A user course reporting intention prediction method based on a hypergraph neural network and related products

Abstract

This application relates to the field of smart education technology and provides a method and related products for predicting user course enrollment intentions based on hypergraph neural networks. The method includes: selecting K similar students based on multidimensional similarity of student static features to construct hyperedges, forming a student hypergraph structure; constraining the feature representations of students at the same sales stage in the student hypergraph structure to tend towards consistency based on JS divergence, outputting a group association representation; processing the student's multi-touchpoint interaction sequence using a Transformer encoder with dynamic weights, and adaptively adjusting it with a sequence mask to output a behavioral sequence representation; fusing the group association representation and the behavioral sequence representation; inputting the fused features into a hybrid expert network adapted to the sales stage to obtain higher-order fused features; and using a dynamic prediction tower to output the course enrollment intention prediction result based on the higher-order fused features. This application can achieve accurate and efficient prediction of education users' real-time course enrollment intentions throughout the entire process.

Description

Technical Field

[0001] This application relates to the field of smart education technology, and in particular to a method for predicting user course enrollment intentions based on hypergraph neural networks and related products. Background Technology

[0002] Predicting real-time course enrollment intentions among users in the education industry is the core of digital marketing for educational institutions. It directly affects enrollment costs, resource allocation, and revenue. Their customer decision-making is characterized by high average order value, long cycle, multiple touchpoints, and strong family decision-making, which distinguishes it from e-commerce or advertising scenarios. The technology in this field has gone through three stages of development: Early stage relied on rule engines and simple statistical methods, characterized by educational institutions relying on sales consultants' experience to judge intent or using simple rule engines (such as "more than three visits + trial lesson completion -> high intent") and statistical methods like the RFM model for tiered operations, unable to handle complex time dynamics and personalized characteristics; Mid-stage adopted machine learning and static profiling, using relevant models and collaborative filtering for intent evaluation, characterized by machine learning methods such as XGBoost / LightGBM for intent evaluation based on static features (grade, subject, family income, etc.) and collaborative filtering for tiered promotion based on similar students' course enrollment behavior; Current stage sees leading institutions introducing deep learning and sequence modeling, using relevant models to process multi-touchpoint interaction sequences to predict real-time intent probabilities, characterized by the introduction of deep learning methods such as LSTM / GRU models to learn user interaction sequences such as online consultations, telephone communication, and in-store visits, and Transformer time-series models to process multi-touchpoint marketing interaction sequences to predict real-time intent probabilities.

[0003] However, existing technologies have the following drawbacks: First, they ignore high-order relationships within student groups, and ordinary graph structures cannot capture many-to-many relationships related to the syndicate effect, resulting in low prediction accuracy for cold-start students. Second, fixed architectures cannot adapt to customers' phased course enrollment decisions, making it difficult to differentiate needs at different stages, dynamically select expert networks, and accurately identify students with zero interest. Third, they cannot handle the uncertainty of multiple touchpoints in educational marketing, and time-series models lack adaptive weighting mechanisms, leading to poor prediction robustness. In summary, existing technologies fail to address the multi-dimensional challenges of structure, time series, and task, necessitating a comprehensive solution for educational scenarios. Summary of the Invention

[0004] This application provides a method and related products for predicting user course enrollment intentions based on hypergraph neural networks. It can be applied to education scenarios and achieve accurate and efficient prediction of education users' course enrollment intentions in real time throughout the entire process based on high-order hypergraph relationships, adaptive uncertainty, and stage dynamic routing.

[0005] In one aspect, this application provides a method for predicting user course registration intention based on a hypergraph neural network, the method comprising:

[0006] Acquire static profile data and dynamic interaction behavior data of students, process the static profile data and the dynamic interaction behavior data respectively to obtain static features of students and multi-touchpoint interaction sequences of students;

[0007] Based on the multidimensional similarity of the student's static features, select K similar students, and construct hyperedges according to the nearest neighbor relationship between each student and its corresponding K similar students to form a student hypergraph structure.

[0008] Based on the JS divergence constraint, the feature representations of students at the same sales stage in the student hypergraph structure tend to be consistent, and the group association representation is output.

[0009] The Transformer encoder with dynamic weights is used to process the multi-touch interaction sequence of the trainees, and the sequence mask is used for adaptive adjustment to output the behavioral sequence representation.

[0010] The group association representation and the behavioral sequence representation are fused to obtain fused features;

[0011] The fusion features are input into a hybrid expert network adapted to the sales stage to obtain high-order fusion features adapted to the current sales stage.

[0012] Personalized network parameters are generated from the higher-order fusion features using a dynamic prediction tower. Based on these personalized network parameters, predictions are made, and the course registration intention prediction results are output.

[0013] On the other hand, this application provides a user course registration intention prediction system based on a hypergraph neural network, the system comprising:

[0014] The data acquisition module is used to obtain static profile data and dynamic interaction behavior data of trainees;

[0015] The feature engineering module is used to process the static profile data and the dynamic interaction behavior data respectively to obtain the student's static features and the student's multi-touchpoint interaction sequence.

[0016] The hypergraph construction module is used to select K similar students based on the multidimensional similarity of the student's static features, and construct hyperedges according to the nearest neighbor relationship between each student and the corresponding K similar students to form a student hypergraph structure.

[0017] The model training module includes a hypergraph learning submodule, an adaptive temporal coding submodule, a sales-stage adaptive hybrid expert submodule, and a dynamic prediction tower submodule; among which:

[0018] The hypergraph learning submodule is used to constrain the feature representations of students at the same sales stage in the student hypergraph structure to be consistent based on JS divergence, and output a group association representation.

[0019] The adaptive time coding submodule is used to process the student's multi-touch interaction sequence using a Transformer encoder with dynamic weights, and to adaptively adjust it in conjunction with a sequence mask to output a behavioral sequence representation.

[0020] The adaptive hybrid expert submodule for the sales stage is used to fuse the group association representation and the behavioral sequence representation to obtain fused features.

[0021] The dynamic prediction tower submodule is used to input the fused features into a hybrid expert network adapted to the sales stage to obtain a high-order fused feature adapted to the current sales stage.

[0022] The prediction service module is used to generate personalized network parameters for the higher-order fusion features using a dynamic prediction tower, perform predictions based on the personalized network parameters, and output the course registration intention prediction results.

[0023] In another aspect, this application also provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements any of the user course registration intention prediction methods based on hypergraph neural networks.

[0024] In another aspect, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the user course registration intention prediction methods based on hypergraph neural networks.

[0025] In another aspect, this application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the user course registration intention prediction method based on hypergraph neural networks described in the above aspects.

