Academic career prediction recommendation method and system based on multi-modal data feature analysis

Abstract

This invention relates to the field of intelligent analysis technology for educational data, and discloses a method and system for predicting and recommending academic trends based on multimodal data feature analysis. The method includes: collecting student data from multiple sources and performing subject alignment; performing multimodal preprocessing on the aligned dataset; extracting multimodal features such as academic foundation, scientific research innovation, competition practice, career inclination, and relational resources; generating a time-series sequence of growth trajectories and extracting time-series encoding vectors through a long short-term memory network; fusing multimodal features to generate a unified student representation vector; outputting prediction results through joint prediction via postgraduate entrance examination and employment channels; and performing joint path discrimination to generate hierarchical recommendation results.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for educational data, specifically to a method and system for predicting and recommending academic trends based on multimodal data feature analysis. Background Technology

[0002] Undergraduate students in their second and third years of university typically face multiple choices, such as postgraduate entrance examination, employment, and studying abroad. Analyzing their academic trajectory is of great significance for both students' personal planning and university training and management.

[0003] Existing methods for analyzing academic trends generally use structured data such as course grades and semester GPA from the academic affairs system as the main input, and employ static scoring or rule-based judgment to predict students' trends.

[0004] However, the above methods have the following shortcomings: First, the data dimensions are limited, and research records, competition practice, text semantics, and relational data are not included in the analysis, making it impossible to comprehensively reflect the students' overall development status. Second, the analysis method is a static snapshot, lacking continuous modeling of students' growth trends across semesters, making it difficult to distinguish between students with upward, stable, or fluctuating growth patterns, and easily leading to misjudgments. Third, postgraduate entrance examination prediction and employment recommendation are independent of each other, failing to provide comprehensive joint decision-making suggestions, resulting in incomplete decision-making. Fourth, the recommendation results lack interpretable evidence, making it difficult for students and tutors to understand the source of the recommendations, thus reducing the credibility of the recommendation results. Summary of the Invention

[0005] This invention provides a method and system for predicting and recommending academic paths based on multimodal data feature analysis, which solves the technical problems in related technologies that are unable to uniformly represent students' multi-source heterogeneous data, and are difficult to simultaneously predict both postgraduate entrance examination and employment paths and generate personalized recommendation lists.

[0006] This invention discloses a method for predicting and recommending academic trends based on multimodal data feature analysis, comprising the following steps: collecting raw data associated with the target student from multiple data sources; performing subject alignment on the raw data from different systems to generate an aligned dataset for the same student subject; performing normalization processing on the structured numerical data in the aligned dataset; and outputting semantic vector representations of the unstructured text data through a pre-trained language model to generate standardized multimodal feature inputs; extracting a multidimensional feature set from the standardized multimodal feature inputs, the multidimensional feature set including at least academic foundation features, scientific research and innovation features, competition practice features, career orientation features, and relationship resource features; and then... The feature set is divided into time slices by semester, arranged chronologically to form a growth trajectory time series. This growth trajectory time series is input into a Long Short-Term Memory (LSTM) network, which outputs a time-series encoding vector. The numerical features, semantic vector representation, and time-series encoding vector are concatenated and input into a fully connected fusion layer, which outputs a unified student representation vector. This unified student representation vector is then input into a postgraduate entrance examination prediction channel and an employment prediction channel, respectively. The postgraduate entrance examination prediction channel outputs the probability of success in the postgraduate entrance examination, while the employment prediction channel outputs the employment competitiveness score. By combining the probability of success in the postgraduate entrance examination, the employment competitiveness score, and the target student's path preference information, the academic direction category is determined. Based on the determination results, a recommendation list is generated by filtering and sorting candidates from the target database.

[0007] Furthermore, raw data associated with the target students was collected from multiple data sources, including structured data, semi-structured data, unstructured data, and relational data from academic affairs systems, student affairs systems, research platforms, competition platforms, employment platforms, and learning platforms. Structured data included course grades, semester GPA, major ranking, scholarship records, competition award records, project participation records, and internship experience records. Unstructured data included personal statements, resumes, tutor comments, and paper abstracts. Relational data included tutor relationships, course selection relationships, project participation relationships, team collaboration relationships, and corporate internship relationships. Subject alignment included mapping the same student's records across different systems to a unified student subject identifier based on student ID, identification document, name standardization results, department identifier, and year of enrollment.

[0008] Furthermore, the normalization process for the structured numerical data in the aligned dataset includes: imputing missing values ​​by combining the statistical mean of the same major and grade group with the historical trend of the target student; extracting numerical sequences of the same field from the target student in previous semesters, calculating trend estimates using a linear interpolation algorithm, and using a weighted average of the group statistical mean and the individual trend estimates as the imputation result; performing amplitude compression on outliers, truncating values ​​outside the normal range to a preset boundary value; and evaluating the credibility level of each data source, with the credibility level determined comprehensively based on the update frequency, data completeness, and historical accuracy of the data source, and assigning low fusion weights to data records with credibility levels below a preset threshold.

[0009] Furthermore, extracting multidimensional feature sets from standardized multimodal feature inputs also includes: extracting behavioral stability features from learning behavior records. Behavioral stability features include attendance continuity indicators, timely assignment submission rate, fluctuation range of learning behavior, and number of interruptions in task completion. Behavioral stability features are used to apply prudent confidence assessments to students with large fluctuations in academic performance in subsequent predictions.

[0010] Furthermore, after outputting the temporal coding vector, the method also includes: calculating the feature change rate between adjacent time slices; when the change rate between adjacent time slices exceeds a preset transition threshold, the time node is marked as a key transition point; wherein, the transition threshold is a preset upper limit of the change rate for each feature dimension; when the absolute value of the change rate of a feature dimension between adjacent time slices exceeds the transition threshold corresponding to that dimension, the marking is triggered; and the key transition point information is concatenated as an additional feature into the temporal coding vector.

[0011] Furthermore, after outputting the unified student representation vector, the process includes: retrieving the top K samples with the highest similarity to the target student from the historical student case database using cosine similarity as the case set; storing the unified student representation vectors of previous students and their corresponding real college admission or employment outcome labels; extracting case enhancement features from the case set, including the college admission success rate, employment destination distribution, and target institution level distribution; concatenating the case enhancement features with the initial fusion vector and re-inputting them into the fully connected fusion layer to output the enhanced unified student representation vector.

