Multi-channel-oriented intelligent matching management method and system for institutional course order

Abstract

The application discloses a multi-channel-oriented institution course order intelligent matching management method and system, belongs to the technical field of data processing, and comprises the following steps: collecting original behavior data, calculating a user intention vector, and clustering to obtain a cluster set; calculating the center vector of each cluster; traversing all courses for the cluster to calculate a comprehensive adaptation score; selecting a preset number of courses from high to low according to the comprehensive adaptation score to obtain a course candidate set and a corresponding comprehensive adaptation score list; calculating the overall matching value of the cluster and the course candidate set through a weighted matching function, calculating the matching error through MSE calculation on the distributed weight and the user registration behavior; and minimizing and optimizing the matching error to obtain the best parameter set. The scheme solves the problems of intention expression deviation, insufficient matching efficiency and resource allocation rigidity in the existing system by proposing a unified modeling structure of multi-channel user behavior sequence, introducing an order cluster construction and center vector expression mechanism.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method and system for intelligent matching and management of institutional course orders across multiple channels. Background Technology

[0002] As educational institutions continue to develop multi-channel enrollment and integrated online-offline operations, the sources of course orders have become increasingly complex and diverse. Users may submit course reservations through WeChat official accounts, mini-programs, telephone customer service, official websites, or enterprise cooperation portals. Each channel exhibits significant structural and temporal differences in user behavior, directly impacting the accuracy and consistency of their interest expression. Currently, most systems employ generic user profiles or static recommendation models, neglecting the moderating role of channel context in intent expression. This makes it difficult to capture short-term fluctuations in user interests across different scenarios, resulting in unstable recommendation quality.

[0003] On the other hand, faced with the contradiction between the increasing number of concurrent order requests and the limited course resources, the traditional strategy of recommending courses based on individual users is gradually showing its bottlenecks in terms of computational efficiency and resource utilization. This is especially true when users flock to register for popular courses, as the system lacks overall allocation capabilities. Furthermore, current mainstream systems lack cluster-level order structure modeling and feedback mechanisms. Course matching strategies often rely on static rules or manual intervention, leading to frequent overload of popular courses and long-term idleness of less popular courses. The lack of a structured understanding of the overall intent behind orders and the inability to automatically adjust severely restricts the order-handling capacity and course supply efficiency of educational institutions.

[0004] To address these issues, we propose a method and system for intelligent matching and management of institutional course orders across multiple channels. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of unstable recommendation quality and unreasonable allocation in the prior art, and to propose an intelligent matching management method and system for institutional course orders for multiple channels.

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

[0007] A multi-channel intelligent matching management method for institutional course orders, including:

[0008] S1: Collect raw behavior data, convert the raw behavior data into fixed-dimensional behavior vectors, combine the behavior vectors with behavior type embedding and time information to form a final behavior vector, arrange different final behavior vectors in time series to obtain a behavior matrix, process the behavior matrix through temporal convolution to obtain a behavior feature sequence, and process the behavior feature sequence through a multi-head attention module to output a fused feature sequence.

[0009] The fused feature sequence is compressed in the time dimension using max pooling to obtain the strongest signal. The strongest signal is then mapped to the user intent space through a fully connected layer network; finally, the user intent vector is output.

[0010] S2: Calculate the weighted similarity of the user intent vector, the weighted similarity being calculated based on the course preferences of different users;

[0011] After extracting the top K feature vectors from the weighted similarity using an adaptive spectral clustering method, a low-dimensional representation space is constructed, and KMeans clustering is performed in the low-dimensional representation space to obtain a cluster set.

[0012] Calculate the center vector of each cluster in the set of clusters;

[0013] S3: Input a structured course resource table, which includes course topic embeddings and course remaining capacity ratio;

[0014] For the cluster, all courses are traversed, and the comprehensive fit score is calculated.

[0015] Based on the comprehensive adaptation score, a preset number of courses are selected from high to low to obtain a set of course candidates and a corresponding list of comprehensive adaptation scores;

[0016] S4: Calculate the overall matching value between the cluster and the course candidate set using a weighted matching function. The weighted matching function introduces distributed weights, which are calculated by the average interest embedding of all users in the cluster and the topic vector of the candidate courses. The distributed weights are the matching weights of the cluster for each course in the course candidate set.