[0026] The method and related products for predicting user course enrollment intention based on hypergraph neural networks provided in this application obtain static profile data and dynamic interaction behavior data of students. The static profile data and dynamic interaction behavior data are processed separately to obtain static features and multi-touchpoint interaction sequences of students. Then, based on the multi-dimensional similarity of the static features, K similar students are selected. Hyperedges are constructed according to the nearest neighbor relationships between each student and its corresponding K similar students to form a student hypergraph structure. Based on JS divergence constraints, the feature representations of students at the same sales stage in the student hypergraph structure tend to be consistent, outputting a group association representation. Furthermore, a Transformer encoder with dynamic weights is used to process the student multi-touchpoint interaction sequences, and sequence masks are used for adaptive adjustment to output behavioral sequence representations. The group association representation and behavioral sequence representations are then fused to obtain fused features. These fused features are input into a hybrid expert network adapted to the current sales stage to obtain higher-order fused features adapted to the current sales stage. Finally, a dynamic prediction tower is used to generate personalized network parameters from the higher-order fused features, and prediction is performed based on these personalized network parameters to output the course enrollment intention prediction result. By constructing a multi-dimensional similarity hypergraph structure for students to achieve high-level information transmission within a group, and relying on adaptive weights and temporal masks to complete dynamic modeling of multi-touchpoint interaction sequences, combined with a hybrid expert reasoning mechanism adapted to all stages of the marketing process, and leveraging a dynamic prediction tower to accurately predict intent intensity and distinguish invalid leads, this approach comprehensively addresses industry pain points such as insufficient group association mining, poor adaptability to the decision-making stage, and weak handling of uncertainties in temporal interactions inherent in traditional intent prediction models. It can efficiently support educational institutions in digital precision marketing, rational allocation of enrollment resources, and refined lead management, enabling accurate and efficient prediction of educational users' real-time course enrollment intentions throughout the entire process. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the architecture of a user course registration intention prediction system based on a hypergraph neural network provided in an embodiment of this application;

[0028] Figure 2 This is a flowchart illustrating the steps of a method for predicting user course registration intention based on a hypergraph neural network, as provided in an embodiment of this application.

[0029] Figure 3 This is a schematic diagram of the closed-loop business application process in online education marketing provided in the embodiments of this application;

[0030] Figure 4 This is a schematic diagram illustrating the implementation process of predicting user course registration intention based on a hypergraph neural network, as provided in an embodiment of this application.

[0031] Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of this application;

[0032] Figure 6 This is a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] To facilitate understanding of the embodiments of this application by those skilled in the art, the technical terms that may be involved are explained as follows:

[0035] Hyper-Temporal Graph Neural Network (HT-GNN) refers to the network architecture proposed in this application, which integrates hypergraph structure learning, adaptive temporal coding, and dynamic task routing.

[0036] Peer effect: This refers to the phenomenon in educational settings where students with similar backgrounds (same grade, same subject, same region) exhibit similar decision-making patterns.

[0037] Mixture of Experts (MoE) refers to a neural network architecture consisting of multiple expert networks and gating mechanisms, where the gating weights are associated with the input data.

[0038] Dynamic Prediction Tower refers to a network whose parameters (weight matrix) are dynamically generated from the input data to create a prediction network, achieving one network per student.

[0039] Jensen-Shannon divergence (JS divergence) measures the similarity between two sets of feature probability distributions. In this embodiment, it brings the feature distributions of trainees at the same sales stage closer together, reducing their differences and achieving group feature alignment.

[0040] Classification Token (CLS): This token is usually placed at the beginning of the sequence. It aggregates the global information of the entire time series and is ultimately used as the global feature of the entire sequence for subsequent classification, feature summarization, and global semantic extraction.

[0041] One-hot encoding is a method that converts discrete categorical data into binary sparse vectors. If the total number of categories is N, an N-dimensional vector is generated, with only the corresponding category position set to 1 and all other positions set to 0. This method can digitize categorical features.

[0042] Scott's Normal Reference Rule: The Scott rule is a classic heuristic formula used in kernel density estimation and histograms to automatically calculate the optimal bandwidth / bin width. It assumes that the data is approximately normally distributed and aims to minimize the integral mean square error of the density estimate.

[0043] Random deactivation / dropout: During the model training phase, some neurons in the network can be temporarily deactivated with a probability of 0.2, preventing them from participating in forward and backward propagation, while the remaining neurons have a probability of 0.8 that they continue to function normally.

[0044] Existing technologies suffer from the drawback of neglecting high-order relationships within student groups, i.e., a lack of structural dimension. While graph neural network techniques exist, they typically employ ordinary graph structures. These structures can only model simple pairwise relationships between students. However, in educational settings, significant peer pressure exists. Students in the same grade, with similar weaknesses in certain subjects, or from the same region face similar academic pressures, leading to highly similar course selection decisions. Furthermore, parents with similar income levels and educational philosophies exhibit consistent price sensitivity and course preferences. Ordinary graph structures cannot capture such many-to-many relationships and lack explicit supervision mechanisms to ensure alignment between learning representations and business objectives, resulting in extremely low prediction accuracy for cold-start students.

[0045] Existing technologies also suffer from limitations due to their fixed architecture, which cannot adapt to multi-stage decision prediction. This is because of the lack of task dimensions, which manifests in various stages of education, such as telephone communication for course registration, in-store visits, trial lessons, and final contract signing. Using a fixed multi-task architecture leads to a phased evolution of decision-making, typically involving initial consultations and the sharing of the same feature extractors across tasks. This makes it difficult to distinguish between the different needs of "early lead nurturing" and "final push to close the deal." Furthermore, the expert routing mechanism is independent of the input, meaning it cannot dynamically select the most suitable expert network based on the student's current stage, and it lacks the ability to accurately identify students with zero interest, thus leading to a waste of sales resources.

[0046] Existing technologies still have limitations in handling the multi-touchpoint uncertainties of education marketing, namely the lack of a time dimension. This manifests as the use of time-series models based on the premise of stable interaction models, the lack of adaptive weighting mechanisms, and the inability to dynamically adjust weights based on the completeness and timeliness of sequence data. This leads to problems such as the prediction results being sensitive to strategy changes and having poor robustness. However, education marketing itself has significant intervention uncertainties, such as the dynamic adjustment of sales consultant follow-up strategies, the pulse-like intervention of marketing activities, and sudden changes in students' status.

[0047] In summary, existing technologies have failed to address the multi-dimensional challenges of customer relationship structure modeling, time series modeling, and task modeling, and there is an urgent need for comprehensive solutions tailored to educational scenarios.

