Artificial intelligence-driven employment path planning and decision-making system for college students

Abstract

The invention relates to the technical field of employment planning management, in particular to an artificial intelligence-driven employment path planning decision-making system for college students. Comprising a data processing unit; a model construction unit; a scheme generation optimization unit which generates an employment path planning scheme based on the data processed by the data processing unit and the employment path planning model trained by the model construction unit; and an interactive display unit. According to the method, multi-target evaluation is performed on four dimensions such as post matching degree and occupational development potential through a non-dominated sorting genetic algorithm, a Pareto leading-edge solution set is generated, and the scheme is optimized, so that the problem of one-sided planning caused by single-dimension recommendation in the prior art is solved, and systematicness and comprehensiveness of an employment path scheme are realized; the employment path planning model parameters are updated in real time based on the time triggering mechanism, the problem that an existing system is insufficient in response to employment market changes is solved, and the adaptability of the model to market dynamics is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of employment planning and management, and in particular to an artificial intelligence-driven employment path planning decision-making system for college students. Background Art

[0002] As the number of college graduates continues to expand, university graduates face challenges in finding jobs, including information asymmetry and a lack of systematic guidance for career path planning. While the application of artificial intelligence technology in the employment sector is gradually gaining momentum, existing systems are often limited to single-dimensional job recommendations or static path planning. They struggle to integrate university students' personal characteristics and academic data with the dynamic demands of the job market. In particular, they face significant deficiencies in multi-objective optimization (such as comprehensive assessments of job matching, career development potential, and salary satisfaction), real-time data processing, and interactive skills training. More efficient intelligent decision-making systems are urgently needed to enhance the scientific nature and adaptability of employment planning.

[0003] For example, Chinese patent CN202510237083.4 discloses an employment matching and career development planning system, including a user information acquisition module, a job information analysis module, a matching algorithm module and a career planning suggestion module, wherein the user information acquisition module, the job information analysis module, the matching algorithm module and the career planning suggestion module are electrically connected in sequence; the present invention comprehensively acquires user information by constructing a user information acquisition module, analyzes job requirements in real time in combination with the job information analysis module, adopts an advanced matching algorithm module to achieve accurate job matching, and introduces a machine learning algorithm through the career planning suggestion module to provide users with personalized career development paths and training improvement plans; this design significantly improves the matching accuracy and user satisfaction, while helping users to achieve accurate career development planning and optimize the user experience. For example, Chinese patent CN202010282323.X discloses an information analysis system based on student employment data, including a student information subsystem, a historical information subsystem, and a career planning subsystem; through such a setting, based on past basic data and through data accumulation, employment plans can be formulated for students, especially international students, in a directional and targeted manner, and employment guidance can be given to international students. However, due to their limited understanding of the employment market, employment planning reports are generated based on the employment follow-up data collected in the system, which is conducive to students learning about similar employment fields of students with the same characteristics, and is convenient for guiding students' learning direction.

[0004] While the aforementioned technical solutions possess their own design advantages, they also suffer from the following technical drawbacks: First, while Chinese patent CN202510237083.4 provides personalized career planning through machine learning, it relies solely on a single matching algorithm for job recommendations and lacks a multi-dimensional collaborative optimization evaluation system, such as job matching, career development potential, and salary satisfaction. This makes it difficult to balance students' diverse needs in complex employment scenarios. Furthermore, it lacks a model parameter update mechanism based on job market dynamics, making it unable to adapt to real-time job demands. Second, while Chinese patent CN202010282323.X can generate employment planning reports based on historical data, it only analyzes static data for international students and does not introduce a time-triggered dynamic data processing process. This makes it insufficiently responsive to real-time fluctuations in the job market. Furthermore, it completely lacks interactive skills training (such as the construction and evaluation of simulated interview scenarios), failing to provide practical support for students to improve their employability. Therefore, we propose an AI-driven career path planning decision-making system for college students. Summary of the Invention

[0005] The purpose of the present invention is to provide an artificial intelligence-driven career path planning decision-making system for college students, so as to solve the problems proposed in the above-mentioned background technology, such as the lack of a multi-objective optimization mechanism leading to one-sided planning schemes, insufficient dynamic data processing capabilities that make it difficult to adapt to market changes, and the lack of an interactive training module that makes it impossible to improve practical skills.

[0006] To solve the above technical problems, the present invention aims to provide an artificial intelligence-driven career path planning decision system for college students, comprising: A data processing unit, which is based on web crawler technology and interfaces with data interfaces to collect and pre-process personal, academic, and job market data of college students, and uses data filtering algorithms and format conversion rules to complete data cleaning and structuring; A model building unit, which uses a convolutional neural network and a long short-term memory network architecture to train a deep learning model based on historical case data and build an employment path planning model; a scheme generation and optimization unit, which generates an employment path planning scheme based on the data processed by the data processing unit and the employment path planning model trained by the model building unit, and uses a non-dominated sorting genetic algorithm to evaluate the employment path planning scheme from four dimensions: job matching, career development potential, salary satisfaction, and regional adaptability to generate a Pareto frontier solution set, and optimizes the employment path planning scheme based on the evaluation results; and at the same time, uses a time trigger mechanism to update the parameters of the employment path planning model based on employment market dynamics; An interactive display unit intuitively presents the optimized employment path planning scheme to college students through visual charts and text reports; receives real-time feedback from college students on the scheme through an interactive interface; pushes personalized employment information based on the optimized employment path planning scheme; and uses virtual reality technology to construct simulated interview scenarios to assist college students in interview skills training and assessment.

[0007] As a further improvement of this technical solution, the data processing unit includes a data acquisition module and a data preprocessing module, wherein: The data collection module connects to the API interface of corporate recruitment websites through web crawler technology to collect college students' personal information, academic performance data, campus activity records, as well as job demand data in the employment market, industry salary level data and corporate recruitment dynamics data; The data preprocessing module is used to perform word segmentation and stop word removal on the collected unstructured text data, normalize the numerical data, and interpolate the missing values.