[0017] S5: The matching error is obtained by performing MSE calculation on the distributed weights and user registration behavior; the optimal parameter set is obtained by minimizing the matching error.

[0018] Preferably, the raw behavioral data is collected from log systems of different channels, and then the time format is unified, anomalies are removed, and the data is sorted.

[0019] Preferably, the weighted similarity calculation incorporates a time offset term and a course clustering suppression term. The time offset term is calculated by the difference in behavior time among different users within a preset time window; the course clustering suppression term is calculated based on the cosine of the angle between the vectors of different users browsing courses.

[0020] Preferably, the value of K is dynamically calculated by combining clustering evaluation indicators and business thresholds.

[0021] Preferably, the overall fit score is calculated by the cosine similarity between the cluster topic and the course topic and the ratio of the remaining capacity of the course; the cluster topic is obtained by the center vector of the cluster.

[0022] Preferably, a capacity penalty coefficient is also introduced into the comprehensive matching score calculation to prioritize courses with remaining slots among highly similar courses.

[0023] Preferably, user registration behavior is collected from real registration data, obtained from the registration management platform logs.

[0024] A smart matching management system for institutional course orders across multiple channels, including:

[0025] An intent recognition module is configured to collect raw behavioral data, convert the raw behavioral data into fixed-dimensional behavioral vectors, combine the behavioral vectors with behavioral type embedding and temporal information to form a final behavioral vector, arrange different final behavioral vectors in a time sequence to obtain a behavioral matrix, process the behavioral matrix through temporal convolution to obtain a behavioral feature sequence, process the behavioral feature sequence through a multi-head attention module to output a fused feature sequence, compress the fused feature sequence in the temporal dimension through a max pooling operation to obtain the strongest signal, map the strongest signal to the user intent space through a fully connected layer network, and finally output the user intent vector.

[0026] A cluster construction module is configured to calculate the weighted similarity of user intent vectors, the weighted similarity being calculated based on different user course preferences; after extracting the top K feature vectors from the weighted similarity using an adaptive spectral clustering method, a low-dimensional representation space is constructed, and KMeans clustering is performed in the low-dimensional representation space to obtain a cluster set; the center vector of each cluster in the cluster set is calculated.

[0027] The course selection module takes a structured course resource table as input, which includes course topic embeddings and course remaining capacity ratios; it iterates through all courses in the cluster and calculates the comprehensive adaptation score; and selects a preset number of courses from high to low based on the comprehensive adaptation score to obtain a course candidate set and a corresponding comprehensive adaptation score list.

[0028] The order matching module is configured to calculate the overall matching value between the cluster and the course candidate set through a weighted matching function. The weighted matching function introduces distributed weights, which are calculated by the average interest embedding of all users in the cluster and the topic vector of the candidate courses. The distributed weights are the matching weights of the cluster for each course in the course candidate set.

[0029] The feedback adjustment module is configured to obtain the matching error by performing MSE calculation on the distributed weights and user registration behavior; and to minimize the matching error to obtain the optimal parameter set.

[0030] In summary, the technical effects and advantages of this invention are as follows: First, by proposing a unified modeling structure for multi-channel user behavior sequences and combining temporal convolution and cross-channel attention mechanisms, this invention constructs a user short-term interest vector with high semantic consistency, supporting the dynamic capture of intent expression. Second, by introducing an order cluster construction and center vector expression mechanism, it realizes the structural aggregation of orders in the semantic space, and constructs a course candidate pool through graph regular spectral clustering and semantic filtering of course resources, effectively improving matching efficiency and reducing computational redundancy. Finally, based on the structured matching distribution between clusters and courses, the system introduces a feedback learning mechanism to automatically optimize matching parameters through real user registration data, achieving proactive adjustment of course resources and cold start suppression capabilities. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method steps in this invention;

[0032] Figure 2 This is a schematic diagram of the system structure in this invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0034] like Figure 1 As shown, the intelligent matching and management method for institutional course orders across multiple channels includes:

[0035] S1: Collect raw behavior data, convert the raw behavior data into fixed-dimensional behavior vectors, combine the behavior vectors with behavior type embedding and time information to form a final behavior vector, arrange different final behavior vectors in time series to obtain a behavior matrix, process the behavior matrix through temporal convolution to obtain a behavior feature sequence, and process the behavior feature sequence through a multi-head attention module to output a fused feature sequence.