[0048] This application's embodiments achieve high-level information transmission within a group by constructing a multi-dimensional similarity hypergraph structure for students. It utilizes adaptive weights and temporal masks to dynamically model multi-touchpoint interaction sequences, combines a hybrid expert reasoning mechanism adapted to all stages of the marketing process, and leverages a dynamic prediction tower to accurately predict intent intensity and distinguish invalid leads. This comprehensively addresses industry pain points in traditional intent prediction models, such as insufficient group association mining, poor adaptability at the decision-making stage, and weak handling of uncertainties in temporal interactions. It can efficiently support educational institutions' digital precision marketing, rational allocation of enrollment resources, and refined lead management, enabling accurate and efficient prediction of educational users' real-time course enrollment intentions throughout the entire process.

[0049] Reference Figure 1 This illustration shows a schematic diagram of the architecture of a user course enrollment intention prediction system based on a hypergraph neural network, as provided in an embodiment of this application. In practical applications, this system can be applied to the enrollment management system of any educational institution. The hardware can be a 4*NVIDIA RTX 3090 GPU server, and the software environment can be based on PyTorch 2.0 and the DGL framework. It is suitable for accurately targeting high-intent users for course enrollment, optimizing marketing resource allocation, and improving enrollment conversion rates in K-12 education, vocational education, and extracurricular education scenarios. This embodiment of the application does not impose any limitations on this.

[0050] like Figure 1 As shown, the system may include a data acquisition module 11, a feature engineering module 12, a hypergraph construction module 13, a model training module 14, and a prediction service module 15. The model training module 14, based on the internal technical strategies of model construction, may include a hypergraph learning submodule 141, an adaptive temporal coding submodule 142, a sales-stage adaptive hybrid expert submodule 143, and a dynamic prediction tower submodule 144.

[0051] Optionally, the data acquisition module 11 can be used to collect static profiles of students, multi-touchpoint interaction sequences, sales stages, and course enrollment intention tags; the feature engineering module 12 can be used to implement categorical feature embedding (such as grade, subject, educational philosophy, etc.), numerical feature discretization (such as total learning motivation score, time interval, etc.), and temporal feature binning; the hypergraph construction module 13 can be used to build a multi-dimensional similarity hypergraph structure for students to realize the transmission of high-order information in the group; the hypergraph learning submodule 141 can be used to address the problem of modeling the structure of student groups by replacing traditional neural networks with hypergraph neural networks; the adaptive temporal coding submodule 142 can be used to design a Transformer encoder with dynamic weights to address the uncertainty of multi-touchpoints in education marketing; the sales stage adaptive hybrid expert submodule 143 can be used to address the problem of different strategies required for different sales stages by equipping each sales stage with a specific expert using the MOE architecture; and the dynamic prediction tower submodule 144 can be used to address the severe long-tail distribution in actual education customer data by dynamically generating prediction network parameters to achieve personalized prediction for each student.

[0052] By employing the above strategies, the embodiments of this application can improve the accuracy of cold start clue prediction by utilizing information from similar learner groups, ensure the consistency between representation learning and business objectives through explicit supervision, adaptively and dynamically adjust model weight parameters based on the completeness and timeliness of the interaction sequence, and select the optimal feature combination and expert system based on the learner's stage.

[0053] In some embodiments of this application, the data acquisition module 11 can be used to acquire static profile data and dynamic interaction behavior data of trainees.

[0054] Feature engineering module 12 can be used to process static profile data and dynamic interactive behavior data separately to obtain student static features and student multi-touchpoint interaction sequences.

[0055] The hypergraph construction module 13 can be used to select K similar students based on the multidimensional similarity of the students' static features, and construct hyperedges according to the nearest neighbor relationship between each student and the corresponding K similar students to form a student hypergraph structure.

[0056] The Hypergraph Learning Submodule 141 can be used to ensure that the feature representations of students at the same sales stage in the student hypergraph structure tend to be consistent based on JS divergence constraints, and output a group association representation.

[0057] The adaptive time coding submodule 142 can be used to process the multi-touch interaction sequence of trainees using a Transformer encoder with dynamic weights, and adaptively adjust it in combination with a sequence mask to output a behavioral sequence representation.

[0058] The sales phase adaptive hybrid expert submodule 143 can be used to fuse group association representations and behavioral sequence representations to obtain fused features.

[0059] The dynamic prediction tower module 144 can be used to input fused features into a hybrid expert network adapted to the sales stage to obtain high-order fused features adapted to the current sales stage.

[0060] The prediction service module 15 can be used to generate personalized network parameters for high-order fusion features using dynamic prediction towers, make predictions based on personalized network parameters, and output the prediction results of course registration intention.

[0061] In this embodiment of the application, the above modules can jointly support system capabilities, and customer acquisition growth in the education scenario can be achieved based on the above algorithm architecture.

[0062] Reference Figure 2 The diagram illustrates a flowchart of a method for predicting user course registration intention based on a hypergraph neural network, as provided in an embodiment of this application. Specifically, it may include the following steps:

[0063] Step S201: Obtain the static profile data and dynamic interaction behavior data of the student, process the static profile data and dynamic interaction behavior data respectively to obtain the student's static features and the student's multi-touchpoint interaction sequence.

[0064] In some embodiments of this application, a data acquisition module can be used to collect static profile data and dynamic interaction behavior data of students, so that the collected data can serve as the basis for predicting users' course enrollment intentions.

[0065] Optionally, the static profile data of trainees consists of fixed attribute data, which may include fixed information across all dimensions such as grade / subject, learning motivation assessment, family background, and sales stage, and can comprehensively depict the basic profile of users.

[0066] For example, student static data can be as follows:

[0067] {

[0068] "Grade": "S11", / / Enumeration values: P1-P6, J1-J3, S1-S3, Adult

[0069] "Strengths in Subjects": "History", / / Enumerated values: Chinese, Mathematics, English, Physics, Chemistry, Biology, History, Geography, Politics

[0070] "Weak subject": "Geography", / / Same as above

[0071] "Have you enrolled in other institutions?": "No", / / Yes / No

[0072] "Education Level": "Associate Degree", / / Junior High School, High School, Associate Degree, Bachelor's Degree, Master's Degree and above (Parents)

[0073] "Household Income Level": "10,001-20,000 yuan", / / Range Code

[0074] "Decision Weight": "Key Decision Maker", / / Key Decision Maker, Participating Decision Maker, Not Participating Decision Maker

[0075] "Educational Philosophy": "Quality-Oriented", / / Exam-Oriented, Quality-Oriented, Balanced, Undefined

[0076] "Key Focus Areas": "Class Time", / / Faculty, Class Time, Price, Reputation, Results