[0012] Furthermore, after outputting the unified student representation vector, the process includes: establishing a heterogeneous relationship graph based on relational data. The node types in the heterogeneous relationship graph include student nodes, course nodes, tutor nodes, project nodes, competition nodes, enterprise nodes, and job nodes. The edge types between nodes include elective relationships, mentorship relationships, participation relationships, collaboration relationships, and internship relationships. The heterogeneous relationship graph is input into a graph neural network, and the node representations are updated through message passing and neighborhood aggregation operations to output the relationship graph embedding vector of the target student node. After performing L2 normalization on the relationship graph embedding vector, it is concatenated with the unified student representation vector, and a relationship-enhanced unified student representation vector is generated through projection through a fully connected layer.

[0013] Furthermore, the postgraduate entrance examination prediction channel and the employment prediction channel share the unified student representation vector as input, and each has independent network parameters. During the training phase, they are jointly trained using a multi-task learning approach. The total loss of the joint training is the weighted sum of the cross-entropy loss of the postgraduate entrance examination prediction channel and the cross-entropy loss of the employment prediction channel. The weights of the two losses are determined according to the ratio of the number of samples with corresponding result labels in the training set for each task to the total number of samples in both classes.

[0014] Furthermore, based on the discrimination results, a recommendation list is generated by filtering and sorting from the candidate target database. The discrimination rules for academic pursuit categories are as follows: when the probability of success in postgraduate entrance examination is higher than the first threshold and the intensity of the desire for further study is higher than the willingness threshold, it is judged as postgraduate entrance examination priority; when the employment competitiveness score is higher than the second threshold and the intensity of the employment willingness is higher than the willingness threshold, it is judged as employment priority; when both the probability of success in postgraduate entrance examination and the employment competitiveness score are in the middle range, it is judged as dual-track approach; a comprehensive score is calculated for each recommended object in the recommendation list. The comprehensive score is a weighted sum of the suitability score, the probability of success in postgraduate entrance examination, and the case support, where the suitability score and the probability of success in postgraduate entrance examination are determined from the following criteria: The prediction results are obtained from the research prediction channel or employment prediction channel. Case support is the proportion of cases in the case set that are consistent with the trend of the recommended object. The sum of the weight coefficients of the three indicators is 1. The candidates in the recommendation list are sorted in descending order according to the comprehensive score, and the candidates with the highest ranking are output as the final recommendation results. For each recommendation result, the prediction process is repeated by applying multiple random perturbations to the student's unified representation vector. The variance of the multiple prediction results is used as the uncertainty index. When the uncertainty index exceeds the preset uncertainty threshold, the comprehensive score of the recommended object is adjusted downward. The adjustment magnitude is proportional to the uncertainty index.

[0015] This invention discloses an academic trajectory prediction and recommendation system based on multimodal data feature analysis. The system is used to execute an academic trajectory prediction and recommendation method based on multimodal data feature analysis. The prediction and recommendation system includes: a data acquisition and alignment module, used to collect raw data associated with the target student from multiple data sources, perform subject alignment on the raw data from different systems, and generate an aligned dataset oriented towards the same student subject; a multimodal preprocessing module, used to perform normalization processing on the structured numerical data in the aligned dataset, and output semantic vector representations of unstructured text data through a pre-trained language model to generate standardized multimodal feature inputs; and a feature extraction module, used to extract a multidimensional feature set from the standardized multimodal feature inputs. The multidimensional feature set includes at least academic foundation features, scientific research and innovation features, competition practice features, and career inclination features. The system includes: a feature set and a relationship resource feature set; a time-series analysis module, which divides the multi-dimensional feature set into time slices by semester, arranges them in chronological order to form a growth trajectory time-series sequence, inputs the growth trajectory time-series sequence into a long short-term memory network, and outputs a time-series encoding vector; a feature fusion module, which concatenates numerical features, semantic vector representations, and time-series encoding vectors and inputs them into a fully connected fusion layer, outputting a unified student representation vector; a dual-channel prediction module, which inputs the unified student representation vector into a postgraduate entrance examination prediction channel and an employment prediction channel respectively, with the postgraduate entrance examination prediction channel outputting the probability of success in the postgraduate entrance examination and the employment prediction channel outputting the employment competitiveness score; and a path discrimination and recommendation module, which comprehensively considers the probability of success in the postgraduate entrance examination, the employment competitiveness score, and the path preference information of the target student to determine the academic direction category, and generates a recommendation list by filtering and sorting from the candidate target database based on the discrimination results.

[0016] The present invention has the following beneficial effects: This invention integrates academic performance, research records, competition practice, text semantics, and relational data in a multimodal manner to generate a unified student representation vector containing multidimensional information. This vector can more comprehensively depict the student's overall development status and overcome the prediction bias caused by relying solely on single data dimensions such as grades and GPA.

[0017] This invention forms a time-series sequence of growth trajectories by arranging the multidimensional features of each semester in chronological order, and extracts the time-series encoding vector using a long short-term memory network. This allows the invention to capture the growth trend, fluctuation level, and stage stability of students across semesters, overcoming the misjudgment problem caused by static snapshot analysis ignoring the development process.

[0018] This invention outputs two types of prediction results simultaneously through postgraduate entrance examination prediction and employment prediction channels, and performs joint path discrimination based on the dual-channel results. It can provide comprehensive decision-making suggestions such as prioritizing postgraduate entrance examination, prioritizing employment, or pursuing both paths in parallel, overcoming the problem of incomplete decision-making caused by the separation of postgraduate entrance examination analysis and employment analysis.

[0019] This invention provides analogical evidence through historical case similarity retrieval and outputs key influencing features and case information through recommendation explanation reports, enabling the recommendation results to have a traceable explanatory path and overcoming the problem that the lack of explanatory basis for recommendation results makes it difficult for students and tutors to accept them.

