A college entrance examination volunteer intelligent recommendation method based on a space-time graph convolution network

Abstract

The application discloses a high school entrance examination volunteer intelligent recommendation method based on a space-time graph convolution network, defines colleges and universities and enrollment majors as graph nodes, collects year-by-year admission ranking data to construct node characteristic vectors, calculates the comprehensive similarity between nodes based on the college level, discipline category, region and ranking change trend, generates a weighted adjacent matrix, constructs an STGCN model stacked by multiple space-time convolution blocks, each space-time convolution block sequentially comprises a time convolution layer, a graph convolution layer and a time convolution layer, and respectively extracts time sequence dependent features and spatial correlation features along the time dimension and the node dimension, predicts the next annual admission ranking with continuous multi-year data, compares the predicted ranking with the examinee ranking, estimates the admission probability in combination with the off-exception residual distribution, and outputs a 'rush','stable' and 'protect' classified volunteer recommendation list after hard constraint filtering. The application realizes the joint modeling of the spatial correlation and the time sequence between colleges and majors, improves the admission ranking prediction accuracy and the accuracy of the volunteer recommendation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recommendation, and specifically relates to an intelligent recommendation method for college entrance examination applications based on spatiotemporal graph convolutional networks. Background Technology

[0002] College application is a crucial step in China's higher education admissions process. Students must choose suitable combinations of majors from thousands of universities within a limited number of choices. The core decision-making criterion is whether the student's college entrance examination ranking meets the admission requirements of their target universities and majors. Currently, there are three main technical solutions in the field of college application recommendation. The first is a ranking prediction method based on statistical regression. This method establishes linear regression, multinomial regression, or backpropagation neural network models for each university and major, uses historical admission ranking data to fit a trend line to predict the admission ranking for the following year, and then combines this with the student's score or ranking to screen and recommend universities. The second is a college application recommendation method based on collaborative filtering. This method finds historical students with similar scores and rankings to the current student, uses their actual admitted universities and majors as recommendations, and combines these with parameters such as the student's subject combination and preferred region to generate a college application plan. The third is a recommendation method based on knowledge graphs or graph embedding. This method constructs a knowledge graph from information on universities, majors, and students, and uses graph embedding or graph attention mechanisms to mine the relationships between entities to explore professional clusters with similar admission difficulty.

[0003] However, the aforementioned existing technical solutions all have significant shortcomings. First, statistical regression-based methods treat each institution and major as an independent entity for prediction, completely ignoring the significant linkage effect between similar institutions and majors in terms of changes in admission rankings. For example, when the ranking of a computer science major at a certain institution rises, the rankings of computer science majors at other institutions at the same level often change in tandem. Independent modeling cannot utilize this spatial correlation information, leading to insufficient prediction accuracy. Furthermore, the regression models used in these methods have limited time-series modeling capabilities. Whether linear regression or LSTM models, they struggle to fully capture the non-linear changes and complex temporal dependencies in the time series of admission rankings, resulting in significantly increased prediction errors in years with large ranking fluctuations. Second, collaborative filtering-based methods are essentially experience-based matching methods, relying on historical applicant admission results for recommendations. They cannot predict future admission rankings, and their accuracy drops sharply when admission policies or application trends change. They also face data sparsity issues and are ineffective for recommending newly added institutions or less popular majors. Furthermore, although some solutions have incorporated graph embedding or knowledge graph techniques, these graph structures are only used to discover similarities between institutions and majors to provide "alternative" options, or directly for recommendation matching, rather than for accurate time-series prediction of admission rankings. Graph structures have never been used to jointly model the spatial association features between institutions and majors and the time-series features of admission rankings. In summary, existing technologies generally lack the ability to jointly model the spatial associations and time-series features between institutions and majors, limiting further improvements in prediction accuracy. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the above shortcomings, the purpose of this invention is to provide an intelligent recommendation method for college entrance examination application based on spatiotemporal graph convolutional networks, which realizes the joint modeling of spatial correlation and time series between colleges and majors, and improves the accuracy of admission ranking prediction and application recommendation.

[0005] Technical Solution: To achieve the above objectives, this invention provides an intelligent college application recommendation method based on spatiotemporal graph convolutional networks, comprising the following steps: Step 1, Graph Structure Construction: Define each combination of institutions and majors as a node in the graph, with a total of N nodes; collect admission ranking data for each node over T consecutive years to construct node feature vectors, and all nodes constitute a feature tensor; calculate the comprehensive similarity s between any two nodes based on institution level, discipline category, region, and admission ranking trends. ij A weighted adjacency matrix A∈R is generated based on the relationship between the comprehensive similarity and a preset threshold. N×NThis forms a sparse graph structure; it explicitly models the spatial relationships between colleges and majors, so that the prediction not only relies on the historical data of a single node, but also integrates the linkage and change information of colleges and majors of the same level, type and discipline, thus solving the defect of existing technology that cannot utilize spatial relationship information when modeling independently. Step 2, STGCN Model Training and Prediction: Construct a spatiotemporal graph convolutional network model consisting of multiple stacked spatiotemporal convolutional blocks. Each spatiotemporal convolutional block sequentially includes a first temporal convolutional layer, a graph convolutional layer, and a second temporal convolutional layer. Input the feature tensor and adjacency matrix into the model and train the model by predicting the admission ranking for the next year using data from the previous T years. Input the data from the most recent T years into the trained model and output the predicted admission ranking for each node in the target year. By extracting features along both the time dimension (temporal convolution) and the node dimension (graph convolution) simultaneously through STGCN, joint modeling of spatial correlation and time series is achieved, rather than treating the two separately. Step 3, Recommending College Applications and Assessing Admission Probability: Obtaining the Candidate's College Entrance Examination Ranking (r) s The system compares the predicted admission rankings at each node with the candidates' rankings, and categorizes the recommendations into three levels: "ambitious," "stable," and "safe," outputting a tiered college application recommendation list. This completes the technical chain from data collection, graph construction, model training, ranking prediction to application recommendation, providing candidates with a directly executable tiered application list. Furthermore, it is not dependent on data distribution in a specific province or year, and can be extended to application recommendation scenarios across different provinces and academic categories.