[0077] Price Sensitivity: Low, / / High, Medium, Low

[0078] "Academic Performance": "Poor", / / Excellent, Good, Average, Poor

[0079] "College Entrance Goals": "College Entrance Examination Goal Decomposition and Path Planning", / / Text Encoding

[0080] "Sales Stage": "Visit but No Contract Signed", / / New Leads, Phone Communication, Visit but No Contract Signed, Trial Lesson but No Conversion, Quotation but No Contract Signed, Contract Signed, Lost

[0081] / / Learning motivation assessment (scale score 1-5)

[0082] "Learning Motivation_Learning Goals": 4,

[0083] "Learning Motivation / Interest": 3,

[0084] "Learning Motivation_Learning Perseverance": 4,

[0085] "Learning Motivation_Learning Confidence": 4,

[0086] "Learning Motivation_Learning Emotions": 3,

[0087] "Learning Motivation - Social Support (Family Support)": 4,

[0088] "Total Learning Motivation Score": 22, / / Sum or weighted average

[0089] / / Learning ability assessment (scale score 1-5)

[0090] "Learning Ability - Attention": 5,

[0091] "Learning Ability / Memory": 5

[0092] "Learning Ability / Analytical Ability": 4

[0093] "Learning Ability_Creativity": 5,

[0094] "Learning ability - Introspection": 5,

[0095] "Total Learning Ability Score": 24

[0096] / / Assessment of learning habits (scale score 1-5)

[0097] "Study Habits - Self-Discipline": 3,

[0098] "Study Habits - Efficiency": 4

[0099] "Study Habits - Consistency": 4

[0100] "Study Habits_Behavioral Correctness": 3,

[0101] "Study Habits - Adjustment": 3,

[0102] Total Score for Study Habits: 17

[0103] }

[0104] In some embodiments of this application, static feature standardization processing can be performed on static portrait data. Specifically, such as... Figure 3 As shown, the feature engineering module can be used to embed categorical features and discretize numerical features in static profile data. Specifically, categorical features (such as grade, subject, and educational philosophy) in static profile data are One-Hot encoded and then embedded, while numerical features in static profile data are discretized according to preset rules, such as the Scott rule, and then uniformly embedded and mapped. This completes the standardization of structured data and yields a unified representation of student static features.

[0105] Dynamic interactive behavior data can be represented by a single complete interactive behavior feature group as an independent interactive sequence. Action frequency can represent the actual number of interactions for the corresponding marketing behavior, with a value of 0, 1, or greater than 1. Marketing behaviors can include, for example, telephone follow-ups, assessment pushes, course trials, and intention surveys. The previous intention serves as the initial baseline value for the sequence, primarily used to construct incomplete interactive sequence samples and improve data diversity. Simultaneously, auxiliary time-series information such as interaction duration, interval duration, and academic cycle are statistically analyzed. It should be noted that the student action interaction sequence is sequence data, and each element in the sequence data must have an initial intention. The previous intention can serve as a label when constructing incomplete sequences for data augmentation. This application's embodiments do not impose limitations on this.

[0106] In practical applications, various interaction features recorded during each interaction can be obtained from dynamic interactive behavior data. These interaction feature groups can be used as independent interaction sequences to obtain the student's multi-touchpoint interaction sequence.

[0107] For example, a dynamic interaction sequence for trainees can be shown below:

[0108] {

[0109] Total number of interactions: 1

[0110] "Hours since first interaction": 0.0,

[0111] "Hours since last interaction": 0.0,

[0112] Average interaction interval: 0.0,

[0113] "This action": "Telephone follow-up", / / Enumeration: event invitation, assessment push, telephone follow-up, private domain chat, course trial, material distribution, in-store visit, price negotiation

[0114] "Action Frequency_Event Invitation": 0, / / Cumulative Number of Invitations

[0115] "Action Frequency_Assessment Push": 0,

[0116] "Action Frequency_Telephone Follow-up": 1,

[0117] "Action Frequency_Private Domain Chat": 0,

[0118] "Movement Frequency_Course Preview": 0,

[0119] "Action Frequency_Data Distribution": 0,

[0120] "Previous Intention": 0.37 / / The model's prediction from the previous time step, initially set to 0.5

[0121] },

[0122] {

[0123] Total number of interactions: 2

[0124] "Hours since first interaction": 48.0

[0125] "Hours since last interaction": 48.0

[0126] Average interaction interval: 48.0

[0127] "This action": "Evaluation push",

[0128] "Action Frequency_Event Invitation": 0,

[0129] "Action Frequency_Assessment Push": 1,

[0130] "Action Frequency_Telephone Follow-up": 1,

[0131] "Action Frequency_Private Domain Chat": 0,

[0132] "Movement Frequency_Course Preview": 0,

[0133] "Action Frequency_Data Distribution": 0,

[0134] "Previous Intention": 0.42

[0135] }

[0136] / / ... Sequence lengths vary, averaging 5-8 steps, with a maximum of 10 steps.

[0137] In some embodiments of this application, the multi-dimensional business data may further include tag data, meaning that tag data can also be collected through a data acquisition module. Specifically, the tag data can be derived based on static profile data and dynamic interaction behavior data. Optionally, the tag data may include business tags such as contract category tags, continuous values ​​of intent intensity, contract amount, and conversion cycle.

[0138] For example, the tag data can be as follows:

[0139] {

[0140] "labels": [0], / / 0 = not signed (within 30 days), 1 = signed

[0141] “Intent Strength”: 0.0, / / 0-1 continuous value, marked by the sales consultant or calculated from behavioral data (e.g., trial lesson completed + quote accepted = 0.8)

[0142] "Contract Amount": 0, / / Actual payment amount, 0 indicates no contract signed.

[0143] "Contracted Subjects": null, / / Enumeration or null

[0144] "Conversion Cycle": null / / Number of days or null

[0145] }

[0146] In this embodiment, the label data can serve as the sole supervision signal for model training, used to calculate the loss function during the process, and to support model convergence and early stopping determination.

[0147] Step S202: Select K similar students based on the multidimensional similarity of the students' static features, and construct hyperedges according to the nearest neighbor relationship between each student and its corresponding K similar students to form a student hypergraph structure.

[0148] The student hypergraph structure refers to grouping a group of similar students into a superedge / supernode, batch aggregating students of the same type, and constructing the overall student association structure.