[0020] This invention performs risk correction on recommendation results by assessing prediction uncertainty and correcting for population distribution bias. This can suppress abnormal recommendations caused by insufficient data or unbalanced sample distribution and improve the stability of recommendation results across different professional and grade groups. Attached Figure Description

[0021] Figure 1 This is a flowchart of the academic trajectory prediction and recommendation method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the growth trajectory of STU2042's basic academic characteristics in each semester, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the STU2042 multidimensional feature set distribution provided in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the dual-channel prediction output results provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the comprehensive scoring and sub-item scores of the postgraduate entrance examination recommendation list provided in this embodiment of the invention; Figure 6 This is a schematic diagram comparing the numerical values ​​before and after preprocessing of the structured numerical field provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the joint training sample distribution and loss weights provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the STU2042 ability gap analysis (compared with the average admission score of University A) provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the uncertainty assessment of the comprehensive score of candidate universities provided in an embodiment of the present invention. Detailed Implementation

[0022] Example 1 In the context of undergraduate education, students typically face multiple choices during their sophomore and junior years, including postgraduate studies, employment, and studying abroad. Existing methods for analyzing academic trends generally rely on static judgments based on limited structured data such as grades and GPA, which have the following shortcomings: the data dimensions are singular, failing to reflect the student's overall development; there is a lack of continuous analysis of students' growth trends across semesters, making it difficult to distinguish between students with upward, stable, or fluctuating growth patterns; postgraduate entrance examination predictions and employment recommendations are independent, failing to provide joint decision-making advice; and the recommendation results lack interpretability, making it difficult for students and mentors to understand the basis for the recommendations.

[0023] In this embodiment, the server executing the academic trajectory prediction and recommendation method is equipped with at least a data acquisition interface, data processing service, feature calculation service, time series analysis service, predictive inference service, and report generation service. The data acquisition interface interfaces with the academic affairs system, student affairs system, research platform, competition platform, employment platform, and learning platform. The server's storage layer maintains a student original database, a standardized feature library, and a case library of students from previous years.

[0024] Please see Figures 1-9 A method for predicting and recommending academic trends based on multimodal data feature analysis includes the following steps: Step 1: Collect student data from multiple sources and perform subject alignment Raw data related to target students were collected from academic affairs systems, student affairs systems, scientific research platforms, competition platforms, employment platforms, learning platforms, and student self-filled materials to form a multi-source raw data set.

[0025] The collected raw data should include at least the following categories: structured data, including course grades, semester GPA, major ranking, scholarship records, competition award records, project participation records, and internship experience records; semi-structured data, including application forms, activity records, tutor evaluations, employment intention forms, and course selection information; unstructured data, including personal statements, resumes, self-summaries, tutor comments, and paper abstracts; and relationship data, including tutor relationships, course selection relationships, project participation relationships, team collaboration relationships, and corporate internship relationships.

[0026] For raw data from different systems, alignment is performed based on student ID, identification document, name standardization result, department ID, and year of enrollment. During the alignment process, records of the same student in different systems are mapped to a unified student identity, generating an aligned dataset oriented towards the same student.

[0027] Step 2: Perform multimodal preprocessing on the aligned dataset Preprocessing operations are performed on different types of data in the aligned dataset to generate standardized multimodal feature inputs.

[0028] Structured numerical data undergoes range-based mean normalization, mapping values ​​with different dimensions to a unified numerical range. Missing values ​​are imputed using the statistical mean of a group within the same major and grade level. Outliers are compressed to truncate values ​​outside the normal range to preset boundary values. Categorical fields undergo unique encoding transformations to generate coded representations suitable for numerical calculations.

[0029] For unstructured text data, personal statements, resumes, mentor comments, and paper abstracts are input into a pre-trained language model, which outputs semantic vector representations for each text. These semantic vector representations contain semantic information such as target intent, ability expression, and attitude inclination.

[0030] It should be noted that, in addition to using the population statistical mean, the above-mentioned missing value imputation can also be combined with the historical trends of individual target students for joint imputation. Specifically, when a target student has a missing field in a certain semester, the numerical sequence of the same field in the previous semester is extracted, the trend estimate of the numerical sequence is calculated using a linear interpolation algorithm, and the weighted average of the population statistical mean and the individual trend estimate is used as the imputation result.

[0031] In this embodiment, to reduce the interference of low-quality data on subsequent predictions, the credibility level of each data source is evaluated. The credibility level is determined comprehensively based on the update frequency, data completeness, and historical accuracy of the data source. Data records with credibility levels below a preset threshold are assigned lower fusion weights to suppress their contribution in subsequent feature fusion stages.

[0032] Step 3: Extract multimodal features from standardized data Based on standardized multimodal feature input, student development features are extracted from different modalities to form a multidimensional feature set.

[0033] We extracted fundamental academic features from structured data, including average course grades, GPA trend slope, pass rate for major courses, and score distribution for core courses. We extracted research innovation features from research-related records, including the number of published papers, patent applications, number of project participations, and contributions to projects led by supervisors. We extracted competition and practice features from competition and practice records, including the highest level of competition, distribution of award levels, cumulative internship duration, and frequency of social practice participation.

[0034] Career preference features are extracted from semantic vector representations, including the intensity of desire for further education, the intensity of desire for employment, regional preference distribution, and industry preference distribution. Relationship resource features are extracted from relational data, including the frequency of mentor guidance, team collaboration participation, and the number of times contacted with companies.

[0035] It should be noted that the aforementioned core course score distribution refers to the statistical distribution characteristics of the scores of subjects marked as core professional courses among the courses taken by the target students, including the mean, standard deviation, and minimum score. The marking is based on course attribute tags in the academic affairs system.

[0036] In this embodiment, to capture stability information at the student behavior level, behavioral stability features are extracted from learning behavior records, including attendance continuity indicators, timely assignment submission rate, fluctuation range of learning behavior, and number of task completion interruptions. These behavioral stability features help identify students with significant fluctuations in academic performance, thereby allowing for a more cautious confidence assessment of predictions for such students in subsequent predictions.

[0037] Step 4: Generate a time series of student growth trajectories The target student's multidimensional feature set is divided into multiple time slices by semester. Each time slice includes academic foundation characteristics, scientific research and innovation characteristics, competition practice characteristics, career orientation characteristics, and relationship resource characteristics within that semester. The time slices are arranged in chronological order by semester to form a time sequence of growth trajectory.

[0038] The time-series sequence of student growth trajectories is input into a Long Short-Term Memory (LSTM) network, which outputs a time-series encoded vector. The LTM network takes as input a sequence of feature vectors from each semester and outputs a time-series encoded vector containing information on growth trends, fluctuation levels, and stage stability. During training, the trend labels of previous students are used as supervision signals, and the cross-entropy loss function is employed, with parameters updated using the Adam optimization algorithm. During training, a linear mapping layer is connected to the output of the LTM network to map the time-series encoded vector to the probability distribution of trend categories for parameter learning. During inference, the linear mapping layer is removed, and the time-series encoded vector is directly used in subsequent fusion operations.