[0006] Furthermore, the comprehensive similarity s ij The calculation formula is: s ij =α·s ij level +β·s ij subject +γ·s ij region +δ·s ij dep Among them, s ij level For similarity in institutional level, s ij subject For similarity of subject categories, s ij region For regional similarity, s ij dep The similarity depends on the historical admission ranking trend; α, β, γ, and δ are non-negative weight coefficients and α+β+γ+δ=1; The generation rule for the adjacency matrix is: when s ij ≥θ and s ij dep≥τ dep At that time, A ij =s ij Otherwise A ij =0, where θ and τ dep Set a preset threshold; and retain the self-loop A. ii =1. By mapping attributes that are difficult to quantify directly, such as the level of the institution, the discipline, the geographical distribution, and the historical trends, to the interval [0,1] and summing them by weight, the related information of different dimensions can be comprehensively evaluated under the same framework, realizing a comprehensive consideration of the complex relationships between institutions and majors.

[0007] Furthermore, the historical admission ranking trend depends on the similarity s ij dep Calculate using the following steps: The first step is to calculate the Pearson correlation coefficient ρ for the normalized admission ranking sequences of nodes i and j. ij Trend slope difference Δβ ij and dynamic time warping distance d DTW,ij ; The second step is to calculate s. ρ,ij =max(0,ρ ij ), s β,ij =1 / (1+Δβ ij ) and s DTW,ij =1 / (1+d DTW,ij ); Third step, press s dep,ij =λ1·s ρ,ij +λ2·s β,ij +λ3·s DTW,ij Calculate the dependency scores, where λ1, λ2, λ3 ≥ 0 and λ1 + λ2 + λ3 = 1; When the number of common valid years is less than 3 or the variance of any series is 0, s ρ,ij Set it to 0 and mark it as low confidence. At the same time, it evaluates the similarity of the historical position change trends of two nodes from three dimensions: correlation (Pearson), change magnitude (slope difference), and morphological similarity (DTW), which is more comprehensive and robust than using the correlation coefficient alone.

[0008] Furthermore, in the spatiotemporal convolutional block: The first temporal convolutional layer performs one-dimensional convolution along the time dimension, and uses left causal padding to keep the time length unchanged. The convolutional kernel parameters are learned through backpropagation of the loss function. Left causal padding ensures that the temporal convolution only uses historical information, avoids data leakage, and makes the model strictly follow the time order when making predictions. The graph convolutional layer performs graph convolution based on the adjacency matrix using Chebyshev polynomial approximation, with the polynomial order being K. It aggregates the feature information of adjacent nodes along the node dimension. By using a K-order Chebyshev polynomial approximation graph convolution kernel, the eigenvalue decomposition of the Laplacian matrix is ​​avoided, resulting in low computational complexity. The second temporal convolutional layer performs one-dimensional convolution again along the temporal dimension; Each spatiotemporal convolutional block also includes residual connections and normalization layers. When the input and output channels are inconsistent, the residual branches are matched to the target number of channels using 1×1 convolutions or linear layers. When the number of input channels is inconsistent with the number of hidden channels, the residual branches are mapped to the target number of channels through 1×1 convolutions, ensuring dimensionality matching for the addition operation.

[0009] Furthermore, the temporal convolutional layer uses GLU gated linear units as the activation function: the number of output channels of the temporal convolution is set to twice the number of target channels, and it is divided into candidate features U and gated features V along the channel dimension, with the output being U⊙sigmoid(V). Compared with recurrent structures such as LSTM and GRU, GLU is based on convolution operations and can be computed in parallel in the temporal dimension, resulting in faster training and inference speeds.

[0010] Furthermore, in step two, continuous data from T years is used to predict the minimum admission ranking for the next year, with each sample corresponding to a complete time window of the target province-target year graph. The dataset is divided according to the target year, and training samples are constructed using a rolling time window. The validation set is used to adjust the model's hyperparameters, and the final test set is evaluated only once after all parameters are fixed. Dividing the training, validation, and test sets according to the target year, without randomly shuffling the year order, ensures the authenticity of the model evaluation, that is, using historical data to predict the future, rather than using future data to backtest.

[0011] Furthermore, in step three, the admission probability is estimated as follows: using rolling verification to generate out-of-bounds residuals. The residual scaling parameter σ is estimated from the extrapolated residuals. i , σ i Let be the standard deviation estimator of the residual distribution; when the residuals are approximately normal and the distribution is stable, the admission probability is... Where Φ is the standard normal distribution function, Let r be the predicted minimum admission rank of node i in the target year. sFor candidate ranking, when the number of out-of-place residual samples for node i is less than 5, the residual scale is estimated by merging the data in the following priority order: First, residuals from nodes within the same institution and subject category are merged for estimation; if still insufficient, this is expanded to nodes within the same institution and similar subject categories; if still insufficient, residuals from nodes within the same province, subject category, batch, level of institution, and subject category are merged for estimation. Using the out-of-place residuals generated by rolling validation instead of the training set residuals to estimate the error distribution avoids overly optimistic probability estimates, making the admission probability more realistic and reliable. Simultaneously, when the historical residual samples for a single node are insufficient, a tiered merging strategy based on the same province, subject category, batch, level of institution, and subject category is used to borrow information from similar nodes to estimate the residual scale, ensuring that even less popular nodes can obtain reasonable probability estimates.