[0149] In some embodiments of this application, a hypergraph construction module can be used to match multidimensional similarity of trainees in real time, and construct nearest-neighbor hyperedges by combining business attributes and spatial distance, thereby quickly mining high-order relationships among trainees in the same circle. Specifically, such as Figure 4 As shown, multiple static attributes of students can be utilized, such as extracting all students' inherent static features to create multi-dimensional feature vectors. Then, a distance formula (Euclidean / cosine, etc.) is used to calculate similarity. For each student, the top K most similar students are selected as neighbors to establish connections, thus constructing hyperedges. A large number of nearest neighbor relationships are interwoven and aggregated to ultimately construct a student hypergraph structure.

[0150] For example, we can abandon the single distance calculation method, integrate Euclidean distance and educational business similarity to complete student matching, and construct hyperedges through the K=10 nearest neighbor rule to form an education-specific student hypergraph.

[0151] Step S203: Based on the JS divergence constraint, the feature representations of students at the same sales stage in the student hypergraph structure tend to be consistent, and the group association representation is output.

[0152] In some embodiments of this application, pre-trained JS divergence constraints can be used to align the feature representations of trainees at the same sales stage in real time, and call upon group similarity information to supplement the shortcomings of cold-start trainee features, thereby improving real-time prediction accuracy. Specifically, JS divergence can be used for representation supervision to constrain the feature representations of trainees at the same sales stage to tend to be consistent and to form clear distinctions between different stages, ultimately outputting a group association representation, i.e., a high-order association representation of the trainee group.

[0153] Step S204: The Transformer encoder with dynamic weights is used to process the trainee's multi-touch interaction sequence, and the sequence mask is used for adaptive adjustment to output the behavior sequence representation.

[0154] In some embodiments of this application, such as Figure 4As shown, the processing of student multi-touch interaction sequences can be represented by using a Transformer encoder with dynamic weights. This encoder discretizes various interaction attributes within the sequence and converts them into 16-dimensional embedded vectors. A learnable CLS tag is then added to the beginning of the interaction sequence following the embedded vector, completing the transformation of behavioral data into a model-recognizable vector. The learnable CLS tag can be used to aggregate global behavioral information, primarily for concatenating it into the entire sequence, thus forming a complete behavioral sequence with the CLS tag.

[0155] Optionally, a random length masking strategy can be adopted to adaptively adjust the sequence mask of the complete behavioral sequence to obtain an effective length mask.

[0156] Specifically, the effective sequence length can be randomly truncated based on the actual number of user interaction steps to simulate different marketing follow-up depth scenarios. Finally, the sequence length is uniformly padded to the maximum sequence length, and an effective length mask is generated to complete the time series data normalization. In the specific implementation, the truncated length can be randomly extracted according to the actual interaction length of a single student interaction sample. The complete behavior sequence is then truncated according to the truncated length, and the truncated sequence is uniformly padded to the preset maximum sequence length to generate an effective length mask.

[0157] For example, it can be based on the actual interaction length L of the independent samples. actual That is, the total number of all marketing interaction actions actually performed by the student, randomly selected from L. mask ~Uniform(1, L actual） That is, after sampling a mask length between 1 and the actual total number of steps, it is equivalent to using that mask length as the truncation length of the sequence, retaining only L. mask The interaction length of each step is determined by retaining only the interaction behavior of the first Lmask steps in the sequence and discarding the subsequent content. This simulates different levels of marketing follow-up depth, such as early termination of marketing follow-up, mid-term pause, shallow follow-up, and other real-world scenarios. This expands the diversity of training data, constructs samples with different interaction depths, and allows the model to adapt to the behavioral characteristics of learners under different follow-up progress, thereby enabling the batch construction of diverse training samples.

[0158] For example, the truncated sequence can be uniformly padded to the maximum length T. MAX =10, the insufficient part can be padded and the actual length mask is recorded. The recorded actual length mask is the effective length mask.

[0159] In practical applications, the interaction sequence embedding vector of the effective length mask and the complete behavior sequence can be fed into the Transformer encoder with dynamic weights for encoding, and the behavior sequence representation can be output.

[0160] Optionally, the Transformer encoder with dynamic weights can specifically use an education-specific Transformer with 4-head attention, 2-layer encoding, 256 hidden layers, and Dropout=0.2, instead of fixed attention weights. The model can dynamically adjust the weight allocation based on temporal information and interaction scenarios, while incorporating education-specific temporal elements such as summer and winter vacations, exam periods (e.g., days until the final exam), and interaction duration to deeply explore the behavioral patterns of students following up on various marketing touchpoints. In other words, it can rely on the 4-head attention and dynamic weight mechanism, combined with education-specific temporal information, to complete the encoding and output.

[0161] For example, a specific architecture can be adopted, consisting of a 4-head attention mechanism, a 256-dimensional hidden layer, and a 2-layer encoder. This architecture integrates educational-specific temporal features such as interaction duration, summer and winter vacations, and exam dates. A dynamic weight generation mechanism adaptively adjusts the attention weights to offset data interference caused by changes in marketing strategies and sudden shifts in student states. The sequence embedding carries core information such as student multi-touchpoint interaction behaviors and educational temporal features, which can be used by the attention mechanism to calculate associations. The temporal effective mask can block invalid bits and define the scope of real interactions. The training end can use random masking rules, and the inference end can use dynamic weights to avoid invalid information interfering with attention allocation. This application does not limit this approach.

[0162] Step S205: The group association representation and the behavioral sequence representation are fused to obtain the fused features.

[0163] In some embodiments of this application, the population association representation output by the hypergraph structure and the behavioral sequence representation output by the Transformer can be vector-concatenated and fused using a feature fusion method to achieve, for example... Figure 4 The features shown are fused.

[0164] In practical applications, the obtained fusion features can be used as unified input features for downstream core networks.

[0165] Step S206: Input the fused features into a hybrid expert network adapted to the sales stage to obtain a high-order fused feature adapted to the current sales stage.

[0166] In some embodiments of this application, such as Figure 4 As shown, the fusion features can be input into a hybrid expert network adapted to the sales stage. The expert network configuration rules are based on lead access, including official website registration, telephone consultation, event reports, channel placement, etc., to determine the current sales stage.

[0167] Each sales stage can be configured with a separate expert network. These stages can include the new lead stage, telephone communication stage, visit-but-not-signed stage, trial-and-not-converted stage, quote-but-not-signed stage, and signed / churned stage. In other words, a separate dedicated expert network can be configured for each of the aforementioned six marketing decision-making stages.

[0168] Optionally, each expert network can access only the core business features of the corresponding stage, filtering out irrelevant and redundant features, and accurately focusing on the key points of stage analysis.

[0169] For example, the phase division and expert configuration can be shown in Table 1 below:

[0170] Table 1. Phase Division and Expert Allocation

[0171]

[0172] In some embodiments of this application, the fusion features can be broken down into stage-specific key features, and the stage-specific key features can be input into the expert network of the corresponding stage to extract stage-adaptive features.