[0036] The fused feature sequence is compressed in the time dimension using max pooling to obtain the strongest signal. The strongest signal is then mapped to the user intent space through a fully connected layer network; finally, the user intent vector is output.

[0037] S2: Calculate the weighted similarity of the user intent vector, the weighted similarity being calculated based on the course preferences of different users;

[0038] After extracting the top K feature vectors from the weighted similarity using an adaptive spectral clustering method, a low-dimensional representation space is constructed, and KMeans clustering is performed in the low-dimensional representation space to obtain a cluster set.

[0039] Calculate the center vector of each cluster in the set of clusters;

[0040] S3: Input a structured course resource table, which includes course topic embeddings and course remaining capacity ratio;

[0041] For the cluster, all courses are traversed, and the comprehensive fit score is calculated.

[0042] Based on the comprehensive adaptation score, a preset number of courses are selected from high to low to obtain a set of course candidates and a corresponding list of comprehensive adaptation scores;

[0043] S4: Calculate the overall matching value between the cluster and the course candidate set using a weighted matching function. The weighted matching function introduces distributed weights, which are calculated by the average interest embedding of all users in the cluster and the topic vector of the candidate courses. The distributed weights are the matching weights of the cluster for each course in the course candidate set.

[0044] S5: The matching error is obtained by performing MSE calculation on the distributed weights and user registration behavior; the optimal parameter set is obtained by minimizing the matching error.

[0045] The specific steps are as follows:

[0046] Step 1: Multi-channel user behavior coding and intent embedding construction

[0047] Step one aims to address the challenge of user intent recognition arising from the heterogeneity of user behavior data across multiple channels. The goal is to unify the encoding of behavioral data from various channels, including WeChat Official Accounts, Mini Programs, telephone customer service, and the official website, generating vectors that accurately reflect the user's current course interests. User behavior data encompasses various types, such as browsing, clicking, inquiries, and order creation. The data is collected through the log systems of each channel, resulting in different formats and inconsistent time granularities. Taking WeChat Mini Programs as an example, the system continuously acquires user click and browsing records through a backend API interface. Telephone customer service behavior is obtained through call center recording transcription and structured annotation. The official website data is collected through webpage tracking technology. All channel data is first standardized in time format, anomalies are removed, and the data is sorted before entering the processing flow.

[0048] Each behavioral event is transformed into a fixed-dimensional behavioral vector. This vector is then combined with a 16-dimensional embedding of the behavioral type and location-encoded temporal information to form the final vector representation. Multiple behaviors are arranged in a time sequence to form a behavioral matrix. For each channel, this matrix is ​​fed as input into an independent temporal convolutional network module. This module consists of two one-dimensional convolutional layers with a kernel size of 3 and an output dimension of 64, specifically designed to capture the temporal dependencies and local patterns of behaviors. For example, the most recent 30 behavioral sequences from the WeChat channel are encoded into a feature sequence reflecting the dynamic changes in user interests through this module.

[0049] Next, the feature sequences from all channels, after temporal convolution processing, are input into a Multi-Head Attention (MHA) mechanism. The purpose of this module is to fuse information from various channels and automatically determine which channel's behavior best represents the user's intent in the current scenario. MHA uses four attention heads, each learning a cross-channel weight combination logic, ultimately outputting a unified-dimensional fused feature sequence. This process is as follows:

[0050] H c =TCN c (B c )

[0051] H fusion =MHA(H1,H2,…,H) C )

[0052] in, This is a sequence matrix of user behavior on channel c, where L = 30 represents the number of behaviors, and d b =32 is the embedding dimension after concatenating behavior type and time information; TCN c This indicates that the temporal convolutional network used in this channel consists of two one-dimensional convolutional layers with a kernel size of 3 and 64 output channels; H c It represents the output sequence of each channel after passing through TCN, with a dimension of L×64; MHA stands for Multi-Head Attention Mechanism, which integrates the sequence features of different channels and outputs a fused sequence H with a unified dimension. fusion The dimension is L×128.