[0012] Furthermore, it also includes the step of joint modeling across multiple provinces: Collect admission ranking data of each node in P provinces, and construct a multi-province feature tensor by using the province dimension as the input channel or task dimension; Construct a province-specific adjacency matrix for each province: A （p） =M （p） ⊙sigmoid(S base +ΔS （p） ), Where S base Based on the edge weight score matrix, ΔS （p） To correct the score matrix for the provinces to be studied, M （p） For candidate edges and valid node masks for provinces; Each graph convolutional layer uses the corresponding A for each province. （p） Graph convolution is performed, with kernel parameters shared among the provinces. Each province's output header shows its predicted admission ranking. This is achieved by sharing the underlying adjacency matrix A. base and the province-specific correction matrix ΔS (p) This approach retains the commonalities across provinces while also taking into account the specificities of each province. At the same time, the graph convolution kernel parameters are shared among provinces, reducing the total number of model parameters and lowering the risk of overfitting in provinces with sparse data.

[0013] Furthermore, the correction matrix of the province-specific adjacency matrix adopts a low-rank form. The embedding dimension is much smaller than the number of nodes; the total loss function is The first regularization term limits the deviation of the province map from the base map, preventing province-specific corrections from overfitting noise; the second regularization term promotes a sparse graph structure and reduces invalid edges.

[0014] Furthermore, in step three, the grading of "ambitious," "stable," and "safe" is carried out after filtering based on hard constraints: first, institutions and majors that do not meet the requirements of the province, batch, subject category, subject selection, physical condition, enrollment plan, and major remarks are filtered; then, the admission probability p is used as the basis for further filtering. i Divide the gradient according to preset business rules: 40%≤p i <60% is considered "rushing", 60%≤p i <80% is considered "stable", p i ≥80% is considered "guaranteed"; when p i If the percentage is less than 40%, the application will not be included in the regular recommendation list or will be marked as a high-risk application. Before probability grading, a strict filter will be applied based on factors such as the province of admission, admission batch, subject category, subject selection requirements, and physical condition to ensure that the recommendation results are objectively applicable to the applicants and to avoid recommending invalid applications.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. This invention discloses an intelligent college application recommendation method based on spatiotemporal graph convolutional networks. In terms of spatial association modeling, each combination of "university + major" is defined as a node in the graph. The comprehensive similarity between nodes is calculated based on multi-dimensional attributes such as university level, discipline category, geographical distribution, and historical admission ranking trends. A weighted adjacency matrix is ​​generated by setting a multi-screening mechanism that combines a comprehensive similarity threshold, a dependency similarity threshold, and a validation error improvement threshold. This organically associates universities and majors with similar admission characteristics into a graph structure. On this basis, graph convolution operations are used to aggregate the feature information of adjacent universities and majors along the node dimension. This makes the prediction of each node not only dependent on its own historical data, but also integrates the linkage change information of related universities of the same level, type, and discipline. Under the same training conditions, the prediction error can be significantly reduced compared with the independent regression model.

[0016] 2. This invention presents an intelligent college entrance examination application recommendation method based on a spatiotemporal graph convolutional network. In terms of temporal feature extraction, it employs a temporal convolutional layer to perform one-dimensional convolution along the time dimension, extracting the magnitude of ranking changes, continuous trends, and short-term fluctuation patterns between adjacent years using consecutive year windows as units. Compared to simple linear extrapolation or recurrent neural networks, this method can more precisely capture local nonlinear trends and trend inflection points in the admission ranking time series, with a particularly significant improvement in prediction accuracy in years with large fluctuations in admission rankings. Furthermore, this scheme organically combines graph convolution and temporal convolution through a spatiotemporal graph convolutional network, achieving joint learning of spatial correlation features between colleges and majors and the time series features of admission rankings. This allows the model to simultaneously utilize the historical information of nodes themselves and information from adjacent nodes for prediction, fully exploring the deep-seated patterns in the admission data.

[0017] 3. This invention provides an intelligent college entrance examination application recommendation method based on spatiotemporal graph convolutional networks. In the application recommendation stage, this scheme estimates the admission probability based on the difference between the predicted admission ranking and the candidate's ranking, combined with the out-of-bounds residual distribution generated by rolling verification. After filtering by hard constraints such as the province of admission, batch, subject category, subject selection requirements, and physical condition, the recommendation results are divided into three gradients: "ambitious", "stable", and "safe" according to the calibrated admission probability. This provides candidates with a structured, hierarchical application recommendation scheme with probability quantification.

[0018] 4. This invention provides an intelligent college entrance examination application recommendation method based on spatiotemporal graph convolutional networks, and also offers an extended implementation method with multi-province joint modeling. By constructing a province-specific adjacency matrix and cross-province trend similarity components, it utilizes the common patterns in the changes of admission rankings in different provinces to provide auxiliary prediction information for colleges and majors with insufficient historical data, further expanding the coverage and applicability of the recommendation system. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the intelligent college application recommendation method based on spatiotemporal graph convolutional networks as described in this invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0021] Example 1 In this embodiment, the present invention discloses an intelligent college entrance examination application recommendation method based on spatiotemporal graph convolutional networks. Taking the application recommendation for science students in a certain province as an example, the process of this embodiment includes six steps: data acquisition and preprocessing (S1), node and feature construction (S2), similarity calculation and adjacency matrix construction (S3), STGCN model construction and training (S4), admission ranking prediction (S5), and student application recommendation (S6). The specific implementation process is as follows: Operating environment: The server is configured with a Linux operating system, Python 3.10 runtime environment, PyTorch 2.0 deep learning framework, and NVIDIA GPU (with at least 16GB of video memory) for accelerating model training.

[0022] S1: Data Acquisition and Preprocessing Data on the minimum admission ranking, enrollment plan, batch, subject category, and subject selection requirements for approximately 500 universities and about 8,000 majors across the country in the target province (taking Henan Province as an example) from 2015 to 2025 were collected from publicly available data from provincial education examination authorities and university admission websites.

[0023] Missing value imputation employs a tiered fallback rule. If data for a specific major at a particular university is missing for a given year, the average of the lowest ranking for that year among majors within the same discipline at the same university is used as the first choice. If all data for the same discipline at the same university is missing, the average of data for similar disciplines at the same university is used. If data is still missing, the median of data from universities at the same level, within the same province, in the same batch, within the same discipline, or within the same discipline is used. If imputation still cannot be achieved, the missing value mask M is retained, and the loss at that time point is not calculated during model training.