[0173] Optionally, the hybrid expert network can also include shared experts and a gating network. In this case, the fused features can be input into the shared experts for cross-stage commonality learning to obtain shared expert features, learn the common behavioral patterns of all stages, and prevent the model from becoming fragmented; and, in order to clarify the stage-adaptive features that the current learner focuses on, the fusion ratio of the aforementioned stage-adaptive features can be determined by the state signals indicated by the gating network.

[0174] The dynamic gating network can take the integrity features of the interaction sequence and the one-hot encoding of the student's current sales stage as input, dynamically calculate the weights of all expert networks and complete the normalization, and automatically adjust the optimal expert combination weights according to the user's actual decision-making stage, so as to achieve dynamic feature extraction that is tailored to each individual and conforms to the student's staged course enrollment decision-making pattern.

[0175] Specifically, the weights of each expert network can be dynamically calculated based on the sequence integrity representation of the student's multi-touchpoint interaction sequence and the current sales stage code. For example, according to the rules shown in Table 1, each sales stage is equipped with an independent expert network, and each expert's input consists only of the key features of that stage, avoiding interference from invalid features. The gating network can dynamically calculate the weights of the stage-appropriate features extracted by each expert network, such as w1, w2, w3, w4, w5, and w6, based on the sequence integrity representation and the current sales stage code, and the sum of these weights should be 1, thereby achieving a one-student-one-expert combination mechanism.

[0176] In practical applications, after the gating network calculates the weights, it can be applied in reverse to the expert output features at each stage. That is, the weights can be used to weight and fuse the stage adaptation features extracted by each expert network to obtain the weighted stage adaptation features. This achieves a one-to-one expert combination ratio for each student, dynamically adapting to different follow-up states.

[0177] In some embodiments of this application, the weighted stage adaptation features can be fused with shared expert features to obtain higher-order fused features adapted to the current sales stage.

[0178] Step S207: Use the dynamic prediction tower to generate personalized network parameters for high-order fusion features, make predictions based on the personalized network parameters, and output the course registration intention prediction results.

[0179] The hybrid expert network is the feature preprocessing layer, while the dynamic prediction tower is the backend dual-task decision output layer. The hybrid expert network is only responsible for scenario adaptation, stage stratification, and business feature refinement, distinguishing the patterns of different sales stages such as new leads, phone calls, visits, trial lessons, and quotations. It does not perform any intention scoring or probability judgment, but only refines features. The dynamic prediction tower only performs quantitative prediction and outputs the course enrollment intention prediction results.

[0180] In some embodiments of this application, a dynamic prediction tower can be used to output course registration intention prediction results based on high-order fusion features.

[0181] Optionally, the dynamic prediction tower may include a classification tower and a regression tower; the course enrollment intention prediction result package may include zero intention identification results, intention intensity prediction values, and sales recommendations.

[0182] Specifically, a dynamic prediction tower can be used to generate personalized network parameters for high-order fusion features, a classification tower can be used to predict zero-intention recognition results based on personalized network parameters, and a regression tower can be used to predict intention intensity values. At the same time, based on the zero-intention recognition results, intention intensity prediction values, and preset business threshold rules, corresponding sales follow-up strategies can be matched.

[0183] In some embodiments of this application, the dynamic prediction tower may contain three fully connected layers. In this case, high-order fusion features can be input into the dynamic prediction tower to dynamically generate the weight matrix corresponding to each fully connected layer. Based on the weight matrix corresponding to each fully connected layer, the high-order fusion features are processed to obtain the personalized network parameters corresponding to each student.

[0184] For example, suppose the weight matrices of the three fully connected layers in the dynamic prediction tower are W1, W2, and W3, respectively, with their dimensions corresponding to the input feature dimension, hidden layer dimension, and output dimension. The 96-dimensional high-order fusion feature output by the hybrid expert network can be used as the input to the dynamic prediction tower. This 96-dimensional high-order fusion feature can include student static features, dynamic behavioral features, stage adaptation features, and shared expert features. The dynamic prediction tower then has three fully connected layers, each corresponding to one set of weights. For example, the first layer has a 96-dimensional input and a 48-dimensional output, with weight W1 having a dimension of 96*48; the second layer has a 48-dimensional input and a 24-dimensional output, with weight W2 having a dimension of 48*24; and the third layer has a 24-dimensional input and a 1-dimensional output, with weight W3 having a dimension of 24*1. It should be noted that the weight matrix is ​​dynamically generated based on the current student's 96-dimensional high-order fusion feature. This high-order fusion feature can be mapped through smaller networks to output W1, W2, and W3 specific to each student, achieving personalized prediction for each student.

[0185] In some embodiments of this application, offline marketing business rule bias terms can also be embedded in the model output layer simultaneously to correct the prediction results and make them more aligned with actual business scenarios. Optionally, before outputting the Sigmoid normalization, business rule bias terms such as "trial completion +0.2" and "quote acceptance +0.3" can be injected to improve the reasonableness of the prediction.

[0186] In practical applications, the implementation strategy can be based on the zero-intent probability output by the classification tower to categorize operational plans. When the zero-intent probability is greater than 0.7, it can be automatically identified as a dormant and invalid lead, triggering the lead activation operation mode; otherwise, proactive sales follow-up is executed, and the regression tower can simultaneously output the intention strength prediction value, ultimately achieving personalized prediction output for dual tasks. When the probability is less than 0.7, it is identified as a valid intention lead, at which point a corresponding sales stage-specific follow-up strategy can be automatically matched, completing a deep linkage between the technical model and the education and training marketing business.

[0187] For example, such as Figure 3 As shown, the output results can include high-interest learners (interest score > 0.8), medium-interest learners (interest score 0.3-0.8), and low-interest learners (interest score < 0.3 or p (zero interest) > 0.7). High-interest learners are prioritized for assignment to senior consultants, personalized plans are developed, and in-depth follow-up is conducted within 48 hours. Medium-interest learners follow up using a standard process, focusing on overcoming identification obstacles and periodically reviewing their interest. Low-interest learners are temporarily suspended from proactive telephone follow-ups and moved to a nurturing pool (content operations), with activation and outreach planned for the end of the month.

[0188] In some embodiments of this application, such as Figure 4As shown, the embodiments of this application can also employ a three-loss joint optimization training mechanism. Specifically, a joint loss function can be used to coordinate different learning objectives. Before formal real-time prediction, the model parameters are converged through offline training, achieving offline multi-loss joint optimization.