[0039] It should be noted that the granularity of the above time slices can be divided by semester or by academic year. When students' time in school is short or the data density is insufficient, the academic year granularity is used to ensure that each time slice contains sufficient feature information.

[0040] In this embodiment, to identify key turning points in a student's development, the feature change rate between adjacent time slices is further calculated based on the temporal coding vector. When the change rate between adjacent time slices exceeds a preset transition threshold, the time node is marked as a key transition point. The transition threshold is a preset upper limit of the change rate for each feature dimension. Marking is triggered when the absolute value of the change rate of a feature dimension between adjacent time slices exceeds the corresponding transition threshold for that dimension. The key transition point information is concatenated as an additional feature into the temporal coding vector to enhance the ability to identify students with stage-specific changes in the subsequent prediction stage.

[0041] Step 5: Fuse multimodal features to generate a unified representation vector for students Numerical features, semantic vector representations, and temporal coding vectors from the multidimensional feature set are concatenated to form an initial fusion vector. This initial fusion vector is then input into a fully connected fusion layer, which outputs a unified student representation vector. The fully connected fusion layer performs weighted mapping on features from different modalities, projecting them into a unified embedding space. The fully connected fusion layer uses the trend labels of past students as supervision signals, employs the cross-entropy loss function, and is trained using the Adam optimization algorithm.

[0042] It should be noted that the above concatenation operation requires that the numerical features, semantic vector representation, and temporal coding vector have all undergone normalization processing before concatenation to eliminate the impact of differences in numerical scales of different modal features on the fusion result. Specifically, the numerical features have already undergone mean normalization based on the range in step 2; the semantic vector representation and temporal coding vector have respectively undergone L2 normalization to ensure that the magnitudes of the semantic vector representation and temporal coding vector are unified before participating in the concatenation.

[0043] In this embodiment of the application, in order to incorporate reference information from historical cases to enhance the discriminative power of the student's unified representation vector, the following steps are also included in step 5: Step 501: Retrieve historical samples similar to the current student from the past student case database. The past student case database stores the unified student representation vectors of previous students and their corresponding real further education or employment outcome labels. During the retrieval, cosine similarity is used to calculate the similarity between the current student's initial fusion vector and each sample in the past student case database. The top K samples with the highest similarity are selected as the similar case set, where K is the preset number of retrievals.

[0044] Step 502: Extract case enhancement features from the set of similar cases. Case enhancement features include the success rate of further education, the distribution of employment destinations, and the distribution of target school levels for each case in the set of similar cases.

[0045] Step 503: Concatenate the enhanced features of the case with the initial fusion vector, re-input it into the fully connected fusion layer, and output the enhanced unified representation vector of the student.

[0046] It should be noted that the K value is preset based on the size of the student case study database from previous years. When the total number of samples in the student case study database from previous years is less than 500, K is set to 5, and when the total number of samples in the student case study database from previous years is greater than or equal to 500, K is set to 10.

[0047] In this embodiment of the application, in order to utilize the relationships between students and courses, tutors, projects, companies, and positions, the following steps are added in addition to step 5: Step 511: Establish a heterogeneous relationship graph based on relational data. The node types in the heterogeneous relationship graph include student nodes, course nodes, tutor nodes, project nodes, competition nodes, company nodes, and job node. The edge types between nodes in the heterogeneous relationship graph include elective relationships, mentorship relationships, participation relationships, collaboration relationships, and internship relationships.

[0048] Step 512: Input the heterogeneous relationship graph into the graph neural network and output the relationship graph embedding vector of the target student node. The graph neural network takes the initial feature vector and edge connections of each node as input, updates the node representation through message passing and neighborhood aggregation operations, and outputs the relationship graph embedding vector of the target student node. During training, the direction labels of previous students are used as supervision signals, the cross-entropy loss function is used, and the Adam optimization algorithm is used to update the parameters. During the training phase, a fully connected output layer is connected to the output of the graph neural network to map the relationship graph embedding vector to the probability distribution of the direction category for parameter learning; during the inference phase, the fully connected output layer is removed, and the relationship graph embedding vector is directly used to participate in subsequent fusion.

[0049] Step 513: After performing L2 normalization on the relationship graph embedding vector, concatenate it with the student unified representation vector, and project it through a fully connected layer to generate a relationship-enhanced student unified representation vector.

[0050] Step 6: Implement a joint prediction for both postgraduate entrance examination and employment. Based on the student's unified representation vector, the prediction results are output through the postgraduate entrance examination prediction channel and the employment prediction channel, respectively.

[0051] The postgraduate entrance exam prediction channel inputs a unified student representation vector into a fully connected network for prediction, outputting the probability of success, target university fit score, and major fit score. The fully connected network consists of several hidden layers and an output layer. The output layer contains three output units, corresponding to the probability of success, target university fit score, and major fit score, respectively. The probability of success is mapped to the 0-1 range using a sigmoid function, while the target university fit score and major fit score are normalized over the candidate university or major set using a softmax function before being output. The employment prediction channel inputs a unified student representation vector into a fully connected network for prediction, outputting the employment competitiveness score, target company fit score, and job fit score. The fully connected network for employment prediction consists of several hidden layers and an output layer. The output layer contains three output units, corresponding to the employment competitiveness score, the target company fit score, and the job fit score, respectively. The employment competitiveness score is mapped to the 0-1 range using the sigmoid function, while the target company fit score and the job fit score are normalized on the candidate company or job set using the softmax function before being output.

[0052] It should be noted that the aforementioned fully connected network for postgraduate entrance exam prediction and fully connected network for employment prediction share the student's unified representation vector as input, but each has independent network parameters. During the training phase, both networks use the actual postgraduate entrance exam results and actual employment results of previous students as supervision labels, and are jointly trained using a multi-task learning approach. During joint training, the total loss function... for:

[0053] in, For the cross-entropy loss of the postgraduate entrance examination prediction channel, For the cross-entropy loss of the employment prediction channel, W1 and W2 are the weights of the two losses, respectively, and their values ​​are determined according to the proportion of the sample size for each task. Specifically, , , where m1 is the number of samples in the training set with postgraduate entrance examination result labels, and m2 is the number of samples in the training set with employment result labels. The Adam optimization algorithm is used to update the parameters of the fully connected network for postgraduate entrance examination prediction and the fully connected network for employment prediction.