[0024] Positional features are preferentially converted to provincial percentiles q. i,t =r i,t / n g,t , where n g,t This represents the upper limit of the number of candidates or comparable rankings for the corresponding province, year, subject category, and batch. This quantity can be determined after the college entrance examination results are released but before the admission results are generated, and therefore can be used for prediction of the target year. i,t Let n be the original value of the lowest admission rank of node i in year t. If n cannot be obtained... g,t If the value is within the range, normalization should be performed using a fixed scale determined only during the training period, and out-of-range values ​​should be truncated; the actual admission results for the target year should not be used to calculate r. min or r max The validation and test sets reuse only the transformation rules determined during the training period.

[0025] The processing rules for different categories of data are as follows: the lowest admission ranking is adopted as the provincial percentile; other continuous fields such as enrollment plan and plan change rate are fitted with Min-Max or Z-score parameters during the training period; discrete fields such as batch, subject category, subject selection requirements, college level and major category are encoded as one-hot vectors or similarity components; missing markers are used as 0 / 1 mask channels input to the model.

[0026] S2: Node and Feature Construction Define N=8000 university and major nodes. Define each combination of "university + major (or major category, major group)" as a stable node in the graph. For example, "Peking University - Computer Science and Technology" is a node. For different provinces, academic categories, and admission batches, use corresponding task channels and validity masks to record whether the node is recruiting.

[0027] Taking the "Peking University - Computer Science and Technology" node as an example, its ranking from 2021 to 2025 is 150, 138, 162, 145, and 140 respectively; let the upper limit of the comparable ranking of the same subject category in the corresponding year be n. 2021 up to n 2025 Then its percentile eigenvector is X. i =[150 / n 2021 ,138 / n 2022 ,162 / n 2023 ,145 / n 2024 140 / n 2025 All nodes constitute X∈R 8000×5×C x The percentiles mentioned above are calculated for each year based on the corresponding number of test takers or the upper limit of comparable rankings, ensuring comparability of rankings across different years and provinces. When predicting 2026, the model first outputs the predicted percentile values. Then, the number of candidates or the upper limit of comparable rankings announced in 2026 are used to convert the predicted minimum admission ranking.

[0028] S3: Similarity Calculation and Adjacency Matrix Construction For any two nodes v i and v j Calculate the overall similarity: s ij =α·s ij level +β·s ij subject +γ·s ij region +δ·s dep,ij Among them, s ij level The similarity score for institutions is calculated as follows: 1 for institutions at the same level, 0.5 for adjacent levels, and 0 for others. Institutional levels are primarily categorized based on labels from relevant authorities or publicly available policies, such as "Double First-Class Universities," "Former 985 Universities," "Former 211 Universities," "Provincial Key Undergraduate Universities," "Regular Undergraduate Universities," and "Vocational Colleges." For institutions not categorized by publicly available labels, levels are determined by the median percentile of the lowest admission ranking in the target province for the same subject category over the past three years, and these grading thresholds are fixed and stored. Two nodes with the same level number are assigned a score of 1; a difference of 1 is assigned a score of 0.5; and a difference greater than 1 is assigned a score of 0.

[0029] s ij subjectThe similarity score is assigned to academic disciplines, with 1 for the same discipline, 0.5 for similar disciplines, and 0 for the rest. "Similar disciplines" are defined as disciplines in the Ministry of Education's undergraduate major catalog that have the same first-level discipline or are pre-defined adjacent majors. The system has a built-in discipline similarity table. For example, computer science and electronic information, automation, etc., can be set as similar; clinical medicine and mechanical engineering are not set as similar.

[0030] s ij region For regional similarity, 1 is assigned to the same province, 0.5 to the same region, and 0.2 to the rest. The "same region" is determined according to the preset provincial regional mapping table, which can be divided into regions such as North China, Northeast China, East China, Central and Southern China, Southwest China, and Northwest China.

[0031] Dependence on similarity s dep,ij The correlation coefficient ρ is calculated using the Pearson correlation coefficient, the difference in trend slopes, and the normalized DTW distance. The first step involves calculating the Pearson correlation coefficient ρ for the normalized minimum admission ranking sequences of nodes i and j. ij Trend slope difference Δβ ij =|β i -β j | and the DTW distance d normalized to the optimal path length DTW,ij The second step is to calculate s. ρ,ij =max(0,ρ ij ), s β,ij =1 / (1+Δβ ij ) and s DTW,ij =1 / (1+d DTW,ij Third step, press s dep,ij =λ1·s ρ,ij +λ2·s β,ij +λ3·s DTW,ij Calculate the dependency scores, where λ1, λ2, λ3 ≥ 0 and λ1 + λ2 + λ3 = 1. If the number of common valid years is less than 3 or the variance of any sequence is 0, then s ρ,ij Set to 0 and mark as low confidence; negative correlations are not written into the positive adjacency matrix and can be used as independent competitive feature channels.

[0032] α=0.25, β=0.25, γ=0.10, and δ=0.40 are the initial weights for the single-province implementation, corresponding to institution level, discipline category, region, and dependency similarity, respectively, with a sum of 1. The weights for each dimension employ a constrained two-stage grid search satisfying non-negativity and a sum of 1: the first stage performs a coarse search with a step size of 0.05, and the second stage, centered on the optimal combination found in the coarse search, performs a fine search within a local range of 0.10 above and below each weight with a step size of 0.02. Each set of weights has a newly generated adjacency matrix, and is trained using at least 3 random seeds under the same rolling partition, training epochs, and early stopping rules; the final test set does not participate in weight selection.

