to teaching principles.

[0132] In summary, this invention achieves accurate semantic alignment of multi-source heterogeneous data by constructing a bidirectional profile tensor with confidence weights. It utilizes cross-wavelet transform to extract phase coherence coefficients and combines conditional information entropy to dynamically penalize and correct the basic complementarity, effectively quantifying the real-time load matching status and uncertainty risk in the teaching interaction process. Simultaneously, it introduces an offline reinforcement learning strategy network to generate a reordering adjustment factor using the skill growth rate per unit time as a delayed reward signal, integrating long-term teaching effect feedback into immediate matching decisions. Then, an improved bilateral matching algorithm with capacity constraints and priority awareness is used to solve for a stable solution. Based on multi-dimensional feedback data, the topological constraint parameters, probability distribution model, and network weights are continuously iterated. Under the premise of ensuring the stability of the matching results, this invention improves the timeliness, effectiveness orientation, and adaptive evolution capability of the system in bidirectional matching between coaches and students.

[0133] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A coach-student intelligent matching system based on two-way user profiles, characterized in that: include: Image alignment module, load monitoring module, complementary calculation module, strategy rearrangement module, game solving module, and parameter update module; The portrait alignment module is used to acquire multi-source data from the coach's end and multi-source data from the student's end, and map the multi-source data from the coach's end and the multi-source data from the student's end into a bidirectional portrait tensor with confidence weights based on preset cross-modal semantic anchor points. The load monitoring module is used to collect real-time behavior time-series signals from the coach's end and real-time response time-series signals from the student's end, perform cross wavelet transform on the real-time behavior time-series signals from the coach's end and real-time response time-series signals from the student's end to extract phase coherence coefficients, and construct a conditional probability distribution model of the phase coherence coefficients based on historical effective teaching data. The complementary calculation module is used to receive bidirectional profile tensors and phase coherence coefficients, calculate the basic complementarity degree in the directional distortion potential space embedded with prior knowledge of teaching styles through an asymmetric metric function, calculate the conditional information entropy of the current candidate pair using a conditional probability distribution model, apply dynamic penalty correction to the basic complementarity degree based on the conditional information entropy to obtain the time-fit score, and generate an initial candidate matching pool based on the time-fit score. The strategy reordering module is used to load an offline reinforcement learning strategy network, use the skill growth rate per unit time as a delayed reward signal to output a reordering adjustment factor to the candidate pairs in the initial candidate matching pool, and generate an effect-oriented adaptation score and a reordering candidate list based on the reordering adjustment factor and the time-effect adaptation score. The game-solving module is used to obtain the coach's real-time remaining capacity and the student's priority weight, and to execute an improved bilateral matching algorithm with the reordered candidate list as the preference input. After an iterative proposal process with capacity constraints and priority awareness, a stable matching scheme is output. The parameter update module is used to collect real-time satisfaction data, process behavior data, and staged skill growth data corresponding to the stable matching scheme, and to adjust the topological constraint parameters and conditional probability distribution model in the complementary calculation module, the offline reinforcement learning policy network weights in the policy rearrangement module, and the capacity decay parameters in the game solving module.

2. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 1, characterized in that: The profile alignment module maps multi-source data from both the coach and student sides into bidirectional profile tensors with confidence weights. The specific steps are as follows: Modal classification processing was performed on the multi-source data from the coach's side and the multi-source data from the student's side to obtain a structured label set, an unstructured text corpus, and a continuous behavior time series. For structured tag sets, one-hot encoding is used to generate basic feature vectors; for unstructured text corpora, a pre-trained semantic encoder is used to extract semantic embedding vectors; and for continuous behavior time-series, a sliding window statistical method is used to generate behavior summary vectors. The basic feature vector, semantic embedding vector, and behavior summary vector are concatenated into the original portrait matrix according to a preset dimension order; Based on the preset cross-modal semantic anchor point system, the correlation strength between each dimension feature and the corresponding semantic anchor point in the original portrait matrix is ​​calculated, and the correlation strength is used as the confidence weight of the corresponding dimension. The confidence weights are multiplied element-wise with the corresponding dimensional features in the original portrait matrix to obtain a bidirectional portrait tensor.

3. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 2, characterized in that: The preset cross-modal semantic anchors include three types of benchmark semantic nodes: teaching style anchors, knowledge domain anchors, and interaction preference anchors. Calculate the association strength between each dimension feature in the original portrait matrix and the corresponding semantic anchor point, including determining the modality category to which each dimension feature belongs, and selecting the subdivided semantic sub-nodes under the benchmark semantic node corresponding to the modality category as the target anchor point; Calculate the cosine similarity between the dimensional feature vector and the target anchor vector in the normalized semantic space, and determine the cosine similarity value as the confidence weight.

4. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 3, characterized in that: The load monitoring module performs cross-wavelet transform on the real-time behavior time-series signal from the coach's end and the real-time response time-series signal from the student's end to extract the phase coherence coefficients. The specific steps are as follows: A continuous wavelet transform is performed on the real-time behavior time-series signal of the coach to obtain the wavelet power spectrum of the coach, and a continuous wavelet transform is performed on the real-time response time-series signal of the student to obtain the wavelet power spectrum of the student. The cross wavelet spectrum is obtained by performing complex conjugate multiplication on the wavelet power spectrum of the coach and the wavelet power spectrum of the student. A smoothed cross spectrum is obtained by performing Gaussian smoothing filtering on the cross wavelet spectrum in the time dimension; The cross-wavelet coherence value is calculated based on the smooth cross spectrum, the wavelet power spectrum at the coach's end, and the wavelet power spectrum at the student's end. The arithmetic mean of the cross wavelet coherence values ​​within the preset effective teaching frequency band is determined as the phase coherence coefficient.

5. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 4, characterized in that: The complementary computing module calculates the conditional information entropy of the current candidate pair using a conditional probability distribution model. The specific steps are as follows: The coach profile features and student profile features of the current candidate pair are extracted from the bidirectional profile tensor. The coach profile features and student profile features are then input into the directed distorted latent space mapping network to obtain the joint representation vector. Using the joint representation vector as the query condition, the matching phase coherence coefficient probability density curve is retrieved in the conditional probability distribution model. The cumulative probability value is obtained by performing a definite integral operation on the probability density curve of the phase coherence coefficient within a preset effective teaching interval; The negative natural logarithm of the cumulative probability value is defined as the conditional information entropy.

6. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 5, characterized in that: The complementary calculation module applies a dynamic penalty correction to the basic complementarity based on conditional information entropy to obtain the timeliness adaptation score. The specific steps are as follows: Set an upper limit and a lower limit for the information entropy threshold; When the conditional information entropy is less than the lower limit of the information entropy threshold, the basic complementarity is directly determined as the timeliness adaptation score. When the conditional information entropy is greater than the upper limit of the information entropy threshold, the basic complementarity is multiplied by the preset minimum retention coefficient and then determined as the timeliness adaptation score. When the conditional information entropy is between the lower limit and the upper limit of the information entropy threshold, the timeliness adaptation score is calculated according to the linear decay formula, which is: ; in, Indicates timeliness-appropriate score. Indicates the basic complementarity. Represents conditional information entropy. This represents the lower limit of the information entropy threshold. This represents the upper limit of the information entropy threshold. This indicates the preset minimum retention factor.

7. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 6, characterized in that: The strategy reordering module uses the skill growth rate per unit time as a delayed reward signal to output a reordering adjustment factor for candidate pairs in the initial candidate matching pool. The specific steps are as follows: Extract a set of historical samples that have the same combination of coach profile features and student profile features as the current candidate pair from historical matching records; The changes in skill test scores and corresponding durations of all samples in the historical sample set within the preset observation period after matching are statistically analyzed, and the ratio of the change in skill test scores to the duration is calculated as the skill growth rate per unit time. Input the skill growth rate per unit time into the value function header of the offline reinforcement learning policy network, and output the state value estimate; The state value estimate is normalized, and the normalized value is used as the reordering adjustment factor.

8. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 4, characterized in that: The load monitoring module constructs a conditional probability distribution model of the phase coherence coefficient based on historical valid teaching data. The specific steps are as follows: Valid teaching samples with post-lesson satisfaction scores higher than the preset passing score were selected from historical teaching records; Extract the phase coherence coefficient time series corresponding to the valid teaching samples; Adaptive bandwidth kernel density estimation is performed on the phase coherence coefficient time series to obtain the original probability density function; The formula for calculating the adaptive bandwidth kernel density estimation is as follows: ; in, This indicates that the original probability density function is in Density value at that location, This represents the total number of valid teaching samples. Indicates the first The phase coherence coefficient of a valid teaching sample. Represents the Gaussian kernel function. Indicates the first The adaptive bandwidth of each sample; the original probability density function is stored as a conditional probability distribution model.

9. The intelligent matching system for coaches and students based on two-way user profiles as described in claim 5, characterized in that: The training process of the directed warp latent spatial mapping network includes: Collect positive and negative sample pairs with labeled matching results; The coach profile features and student profile features of positive sample pairs and negative sample pairs are respectively input into the mapping network to be trained to obtain the joint representation vector of positive samples and the joint representation vector of negative samples. A contrastive loss function is constructed with the objective of maximizing the difference between the joint representation vectors of positive samples and the joint representation vectors of negative samples. The contrastive loss function is used to perform gradient descent updates on the mapping network to be trained until the loss function converges. The converged network parameters are then determined as the final parameters of the directional warped latent space mapping network.

10. The intelligent matching system for coaches and students based on bidirectional user profiles as described in claim 6, characterized in that: The parameter update module adjusts the topological constraint parameters and conditional probability distribution model in the complementary calculation module. The specific steps are as follows: The overall effectiveness score for this match is calculated based on real-time satisfaction data and phased skill growth data. When the overall performance score is lower than the preset update trigger threshold, the conditional information entropy and basic complementarity corresponding to this match are extracted as abnormal samples. Incremental correction of the adaptive bandwidth parameter of the conditional probability distribution model is performed using outlier samples; The upper and lower limits of the information entropy threshold are adjusted based on the degree of deviation between the conditional information entropy and the basic complementarity in abnormal samples. Write the corrected adaptive bandwidth parameters and the adjusted upper and lower limits of the information entropy threshold back to the configuration storage area of ​​the complementary computing module.

Images