[0008] As a further improvement of the present technical solution, the data preprocessing module performs word segmentation and stop word removal on the collected unstructured text data, normalizes the numerical data, and interpolates missing values, including the following steps: S130.1. Perform outlier filtering on the collected raw data and dynamically calculate the neighborhood radius parameter range based on the minimum distance and average distance between samples in the dataset. ; At the same time, according to the total number of samples N, determine the minimum sample number range ; Traversal via grid search and The optimal parameter that can make the core point ratio (number of core points / total number of samples) adapt to the density distribution of the data set is selected by combining the values ​​of For each sample point , statistics of Number of samples in the neighborhood , calculate the abnormality : ; When the abnormality When the value is lower than the statistical distribution threshold of the core point abnormality under the optimal parameters, it is determined to be an outlier and removed; S130.2. For cleaned data containing text content (such as resume descriptions and job requirements), perform word segmentation, denoising, and semantic mapping, where: Word segmentation and denoising: Use word segmentation tools to break down text into a set of terms , filter stop words and punctuation; Semantic mapping: Based on the professional classification standards, the professional terminology database is built (such as integrating the National Occupational Classification Dictionary and high-frequency skill words on the recruitment platform), Perform the following matches: like Concepts in the terminology database Completely consistent, matching score =1; If there are character differences, calculate the Edit distance (allowing ≤ 2 character differences), matching score =0.6; When there is no match =0; By formula: , mapping the text to a structured professional knowledge graph; where, For term Term frequency - inverse document frequency; S130.3. Process numerical data (e.g., GPA, internship duration) and categorical data (e.g., education level, major), including: Numerical data standardization: Z-score standardization method is used to eliminate dimensional differences. The formula is: ;in, For the The sample in The original value of the numerical feature, is the mean of all samples of this feature, is the standard deviation; It is the absolute value of the numerical feature after Z-score standardization; Data coding by type: for the characteristics of academic qualifications, majors, and school types, set them to include Mutually exclusive categories , the vector dimension after one-hot encoding is ; If the i-th sample belongs to the category , then generate -dimensional binary vector , and satisfy: ; in, is the dimension index of the binary vector generated by one-hot encoding; is the category index; Structured feature matrix construction: Assume that the data matrix after outlier filtering is , the text semantic mapping feature matrix is , the standardized numerical feature matrix is , the one-hot encoding classification feature matrix is ;in, is the feature dimension retained after filtering, is the text feature dimension, is the number of numerical features, is the sum of the encoding dimensions of all categorical features; the above matrices are integrated through column concatenation to obtain the final structured feature matrix: ; in, is the total number of samples, and this matrix serves as the input feature matrix of the employment path planning model.

[0009] As a further improvement of this technical solution, the model building unit includes a case preprocessing module, a feature adaptation module, a dual network fusion module and a prediction output module, wherein: The case pre-processing module is used to clean and remove duplicate historical employment case data and mark key nodes in the career development stage (such as internship and employment, campus recruitment); The feature adaptation module is used to convert the labeled data into a feature sequence that is adapted to the input format of the convolutional neural network and the long short-term memory network; The dual network fusion module extracts the spatial correlation between features through a convolutional neural network and captures the temporal laws of career development through a long short-term memory network; The prediction output module outputs prediction results of job matching and career development cycle based on the fusion features.

[0010] As a further improvement of this technical solution, the model building unit training the deep learning model includes the following steps: S250.1. Stratify historical case data by profession, education level, and region to ensure that each layer of data is representative of the corresponding group and construct a training sample set; S250.2. Use a cross-validation strategy to divide the training set into a validation set, iteratively adjust the model parameters, and dynamically modify the training strategy based on the performance changes of the validation set to enhance the model's generalization ability. S250.3. Based on the differences in the number of samples in different professional fields, dynamically adjust the weight of each field in the training loss calculation to balance the impact of uneven sample distribution on model training.

[0011] As a further improvement of the present technical solution, the solution generation optimization unit includes an initial solution generation module, a multi-objective evaluation module, a Pareto solution set optimization module and a dynamic update module, wherein; The initial solution generation module generates an initial employment path solution set including job sequence, skill improvement plan and time nodes based on the data processed by the data processing unit and the employment path planning model trained by the model building unit. , is the number of options; The multi-objective evaluation module uses a non-dominated sorting genetic algorithm to perform a multi-objective evaluation of the initial employment path plan from four dimensions: job matching, career development potential, salary satisfaction, and regional adaptability, through fast non-dominated sorting and crowding distance calculation; The Pareto solution set optimization module selects the Pareto optimal solution combination from the non-dominated frontier solution set based on the congestion distance; The dynamic update module updates the employment path planning model parameters based on dynamic employment market data through a time trigger mechanism to maintain the adaptability of the employment path planning model.

[0012] As a further improvement of the present technical solution, the multi-objective evaluation module adopts a non-dominated sorting genetic algorithm to perform multi-objective evaluation, which includes the following steps: S320.1. Quantification of evaluation indicators: Convert the four evaluation dimensions into calculable numerical indicators, including job matching , career development potential , salary satisfaction and regional adaptability ;in: Job matching: ,in Output of the employment path planning model The probability of matching with the target position (value range [0,1]); Career Development Potential: ,in The solution predicted by the model Corresponding career advancement cycle; Salary satisfaction: ,in For the plan Expected salary, is the median salary in the target industry; Regional adaptability: ,in For the plan Job demand index of target cities, It is the sum of the job demand index of the entire market; S320.1, Fast Non-Dominated Sorting: For any two solutions 、 , define the dominance relationship: ; in, For the plan In the goals; Representation scheme Control Plan ; Indicates that for all ; Indicates that there is at least one ; like Dominate ,but Domination count Add 1, The dominated set Include ; Divide the Pareto frontier layer by layer according to the dominance count ,in is the set of non-dominated solutions, for The next frontier of the dominated solution; The frontier set of the last layer of non-dominated sorting layer; S320.3. Congestion distance calculation: For the same frontier set The solution in , sorting by each dimension and then calculating the crowding distance to maintain the diversity of the solution set: ; in, For the plan The crowding distance; 、 for In dimension Adjacent solutions on is the specific scheme for which the congestion degree is to be calculated; is the index of the target dimension; 、 For the The maximum and minimum values ​​of all solutions under the target dimension; like is a boundary solution (i.e., there is no adjacent solution in a certain dimension), and its congestion distance is set to To ensure that the boundary solutions are prioritized in the screening; S320.4. Algorithm parameter settings: Setting population size , crossover probability , mutation probability , number of iterations , so that the non-dominated sorting genetic algorithm can efficiently generate high-quality Pareto frontier solution sets in employment path planning, achieving solution set diversity, convergence efficiency and computational feasibility; A new population is generated by genetic crossover (such as simulated binary crossover) and polynomial mutation operations, combined with an elite retention strategy (directly incorporating the previous frontier set solution into the new population) to ensure that the solution set converges quickly to the Pareto frontier.

[0013] As a further improvement of this technical solution, the interactive display unit includes a visualization output module, an interactive feedback module, an information push module and a VR interview training module, wherein: The visual output module outputs the optimized employment path planning scheme in the form of visual charts and text reports; The interactive feedback module collects feedback information of college students on the program through the interactive interface; The information push module pushes personalized employment information based on the optimized employment path planning scheme; The VR interview training module uses virtual reality technology to construct simulated interview scenarios to assist in interview skills training and assessment.