[0033] The initial thresholds for candidate edge generation are set to θ = 0.6 and τ. dep =0.6. Only when s ij ≥θ、s dep,ij ≥τ dep And ΔRMSE ij When ≥ε, let A ij =s ij Otherwise let A ij =0, and retain the self-loop A ii =1. Where ΔRMSE ij =RMSE(A -ij )-RMSE(A), A -ij This represents the adjacency matrix after deleting candidate edge (i,j). ε>0 is determined by the inner rolling verification and cannot be retained solely based on "no increase in error". The combined threshold θ and dependency threshold τ are used. dep As an adjustable hyperparameter, it can be searched in steps of 0.05 within the range of 0.30 to 0.80; the improvement threshold is verified by taking 0.5% of the baseline RMSE as the initial value, i.e., ε = 0.005 × RMSE. base And adjust it in the inner rolling verification.

[0034] S4: STGCN Model Construction and Training The model structure contains two spatiotemporal convolutional blocks (ST-ConvBlock). Each spatiotemporal convolutional block performs first-time convolution, Chebyshev graph convolution, second-time convolution, residual connection, and normalization in sequence.

[0035] The model input tensor is defined as X∈R B×T×N×C x Each ST-ConvBlock sequentially contains a first temporal convolutional layer, a graph convolutional layer, a second temporal convolutional layer, a residual connection, and a normalization layer. The temporal convolution extracts annual variations along the T dimension, and the graph convolution aggregates features from adjacent nodes along the N dimension. When the input and output channels are inconsistent, the residual branch uses a 1×1 convolution or a linear layer to match the number of hidden channels C. h The first block receives the original feature X, and the k-th block outputs H.k As input to the (k+1)th block.

[0036] First-time convolutional layer: kernel width K t =3, the number of input channels in the first block is C x The number of hidden channels in subsequent blocks is C. h =64; The convolutional layer outputs twice the target channels and passes through GLU to obtain C. h The GLU is used to perform gating filtering after temporal convolution: the number of output channels of the temporal convolution is set to twice the number of target channels, and it is split into two parts, A and B, along the channel dimension. The final output is A⊙sigmoid(B). The temporal convolution uses left-hand causal padding to preserve the time length and avoid using information about the year to be predicted. For node i, the convolution output is h. i,t,c' =Σ q,c θ q,c,c' x i,t-q,c +b c Where q=0,...,K t-1 When tq < 1, use left-side zero padding.

[0037] Graph convolutional layer: Let A~=A+I, L=ID -1 / 2 A~D -1 / 2 And let the scaled Laplacian matrix L ~ = 2L / λ max -I; graph convolution is defined as... Where T0=I, T1=L~, T k =2L~T k-1 -T k-2 The graph has a convolution order of K=3 and a number of hidden channels C. h =64.

[0038] Second temporal convolutional layer: kernel width K t =3, the number of hidden input and output channels is C. h =64, using GLU gating. The GLU channel segmentation rule is: the target hidden channel number is C. h When the convolution frequency is 64, the temporal convolution actually outputs 128 channels. The first 64 channels are used as candidate features U, and the last 64 channels are used as gated features V, with the output being U⊙sigmoid(V). The residual branch is adjusted to C through 1×1 convolution or linear mapping. h One channel.

[0039] The model is terminated by a fully connected output layer, which maps features to a single predicted value. The output of the final ST-ConvBlock is either taken at the last time step or subjected to attention pooling to obtain the node representation h. i Fully connected layer output Limited to [0,1]. The predicted minimum admission ranking for the target year is... , where n g,target The number of applicants or the upper limit of comparable rankings published for the target year is used, not the admission results for the target year.

[0040] Training configuration: Predict the minimum admission ranking for the next year using data from T consecutive years. Taking T=5 as an example, the minimum admission ranking percentile for the next year is predicted using percentiles, enrollment plans, batches, subject categories, subject selection, and node graph features from the past 5 years. The dataset is divided according to the target year, without randomizing the years. If data from 2015 to 2025 is available, the window of 2020 to 2022 (target years) can be used as the basic training set, 2023 and 2024 as rolling validation years, and 2025 as the independent test year; rolling validation cycles such as "train to 2021, validate to 2022", "train to 2022, validate to 2023", and "train to 2023, validate to 2024" can also be constructed.

[0041] Each rolling validation layer performs data processing independently. Missing value imputation parameters, percentile conversion rules, mean and standard deviation of other continuous fields, category mapping, node similarity, and adjacency matrix can only be determined from the training period data of that layer or the candidate size information available before admission; only fixed conversion rules are used for the validation year and the test year. Admission results from future years must not be used to generate the training period adjacency matrix.

[0042] The validation set is used to adjust the overall similarity weight, adjacency matrix threshold, number of STGCN blocks, graph convolution order K, and temporal convolution kernel size K. t The parameters to be considered include the number of hidden channels, learning rate, weight decay, batch size, and early stopping rounds. For each set of candidate parameters, the adjacency matrix is ​​regenerated using the training data, and the model is retrained. Each rolling validation fold is trained repeatedly with at least three random seeds. As selection criteria, F represents the rolling validation folds, and S represents the number of random seeds. The parameter combination with the lowest average validation RMSE is selected; when the differences are not significant, combinations with sparser adjacency matrices and smaller model sizes are preferred.

[0043] The final test set does not participate in missing value handling parameter estimation, feature transformation, adjacency matrix construction, weight search, model hyperparameter selection, early stopping or probability threshold calibration. It is only evaluated once after all parameters are fixed and the MAE, RMSE and MSE are reported.

[0044] Training configuration example: batch size size =2, learning rate lr=0.001, optimizer is Adam, maximum training epochs=200, and early stopping is implemented based on rolling validation loss. Batch size is adjusted according to the number of effective graph time windows per training fold.

[0045] S5: Predicted Admission Ranking After completing independent testing and fixing the parameters, the model is retrained using all data up to 2025. Then, the features from the most recent 5 years (2021-2025) are input into STGCN, and the predicted percentile rank of node i in the target year is output first. Then convert it into the predicted minimum admission ranking It outputs the predicted minimum admission ranking for 2026 for approximately 8,000 colleges and majors in Henan Province.