[0053] For example, if a user frequently inquires about "high school senior evening self-study classes" via telephone, but does not exhibit any significant behavior on other channels, MHA will automatically assign higher weight to the output of the telephone behavior channel, making the final intent vector more accurately reflect the true enrollment intent.

[0054] Next, H will be processed through a max pooling operation. fusion Compression is applied along the time dimension, and the strongest signal is extracted as the overall representation of the behavior, forming a 128-dimensional vector h.pool The goal of pooling is to identify the most prominent patterns of interest in a sequence of behaviors.

[0055] Finally, h pool By passing through a fully connected network layer (FC layer), the vector is mapped to the user intent space to obtain the final user intent vector:

[0056] u intent =FC(Pool(H) fusion ))

[0057] Where Pool represents the max pooling operation; FC is a fully connected mapping network with an output dimension of 128; It is the final output user intent vector, representing the current user's course interest across all channels.

[0058] The intent vector mechanism built in this step effectively addresses the real-world problem of "inconsistent behavior of the same user across different channels" in the education and training business. It enables the system to accurately determine the user's recent enrollment intent and forms the foundation of the entire multi-channel institutional course order intelligent matching system.

[0059] Step 2: Construct order clusters based on intent vectors

[0060] The core objective of step two is to transform the user intent vector u output in step one into... intent While fully preserving its semantic expressive power, an order cluster with semantic consistency, clear structure, and business controllability is constructed. This is used for subsequent course matching and scheduling. Unlike general user interest clustering scenarios, the "intelligent matching management of institutional course orders" addressed by the patent has several important practical constraints: First, user behavior comes from multiple channels, and the expression of interests from different channels may have time lags or style differences, leading to asymmetry in intent distribution; second, there are short-term high-concurrency characteristics within the same time window, requiring clustering to have online update capabilities; third, the clustering results must meet certain business structure requirements, such as the number of orders in each cluster cannot be too small, otherwise it will increase course scheduling costs. Therefore, it is necessary to design a controllable, adaptive clustering method that is suitable for educational business.

[0061] The input is the set of user intent vectors output from step one. Each of them This vector represents the course preference representation generated by the i-th user based on their multi-channel behavior sequence, where N is the number of orders in the current window. This vector encodes behaviors such as clicks, browsing, and inquiries, and integrates channel weights and time-series information, resulting in a highly semantic, low-dimensional, dense representation.

[0062] Considering that the intention distribution in the patent scenario may have a non-spherical structure and non-uniform density distribution, we introduce a spectral clustering method with a sparse graph regularization term. The core innovation of this method is to add two patent scenario-specific improvement terms to the similarity matrix of traditional spectral clustering: (1) regularization of the intention "adjacency time difference" to reflect the difference in interest update rate between channels; (2) constructing a "course proximity inhibition term" through institutional strategy adjustment target, so that orders with overly concentrated interest distribution will not all enter the same cluster, preventing a certain type of popular courses from causing resource overload.

[0063] First, calculate the weighted similarity for each pair of user intent vectors:

[0064]

[0065] Where ||·||2 is the Euclidean distance, and σ is the temperature parameter;

[0066] Δt ij It is the difference in the most recent behavior time between users i and j (normalized to between 0 and 1), reflecting the similarity of user interest updates;

[0067] cos(θ ij ) is the cosine of the angle between the vectors of the courses last viewed by the two users, used to determine the semantic proximity of their target courses;

[0068] λ t and λ c These are two regularization coefficients that control the influence of the time offset term and the course aggregation suppression term, respectively, and are generally set to the range of 0.1 to 0.3.

[0069] The similarity design embodies three innovations: (1) asymmetric interest intensity adjustment to avoid misclustering of users who are "close in distance but lag in behavior"; (2) consideration of the similarity of course goal dimensions to avoid everyone recommending the same popular course; and (3) incorporating the time sensitivity of interest into the structural similarity calculation, which is a one-dimensional innovation that is rarely handled in current educational clustering systems.

[0070] After constructing the weighted graph, we use an adaptive spectral clustering method to extract the top K feature vectors from the graph, construct a low-dimensional representation space, and then perform KMeans clustering in this space to finally form a set of clusters. The K value is not statically set, but dynamically calculated by combining clustering evaluation metrics (such as the Davies-Bouldin Index) with business thresholds to ensure that each cluster achieves a balance between interest consistency and order quantity.