[0054] Step 7: Perform joint path discrimination and generate hierarchical recommendation results By comprehensively considering the probability of success in postgraduate entrance examinations, employment competitiveness scores, target students' path preference information, and constraints, the appropriate academic path category for the target students can be determined.

[0055] Path preference information was obtained from students' self-completed employment intention forms and career inclination characteristics. Constraints included the number of remaining semesters on campus and the number of research achievements already accumulated.

[0056] The discrimination rules are as follows: When the probability of success in the postgraduate entrance examination is higher than the first threshold and the target student's willingness to pursue further studies is higher than the willingness threshold, the student is judged to prioritize postgraduate studies; when the employment competitiveness score is higher than the second threshold and the target student's willingness to work is higher than the willingness threshold, the student is judged to prioritize employment; when both the probability of success in the postgraduate entrance examination and the employment competitiveness score are in the middle range, the student is judged to pursue both paths simultaneously. The first and second thresholds are preset discrimination boundaries for the probability of success in the postgraduate entrance examination and the employment competitiveness score, respectively. The willingness threshold is a preset discrimination boundary for the intensity of willingness to pursue further studies and the intensity of willingness to work. The middle range is the range in which both the probability of success in the postgraduate entrance examination and the employment competitiveness score are not higher than the first and second thresholds, and neither are lower than their respective lower limits. All of these thresholds are preset based on the verification results of historical data.

[0057] Based on the judgment results, a corresponding candidate set is selected from the candidate target database. If the judgment result prioritizes postgraduate entrance examination, the candidate universities are sorted according to the hierarchical order of university level, school name, discipline category, and major, generating a postgraduate entrance examination recommendation list. If the judgment result prioritizes employment, the candidate positions are sorted according to the hierarchical order of industry category, company name, job type, and work location, generating an employment recommendation list. If the judgment result is a dual-track process, both the postgraduate entrance examination recommendation list and the employment recommendation list are generated simultaneously.

[0058] Calculate a comprehensive score for each recommended item. The formula for calculating the comprehensive score R is:

[0059] Where R is the overall score, S fit The fit score, obtained from the prediction results in step 6, represents the degree of matching between the target student and the recommended candidate; P succ The probability of success in the postgraduate entrance examination is obtained from the prediction results in step 6, representing the likelihood of the target student succeeding in the postgraduate entrance examination on the path corresponding to the recommended object; r case Case support is the percentage of cases in the case set that align with the trend of the recommended object. ,β, These are the weighting coefficients of the three indicators, satisfying... Its value is preset based on the verification results of previous data. Because S fit、 P succ and r case All three indicators have been normalized to the 0-1 range, and their dimensions are consistent, allowing for direct weighted summation.

[0060] The candidates in the recommendation list are sorted in descending order based on the comprehensive score, and the top-ranked candidates are output as the final recommendation results to the report generation service.

[0061] In this embodiment of the application, to enhance the interpretability of the recommendation results, a recommendation explanation report is generated simultaneously with the recommendation list. The recommendation explanation report includes the following: recommendation basis, listing the top features that contribute the most to the prediction results and their values; similar historical cases, listing the basic information and outcome trends of the top-ranked cases in the case set; and risk warnings, marking recommendation items whose prediction confidence is lower than a preset confidence threshold.

[0062] In this embodiment of the application, in order to perform risk correction on the recommendation results, the following steps are also included in addition to step 7: Step 701: Calculate the prediction uncertainty for each recommendation result. The prediction uncertainty is obtained by repeatedly performing the prediction process in step 6 after applying multiple random perturbations to the student's uniform representation vector, and the variance of the multiple prediction results is used as the uncertainty index.

[0063] Step 702: When the uncertainty index exceeds the preset uncertainty threshold, the comprehensive score of the recommended object is adjusted downward, and the adjustment magnitude is proportional to the uncertainty index.

[0064] Step 703: For student groups of different majors and grades, statistically analyze the distribution characteristics of the recommendation results. When the deviation of the recommendation distribution of a certain group from the overall distribution exceeds the deviation threshold, perform distribution alignment correction on the prediction results of that group to reduce the impact of group bias on the recommendation results.

[0065] In this embodiment of the application, in order to provide improvement suggestions when there is a gap between the ability of the target student and the target path, the following steps are added in addition to step 7: Step 711: Calculate the difference between the multidimensional feature set of the target student and the average features of the students who were admitted or hired in previous years corresponding to the top-ranked recommended student in the recommendation list, and generate an ability gap vector.

[0066] Step 712: Identify the top few feature dimensions with the largest absolute difference in the capability gap vector, and take the feature categories corresponding to the top few feature dimensions with the largest absolute difference in the capability gap vector as the areas to be improved.

[0067] Step 713: Based on the feature categories of the area to be improved, match corresponding improvement suggestions from the preset improvement measure mapping table. The improvement measure mapping table stores the correspondence between feature categories and specific improvement behaviors, including taking supplementary courses, participating in competitions at specific levels, increasing internship and practical experience, and strengthening scientific research training.

[0068] In this embodiment of the application, in order to enable the academic trajectory prediction and recommendation method to have continuous update capabilities, after outputting the recommendation results in step 7, the following steps are also included: Step 721: In subsequent semesters, collect new semester data for the target students and update the multidimensional feature set and growth trajectory time series of the target students.

[0069] Step 722: After obtaining the actual postgraduate entrance examination admission results or employment contract results, store the unified student representation vector and the actual result label of the target student into the case library of previous students to expand the sample size of the case library of previous students.

[0070] Step 723: When the number of new samples added to the past student case library reaches the preset update trigger number, the parameters of the postgraduate entrance examination prediction fully connected network and the employment prediction fully connected network are fine-tuned using the updated past student case library to complete one iteration update.

[0071] Technical effects of this embodiment This implementation method integrates academic performance, research records, competition practice, text semantics, and relational data in a multimodal manner to generate a unified student representation vector containing multidimensional information. Therefore, it can more comprehensively depict the student's overall development status and overcome the prediction bias caused by relying solely on single data dimensions such as grades and GPA.

[0072] This implementation method forms a time-series sequence of growth trajectories by arranging the multidimensional features of each semester in chronological order, and extracts the time-series encoding vector using a long short-term memory network. Therefore, it can capture the growth trend, fluctuation level and stage stability of students across semesters, and overcome the misjudgment problem caused by static snapshot analysis ignoring the development process.