[0046] S6: Student Application Recommendations Get the candidate's college entrance examination ranking r s Suppose that a science student from Henan Province ranks r in the college entrance examination. s =p=15000. The system first filters out colleges and majors that do not meet the requirements of the province, batch, subject category, subject selection, physical condition, enrollment plan, and major remarks.

[0047] For each institution and major node i, the minimum admission ranking will be predicted. With the candidate's rank r s Comparison. If (The smaller the rank value, the higher the ranking), then the candidate has reached the predicted admission requirements for that node; the final judgment is based on the calibrated admission probability and the results of hard constraint filtering.

[0048] The system calculates the admission probability based on the distribution of out-of-bounds residuals. Out-of-bounds residuals are generated using rolling validation. When there are fewer than 5 residual samples at the node level, the residual scale σ is estimated by merging samples according to the same province, subject category, batch, similar institution level, and discipline. i >0. When the residuals are approximately normal and the distribution is stable, the probability of admission is , interval This is the prediction interval. Candidates whose ranking falls within this interval are marked as having boundary risk; when... The time stamp indicates a higher probability of admission.

[0049] Divide the gradient according to the following business rules: 40%≤p i <60% is considered "rushing", 60%≤p i <80% is considered "stable", p i ≥80% is considered "guaranteed"; when p i If the percentage is less than 40%, the volunteer will not be included in the regular recommendation list or will be marked as a high-risk volunteer. The following percentage range is only a heuristic example when probability calibration has not been completed: "Going for it" meets the requirements. "Stability" satisfies "Guarantee" satisfies The final output is a list of recommended college choices arranged in three tiers: "ambitious," "stable," and "safe."

[0050] Example 2 In this embodiment, the present invention discloses an intelligent college entrance examination application recommendation method based on spatiotemporal graph convolutional networks. Taking application recommendation for multi-province joint modeling as an example, the process of this embodiment includes six steps: multi-province data collection (S1), multi-dimensional feature vector construction (S2), cross-province similarity enhancement adjacency matrix construction (S3), multi-channel STGCN model training (S4), provincial admission ranking prediction (S5), and application recommendation for candidates in designated provinces (S6). The specific implementation process is as follows: Operating environment: Same as in Example 1, GPU memory of at least 32GB is recommended to support larger-scale model training.

[0051] S1: Data Collection from Multiple Provinces Data on admission rankings, enrollment plans, batches, subject categories, and subject selection requirements for approximately 8,000 university and major nodes across 30 selected provincial enrollment regions (P=30) from 2015 to 2025 were collected. Rankings for each province were converted to provincial percentiles, and a province-node-year validity mask was retained.

[0052] Before calculating cross-provincial correlation, intra-provincial standardization must be performed. For differences in the number of candidates in different provinces, the original rankings are first converted to intra-provincial percentiles q=r / n or normalized according to the Min-Max normalization of the same subject category in the same province, year, and year, where n is the number of candidates in that subject category in that province or the upper limit of comparable rankings; then the cross-provincial trend correlation coefficient is calculated to avoid bias caused by directly comparing the original rankings of different provinces.

[0053] S2: Construction of Multidimensional Feature Vectors For each target year, a multi-province time window sample is constructed from the enrollment data of 30 provincial-level regions over the previous T=5 consecutive years. A single sample is X. m ∈R T×N×P×C x After adding the batch dimension, X∈R B×T×N×P×C x Provinces, nodes, and year locations that did not enroll or were missing are marked with a mask. Alternatively, the province and input channel can be expanded as X∈R. T×N×C x' C x' =P·C x .

[0054] The multi-province implementation uses a unified set of college and major nodes, with province information processed as an explicit task dimension. The same pair of nodes can have different association strengths in different provinces; therefore, the adjacency matrix is ​​divided into a shared basic adjacency matrix A. base And the province-specific adjacency matrix A (p) Among them, A baseA represents a stable relationship formed by institutional level, academic discipline, region, and cross-year common trends. (p) Let p represent the corrected complete adjacency matrix for province p.

[0055] When the province dimension is explicitly retained, the input is split into H parts by province. (l,p) ∈R B×T×N×C l The l-th graph convolutional layer uses Azq for each province. (p) Its Chebyshev diagram convolution is Among them, L~ (p) By A (p) Calculate the convolution parameters θ k (l) It is shared among provinces. The outputs of each province are re-stacked along the P-dimensional plane and cross-province information can be exchanged through 1×1 convolutions or province attention layers.

[0056] When the province dimension is expanded to C x' =P·C x If a single adjacency matrix A is still used on N nodes... base Graph convolution can only ensure that all provinces share the same node neighborhood; provincial differences are expressed only through channel weights. Therefore, this flattened form is only equivalent to the explicit province-dimensional form when the adjacency matrix is ​​shared. If it is necessary to preserve province-specific edge relationships, graph convolution should be performed separately for each province before concatenating the channels; or a province-node hypergraph should be constructed, rewriting the input as X. sup ∈R B×T×(P·N)×C x And adopt A sup =blockdiag(A (1) ,...,A (P) )+A cross A cross Used to connect the same college / major nodes in different provinces.

[0057] S3: Construction of Adjacency Matrix with Enhanced Cross-Provincial Similarity Based on the similarity calculation in Example 1, two additional components are added: cross-provincial trend and provincial similarity. The comprehensive similarity formula is expanded as follows: s ij =α·s ij level +β·s ij subject +γ·s ij region +δ·s ij trend,cross +η·s ij province Where s ij trend,crossFor the average dependency similarity across provinces, s ij province The similarity is between provinces; each weight is non-negative and satisfies α+β+γ+δ+η=1.