[0014] As a further improvement of the present technical solution, the visualization output module includes a multi-dimensional chart generation submodule, a dynamic interactive rendering submodule, and a report structured generation submodule, wherein: The multi-dimensional chart generation submodule is used to generate a job matching heat map, a three-dimensional Sankey diagram of career development paths (with nodes labeled with job titles and skill keywords), and a salary expectation probability distribution diagram; The dynamic interactive rendering submodule is used to perform three types of interactive operations: chart zooming, path node highlighting, and timeline dragging and browsing; The report structured generation submodule generates a text report according to the structure of plan priority ranking-key time node Gantt chart-skill improvement roadmap.

[0015] As a further improvement of this technical solution, the VR interview training module includes, among others: The dynamic scenario generation submodule loads the corresponding interview scenario based on the target position type (technical position / management position / marketing position) from the preset scenario template library constructed by public simulation cases of university employment guidance centers and common interview scenario materials of enterprises; The intelligent interviewer submodule extracts core skill keywords based on the target job description, builds a follow-up question rule library, and generates follow-up questions in real time based on the match between the interviewee's answers and the follow-up question rule library; The multi-dimensional assessment submodule generates a quantitative assessment report based on three dimensions: language expression, professional knowledge, and stress coping; The training file management submodule is used to store historical interview records, generate ability improvement curves, and compare and analyze interview performance at different stages.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses a non-dominated sorting genetic algorithm to perform a multi-objective evaluation of four dimensions, including job matching and career development potential, to generate a Pareto frontier solution set and optimize the plan. This solves the problem of one-sided planning caused by single-dimensional recommendations in existing technologies, and achieves a systematic and comprehensive employment path plan. 2. This invention uses a time-triggered mechanism to update the parameters of the employment path planning model in real time. Combined with the dynamic data collection and cleaning technology of the data processing unit, it solves the problem of the existing system's insufficient response to changes in the employment market and improves the model's adaptability to market dynamics. 3. This invention uses virtual reality technology to construct simulated interview scenarios, integrating intelligent interviewers and multi-dimensional assessment modules, addressing the lack of practical training in traditional systems and providing interactive support for interview skills training and assessment for college students. 4. This invention collects multi-source data through web crawler technology and data interface docking. Combining word segmentation, semantic mapping and standardization processing, it solves the problems of data fragmentation and unstructured data and realizes the efficient integration of college students' personal data and employment market data. 5. This invention adopts a fusion architecture of convolutional neural networks and long short-term memory networks, combined with stratified sampling, cross-validation and dynamic adjustment of sample weights, to improve the model's adaptability to groups with different majors and educational backgrounds, and ensure the accuracy of job matching and career development cycle predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a system framework diagram of the present invention; The meaning of each number in the figure is: 100, data processing unit; 110, data acquisition module; 120, data preprocessing module; 200, model building unit; 210, case preprocessing module; 220, feature adaptation module; 230, dual network fusion module; 240, prediction output module; 300, solution generation optimization unit; 310, initial solution generation module; 320, multi-objective evaluation module; 330, Pareto solution set optimization module; 340, dynamic update module; 400. Interactive display unit; 410. Visual output module; 411. Multi-dimensional chart generation sub-module; 412. Dynamic interactive rendering sub-module; 413. Report structured generation sub-module; 420. Interactive feedback module; 430. Information push module; 440. VR interview training module; 441. Scene dynamic generation sub-module; 442. Intelligent interviewer sub-module; 443. Multi-dimensional evaluation sub-module; 444. Training file management sub-module. DETAILED DESCRIPTION

[0018] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment provides an artificial intelligence-driven career path planning decision system for college students, including: Data processing unit 100, which is based on web crawler technology and connects to the data interface, is used to collect and pre-process college students' personal, academic and employment market data, and use data filtering algorithms and format conversion rules to complete data cleaning and structured processing; In this embodiment, the data processing unit 100 includes a data acquisition module 110 and a data preprocessing module 120, wherein: The data collection module 110 connects to the API interface of the enterprise recruitment website through web crawler technology to collect college students' personal information, academic performance data, campus activity records, as well as job demand data in the employment market, industry salary level data and enterprise recruitment dynamic data; In this example, the web crawler uses the Python Scrapy framework to implement a web crawler. By configuring a User-Agent pool and an IP proxy pool, anti-crawling mechanisms are circumvented. XPath parsing rules are designed for corporate recruitment websites to extract data such as job titles and salary ranges. API integration utilizes OAuth 2.0 authentication to retrieve job data via HTTP requests.

[0020] Furthermore, this embodiment employs an incremental collection and error handling strategy. Specifically, this involves storing the last_update_time timestamp in a MySQL table to implement incremental collection, retrieving only updated data. When an API request returns a 429 or 500 error, an exponential backoff retry mechanism is activated, with an initial delay of 5 seconds. Each subsequent retry delay is doubled and a random value is added, for a maximum of three retries.

[0021] The data preprocessing module 120 is used to perform word segmentation and stop word removal on the collected unstructured text data, normalize the numerical data, and interpolate missing values.

[0022] In this embodiment, the data preprocessing module 120 utilizes the Jieba word segmentation tool and loads a custom dictionary containing over 3,000 professional terms. This is combined with the Harbin Institute of Technology stop word list and manually supplemented industry-related terms for text denoising. During the word segmentation process, a Hidden Markov Model (HMM) is used to identify unregistered words and filter out invalid words with a length of less than 2. Furthermore, for interpolation of missing values, mean interpolation is used for numerical data (such as GPA), by calculating the mean of the non-missing values ​​in the column. Mode interpolation is used for categorical data (such as "education level"), by filling in the most frequently occurring category in the sample.