[0071] Calculate the center vector for each cluster:

[0072]

[0073] The center vector c i It is subsequently used in the candidate course selection stage as a "collective expression" of the cluster's intentions.

[0074] The output of this step is as follows:

[0075] The clustering results show that each cluster contains a group of user orders that are similar in the multidimensional semantic space;

[0076] C = {c1, ..., c} K}: The center vector of each cluster, which serves as the input for matching subsequent courses.

[0077] In summary, this step fully utilizes the high-quality intent vector from Step One, introducing two regularization factors—time difference and course proximity—to form a structured clustering result with multi-dimensional semantic judgment. This not only enhances the expressive power of the clustering structure but also aligns with educational institutions' management strategy of "avoiding overfitting to recommend popular courses." Furthermore, it possesses strong engineering practicality and controllability, serving as a core bridging step from perception to structural understanding in the entire patented solution.

[0078] Step 3: Filter the course candidate pool based on order cluster intent

[0079] The core task of step three is to analyze each order cluster generated in step two. Based on its central intention vector c i From the structured course resource library The AI ​​selected a pool of candidate courses that met the cluster's needs, were feasible for business operations, and had reasonable scheduling resources. The essence of this step is to achieve a high-quality initial screening of "user group needs → course resource response", which is the first step in the entire patent resource scheduling chain.

[0080] This step not only ensures that the selected courses are highly relevant to the cluster's intent in terms of thematic semantics, but also fully incorporates actual constraints of the teaching business, such as the course's capacity, the match between class time slots and the active time of ordering users, geographical accessibility, and teaching format preferences. More importantly, we hope that this step can demonstrate strategic intelligence, such as favoring low-load courses when resources are scarce, or prioritizing courses from high-value institutions, truly reflecting the "controllability" and "business affinity" of the matching system.

[0081] enter

[0082] Cluster center intention vector (Step 2 output represents the cluster’s thematic interest).

[0083] Structured Course Resource Table Only the following are retained:

[0084] Course Theme Embedding (Maintain the same source mapping method as in step one, and keep the units consistent);

[0085] Course remaining capacity ratio γ j ∈[0,1] (Real-time statistics from the database).

[0086] The system is a cluster Travel through all courses A concise and business-specific "penalized similarity" score is calculated:

[0087]

[0088] Cosine similarity between cluster themes and course themes, with a range of [-1,1].

[0089] γ j : Remaining capacity ratio of courses, range [0,1];

[0090] λ: Capacity penalty coefficient, empirically set to 0.2-0.3, used to prioritize courses with available spots among highly similar courses to avoid overloading popular courses;

[0091] s ij Overall fit score: The higher the value, the better the fit for the cluster.

[0092] After the score is calculated, the system presses s ij From high to low The top K courses (K is usually 5) are selected to form a candidate pool. The entire process depends only on c. i This avoids mixing of dimensions and redundancy of variables with two existing resource fields.

[0093] Output

[0094] Cluster The course candidate pool contains K high-scoring courses;

[0095] The corresponding score list is provided for use in the next step of fine-tuning.

[0096] By incorporating a capacity penalty into cosine similarity, this scheme achieves collaborative filtering of interest matching and resource availability without adding a large amount of new data: courses with more relevant themes and more remaining slots receive higher scores, effectively alleviating congestion of popular courses and increasing exposure opportunities for less popular courses. This lightweight design maintains variable consistency and dimensional uniformity while meeting the real-time operability and business reproducibility requirements of the patent, providing a high-quality, low-redundancy candidate set for subsequent fine-grained matching.

[0097] Step 4: Cluster-level matching and allocation between order clusters and course candidate pools

[0098] The core task of this step is to optimize the course candidate pool built in the previous step. For its order cluster The modeling of cluster-level matching relationships between clusters is used to explicitly express the overall matching tendency between a cluster and its candidate courses, rather than directly outputting the final assignment result.