[0058] For nodes i and j, first calculate s within each valid province p according to the aforementioned Pearson, slope difference, and normalized DTW rules. dep,ij,p Redefine cross-provincial average dependency similarity , where v ij,p ∈{0,1} indicates whether the node pair of the province is valid in the sequence, and the denominator must be greater than 0. The province weight is w. p =T valid,p / T, where T valid,p Let P be the number of common valid years for node pairs within province p, and T be the time window length. Only data that truly exists and meets the criteria of being from the same province, year, subject category, and batch are counted as valid; if imputed values ​​are used, the missing mask should be retained and the corresponding weight reduced. If the number of common valid years is less than 3 or the variance of any sequence is 0, the province is not included in P. valid .

[0059] The province-specific adjacency matrix is ​​constructed according to the following formula: A (p) =M (p) ⊙sigmoid(S base +ΔS (p) ). Among them, S base =logit(clip(A base M is the score matrix with basic edge weights, where ζ is a truncation constant to prevent numerical overflow; (p) ∈{0,1} N×N M is a symmetric candidate edge and effective node mask for provinces. ij (p) =m i (p) ·m j (p) ·M~ ij (p) , where m i (p) ∈{0,1} indicates whether node i is valid in province p; ΔS (p) Adjust the score matrix for the provinces to be studied. Initialize ΔS. (p) =0, causing the model to start training from a shared base graph. To reduce the number of parameters, a low-rank form can be used. The embedding dimension is much smaller than the number of nodes.

[0060] The correction matrix and model parameters are learned through backpropagation using the same prediction loss. The total loss is defined as... The first regularization term limits the deviation of the province map from the base map, while the second regularization term promotes a sparse graph structure. Symmetry transformation, candidate edge masking, and self-loop processing are performed each time the adjacency matrix is ​​calculated, i.e., A (p) ←M (p) ⊙(A (p) +A (p)T ) / 2 and let A ii (p) =m i (p) That is, valid nodes retain self-loops, while invalid nodes do not. Edges corresponding to nodes that are not enrolled or are missing are defined by M. (p) If the value is set to zero, the predicted position will not be included in the loss.

[0061] The multi-province implementation adopts a constrained two-stage grid search. The search range for cross-province trend weights is 0.25 to 0.55, for institution level weights it is 0.10 to 0.35, for academic discipline weights it is 0.10 to 0.35, for regional weights it is 0.05 to 0.20, and for provincial similarity weights it is 0 to 0.15. All weights are non-negative and sum to 1. The default combinations are 0.45, 0.20, 0.20, 0.10, and 0.05 respectively. The coarse search step size is 0.10, and the fine search step size is 0.02. The final weights are determined based on the average RMSE of the validation set under at least 3 random seeds; the test set is not used for parameter tuning.

[0062] S4: Multi-channel STGCN model training The model structure primarily employs an explicit province dimension as its implementation method. Each ST-ConvBlock uses a corresponding A for each province. (p) Graph convolution is performed while sharing Chebyshev kernel parameters; the output is then stacked along the provincial dimensions. The model contains 3 ST-ConvBlocks with C hidden channels. h =128, graph convolution order K=3, time convolution kernel width K t =3; The output layer adopts a province multi-head structure and uses a validity mask to block unenrolled nodes.

[0063] The multi-province model first converts the original rankings of each province into provincial percentiles, and then utilizes the shared STGCN convolution parameters and the province-specific adjacency matrix A. (p) The joint learning model exhibits common patterns and provincial differences. The provincial output heads output the percentile ranking for the target year, which is then converted to the minimum admission ranking based on the published number of test-takers or the upper limit of comparable rankings for that year. The base graph, correction matrix, and network parameters are all learned using only training fold data. Rolling validation is used to select the regularization strength, low-rank dimension, and sparsity threshold.

[0064] Training configuration example: batch size size=2, learning rate lr=0.0005, optimizer is AdamW, weight decay decay =0.0001, maximum training epochs=300; adjust the batch size according to the number of effective graph time windows for each training fold, and use rolling validation to stop early.

[0065] The weighted mean square error of all valid provinces, nodes, and samples is: M m,p,i ∈{0,1} is the validity mask, ω p ≥0 represents the province loss weight, and the denominator must be greater than 0. After calculating the loss, backpropagation is performed, and the final test set is only used for a one-time report on generalization performance.

[0066] S5: Predicted Admission Rankings by Province After completing independent testing and fixing the parameters, the model was retrained using all multi-province data up to 2025. Then, the features from the past five years from 2021 to 2025 were input, and the predicted minimum admission ranking for approximately 8,000 college and major nodes in 30 provincial enrollment regions was output for 2026.

[0067] S6: Recommendations for Candidates from Specific Provinces Based on the province where the examinee is located, the prediction results for that province are extracted. For example, if a science examinee in Guangdong Province has a ranking p=20000, the predicted admission ranking data for Guangdong Province is extracted from the model output. First, colleges and majors that do not meet the hard constraints such as the province, batch, subject category, and subject selection requirements are filtered out. Then, a list of recommended colleges and majors is generated according to the tiered rules of "ambitious / stable / safe" in Example 1.

[0068] The advantage of multi-province joint modeling is that there are certain common patterns in the changes of admission rankings in different provinces (such as the nationwide rise in popularity of a certain major). Joint training of data from multiple provinces enables the model to capture such cross-province common trends, thereby improving the accuracy of predictions from a single province. In particular, for colleges and majors with a short enrollment period and insufficient historical data in some provinces, data from other provinces can be used to assist in prediction.

[0069] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent college application recommendation based on spatiotemporal graph convolutional networks, characterized in that, Includes the following steps: Step 1, Graph Structure Construction: Define each combination of colleges and majors as a node in the graph, with a total of N nodes; collect admission ranking data for each node over T consecutive years to construct node feature vectors, and all nodes constitute a feature tensor; Calculate the comprehensive similarity s between any two nodes based on the trends of changes in institution level, academic discipline, region, and admission ranking. ij A weighted adjacency matrix A∈R is generated based on the relationship between the comprehensive similarity and a preset threshold. N×N This forms a sparse graph structure; Step 2, STGCN model training and prediction: Construct a spatiotemporal graph convolutional network model consisting of multiple spatiotemporal convolutional blocks stacked together. Each spatiotemporal convolutional block includes a first temporal convolutional layer, a graph convolutional layer, and a second temporal convolutional layer in sequence. Input the feature tensor and adjacency matrix into the model, train the model by predicting the admission ranking for the next year using data from the last T years, and input the data from the most recent T years into the trained model to output the predicted admission ranking for each node in the target year. Step 3, Recommending College Applications and Assessing Admission Probability: Obtaining the Candidate's College Entrance Examination Ranking (r) s The predicted admission rankings of each node are compared with the candidates' rankings. Based on the comparison results, the recommendation results are divided into three tiers: "ambitious", "stable", and "safe", and a tiered volunteer recommendation list is output.