[0023] In this embodiment, the data preprocessing module 120 performs word segmentation and stop word removal on the collected unstructured text data, normalizes the numerical data, and interpolates missing values, including the following steps: S130.1. Perform outlier filtering on the collected raw data and dynamically calculate the neighborhood radius parameter range based on the minimum distance and average distance between samples in the dataset. ; At the same time, according to the total number of samples N, determine the minimum sample number range ; Traversal via grid search and The optimal parameter that can make the core point ratio (number of core points / total number of samples) adapt to the density distribution of the data set is selected by combining the values ​​of For each sample point , statistics of Number of samples in the neighborhood , calculate the abnormality : ; When the abnormality When the value is lower than the statistical distribution threshold of the core point abnormality under the optimal parameters, it is determined to be an outlier and removed; S130.2. For cleaned data containing text content (such as resume descriptions and job requirements), perform word segmentation, denoising, and semantic mapping, where: Word segmentation and denoising: Use word segmentation tools to break down text into a set of terms , filter stop words and punctuation; Semantic mapping: Based on the professional classification standards, the professional terminology database is built (such as integrating the National Occupational Classification Dictionary and high-frequency skill words on the recruitment platform), Perform the following matches: like Concepts in the terminology database Completely consistent, matching score =1; If there are character differences, calculate the Edit distance (allowing ≤ 2 character differences), matching score =0.6; When there is no match =0; By formula , mapping the text to a structured professional knowledge graph; where, For term Term frequency - inverse document frequency; S130.3. Process numerical data (e.g., GPA, internship duration) and categorical data (e.g., education level, major), including: Numerical data standardization: Z-score standardization method is used to eliminate dimensional differences. The formula is: ;in, For the The sample in The original value of the numerical feature, is the mean of all samples of this feature, is the standard deviation; It is the absolute value of the numerical feature after Z-score standardization; Data coding by type: for the characteristics of academic qualifications, majors, and school types, set them to include Mutually exclusive categories , the vector dimension after one-hot encoding is ; If the i-th sample belongs to the category , then generate dimensional binary vector , and satisfy: ; in, is the dimension index of the binary vector generated by one-hot encoding; is the category index; Structured feature matrix construction: Assume that the data matrix after outlier filtering is , the text semantic mapping feature matrix is , the standardized numerical feature matrix is , the one-hot encoding classification feature matrix is ;in, is the feature dimension retained after filtering, is the text feature dimension, is the number of numerical features, is the sum of the encoding dimensions of all categorical features; the above matrices are integrated through column concatenation to obtain the final structured feature matrix: ; in, is the total number of samples, and this matrix serves as the input feature matrix of the employment path planning model.

[0024] As a further explanation of the steps, the improvement of outlier filtering in this embodiment is based on the DBSCAN algorithm. First, the parameter range is dynamically derived around the total number of samples N: , taking into account both the minimum cluster size and local structure identification; The minimum distance of the dataset is used as the lower limit and 1.5 times the average distance is used as the upper limit to balance the coverage of local dense and weakly correlated samples. Through grid search, parameter combinations with a core point ratio of 20% to 50% are selected to avoid the problem of missed detection of outliers due to a high core point ratio or the problem of mistaken deletion of normal points due to a low core point ratio. On this basis, the concept of "density" of DBSCAN is expanded to define the anomaly degree formula , the number of neighborhood samples and the radius are combined to quantify the local sparsity, and then the mean of the abnormality of the core points is calculated and standard deviation ,set up The threshold is used to determine the outliers, thus realizing the coordination between parameter screening and anomaly degree calculation.

[0025] Furthermore, semantic mapping integrates edit distance and TF-IDF to build a collaborative logic of "surface similarity + semantic importance": edit distance measures the character difference between candidate terms and library concepts, with an exact match (edit distance = 0) scoring 1 point, a difference of 1 to 2 characters scoring 0.6 points, and a difference of more than 2 resulting in no match; TF-IDF uses term frequency (TF, reflecting local importance) and inverse document frequency ( , reflecting global rarity) dimension, quantifying the semantic weight of the term in the document. The association between the two is achieved through the mapping score formula , that is, terms that are superficially similar and semantically important receive higher scores. When executing, candidate terms are first extracted and the matching score is calculated. After pairwise operation with TF-IDF value, the library concept with the highest score is taken as the mapping result.

[0026] It can be understood that the outlier detection in this embodiment is based on DBSCAN, completing the complete process of "original algorithm - dynamic parameter derivation - grid search screening - anomaly formula expansion - statistical threshold determination": starting from the density clustering logic of the classic algorithm, through parameter range derivation, proportion constraint and formula expansion, quantitative identification of outliers is achieved; semantic mapping relies on the surface matching of edit distance and the semantic weighting of TF-IDF, and realizes the collaborative mechanism of "character difference measurement - semantic weight calculation - score weighted matching" through formula concatenation.

[0027] It should be added that, in terms of the verification and storage of data processing in this embodiment, the original data is stored in MongoDB, and documents are stored according to student ID and job record. The structured feature matrix is ​​saved in Parquet format to adapt to the reading of the distributed computing framework. At the same time, the changes in the interquartile range before and after outlier filtering are verified by box plot comparison, and 200 samples are randomly selected to manually verify the text mapping results, requiring the edit distance matching accuracy to be ≥90%.

[0028] The model building unit 200 uses a convolutional neural network and a long short-term memory network architecture to train a deep learning model based on historical case data and build an employment path planning model. In this embodiment, the model building unit 200 includes a case preprocessing module 210, a feature adaptation module 220, a dual network fusion module 230, and a prediction output module 240, wherein: The case pre-processing module 210 is used to clean and remove duplicate historical employment case data and mark key nodes in the career development stage (such as internship and employment, campus recruitment); As a further explanation of this embodiment, the case preprocessing module 210 in this embodiment cleans and removes duplicates from the historical employment case data, identifies duplicate records based on the triple of "student ID + company name + job title" and retains the latest entry of "last updated time", while marking missing values ​​such as "internship duration" as "unknown" and retaining null values ​​for numerical missing values; in addition, when marking key nodes of career development, internship entry is defined as the date of signing the internship agreement, and school recruitment is defined as the date of issuing the first formal offer, and stratified sampling formula is used according to the dimensions of major, education level and region. , construct a training set, where is the number of samples extracted from the h-th layer, is the total sample size, For the For example, if the total sample size is 10,000, and the “Computer Science + Bachelor’s Degree + First-tier” layer has 1,500 original data items, then 1,500 items will be sampled to ensure the representativeness of the stratification.

[0029] The feature adaptation module 220 is used to convert the labeled data into a feature sequence that is adapted to the input format of the convolutional neural network and the long short-term memory network; As a further explanation of this embodiment, the feature adaptation module 220 in this embodiment converts the labeled data into a network input format, uses one-hot encoding for spatial features such as academic qualifications and majors (such as "Master" is mapped to [0,1,0]), and calculates the day-level time difference sequence for temporal features such as internship intervals. It finally constructs a CNN two-dimensional feature map input with a dimension of [B,180,12] (B is the batch size, 180 is the time window, and 12 is the spatial feature dimension) and an LSTM three-dimensional time series sequence input of [B,3,5] (3 is the number of key nodes, and 5 is the temporal feature dimension).

[0030] The dual network fusion module 230 extracts the spatial correlation between features through a convolutional neural network and captures the temporal laws of career development through a long short-term memory network; As a further explanation of this embodiment, the dual network fusion module 230 in this embodiment extracts spatial correlation features through a two-layer CNN: The first layer uses 32 3×3 convolution kernels and 2×2 pooling to output [B, 178, 10]; The second layer uses 64 3×3 convolution kernels and flattens to 1408 dimensions after pooling; At the same time, a two-layer bidirectional LSTM (64 hidden units) is used to capture the temporal regularity, and after bidirectional splicing, the last time step is taken to output a 128-dimensional vector, and finally the , concatenate the fusion features, and output a 256-dimensional fusion vector through a 128-dimensional fully connected layer (ReLU activation).