[0099] enter

[0100] Cluster (Composed of multiple user orders, originating from step two);

[0101] Candidate Course Collection (From step three);

[0102] Matching score set S i ={s ij (From step three, a semantic-capacity comprehensive score);

[0103] To model clusters To determine the overall preference distribution relationship between the course and its candidate course set, we construct a weighted matching function with a regularization term, as follows:

[0104]

[0105] in:

[0106] α ij For clusters of courses r j The matching weights satisfy ∑α ij =1;

[0107] s ij It is the matching score (semantic + capacity combined score) provided in the previous step;

[0108] Regular terms Control the sparsity of weights to prevent excessive concentration on a single course;

[0109] hyperparameter λ c>0 controls the degree of balance, and can usually be taken as [0.01, 0.1].

[0110] The modeling function Essentially, it models the matching quality between clusters and the course candidate pool, and uses adjustable distributed weights α. ij This reflects the degree of preference for different courses. The system can be solved using gradient descent or closed-form solutions for α. i Solving this problem yields a stable cluster preference distribution.

[0111] Specifically, to improve model interpretability and embed actual order features, we adjusted α. ij Configure the structure as follows:

[0112]

[0113] in:

[0114] For clusters The average interest embedding of all users (obtained by aggregating the user interest vectors output by the model in step one);

[0115] It is a course r j The topic vector;

[0116] β is a semantic adjustment coefficient that enhances the modeling of the fit between cluster interests and course topics.

[0117] Example explanation: Let It includes 60 users, and the mean of their user interest embeddings is used to construct the data.

[0118] Assume that candidate courses r1, r2, and r3 each have a score s. i1 =0.72,s i2 =0.65,s i3 =0.49, at the same time Then α ij The result calculated using the softmax structure could be {0.30, 0.35, 0.35}, reflecting the cluster's true overall preference for different courses.

[0119] Output

[0120] α i ={α ij}: The weight distribution of candidate courses by clusters;

[0121] The overall matching value of the cluster-course pair can be used as input for the next step of scheduling, ranking, or reinforcement learning feedback.

[0122] Step 5: Feedback Data Collection and Recommendation Parameter Adjustment Mechanism

[0123] The core task of this step is to output the result α based on the cluster-course matching model constructed in step four. i and Design a learnable feedback modeling mechanism to express and update the recommendation preference parameters of the matching model. It is important to emphasize that this step is not for policy execution or post-processing, but rather to further model the structural relationship between user behavior feedback and matching recommendations. The goal is to form a feedback-driven recommendation parameter optimization model, providing a stable and structured feedback expression for subsequent long-term iterative learning, and ensuring that the entire system possesses a closed-loop capability of "model-recommendation-feedback".

[0124] The input must strictly originate from the previous step, specifically including:

[0125] α i ={α ij}: Cluster Course candidate pool Each course in the middle j The matching weights are modeled and output by the softmax structure in step four;

[0126] The overall matching strength of the cluster is represented by the formula in step four;

[0127] y i ={y ij}: Collected from real registration data, y ij =1 indicates that at least one user within the cluster ultimately enrolled in the course r. j y ij =0 Otherwise, this data can be obtained directly from the registration management platform logs;

[0128] θ: Used in the matching weight model to generate s ij With α ij The parameter set.

[0129] Based on this, instead of performing external execution operations, we introduce a feedback modeling function. This is used to measure the matching error between predicted recommendation preferences and actual registration behavior. This error can be defined as:

[0130]

[0131] Formula explanation:

[0132] α ij : For the model of course r j Recommendation preferences;

[0133] This represents the true registration distribution, and its vector is related to α. i Perform MSE calculation;

[0134] ∈ is a very small positive number, to prevent the denominator from being 0 when there is no registration behavior;

[0135] This structure is a fully learnable preference feedback modeling objective that does not involve specific operations but only describes "how user behavior guides parameter fine-tuning".

[0136] Based on this, we construct a generalizable recommendation parameter update structure and define the following optimization objective function:

[0137]

[0138] in:

[0139] θ represents the recommendation preference parameters to be optimized, such as semantic score weights and capacity weights.

[0140] ρ is the regularization coefficient;

[0141] The optimization objective is to minimize the total matching bias across all clusters, ultimately α i It more closely reflects the distribution of actual registration behavior.

[0142] The innovation of this scheme lies in: establishing a cluster recommendation distribution α i To user registration behavior y i The error modeling structure enables the recommender system to have self-supervised fine-tuning capabilities, which is a structure-level feedback expression method rarely seen in traditional educational recommender systems.