[0073] This implementation method outputs two types of prediction results simultaneously through the postgraduate entrance examination prediction channel and the employment prediction channel, and performs joint path discrimination based on the dual channel results. Therefore, it can provide comprehensive decision-making suggestions such as prioritizing postgraduate entrance examination, prioritizing employment, or parallel development of both channels, overcoming the problem of incomplete decision-making caused by the separation of postgraduate entrance examination analysis and employment analysis.

[0074] This implementation provides analogical evidence through historical case similarity retrieval and outputs key influencing factors and case information through recommendation explanation reports. Therefore, the recommendation results have a traceable explanation path, overcoming the problem that the lack of explanatory basis for the recommendation results makes it difficult for students and tutors to accept them.

[0075] This implementation method performs risk correction on the recommendation results by assessing the uncertainty of prediction and correcting for population distribution bias. Therefore, it can suppress abnormal recommendations caused by insufficient data or unbalanced sample distribution and improve the stability of the recommendation results in different professional and grade groups.

[0076] A third-year computer science and technology student (student ID: STU2042) from the School of Engineering at a university applied for an academic trajectory assessment at the end of the fall semester of 20XX. The student had completed five semesters of study, accumulating data on course grades, competition participation, project experience, and internship records, and also submitted a personal statement and employment intention form. The server collected the student's raw data from the academic affairs system, student affairs system, research platform, competition platform, employment platform, and learning platform, and initiated an academic trajectory prediction and recommendation process.

[0077] Example of Step 1: Multi-source data acquisition and subject alignment The server collects raw records associated with STU2042 from six platforms. The academic affairs system stores grades and GPA data using student ID as the primary key, the research platform associates project participation records with identification documents, and the employment platform associates internship experience with standardized name results and department identifiers. Due to the inconsistency of primary keys across platforms, after subject alignment, the above records are uniformly mapped to the student subject identifier STU2042, generating an aligned dataset.

[0078] Table 1. Results of Multi-Source Raw Data Acquisition and Alignment

[0079] After alignment, the records of STU2042 were merged from independent entries in six systems into a unified aligned dataset, eliminating the data silos caused by differences in primary keys.

[0080] Example of Step 2: Multimodal Preprocessing Preprocessing was performed on different data types in the aligned dataset. Numerical fields such as course grades, GPA, and major rankings underwent mean normalization based on the range. STU2042's professional English grade field in the third semester was missing; therefore, the numerical sequences for this field from the previous two semesters were extracted, and the trend value was estimated using linear interpolation. This value was then weighted and averaged with the mean of the same major and year group to fill the gap. STU2042's submitted personal statement and tutor comments were input into a pre-trained language model, outputting semantic vectors containing information such as desire for further study, ability expression, and attitude. The reliability level of each data source was comprehensively evaluated based on update frequency, completeness, and historical accuracy. Data from the employment platform was assigned a lower fusion weight due to its lower completeness.

[0081] Table 2. Examples of preprocessing structured numeric fields

[0082] Example of Step 3: Multimodal Feature Extraction Based on the preprocessing results, developmental features of STU2042 were extracted from each modality to form a multidimensional feature set. Academic foundation features included the mean course grades and GPA trend slope; research and innovation features included the number of project participations and the contribution of supervisors to projects; behavioral stability features included attendance continuity indicators and timely assignment submission rate. Career tendency features extracted from the semantic vectors showed that STU2042's intention to pursue further studies was 0.71, and their intention to work was 0.48, indicating a strong inclination towards postgraduate studies.

[0083] Table 3. STU2042 Multidimensional Feature Set (Partial)

[0084] Example of Step 4: Generation of a Time Series of Growth Trajectory The multidimensional feature sets of STU2042 across five semesters were arranged sequentially by semester, forming five time slices that constituted a time-series sequence of growth trajectory. This sequence was input into a Long Short-Term Memory (LSTM) network, outputting a time-series encoding vector. Further calculation of the feature change rate between adjacent time slices revealed that the change rate of the GPA trend slope between semesters 3 and 4 was 0.38, exceeding the transition threshold of 0.30 for this dimension. This time point was marked as a key transition point (corresponding to a phased improvement in STU2042's academic performance after participating in the research project). The information of the key transition point was concatenated into the time-series encoding vector.

[0085] Table 4. Time-series slices of STU2042's growth trajectory (academic foundation characteristics)

[0086] Example of Step 5: Multimodal Fusion and Unified Representation Vector Generation The numerical features of STU2042 (all of which have been normalized to mean), semantic vector (after L2 normalization), and temporal coding vector (after L2 normalization, including key transition point information) are concatenated into an initial fusion vector, which is then input into a fully connected fusion layer to output the initial unified student representation vector.

[0087] Subsequently, case enhancement was performed: using the initial fusion vector of STU2042, cosine similarity was searched in the historical student case database (620 samples in total, K=10) to obtain 10 similar cases. The enhanced case features showed that the success rate of further education in these 10 similar cases was 0.70%, the proportion of employment in the internet industry was 0.20%, and the target university level was mainly Double First-Class universities (0.65%). The enhanced case features were then concatenated with the initial fusion vector and re-input into the fully connected fusion layer to output the enhanced unified student representation vector.

[0088] Further relationship graph enhancement is performed: a heterogeneous relationship graph is built based on the relationship data of STU2042, with nodes including STU2042 himself, 3 tutors, 5 core courses, 2 participating projects, and 1 internship company. The graph neural network outputs the relationship graph embedding vector of STU2042 nodes through message passing and neighborhood aggregation. After L2 normalization, it is concatenated with the student unified representation vector, and then projected through a fully connected layer to generate a student unified representation vector with enhanced relationships, which is used for subsequent dual-channel prediction.

[0089] Example of Step 6: Joint Prediction of Postgraduate Entrance Examination and Employment Channels The unified representation vector of students with enhanced relationships is simultaneously input into both the postgraduate entrance examination prediction channel and the employment prediction channel. The postgraduate entrance examination prediction channel outputs: a success rate of 0.73, target university suitability score (top three candidate universities: 0.61, 0.52, and 0.41 respectively), and major suitability score (Artificial Intelligence: 0.69, Software Engineering: 0.55). The employment prediction channel outputs: employment competitiveness score of 0.61, target company suitability score (major internet companies: 0.58, technology-based SMEs: 0.49), and job suitability score (algorithm engineer: 0.63, backend developer: 0.47).