[0031] The prediction output module 240 outputs prediction results of job matching and career development cycle based on the fusion features.

[0032] As a further explanation of this embodiment, the prediction output module 240 in this embodiment performs dual-task prediction based on the fusion feature: the calculation of the job matching degree is performed by ,in: is the Sigmoid function, is a 256×1 weight matrix, To offset the bias, career development cycle prediction is done through ,in: is a 256×1 weight matrix, For bias.

[0033] In this embodiment, the model building unit 200 trains the deep learning model including the following steps: S250.1. Stratify historical case data by profession, education level, and region to ensure that each layer of data is representative of the corresponding group and construct a training sample set; S250.2. Use a cross-validation strategy to divide the training set into a validation set, iteratively adjust the model parameters, and dynamically modify the training strategy based on the performance changes of the validation set to enhance the model's generalization ability. S250.3. Based on the differences in the number of samples in different professional fields, dynamically adjust the weight of each field in the training loss calculation to balance the impact of uneven sample distribution on model training.

[0034] As a further explanation of this embodiment, the model training in this embodiment uses a 5-fold cross-validation strategy to divide the training set and the validation set, and dynamically adjusts the learning rate (0.001-0.01) and the batch size (32-128) according to the matching degree AUC and the cycle MAE; in view of the differences in sample size in different professional fields, , calculate the domain weight, where For the field The number of samples; the final loss function is , fusing binary cross entropy and mean square error to balance sample distribution and enhance generalization ability.

[0035] The scheme generation and optimization unit 300 generates an employment path planning scheme based on the data processed by the data processing unit 100 and the employment path planning model trained by the model building unit 200. The scheme generation and optimization unit 300 also uses a non-dominated sorting genetic algorithm to evaluate the employment path planning scheme from four dimensions: job matching, career development potential, salary satisfaction, and regional adaptability, to generate a Pareto frontier solution set, and optimizes the employment path planning scheme based on the evaluation results. At the same time, according to the employment market dynamics, the time trigger mechanism is used to update the parameters of the employment path planning model. In this embodiment, the solution generation and optimization unit 300 includes an initial solution generation module 310, a multi-objective evaluation module 320, a Pareto solution set optimization module 330 and a dynamic update module 340, wherein; The initial solution generation module 310 generates an initial employment path solution set including job sequence, skill improvement plan, and time nodes based on the data processed by the data processing unit 100 and the employment path planning model trained by the model building unit 200. , is the number of options; As a further explanation of this embodiment, the initial plan generation module 310 in this embodiment uses a greedy strategy to generate an initial plan set based on the job matching probability matrix (dimension is [target number of jobs, 1]) and the career promotion cycle matrix (dimension is [number of jobs, number of jobs]) output by the model: jobs with a matching probability greater than 0.6 are preferentially selected as the starting point, and subsequent job selection requires that the promotion cycle is less than the industry average and the matching probability decreases by no more than 0.2. At the same time, core skill gaps are extracted based on the job-skill knowledge graph, and a "skill-learning time-resource link" triple plan is generated in descending order according to the TF-IDF weighted gap value. The critical path method (CPM) is used to formulate the timeline, and the internship start time is reversed according to the "internship-campus recruitment conversion cycle" predicted by the model, and the skill learning period is allocated according to the "job sequence promotion cycle × 30%".

[0036] The multi-objective evaluation module 320 uses a non-dominated sorting genetic algorithm to perform a multi-objective evaluation of the initial employment path plan from four dimensions: job matching, career development potential, salary satisfaction, and regional adaptability, through fast non-dominated sorting and crowding distance calculation; The Pareto solution set optimization module 330 selects the Pareto optimal solution combination from the non-dominated frontier solution set based on the congestion distance; As a further explanation of this embodiment, the Pareto solution set optimization module 330 in this embodiment performs dual screening according to "frontier level + crowding distance": the F1 frontier set solution is preferentially retained, and the top 30% solutions within the same frontier set are selected in descending order of crowding distance (for example, 30 solutions are selected when N=100), and simulated binary crossover (SBX, distribution index ηc=20, probability pc=0.8) is used to perform a crossover operation on the job sequence gene bits. For example, the parent generation [algorithm engineer-senior engineer] and [algorithm engineer-architect] are crossed to generate [algorithm engineer-architect-senior engineer], and polynomial mutation (distribution index ηm=20, probability pm=0.2) is combined to mutate the skill plan gene bits (for example, "Python" mutates to "Python+TensorFlow"), and the elite retention strategy is used to directly incorporate the previous frontier set solution into the new population.

[0037] Furthermore, the aforementioned elite retention strategy involves directly incorporating a certain proportion of the non-dominated frontier set solutions obtained in the previous iteration into the new population during the iteration of the non-dominated sorting genetic algorithm (NSGA-II). This strategy aims to prevent the loss of high-quality solutions during evolution and ensure rapid convergence to the Pareto front. Specifically, after each crossover and mutation to generate a new population, solutions representing 10%-20% of the population size are selected from the partitioned non-dominated frontier set and added directly to the new population. Together with the new solutions generated by crossover and mutation, they form the next generation population. This strategy works in conjunction with fast non-dominated sorting and crowding distance calculation. First, the frontier set hierarchy is determined and non-dominated solutions are selected through sorting. Then, the crowding distance is used to preserve diversity, and finally, the high-quality solutions are directly passed on to the next generation. For example, in career path planning, this ensures that the generated Pareto solution set consistently contains solutions with high matching and development potential, avoiding the degradation of solution quality caused by random genetic operations.

[0038] The dynamic update module 340 updates the employment path planning model parameters based on the employment market dynamic data through a time trigger mechanism to maintain the adaptability of the employment path planning model.

[0039] As a further explanation of this embodiment, the dynamic update module 340 in this embodiment adopts a triple trigger mechanism: time-triggered update is automatically started every 7 days, data trigger is started when the amount of employment market data update exceeds 15% of the total data volume (such as the number of new positions is greater than 5,000), and the model is forced to trigger when the verification set matching degree AUC drops by more than 5%; a hot start strategy is adopted during incremental learning, the original model weights are loaded as initial values, and after executing the S130 data processing flow on the new data, only the parameters of the last 3 fully connected layers are fine-tuned (learning rate 0.001), the CNN and LSTM underlying feature extraction layers are frozen, and the data of the last 12 months (about 500,000) are retained through incremental feature matrix updates to ensure that the parameter update takes ≤30 minutes.