2. The method according to claim 1, characterized in that, The comprehensive similarity s ij The calculation formula is: s ij =a·s ij level +β·s ij subject +γ·s ij region +δ·s ij dep Among them, s ij level For similarity in institutional level, s ij subject For similarity of subject categories, s ij region For regional similarity, s ij dep The historical admission ranking trend depends on similarity; α, β, γ, and δ are non-negative weight coefficients and α+β+γ+δ=1; The generation rule for the adjacency matrix is: when s ij ≥θ and s ij dep ≥τ dep At that time, A ij =s ij Otherwise A ij =0, where θ and τ dep Set a preset threshold; and retain the self-loop A. ii =1.

3. The method according to claim 2, characterized in that, The historical admission ranking trend depends on similarity s ij dep Calculate using the following steps: The first step is to calculate the Pearson correlation coefficient ρ for the normalized admission ranking sequences of nodes i and j. ij Trend slope difference Δβ ij and dynamic time warp distance d DTW,ij ; The second step is to calculate s. ρ,ij =max(0,ρ ij ), s β,ij =1 / (1+Δβ ij ) and s DTW,ij =1 / (1+d DTW,ij ); Third step, press s dep,ij =λ1·s ρ,ij +λ2·s β,ij +λ3·s DTW,ij Calculate the dependency scores, where λ1, λ2, λ3 ≥ 0 and λ1 + λ2 + λ3 = 1; When the number of common valid years is less than 3 or the variance of any series is 0, s ρ,ij Set it to 0 and mark it as low confidence.

4. The method according to claim 1, characterized in that, In the spatiotemporal convolutional block: The first temporal convolutional layer performs one-dimensional convolution along the temporal dimension, uses left-side causal padding to keep the time length unchanged, and the convolutional kernel parameters are learned through backpropagation of the loss function. The graph convolutional layer performs graph convolution based on the adjacency matrix using Chebyshev polynomial approximation, with the polynomial order being K, and aggregates the feature information of adjacent nodes along the node dimension. The second temporal convolutional layer performs one-dimensional convolution again along the temporal dimension; Each spatiotemporal convolutional block also includes residual connections and normalization layers. When the input and output channels are inconsistent, the residual branches use 1×1 convolutions or linear layers to match the target number of channels.

5. The method according to claim 4, characterized in that, The temporal convolutional layer uses GLU gated linear units as activation functions: the number of output channels of the temporal convolution is set to twice the number of target channels, and it is divided into candidate features U and gated features V along the channel dimension, with the output being U⊙sigmoid(V).

6. The method according to claim 1, characterized in that, In step two, the minimum admission ranking for the next year is predicted using data from T consecutive years. Each sample corresponds to a complete time window of the target province-target year. The dataset is divided according to the target year and training samples are constructed using a rolling time window. The validation set is used to adjust the model hyperparameters, and the final test set is evaluated only once after all parameters are fixed.

7. The method according to claim 1, characterized in that, In step three, the admission probability is estimated as follows: using rolling verification to generate out-of-bounds residuals. The residual scaling parameter σ is estimated from the extrapolated residuals. i , σ i Let be the standard deviation estimator of the residual distribution; when the residuals are approximately normal and the distribution is stable, the admission probability is... Where Φ is the standard normal distribution function, Let r be the predicted minimum admission rank of node i in the target year. s For the candidate's ranking; when the number of out-of-place residual samples for node i is less than 5, the residual scale is estimated by merging according to the following priority levels: priority is given to merging the residuals of nodes of the same institution and the same discipline category; if it is still insufficient, it is expanded to similar discipline categories of the same institution; if it is still insufficient, the residuals of nodes of the same province, same discipline category, same batch, same level of institution, and same discipline category are merged for estimation.

8. The method according to claim 1, characterized in that, It also includes the step of joint modeling across multiple provinces: Collect admission ranking data of each node in P provinces, and construct a multi-province feature tensor by using the province dimension as the input channel or task dimension; Construct a province-specific adjacency matrix for each province: A （p） =M （p） ⊙sigmoid(S base +ΔS （p） ), Where S base Based on the edge weight score matrix, ΔS （p） To correct the score matrix for the provinces to be studied, M （p） For candidate edges and valid node masks for provinces; Each graph convolutional layer uses the corresponding A for each province. （p） Perform graph convolution, with the kernel parameters shared among all provinces, and each province's output header outputs the predicted admission ranking for that province.

9. The method according to claim 8, characterized in that, The correction matrix of the province-specific adjacency matrix is ​​in low-rank form. The embedding dimension is much smaller than the number of nodes; the total loss function is The first regularization term limits the deviation of the provincial map from the base map, while the second regularization term promotes sparse map structure.

10. The method according to claim 1, characterized in that, In step three, the classification of "rushing," "stability," and "safety" is carried out after the hard constraint filtering: first, filtering out colleges and majors that do not meet the requirements of the province, batch, subject category, subject selection, physical condition, number of enrollment plans, and professional remarks. Then based on the admission probability p i Divide the gradient according to preset business rules: 40%≤p i <60% is considered "rushing", 60%≤p i <80% is considered "stable", p i ≥80% is considered "guaranteed"; when p i If the percentage is less than 40%, the volunteer will not be included in the regular recommendation list or will be marked as a high-risk volunteer.

Images