[0189] Optionally, the similarity between student feature space and actual course registration business can be aligned based on JS divergence loss to optimize group association representation and solve the real-time prediction bias problem of cold start; the features of zero-intent students can be accurately learned based on cross-entropy loss to accurately complete the binary classification judgment of signing and realize the rapid identification of invalid clues; the Huber loss can be used to regress and predict the continuous value of intent intensity in the 0-1 interval, stably fit the continuous value of intent intensity, have strong tolerance for abnormal data errors, reduce the interference of abnormal behavior data on real-time prediction, and improve the overall prediction stability.

[0190] For example, the joint optimization strategy of the three losses can be expressed as:

[0191] L total =β1L JS +β2L CE +β3L LH

[0192] Where L JS L CE L LH These are the JS divergence loss, cross-entropy loss, and Huber loss, respectively; β1, β2, and β3 are the weights of each type of loss.

[0193] In this embodiment, the model can also be trained using the AdamW optimizer for small sample educational data, a cosine annealing learning rate (e.g., an initial learning rate of 5e-5), a batch training strategy (e.g., 256 students / batch), class-balanced data augmentation (e.g., oversampling for minority classes that have signed contracts and random masking for majority classes that have not signed contracts), and differential loss weights (e.g., β1=0.2, β2=2.0, β3=1.0, emphasizing zero-intent recognition and reducing sales harassment), combined with an early stopping mechanism (e.g., 100 rounds, early stopping patience=10), to complete model training and early stopping convergence. This embodiment does not impose any limitations on these aspects.

[0194] In some embodiments of this application, such as Figure 3As shown, after feature processing, including static profile construction, interaction sequence organization, sales stage annotation, and hypergraph neighbor retrieval, as well as hypergraph learning (group similarity), temporal encoding (interaction history), stage adaptation, and dynamic prediction tower output, feedback can be collected on actual signing results, sales consultant ratings, follow-up record feedback, and model prediction calibration. At this point, the model can be iteratively updated based on the access and feedback of new leads, i.e., the model can be learned online, including expert network fine-tuning, hypergraph structure updates, and loss weight adaptation. Through the solution provided in this application embodiment, business performance indicators such as lead utilization, sales efficiency, conversion rate, and student satisfaction are improved, while customer acquisition costs are reduced.

[0195] In this embodiment, based on multi-dimensional business data, dual-path feature encoding is split after standardized feature processing, fused and then connected to an adaptive expert network in the access stage. Model optimization is completed by relying on multi-loss joint training, and finally, the dual-branch output prediction results are linked to the implementation of business strategies.

[0196] Alternatively, for the hypergraph learning submodule, an attention-based hypergraph construction can be used to dynamically determine hyperedge members, which is suitable for data with sparse and dynamically changing student relationships. In addition, Jensen-Shannon divergence can be replaced by Wasserstein distance and maximum mean difference.

[0197] Alternatively, as an alternative to adaptive temporal coding, causal convolution and dilated convolution can be used to capture multi-scale temporal patterns, thereby improving computational efficiency.

[0198] Alternatively, an alternative to sales-stage adaptive MOE could be to use a reinforcement learning-based optimal expert portfolio strategy to obtain more accurate predictive assessments than a single expert when the sales process is complex or the stage boundaries are ambiguous.

[0199] Alternatively, as an alternative to dynamic prediction towers, model upgrades can be achieved by using an independent hypernetwork to generate master network parameters.

[0200] This application does not limit the scope of the embodiments.

[0201] In this embodiment, static profile data and dynamic interaction behavior data of students are acquired. The static profile data and dynamic interaction behavior data are processed to obtain static features of students and multi-touchpoint interaction sequences of students. Then, based on the multi-dimensional similarity of the static features of students, K similar students are selected. Hyperedges are constructed to form a student hypergraph structure according to the nearest neighbor relationship between each student and the corresponding K similar students. Based on the JS divergence constraint, the feature representations of students in the same sales stage in the student hypergraph structure tend to be consistent, and a group association representation is output. In addition, a Transformer encoder with dynamic weights is used to process the multi-touchpoint interaction sequences of students and adaptively adjust them with sequence masks to output behavioral sequence representations. Then, the group association representations and behavioral sequence representations are fused to obtain fused features. The fused features are input into a hybrid expert network adapted to the sales stage to obtain high-order fused features adapted to the current sales stage. Finally, a dynamic prediction tower is used to generate personalized network parameters for the high-order fused features, and prediction is performed based on the personalized network parameters to output the course enrollment intention prediction result. By constructing a multi-dimensional similarity hypergraph structure for students to achieve high-level information transmission within a group, and relying on adaptive weights and temporal masks to complete dynamic modeling of multi-touchpoint interaction sequences, combined with a hybrid expert reasoning mechanism adapted to all stages of the marketing process, and leveraging a dynamic prediction tower to accurately predict intent intensity and distinguish invalid leads, this approach comprehensively addresses industry pain points such as insufficient group association mining, poor adaptability to the decision-making stage, and weak handling of uncertainties in temporal interactions inherent in traditional intent prediction models. It can efficiently support educational institutions in digital precision marketing, rational allocation of enrollment resources, and refined lead management, enabling accurate and efficient prediction of educational users' real-time course enrollment intentions throughout the entire process.

[0202] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of this application.

[0203] This application also provides an electronic device, see embodiments thereof. Figure 5 The provided electronic device 500 includes a memory 510, a processor 520, and a computer program 511 stored in the memory 510 and capable of running on the processor 520. When the computer program 511 is executed by the processor, it implements the various processes of the above-described embodiment of the user course booking intention prediction method based on hypergraph neural network and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0204] This application also provides a computer-readable storage medium, see embodiments thereof. Figure 6 The computer-readable storage medium 600 provided stores a computer program 511. When the computer program 511 is executed by the processor, it implements the various processes of the above-described embodiment of the user course booking intention prediction method based on hypergraph neural network and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0205] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0206] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of the embodiments of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. The division of modules in the embodiments of this application is merely a logical division; in actual applications, there may be other division methods. For example, multiple modules may be combined into or integrated into another system, or some features may be ignored or not performed. Additionally, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interface, and the indirect coupling or communication connection between modules may be electrical or other similar forms, none of which are limited in the embodiments of this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed among multiple circuit modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of this application.

[0207] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0208] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0209] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or modules through some interfaces, and may be electrical, mechanical, or other forms.

[0210] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0211] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0212] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0213] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0214] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0215] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes; these computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0216] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0217] Finally, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0218] The technical solutions provided in the embodiments of this application have been described in detail above. Specific examples have been used in the embodiments of this application to illustrate the principles and implementation methods of the embodiments of this application. The description of the above embodiments is only for the purpose of helping to understand the methods and core ideas of the embodiments of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the embodiments of this application. Therefore, the content of this specification should not be construed as a limitation on the embodiments of this application.