[0040] In this embodiment, the multi-objective evaluation module 320 uses the non-dominated sorting genetic algorithm to perform multi-objective evaluation, including the following steps: S320.1. Quantification of evaluation indicators: Convert the four evaluation dimensions into calculable numerical indicators, including job matching , career development potential , salary satisfaction and regional adaptability ;in: Job matching: ,in Output of the employment path planning model The probability of matching with the target position (value range [0,1]); Career Development Potential: ,in The solution predicted by the model Corresponding career advancement cycle; Salary satisfaction: ,in For the plan Expected salary, is the median salary in the target industry; Regional adaptability: ,in For the plan Job demand index of target cities, It is the sum of the job demand index of the entire market; S320.1, Fast Non-Dominated Sorting: For any two solutions 、 , define the dominance relationship: ; in, For the plan In the goals; Representation scheme Control Plan ; Indicates that for all ; Indicates that there is at least one ; like Dominate ,but Domination count Add 1, The dominated set Include ; Divide the Pareto frontier layer by layer according to the dominance count ,in is the set of non-dominated solutions, for The next frontier of the dominated solution; The frontier set of the last layer of non-dominated sorting layer; S320.3. Congestion distance calculation: For the same frontier set The solution in , sorting by each dimension and then calculating the crowding distance to maintain the diversity of the solution set: ; in, For the plan The crowding distance; 、 for In dimension Adjacent solutions on is the specific scheme for which the congestion degree is to be calculated; is the index of the target dimension; 、 For the The maximum and minimum values ​​of all solutions under the target dimension; like is a boundary solution (i.e., there is no adjacent solution in a certain dimension), and its congestion distance is set to To ensure that the boundary solutions are prioritized in the screening; S320.4. Algorithm parameter settings: Setting population size , crossover probability , mutation probability , number of iterations , so that the non-dominated sorting genetic algorithm can efficiently generate high-quality Pareto frontier solution sets in employment path planning, achieving solution set diversity, convergence efficiency and computational feasibility; A new population is generated by genetic crossover (such as simulated binary crossover) and polynomial mutation operations, combined with an elite retention strategy (directly incorporating the previous frontier set solution into the new population) to ensure that the solution set converges quickly to the Pareto frontier.

[0041] As a further illustration of this embodiment, when using the NSGA-II algorithm to perform four-dimensional evaluation, this embodiment performs boundary processing on the evaluation indicators: When the model predicts a promotion cycle T = 0, it is automatically assigned the industry minimum cycle (e.g., the default for technical positions is 12 months). If the expected salary S of a solution exceeds the 90th percentile of the industry, it is calculated according to the 90th percentile. The OpenMP parallel framework is used to accelerate fast non-dominated sorting, dividing the solution set into four subsets based on the population size N = 100 to process dominance relationships in parallel. When calculating the congestion distance, the boundary solution is assigned the average distance of all solutions in that dimension × 2 (e.g., if the average distance in a dimension is 0.3, the boundary solution distance is 0.6) to prevent numerical overflow.

[0042] Interactive display unit 400, unit 400 intuitively presents the optimized employment path planning plan to college students through visual charts and text reports; receives feedback from college students on the plan in real time through an interactive interface; pushes personalized employment information based on the optimized employment path planning plan; and uses virtual reality technology to build simulated interview scenarios to assist college students in interview skills training and evaluation.

[0043] In this embodiment, the interactive display unit 400 includes a visual output module 410, an interactive feedback module 420, an information push module 430, and a VR interview training module 440, wherein: The visual output module 410 outputs the optimized employment path planning plan in the form of visual charts and text reports; The interactive feedback module 420 collects feedback information on the program from college students through the interactive interface; The information push module 430 pushes personalized employment information based on the optimized employment path planning scheme; As a further explanation of this embodiment, the information push module 430 in this embodiment implements personalized push based on collaborative filtering, specifically including the following steps: First, a scoring matrix is ​​constructed that combines user profiles (including job preferences, skill levels, and regional preferences) with job characteristics. Then, by the formula , calculate the user-job similarity; where, For users to Rating, is the user mean; For users and positions The Pearson correlation coefficient between them measures the similarity of the feature matching between the two; For users average preference ratings across all trait dimensions; For the position In the Attribute values ​​on feature dimensions; For the position The average attribute value across all feature dimensions; Finally, the top 5 most similar jobs are pushed, and the content is dynamically adjusted based on the plan revision request (for example, if the user feedback is "region mismatch", jobs in non-target regions are filtered out).

[0044] VR interview training module 440 uses virtual reality technology to construct simulated interview scenarios to assist in interview skills training and assessment.

[0045] In this embodiment, the visualization output module 410 includes a multi-dimensional chart generation submodule 411, a dynamic interactive rendering submodule 412, and a report structure generation submodule 413, wherein: The multi-dimensional chart generation submodule 411 is used to generate a job matching heat map, a three-dimensional Sankey diagram of career development paths (with nodes labeled with job names and skill keywords), and a salary expectation probability distribution diagram; As a further illustration of this embodiment, the multi-dimensional chart generation submodule 411 in this embodiment is based on a classic visualization algorithm adapted to the employment scenario, specifically including: Job matching heat map: through normalization formula , mapping colors (red high and blue low), where is the job matching probability output by the model, is the minimum value of the matching probability of all positions, The maximum matching probability of all positions, The color value of the node in the job matching heat map; the ForceAtlas2 algorithm is used to layout the nodes, setting the repulsion coefficient to 10, the attraction coefficient to 0.1, and the number of iterations to 500 to balance the node distribution; Career Development Path 3D Sankey Diagram: Node Size Definition ,in It is the career development potential, which is derived from the inverse of the promotion cycle; the edge width is directly related to the job matching degree , intuitively reflects the strength of association; Salary expectation probability distribution diagram: kernel density estimation is used to fit the distribution, and the bandwidth is adjusted to , is the standard deviation of wages, is the sample size, which is adapted to the discrete salary characteristics of the employment scenario.

[0046] The dynamic interactive rendering submodule 412 is used to perform three types of interactive operations: chart zooming, path node highlighting, and timeline dragging and browsing; As a further explanation of this embodiment, the dynamic interaction rendering submodule 412 in this embodiment defines the interaction rules as follows: Chart zoom: wheel scroll distance With scaling factor satisfy , linear mapping of scrolling amplitude and zoom factor; Node highlight: When the mouse hovers, the node transparency , the closer the distance, the higher the transparency, ranging from 0.5-1, focusing on the target node; Timeline dragging: monitor mouse movement events, update the timeline interval in real time, and link the display status of the Sankey chart and Gantt chart.

[0047] The report structured generation submodule 413 generates a text report according to the structure of plan priority sorting - key time node Gantt chart - skill improvement roadmap.

[0048] As a further illustration of this embodiment, the report structure generation submodule 413 in this embodiment uses the TOPSIS method to quantify the priority of the solutions, specifically including: First, extract job matching, development potential, salary satisfaction, and skill acquisition As an evaluation indicator; Then, the positive ideal solution is calculated , negative ideal solution ;in, For the The specific value of an evaluation indicator in a certain scheme; For the The operation of taking the maximum value of all the evaluation indicators; For the The operation of taking the minimum value of all the schemes for the evaluation index; Then, by , calculate the distance between the solution and the positive and negative ideal solutions, and the final score , sort the solutions in descending order of scores; Finally, the report structure follows: solution priority ranking (TOPSIS score) - key time node Gantt chart (generated by CPM algorithm) - skill improvement roadmap (gap value descending) to ensure logical coherence.