[0143] For example, for clusters The recommended distribution is {0.5, 0.3, 0.2}, but the actual number of users registering is {0, 1, 0}, so the feedback modeling error is concentrated in α. i2 With y i2 The deviation will increase the semantic weight of such courses or reduce the sparsity regularization term when optimizing θ, so as to achieve a more flexible response to cold start courses or short-term popular courses.

[0144] Output result:

[0145] Updated recommended parameter set θ (t+1) ;

[0146] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: First, by proposing a unified modeling structure for multi-channel user behavior sequences, and combining temporal convolution and cross-channel attention mechanisms, this invention constructs a user short-term interest vector with high semantic consistency, supporting the dynamic capture of intent expression. Second, by introducing an order cluster construction and center vector expression mechanism, it realizes the structural aggregation of orders in the semantic space, and constructs a course candidate pool through graph regular spectral clustering and semantic filtering of course resources, effectively improving matching efficiency and reducing computational redundancy. Finally, based on the structured matching distribution between clusters and courses, the system introduces a feedback learning mechanism, automatically optimizing matching parameters through real user registration data, and realizing the ability to actively adjust course resources and suppress cold starts.

[0147] This application also provides an intelligent matching management system for institutional course orders across multiple channels, such as... Figure 2 As shown, it includes:

[0148] An intent recognition module is configured to collect raw behavioral data, convert the raw behavioral data into fixed-dimensional behavioral vectors, combine the behavioral vectors with behavioral type embedding and temporal information to form a final behavioral vector, arrange different final behavioral vectors in a time sequence to obtain a behavioral matrix, process the behavioral matrix through temporal convolution to obtain a behavioral feature sequence, process the behavioral feature sequence through a multi-head attention module to output a fused feature sequence, compress the fused feature sequence in the temporal dimension through a max pooling operation to obtain the strongest signal, map the strongest signal to the user intent space through a fully connected layer network, and finally output the user intent vector.

[0149] A cluster construction module is configured to calculate the weighted similarity of user intent vectors, the weighted similarity being calculated based on different user course preferences; after extracting the top K feature vectors from the weighted similarity using an adaptive spectral clustering method, a low-dimensional representation space is constructed, and KMeans clustering is performed in the low-dimensional representation space to obtain a cluster set; the center vector of each cluster in the cluster set is calculated.

[0150] The course selection module takes a structured course resource table as input, which includes course topic embeddings and course remaining capacity ratios; it iterates through all courses in the cluster and calculates the comprehensive adaptation score; and selects a preset number of courses from high to low based on the comprehensive adaptation score to obtain a course candidate set and a corresponding comprehensive adaptation score list.

[0151] The order matching module is configured to calculate the overall matching value between the cluster and the course candidate set through a weighted matching function. The weighted matching function introduces distributed weights, which are calculated by the average interest embedding of all users in the cluster and the topic vector of the candidate courses. The distributed weights are the matching weights of the cluster for each course in the course candidate set.

[0152] The feedback adjustment module is configured to obtain the matching error by performing MSE calculation on the distributed weights and user registration behavior; and to minimize the matching error to obtain the optimal parameter set.

[0153] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent matching and management of institutional course orders across multiple channels, characterized in that: include: S1: Obtain raw behavioral data through log systems from multiple channels, after unifying the time format, removing anomalies, and sorting. The multiple channels include at least one of WeChat official accounts, mini programs, telephone customer service, and official websites. Convert the raw behavioral data into fixed-dimensional behavioral vectors. Combine the behavioral vectors with behavioral type embedding and time information to form a final behavioral vector. Arrange the different final behavioral vectors in time sequence to obtain a behavioral matrix. Process the behavioral matrix through temporal convolution to obtain a behavioral feature sequence. Input the behavioral feature sequences from each channel into a multi-head attention module for attention processing, and then fuse the multi-channel features to obtain a fused feature sequence. The fused feature sequence is compressed in the time dimension using max pooling to obtain the strongest signal. The strongest signal is then mapped to the user intent space through a fully connected layer network; finally, the user intent vector is output. S2: Calculate the weighted similarity of the user intent vector, the weighted similarity being calculated based on the course preferences of different users; After extracting the top K feature vectors from the weighted similarity using an adaptive spectral clustering method, a low-dimensional representation space is constructed, and KMeans clustering is performed in the low-dimensional representation space to obtain a cluster set. Calculate the center vector of each cluster in the set of clusters; S3: Input a structured course resource table, which includes course topic embeddings and course remaining capacity ratio; For the cluster, all courses are traversed, and the comprehensive fit score is calculated. Based on the comprehensive adaptation score, a preset number of courses are selected from high to low to obtain a set of course candidates and a corresponding list of comprehensive adaptation scores; S4: Calculate the overall matching value between the cluster and the course candidate set using a weighted matching function. The weighted matching function introduces distributed weights, which are calculated by the average interest embedding of all users in the cluster and the topic vector of the candidate courses. The distributed weights are the matching weights of the cluster for each course in the course candidate set. S5: The matching error is obtained by performing MSE calculation on the distributed weights and user registration behavior; the optimal parameter set is obtained by minimizing the matching error.