Claims

1. A method for predicting user course registration intention based on hypergraph neural networks, characterized in that, The method includes: Acquire static profile data and dynamic interaction behavior data of students, process the static profile data and the dynamic interaction behavior data respectively to obtain static features of students and multi-touchpoint interaction sequences of students; Based on the multidimensional similarity of the student's static features, select K similar students, and construct hyperedges according to the nearest neighbor relationship between each student and its corresponding K similar students to form a student hypergraph structure. Based on the JS divergence constraint, the feature representations of students at the same sales stage in the student hypergraph structure tend to be consistent, and the group association representation is output. The Transformer encoder with dynamic weights is used to process the multi-touch interaction sequence of the trainees, and the sequence mask is used for adaptive adjustment to output the behavioral sequence representation. The group association representation and the behavioral sequence representation are fused to obtain fused features; The fusion features are input into a hybrid expert network adapted to the sales stage to obtain high-order fusion features adapted to the current sales stage. Personalized network parameters are generated from the higher-order fusion features using a dynamic prediction tower. Based on these personalized network parameters, predictions are made, and the course registration intention prediction results are output.

2. The method according to claim 1, characterized in that, The process of processing the static profile data and the dynamic interaction behavior data to obtain student static features and student multi-touchpoint interaction sequences includes: After one-hot encoding of the categorical features in the static portrait data, embedding representation is performed. After discretization of the numerical features in the static portrait data according to preset rules, embedding mapping is performed uniformly to obtain the static features of the trainees with unified structured data representation. The interaction features recorded during each interaction are obtained from the dynamic interaction behavior data. Each group of interaction features is used as an independent interaction sequence to obtain the student's multi-touch interaction sequence.

3. The method according to claim 1 or 2, characterized in that, characterized in that, The Transformer encoder with dynamic weights processes the student's multi-touch interaction sequence and adaptively adjusts it using a sequence mask to output a behavioral sequence representation, including: Using the Transformer encoder under the dynamic weights, the various interaction attributes in the student's multi-touch interaction sequence are discretized and uniformly converted into 16-dimensional exclusive embedding vectors. A global aggregation marker is added to the beginning of the interaction sequence after the embedding vector to form a complete behavior sequence with the global aggregation marker. The complete behavioral sequence is adaptively adjusted using a sequence mask to obtain an effective length mask; The interaction sequence embedding vector of the effective length mask and the complete behavior sequence is fed into the Transformer encoder under the dynamic weights for encoding, and the behavior sequence representation is output.

4. The method of claim 3, wherein, The adaptive adjustment of the sequence mask of the complete behavior sequence to obtain an effective length mask includes: Based on the actual interaction length of a single student interaction sample, a truncated length is randomly selected, and the complete behavior sequence is truncated according to the truncated length. The truncated sequence is padded to the preset maximum sequence length to generate an effective length mask.

5. The method of claim 2, wherein, A separate expert network is configured for each sales stage, and the hybrid expert network includes shared experts and a gated network; the process of inputting the fused features into the hybrid expert network adapted to the sales stage to obtain high-order fused features adapted to the current sales stage includes: The fused features are broken down into stage-specific key features, and the stage-specific key features are input into the expert network of the corresponding stage to extract stage-adaptive features; and the fused features are input into shared experts to obtain shared expert features. The gated network dynamically calculates the weights of each expert network based on the sequence integrity representation of the student's multi-touch interaction sequence and the current sales stage code. The weights are then used to weight and fuse the stage adaptation features extracted by each expert network to obtain the weighted stage adaptation features. The weighted stage adaptation features are fused with the shared expert features to obtain a high-order fused feature that adapts to the current sales stage.

6. The method of claim 1, wherein, The dynamic prediction tower includes a classification tower and a regression tower. The course registration intention prediction result includes a zero-intention identification result, an intention strength prediction value, and a sales suggestion. The process involves using the dynamic prediction tower to generate personalized network parameters from the higher-order fusion features, performing predictions based on these personalized network parameters, and outputting the course registration intention prediction result, including: The dynamic prediction tower is used to generate personalized network parameters for the higher-order fusion features; The zero-intention recognition result is obtained by using the classification tower based on the personalized network parameters, and the intention intensity prediction value is obtained by using the regression tower. Based on the zero-intention identification result, the intention strength prediction value, and the preset business threshold rules, a corresponding sales follow-up strategy is obtained.

7. The method of claim 6, wherein, The dynamic prediction tower contains three fully connected layers; the generation of personalized network parameters from the higher-order fusion features using the dynamic prediction tower includes: The higher-order fusion features are input into the dynamic prediction tower to dynamically generate the weight matrix corresponding to each fully connected layer. The higher-order fusion features are processed based on the weight matrices corresponding to each fully connected layer to obtain personalized network parameters for each student. 8.A user course reporting intention prediction system based on a hypergraph neural network, characterized in that, The system includes: The data acquisition module is used to obtain static profile data and dynamic interaction behavior data of trainees; The feature engineering module is used to process the static profile data and the dynamic interaction behavior data respectively to obtain the student's static features and the student's multi-touchpoint interaction sequence. The hypergraph construction module is used to select K similar students based on the multidimensional similarity of the student's static features, and construct hyperedges according to the nearest neighbor relationship between each student and the corresponding K similar students to form a student hypergraph structure. The model training module includes a hypergraph learning submodule, an adaptive temporal coding submodule, a sales-stage adaptive hybrid expert submodule, and a dynamic prediction tower submodule; among which: The hypergraph learning submodule is used to constrain the feature representations of students at the same sales stage in the student hypergraph structure to be consistent based on JS divergence, and output a group association representation. The adaptive time coding submodule is used to process the student's multi-touch interaction sequence using a Transformer encoder with dynamic weights, and to adaptively adjust it in conjunction with a sequence mask to output a behavioral sequence representation. The adaptive hybrid expert submodule for the sales stage is used to fuse the group association representation and the behavioral sequence representation to obtain fused features. The dynamic prediction tower submodule is used to input the fused features into a hybrid expert network adapted to the sales stage to obtain a high-order fused feature adapted to the current sales stage. The prediction service module is used to generate personalized network parameters for the higher-order fusion features using a dynamic prediction tower, perform predictions based on the personalized network parameters, and output the course registration intention prediction results.

9. An electronic device, comprising: include: A processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the user course registration intention prediction method based on any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the user course registration intention prediction method based on a hypergraph neural network as described in any one of claims 1 to 7.

Images