[0049] In this embodiment, the VR interview training module 440 includes, among others: The scenario dynamic generation submodule 441 loads the corresponding interview scenario based on the target position type (technical position / management position / marketing position) from a preset scenario template library constructed from public simulation cases of university career guidance centers and common interview scenario materials of enterprises; As a further explanation of this embodiment, the scene dynamic generation submodule 441 in this embodiment loads a preset scene based on the job type: Build a template library for technical (60%), management (30%), and marketing (10%) positions (derived from public cases and corporate materials from university career guidance centers); After the user selects the target position, 1 main scenario + 2 sub-scenarios are randomly selected (such as the required "algorithm question whiteboard deduction" for technical positions, supplemented by "code debugging" and "model deployment" sub-scenarios) to ensure training diversity.

[0050] The intelligent interviewer submodule 442 extracts core skill keywords based on the target job description, builds a follow-up question rule library, and generates follow-up questions in real time based on the matching between the interviewee's answers and the follow-up question rule library; As a further illustration of this embodiment, the intelligent interviewer submodule 442 in this embodiment implements intelligent follow-up questions through text analysis: Use the TF-IDF algorithm to extract the top five high-TF-IDF words in job descriptions (such as "deep learning" and "model optimization") and build a follow-up rule library; Calculate the edit distance between the answer and the rule base keyword. When the question is triggered (for example, if the answer contains "deep model" and the keyword is "deep learning", =1, triggering the “Model Optimization Method” follow-up question).

[0051] The multi-dimensional assessment submodule 443 generates a quantitative assessment report based on three dimensions: language expression, professional knowledge, and stress coping; As a further illustration of this embodiment, the multi-dimensional evaluation submodule 443 in this embodiment performs a quantitative evaluation from three dimensions, specifically including: Language expression: , where the weight 、 Based on the experience of language assessment experts; Expertise: , is the keyword weight, such as "deep learning" accounts for 0.3; Coping with stress: ,The hesitation duration can be detected by speech pauses; Comprehensive score: , weight 、 、 Statistics based on corporate interview scoring standards.

[0052] The training file management submodule 444 is used to store historical interview records, generate capability improvement curves, and compare and analyze interview performance at different stages.

[0053] As a further illustration of this embodiment, the training file management submodule 444 in this embodiment is used to record and analyze historical data: Interview videos, assessment reports, and voice texts are stored by the dimensions of "user ID + job type + training time"; Using linear regression model ,in, is the number of training times, Comprehensive score Generate capacity improvement curve, through the slope Determine the progress trend ( >0 indicates improvement); Extract language expression and professional knowledge scores at different stages, and generate a radar chart to compare skill changes.

[0054] Those skilled in the art will appreciate that the process of implementing all or part of the steps of the above embodiments may be accomplished by hardware, or by instructing related hardware through a program.

[0055] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven career path planning decision system for college students, characterized by: include: A data processing unit (100), the data processing unit (100) is connected to a data interface based on web crawler technology, is used to collect and pre-process personal, academic and employment market data of college students, and uses data filtering algorithms and format conversion rules to complete data cleaning and structured processing; A model building unit (200), wherein the model building unit (200) uses a convolutional neural network and a long short-term memory network architecture based on historical case data to train a deep learning model and build an employment path planning model; A scheme generation and optimization unit (300) generates an employment path planning scheme based on the data processed by the data processing unit (100) and the employment path planning model trained by the model building unit (200), and uses a non-dominated sorting genetic algorithm to evaluate the employment path planning scheme from four dimensions: job matching, career development potential, salary satisfaction, and regional adaptability to generate a Pareto frontier solution set, and optimizes the employment path planning scheme based on the evaluation results; and at the same time, uses a time trigger mechanism to update the parameters of the employment path planning model based on employment market dynamics; An interactive display unit (400), wherein the unit (400) intuitively displays the optimized employment path planning scheme to college students through visual charts and text reports; and receives feedback from college students on the scheme in real time through an interactive interface; Based on the optimized employment path planning plan, personalized employment information is pushed; with the help of virtual reality technology, simulated interview scenarios are constructed to assist college students in interview skills training and evaluation.

2. The artificial intelligence-driven career path planning decision system for college students according to claim 1 is characterized in that The data processing unit (100) includes a data acquisition module (110) and a data pre-processing module (120), wherein: The data collection module (110) is connected to the API interface of the enterprise recruitment website through the web crawler technology, and is used to collect college students' personal information, academic performance data, campus activity records, as well as job demand data in the employment market, industry salary level data and enterprise recruitment dynamic data; The data preprocessing module (120) is used to perform word segmentation and stop word removal on the collected unstructured text data, normalize the numerical data, and interpolate missing values.

3. The artificial intelligence-driven career path planning decision system for college students according to claim 2 is characterized in that: The data preprocessing module (120) performs word segmentation and stop word removal on the collected unstructured text data, normalizes the numerical data, and interpolates the missing values, including the following steps: S130.

1. Perform outlier filtering on the collected raw data and dynamically calculate the neighborhood radius parameter range based on the minimum distance and average distance between samples in the dataset. ; At the same time, according to the total number of samples N, determine the minimum sample number range ; Traversal via grid search and The optimal parameters that can make the core point ratio adapt to the density distribution of the data set are selected by combining the values ​​of For each sample point , statistics of Number of samples in the neighborhood , calculate the abnormality : ； When the abnormality When the value is lower than the statistical distribution threshold of the core point abnormality under the optimal parameters, it is determined to be an outlier and removed; S130.

2. Perform word segmentation, denoising, and semantic mapping on the cleaned data containing text content, where: Word segmentation and denoising: Use word segmentation tools to break down text into a set of terms , filter stop words and punctuation; Semantic mapping: A professional terminology database based on public professional classification standards. Perform the following matches: like Concepts in the terminology database Completely consistent, matching score =1; If there is a character difference, calculate the Edit distance, matching score =0.6; When there is no match =0; By formula: , mapping the text to a structured professional knowledge graph; where, For term Term frequency - inverse document frequency; S130.

