Particle swarm big language model-based higher education vocational planning information pushing method and system

Abstract

The invention provides a higher education occupational planning information pushing method and system based on a particle swarm big language model, and relates to the technical field of occupational planning. According to the method, multi-source heterogeneous information such as school status data, post requirements and industry development is fused, node representation vectors are constructed, and student portraits, occupational features and college resource vectors are generated based on improved BERT-DeepWalk. And dynamic optimization of a parameter space is realized by introducing a large language model of quantum particle swarm optimization. Finally, model output and node representation vector reasoning are combined, and a self-adaptive and personalized occupational planning recommendation strategy is formed. According to the invention, scientificity, precision and timeliness of professional planning information pushing are improved, and intelligent employment guidance requirements of college students are met.

Description

Technical Field

[0001] This invention relates to the field of career planning technology, and in particular to a method and system for pushing career planning information in higher education based on a particle swarm optimization large language model. Background Technology

[0002] Current career planning systems in higher education mostly employ recommendation methods based on rule engines or traditional machine learning models (such as collaborative filtering and decision trees). Typical systems achieve information delivery by matching user profiles with static databases, while some advanced systems incorporate NLP technology to parse career description text. However, existing technologies suffer from three major bottlenecks: 1) the lag in updating dynamic market information leads to insufficient timeliness of recommendations; 2) multi-dimensional decision factors (ability matching, development potential, market trends) are difficult to effectively integrate; and 3) the ability to predict the continuity of long-term career development paths is lacking.

[0003] Traditional particle swarm optimization (PSO) algorithms are prone to getting trapped in local optima during parameter optimization and cannot adapt to the high-dimensional parameter space of large language models. Existing large language models suffer from three major application shortcomings in career planning: 1) insufficient integration of domain knowledge leads to a lack of professional advice; 2) the absence of a dynamic adjustment mechanism makes the recommendations rigid; and 3) weak multi-objective optimization capabilities make it difficult to balance student individual characteristics with actual market demands. Typical systems such as CN119273197A, which use a career matching model, have not addressed the problem of adaptability to dynamic environments. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for pushing career planning information in higher education based on a particle swarm optimization large language model. This method achieves multi-source data fusion, dynamic optimization, accurate matching, and adaptive pushing for career planning in higher education, significantly improving the scientific nature, personalization, and timeliness of career planning information recommendations.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for pushing career planning information in higher education based on a particle swarm optimization large language model includes:

[0007] Real-time acquisition of student registration data from the Ministry of Education, job posting data from recruitment platforms, industry association development reports, and enterprise hiring standards databases yields multi-source heterogeneous datasets.

[0008] The multi-source heterogeneous datasets are correlated to establish node representation vectors;

[0009] Based on the node representation vector, the improved BERT-DeepWalk algorithm is used to generate student profile vectors, occupational feature vectors, and school resource vectors.

[0010] The student profile vector, the occupational feature vector, and the college resource vector are used as the input embedding layer of the preset large language model. Random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions are set simultaneously in the parameter space of the preset large language model to obtain the optimized large language model.

[0011] The output of the optimized large language model is combined with the node representation vector for joint reasoning to obtain the recommendation strategy.

[0012] Preferably, the multi-source heterogeneous datasets are associated to establish node representation vectors, including:

[0013] The multi-source heterogeneous dataset is constructed into a heterogeneous network containing three types of nodes: student nodes, job nodes, and industry event nodes. The student nodes are associated with acquired skills and the time they have mastered them. The job nodes are associated with required skills and their weights in job descriptions. The industry event nodes are associated with the affected industry sectors and their duration of influence.

[0014] The heterogeneous network is defined into three types of connections: the mastery relationship between students and skills, the demand relationship between job positions and skills, and the influence relationship between industry events and job positions.

[0015] Data time slices are divided into quarters, and time-series connections are established for the same nodes across data time slices. The connection weight decreases exponentially as the time interval increases.

[0016] Subnetworks are constructed under each industry category, and spatial attention weights between nodes are calculated to generate node representation vectors with spatiotemporal coordinates; the attention weights are determined by the similarity of node features; the temporal dimension value of the node representation vector is calculated based on the logarithmic decay relationship between the node creation time and the current time; the spatial dimension value is generated through a hash mapping of the industry category code.

[0017] Preferably, the formula for calculating the node representation vector is:

[0018]

[0019] Where K is the total number of related skills, w k Let φ(s) be the composite weight of the k-th skill. k ) represents the skill semantic encoding value, σ 2 Let be the mean variance of the skill vectors within the industry cluster, ∈ be the smoothing constant to prevent zero, η be the dynamic scaling factor, Δt be the node survival time, τ be the industry benchmark update cycle, λ be the decay rate of market popularity correlation, C be the number of related industry events, and I be the mean variance of the skill vectors within the industry cluster. c Let ψ(h) represent the influence intensity of the c-th event. c) represents the industry fractal coding value, d represents the distance between industry cluster centers, and γ represents the spatial decay index.

[0020] Preferably, based on the node representation vector, an improved BERT-DeepWalk algorithm is used to generate student profile vectors, occupational feature vectors, and school resource vectors, including:

[0021] The node representation vectors are divided into three types of entities: student entities, occupational entities, and school entities. The student entities carry student ID numbers and skill mastery time series. The occupational entities include job codes, competency requirement matrices, and industry-related paths. The school entities include associated course codes, teaching resource distribution, and employment matching index.