2. The intelligent matching and management method for multi-channel institutional course orders according to claim 1, characterized in that, The weighted similarity calculation incorporates a time offset term and a course clustering suppression term. The time offset term is calculated by the difference in behavior time among different users within a preset time window. The course clustering suppression term is calculated based on the cosine of the angle between the vectors of different users browsing courses.

3. The intelligent matching and management method for multi-channel institutional course orders according to claim 1, characterized in that, The value of K is dynamically calculated by combining clustering evaluation indicators and business thresholds.

4. The intelligent matching and management method for multi-channel institutional course orders according to claim 1, characterized in that, The comprehensive fit score is calculated by the cosine similarity between the cluster topic and the course topic and the ratio of the remaining capacity of the course; the cluster topic is obtained by the center vector of the cluster.

5. The intelligent matching and management method for multi-channel institutional course orders according to claim 4, characterized in that, The comprehensive adaptation score calculation also introduces a capacity penalty coefficient, which is used to prioritize courses with higher cosine similarity rankings and still available slots from the course candidate set corresponding to the cluster.

6. The intelligent matching and management method for multi-channel institutional course orders according to claim 1, characterized in that, The user registration behavior data is collected from real registration data and obtained from the registration management platform logs.

7. A smart matching management system for institutional course orders across multiple channels, characterized in that: include: An intent recognition module is configured to obtain raw behavioral data from log systems across multiple channels, formatting it in a unified time format, removing anomalies, and sorting the data. These channels include at least one of WeChat official accounts, mini-programs, telephone customer service, and official websites. The raw behavioral data is converted into fixed-dimensional behavioral vectors. These vectors are then combined with behavioral type embeddings and time information to form a final behavioral vector. Different final behavioral vectors are arranged in a time sequence to obtain a behavioral matrix. This matrix is ​​then processed by temporal convolution to obtain a behavioral feature sequence. The behavioral feature sequences from each channel are input into a multi-head attention module for attention processing, and then fused with the multi-channel features to obtain a fused feature sequence. This fused feature sequence is compressed in the time dimension using max pooling to obtain the strongest signal. This strongest signal is then mapped to the user intent space through a fully connected layer network. Finally, a user intent vector is output. A cluster construction module is configured to calculate the weighted similarity of user intent vectors, the weighted similarity being calculated based on different user course preferences; after extracting the top K feature vectors from the weighted similarity using an adaptive spectral clustering method, a low-dimensional representation space is constructed, and KMeans clustering is performed in the low-dimensional representation space to obtain a cluster set; the center vector of each cluster in the cluster set is calculated. The course selection module takes a structured course resource table as input, which includes course topic embeddings and course remaining capacity ratios; it iterates through all courses in the cluster and calculates the comprehensive adaptation score. Based on the comprehensive adaptation score, a preset number of courses are selected from high to low to obtain a set of course candidates and a corresponding list of comprehensive adaptation scores; The order matching module is configured to calculate the overall matching value between the cluster and the course candidate set through a weighted matching function. The weighted matching function introduces distributed weights, which are calculated by the average interest embedding of all users in the cluster and the topic vector of the candidate courses. The distributed weights are the matching weights of the cluster for each course in the course candidate set. The feedback adjustment module is configured to obtain the matching error by performing MSE calculation on the distributed weights and user registration behavior; and to minimize the matching error to obtain the optimal parameter set.

Images