3. Process numerical and categorical data, including: Numerical data standardization: Z-score standardization method is used to eliminate dimensional differences. The formula is: ;in, For the The sample in The original value of the numerical feature, is the mean of all samples of this feature, is the standard deviation; It is the absolute value of the numerical feature after Z-score standardization; Data coding by type: for the characteristics of academic qualifications, majors, and school types, set them to include Mutually exclusive categories , the vector dimension after one-hot encoding is ; If the i-th sample belongs to the category , then generate -dimensional binary vector , and satisfy: ； in, is the dimension index of the binary vector generated by one-hot encoding; is the category index; Structured feature matrix construction: Assume that the data matrix after outlier filtering is , the text semantic mapping feature matrix is , the standardized numerical feature matrix is , the one-hot encoding classification feature matrix is ;in, is the feature dimension retained after filtering, is the text feature dimension, is the number of numerical features, is the sum of the encoding dimensions of all categorical features; the above matrices are integrated through column concatenation to obtain the final structured feature matrix: ； in, is the total number of samples.

4. The artificial intelligence-driven career path planning decision system for college students according to claim 1 is characterized in that: The model building unit (200) includes a case preprocessing module (210), a feature adaptation module (220), a dual network fusion module (230) and a prediction output module (240), wherein: The case pre-processing module (210) is used to clean and remove duplicate historical employment case data and mark key nodes of career development stages; The feature adaptation module (220) is used to convert the labeled data into a feature sequence adapted to the input format of the convolutional neural network and the long short-term memory network; The dual network fusion module (230) extracts spatial correlations between features through a convolutional neural network and captures the temporal laws of career development through a long short-term memory network; The prediction output module (240) outputs prediction results of job matching and career development cycle based on the fusion features.

5. The artificial intelligence-driven career path planning decision system for college students according to claim 4 is characterized in that: The model building unit (200) trains the deep learning model including the following steps: S250.

1. Stratify historical case data by profession, education level, and region to ensure that each layer of data is representative of the corresponding group and construct a training sample set; S250.

2. Use a cross-validation strategy to divide the training set into a validation set, iteratively adjust the model parameters, and dynamically modify the training strategy based on the performance changes of the validation set to enhance the model's generalization ability. S250.

3. Based on the differences in the number of samples in different professional fields, dynamically adjust the weight of each field in the training loss calculation to balance the impact of uneven sample distribution on model training.

6. The artificial intelligence-driven career path planning decision system for college students according to claim 1 is characterized in that: The solution generation optimization unit (300) includes an initial solution generation module (310), a multi-objective evaluation module (320), a Pareto solution set optimization module (330) and a dynamic update module (340), wherein; The initial solution generation module (310) generates an initial employment path solution set including job sequence, skill improvement plan, and time nodes based on the data processed by the data processing unit (100) and the employment path planning model trained by the model building unit (200). , is the number of options; The multi-objective evaluation module (320) uses a non-dominated sorting genetic algorithm to perform a multi-objective evaluation on the initial employment path plan from four dimensions: job matching, career development potential, salary satisfaction, and regional adaptability, through fast non-dominated sorting and crowding distance calculation; The Pareto solution set optimization module (330) selects the Pareto optimal solution combination from the non-dominated frontier solution set based on the congestion distance; The dynamic update module (340) updates the employment path planning model parameters through a time trigger mechanism based on employment market dynamic data to maintain the adaptability of the employment path planning model.

7. The artificial intelligence-driven career path planning decision system for college students according to claim 6 is characterized in that: The multi-objective evaluation module (320) uses a non-dominated sorting genetic algorithm to perform multi-objective evaluation, which includes the following steps: S320.

1. Quantification of evaluation indicators: Convert the four evaluation dimensions into calculable numerical indicators, including job matching , career development potential , salary satisfaction and regional adaptability ; S320.1, Fast Non-Dominated Sorting: For any two solutions 、 , define the dominance relationship: ； in, For the plan In the goals; Representation scheme Control Plan ; Indicates that for all ; Indicates that there is at least one ; like Dominate ,but Domination count Add 1, The dominated set Include ; Divide the Pareto frontier layer by layer according to the dominance count ,in is the set of non-dominated solutions, The frontier set of the last layer of non-dominated sorting stratification; S320.

3. Congestion distance calculation: For the same frontier set The solution in , sorting by each dimension and then calculating the crowding distance to maintain the diversity of the solution set: ； in, For the plan The crowding distance; 、 for In dimension Adjacent solutions on is the specific scheme for which the congestion degree is to be calculated; is the index of the target dimension; 、 For the The maximum and minimum values ​​of all solutions under the target dimension; like is the boundary solution, and its congestion distance is set to To ensure that the boundary solutions are prioritized in the screening; S320.

4. Algorithm parameter settings: Setting population size , crossover probability , mutation probability , number of iterations , so that the non-dominated sorting genetic algorithm can efficiently generate high-quality Pareto frontier solution sets in employment path planning, achieving solution set diversity, convergence efficiency and computational feasibility; A new population is generated through genetic crossover and polynomial mutation operations, combined with an elite retention strategy to ensure that the solution set converges quickly to the Pareto frontier.

8. The artificial intelligence-driven career path planning decision system for college students according to claim 1 is characterized in that: The interactive display unit (400) includes a visual output module (410), an interactive feedback module (420), an information push module (430) and a VR interview training module (440), wherein: The visual output module (410) outputs the optimized employment path planning scheme in the form of visual charts and text reports; The interactive feedback module (420) collects feedback information on the program from college students through an interactive interface; The information push module (430) pushes personalized employment information based on the optimized employment path planning scheme; The VR interview training module (440) uses virtual reality technology to construct a simulated interview scene to assist in interview skills training and assessment.

9. The artificial intelligence-driven career path planning decision system for college students according to claim 8 is characterized in that: The visualization output module (410) includes a multi-dimensional chart generation submodule (411), a dynamic interactive rendering submodule (412), and a report structure generation submodule (413), wherein: The multi-dimensional chart generation submodule (411) is used to generate a job matching heat map, a three-dimensional Sankey diagram of career development paths, and a salary expectation probability distribution map; The dynamic interactive rendering submodule (412) is used to perform three types of interactive operations: chart zooming, path node highlighting, and timeline dragging and browsing; The report structured generation submodule (413) generates a text report according to the structure of program priority sorting-key time node Gantt chart-skill improvement roadmap.

10. The artificial intelligence-driven career path planning decision system for college students according to claim 9 is characterized in that: The VR interview training module (440) includes, among others: The dynamic scene generation submodule (441) loads the corresponding interview scene from a preset scene template library constructed from public simulation cases of university employment guidance centers and common interview scene materials of enterprises based on the target position type; The intelligent interviewer submodule (442) extracts core skill keywords based on the target job description, constructs a follow-up question rule library, and generates follow-up questions in real time according to the matching between the interviewee's answer and the follow-up question rule library; The multi-dimensional assessment submodule (443) generates a quantitative assessment report based on three dimensions: language expression, professional knowledge, and stress coping; The training file management submodule (444) is used to store historical interview records, generate capability improvement curves, and compare and analyze interview performances at different stages.

Images