[0090] The two channels share a unified student representation vector, but their network parameters are independent. During joint training, the training set contains 380 samples m1 labeled with postgraduate entrance examination and 240 samples m2 labeled with employment, for a total of 620 samples M. Therefore, the weights w1 = m1 / M = 380 / 620 ≈ 0.61 and w2 = m2 / M = 240 / 620 ≈ 0.39. The weighted total loss function gives the postgraduate entrance examination task a greater proportion of gradient updates, consistent with the actual distribution of a high proportion of postgraduate entrance examination applicants in the engineering school of STU2042.

[0091] Table 5 Dual-channel prediction output results

[0092] Example of Step 7: Joint Path Discrimination and Hierarchical Recommendation Generation Based on the combined results of the dual-channel analysis, the probability of STU2042 succeeding in the postgraduate entrance examination is 0.73, which is higher than the first threshold of 0.65. The intensity of the desire to pursue further studies is 0.71, which is higher than the desire threshold of 0.60. Therefore, the result is that the postgraduate entrance examination is given priority.

[0093] A list of recommended postgraduate entrance examination institutions is generated in tiered order by institution level, academic discipline, and specialization. A comprehensive score R is calculated for each candidate institution. The comprehensive score formula is as follows: S fit P represents the score for suitability between institutions. succ To calculate the probability of success in the postgraduate entrance examination, r case For case support, ,β, For the corresponding weight and Taking candidate university A (a Double First-Class university, specializing in Artificial Intelligence) as an example: S fit =0.61, P succ =0.73, r case =0.65, weight =0.40, β=0.35, =0.25, then R=0.40×0.61+0.35+0.73+0.25×0.65=0.244+0.256+0.163≈0.66.

[0094] For each recommended candidate, a prediction uncertainty assessment is performed. After applying 20 random perturbations to the unified representation vector of STU2042, step 6 is repeated. The multiple prediction variance of candidate institution A is 0.031, which is lower than the uncertainty threshold of 0.05. Therefore, no downward adjustment is performed on the comprehensive score.

[0095] Table 6. Comprehensive Scoring of the Postgraduate Entrance Examination Recommendation List

[0096] It should be noted that while Institution A and Institution C have similar overall scores, Institution A is a Double First-Class university, which has priority in terms of university level, and therefore it is ranked higher in the recommendation list.

[0097] Simultaneously, a recommendation explanation report is generated: listing the top three features that contribute the most to the prediction results (GPA trend slope 0.12, mentor project contribution 0.68, and intensity of desire for further study 0.71) and their values; listing the basic information of the top three historical cases in the case set (all students in the same major who were eventually admitted to Double First-Class universities for postgraduate studies); and the risk warnings include no recommendation items with confidence levels below the threshold.

[0098] Execution capability gap analysis: The difference between the multidimensional feature set of STU2042 and the average features of students admitted to University A in previous years was calculated to generate a capability gap vector. The two feature dimensions with the largest absolute differences were the number of published papers (difference -1.2) and the highest level of competitions (difference -0.18), corresponding to the areas to be improved: scientific research training and competition participation. Recommendations were matched from the improvement measure mapping table: strengthen scientific research training (strive to publish conference papers as the first author) and participate in national-level competitions to improve the competition level.

[0099] The entire data flow process demonstrates a tight logical connection between each step: the aligned dataset in step 1 provides a unified input for the preprocessing in step 2; the standardized output in step 2 provides standardized numerical values ​​for feature extraction in step 3; the multidimensional feature set in step 3 is arranged in time sequence and encoded into a time-series encoding vector in step 4; step 5 gradually merges numerical features, semantic vectors, time-series encoding vectors, case enhancement features, and relationship graph embedding vectors into a unified student representation vector with enhanced relationships; step 6 uses this vector to simultaneously drive dual-channel prediction; step 7 integrates the dual-channel results with path preference information to complete discrimination, scoring, interpretation, and ability gap analysis, and finally outputs a postgraduate entrance examination priority recommendation result with a traceable explanatory path for STU2042.

[0100] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for academic career path prediction recommendation based on multi-modal data feature analysis, characterized in that, Includes the following steps: Raw data associated with the target student is collected from multiple data sources. Subject alignment is performed on the raw data from different systems to generate an aligned dataset for the same student subject. Normalization is performed on the structured numerical data in the aligned dataset, and semantic vector representations are output from the pre-trained language model for the unstructured text data to generate standardized multimodal feature inputs; Extract a multidimensional feature set from standardized multimodal feature inputs. The multidimensional feature set includes at least academic foundation features, scientific research and innovation features, competition practice features, career orientation features, and relationship resource features. The multidimensional feature set is divided into time slices by semester, and arranged in chronological order to form a growth trajectory time sequence. The growth trajectory time sequence is input into a long short-term memory network and outputs a time sequence encoding vector. The numerical features, semantic vector representations, and temporal coding vectors are concatenated and input into a fully connected fusion layer, which outputs a unified student representation vector. Input the student's unified representation vector into the postgraduate entrance examination prediction channel and the employment prediction channel respectively. The postgraduate entrance examination prediction channel outputs the probability of success in the postgraduate entrance examination, and the employment prediction channel outputs the employment competitiveness score. The system comprehensively considers the probability of success in postgraduate entrance examinations, employment competitiveness scores, and the path preference information of target students to determine the academic path category. Based on the determination results, it filters and sorts the candidate target database to generate a recommendation list.

2. The academic track prediction recommendation method based on multi-modal data feature analysis according to claim 1, characterized in that, Raw data related to the target students was collected from multiple data sources, including structured data, semi-structured data, unstructured data, and relational data from academic affairs systems, student affairs systems, scientific research platforms, competition platforms, employment platforms, and learning platforms. The structured data includes course grades, semester GPA, major ranking, scholarship records, competition awards, project participation records, and internship experience records; the unstructured data includes personal statements, resumes, tutor comments, and paper abstracts; and the relationship data includes tutor relationships, course selection relationships, project participation relationships, team collaboration relationships, and corporate internship relationships. The alignment of the execution subject includes mapping the records of the same student in different systems to a unified student subject identifier based on student ID, ID card identifier, name standardization result, department identifier, and year of enrollment.

3. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, Performing normalization on structured numerical data in an aligned dataset also includes: For missing values, the statistical mean of the same major and grade group and the historical trend of the target student were used for imputation. Specifically, the numerical sequence of the same field of the target student in the previous semester was extracted, the trend estimate was calculated by linear interpolation algorithm, and the weighted average of the group statistical mean and the individual trend estimate was used as the imputation result. Amplitude compression is performed on outliers, truncating values ​​that exceed the normal range to preset boundary values; The credibility level of each data source is assessed separately. The credibility level is determined by a combination of the data source's update frequency, data integrity rate, and historical accuracy. Data records with credibility levels below a preset threshold are assigned a low fusion weight.

4. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, Extracting multidimensional feature sets from standardized multimodal feature inputs also includes: Behavioral stability features are extracted from learning behavior records. These features include attendance continuity indicators, timely assignment submission rate, fluctuation range of learning behavior, and number of interruptions in task completion. Behavioral stability features are used to apply a cautious confidence assessment to students with significant fluctuations in academic performance in subsequent predictions.

5. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, After outputting the timing-coded vector, the following is also included: Calculate the rate of change of features between adjacent time slices. When the rate of change between adjacent time slices exceeds a preset transition threshold, mark the time node as a critical transition point. Among them, the transition threshold is a preset upper limit of the rate of change for each feature dimension. When the absolute value of the rate of change of a certain feature dimension between adjacent time slices exceeds the corresponding transition threshold of that dimension, a flag is triggered. Key transition point information is concatenated as an additional feature into the temporal coding vector.

6. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, After outputting the student uniform representation vector, the following is also included: The case set is obtained by using cosine similarity to retrieve the top K samples with the highest similarity to the target student from the case database of previous students. The case database of previous students stores the unified student representation vector of previous students and their corresponding real admission or employment result labels. Case enhancement features are extracted from the case set, including the success rate of college admission, the distribution of employment destinations, and the distribution of target school levels. The enhanced features of the case are concatenated with the initial fusion vector and then re-input into the fully connected fusion layer to output the enhanced unified representation vector of the student.

7. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, After outputting the student uniform representation vector, the following is also included: Heterogeneous relational graphs are built based on relational data. The node types in the heterogeneous relational graphs include student nodes, course nodes, tutor nodes, project nodes, competition nodes, enterprise nodes, and job nodes. The edge types between nodes include elective relationships, mentorship relationships, participation relationships, collaboration relationships, and internship relationships. The heterogeneous relation graph is input into the graph neural network, and the node representation is updated through message passing and neighborhood aggregation operations. The relation graph embedding vector of the target student node is output. After performing L2 normalization on the relationship graph embedding vector, it is concatenated with the student unified representation vector, and the relationship-enhanced student unified representation vector is generated by projection through a fully connected layer.

8. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, The postgraduate entrance examination prediction channel and the employment prediction channel share the student's unified representation vector as input, and each has independent network parameters. They are jointly trained using a multi-task learning approach during the training phase. The total loss of the joint training is the weighted sum of the cross-entropy loss of the postgraduate entrance examination prediction channel and the cross-entropy loss of the employment prediction channel. The weights of the two losses are determined by the ratio of the number of samples with corresponding result labels in the training set for each task to the total number of samples in the two classes.

9. The academic trajectory prediction and recommendation method based on multimodal data feature analysis according to claim 1, characterized in that, Based on the discrimination results, a recommended list is generated by filtering and sorting candidates from the database, including: The criteria for determining academic path categories are as follows: when the probability of success in postgraduate entrance examination is higher than the first threshold and the intensity of the desire for further study is higher than the intention threshold, it is determined as postgraduate entrance examination priority; when the employment competitiveness score is higher than the second threshold and the intensity of the employment desire is higher than the intention threshold, it is determined as employment priority; when both the probability of success in postgraduate entrance examination and the employment competitiveness score are in the middle range, it is determined as dual-track approach. A comprehensive score is calculated for each recommended object in the recommendation list. The comprehensive score is a weighted sum of the suitability score, the probability of success in the postgraduate entrance examination, and the case support. The suitability score and the probability of success in the postgraduate entrance examination are obtained from the prediction results of the postgraduate entrance examination prediction channel or the employment prediction channel. The case support is the proportion of cases in the case set that are consistent with the trend of the recommended object. The sum of the weight coefficients of the three indicators is 1. The candidates in the recommendation list are sorted in descending order based on the comprehensive score, and the top-ranked candidates are output as the final recommendation results. For each recommendation result, the prediction process is repeated by applying multiple random perturbations to the student's unified representation vector. The variance of the multiple prediction results is used as an uncertainty index. When the uncertainty index exceeds the preset uncertainty threshold, the comprehensive score of the recommended object is adjusted downwards, and the adjustment magnitude is proportional to the uncertainty index.

10. An academic trajectory prediction and recommendation system based on multimodal data feature analysis, used to execute the academic trajectory prediction and recommendation method based on multimodal data feature analysis according to any one of claims 1 to 9, characterized in that, Predictive recommendation systems include: The data acquisition and alignment module is used to collect raw data associated with the target student from multiple data sources, perform subject alignment on raw data from different systems, and generate an aligned dataset for the same student subject. The multimodal preprocessing module is used to perform normalization processing on structured numerical data in the aligned dataset, output semantic vector representations of unstructured text data through a pre-trained language model, and generate standardized multimodal feature inputs. The feature extraction module is used to extract a multidimensional feature set from standardized multimodal feature inputs. The multidimensional feature set includes at least academic foundation features, scientific research and innovation features, competition practice features, career orientation features, and relationship resource features. The time series analysis module is used to divide the multidimensional feature set into time slices by semester, arrange them in chronological order to form a growth trajectory time series, input the growth trajectory time series into the long short-term memory network, and output the time series encoding vector. The feature fusion module is used to concatenate numerical features, semantic vector representations and temporal coding vectors and input them into the fully connected fusion layer to output a unified student representation vector. The dual-channel prediction module is used to input the student's unified representation vector into the postgraduate entrance examination prediction channel and the employment prediction channel respectively. The postgraduate entrance examination prediction channel outputs the probability of success in the postgraduate entrance examination, and the employment prediction channel outputs the employment competitiveness score. The path identification and recommendation module is used to identify the academic path category by comprehensively considering the probability of success in postgraduate entrance examination, employment competitiveness score and path preference information of target students. Based on the identification results, it filters and sorts the candidate target database to generate a recommendation list.

Images