[0022] A type identifier code is created for each entity type, and six heterogeneous path patterns are defined. These six heterogeneous path patterns include: SSO path, CIC path, CEC path, SFC path, CTC path, and CAC path. The SSO path's transmission direction is: Student → Skill ← Career; the weight of the SSO path is the product of skill matching degree and time decay factor. The CIC path's transmission direction is: Career → Industry → Institution; the weight of the CIC path is the product of industry relevance and resource abundance index. The CEC path's transmission direction is: Institution → Course → Career; the weight of the CEC path is the product of course articulation coefficient and employment rate. The SFC path's transmission direction is: Student ← Feedback → Career; the weight of the SFC path is the product of satisfaction score and feedback timeliness coefficient. The CTC path's transmission direction is: Career ← Trend → Career; the weight of the CTC path is job evolution similarity. The CAC path's transmission direction is: Institution → Alumni → Career; the weight of the CAC path is alumni achievement index.

[0023] The semantic similarity between nodes is calculated based on the node representation vectors, and a three-layer weight matrix is ​​generated. The three-layer weight matrix includes a structural weight matrix, a semantic weight matrix, and a time-dependent weight matrix. The structural weight matrix is ​​constructed based on the occurrence frequency of meta-paths. The semantic weight matrix is ​​constructed based on the node similarity calculation results. The time-dependent weight matrix is ​​constructed based on the exponential decay value of the time axis coordinate difference of the node representation vectors. Based on the defined six types of heterogeneous meta-path patterns, a heterogeneous random walk is performed to obtain a walk sequence.

[0024] The walking sequence is input into the improved BERT model to obtain the student profile vector, the occupational feature vector, and the school resource vector.

[0025] Preferably, the heterogeneous random walk specifically involves: when encountering a student node, selecting the SSO and SFC paths; when encountering a professional node, emphasizing the CTC and CEC paths; and when encountering a college node, strengthening the CIC and CAC paths.

[0026] Preferably, the improved BERT model adds spatiotemporal coordinate attention gating to the Transformer layer; the improved BERT model introduces multi-task learning in the pre-training stage; the multi-task learning includes mask node type prediction, meta-path pattern classification, and spatiotemporal coordinate regression.

[0027] Preferably, the student profile vector, the occupational feature vector, and the institution resource vector are used as the input embedding layer of a preset large language model. Simultaneously, random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions are set within the parameter space of the preset large language model to perform quantum particle swarm optimization, resulting in an optimized large language model, including:

[0028] The student profile vector, the occupational feature vector, and the school resource vector are concatenated according to their dimensions to form a joint feature vector;

[0029] An adjustable parameter interface is inserted between the input embedding layer and the attention layer of the preset large language model. The interface dimension is aligned with the particle swarm position encoding dimension. N particle positions are sampled in a uniform distribution within the parameter space of the fully connected layer of the preset large language model.

[0030] Extract a set of particles that meet preset conditions from historical optimization records; the preset conditions include: elite particles with the top 10% fitness ranking, particle distribution covering the parameter space of the most recent 3 training cycles, and particles carrying a time decay factor.

[0031] The cosine similarity between the student profile vector and the occupation feature vector is calculated to obtain the ability matching degree. Based on the temporal attention weight, the area under the demand growth curve of the target position in the next 36 months is predicted by LSTM to obtain the development potential value. The market heat index is obtained by statistically analyzing the job increment and the rate of change of industry event influence intensity in the target industry in the last 24 hours.

[0032] A time-varying fitness function is constructed based on the capability matching degree, the development potential value, and the market popularity index: The calculation formula of the time-varying fitness function F(t) is: F(t)=α(t)·Capability matching degree +β(t)·Development potential value +γ(t)·Market popularity index; where α(t), β(t), and γ(t) are preset dynamic weight coefficients;

[0033] The position and velocity of each particle are encoded into Bloch spherical coordinates, and quantum rotation gate operation and quantum entanglement optimization are performed to establish entanglement pairs between the top 5% of the fitness particles in the guided particle swarm and the bottom 20% of the random particle swarm in each iteration, sharing the position update gradient.

[0034] When the rate of change of the global optimal fitness is less than 1% and the proportion of guiding particles is greater than 60% in three consecutive iterations, it is determined to be converged, and the optimized large language model is obtained.

[0035] Preferably, the output of the optimized large language model is jointly inferred with the node representation vector to obtain an initial recommendation strategy, including:

[0036] The semantic vector set and node representation vector set output by the optimized large language model are transformed to a common vector space of the same dimension through linear mapping.

[0037] Within the common vector space, using the set of semantic vectors as the query and the set of node representation vectors as the key / value pair, the cross-attention weight matrix is ​​calculated to obtain the fused semantic representation.

[0038] Multi-head attention is applied to the fused semantic representation, context dependencies are preserved, and the inter-node structural weight matrix is ​​invoked to perform graph convolution update on the set of node representation vectors;

[0039] The set of node representation vectors and the updated set of node representation vectors are concatenated as column vectors, and then passed through a fully connected layer to output the fused vector F_fuse. i ;

[0040] For the candidate target set, a scoring function is used. i =Softmax(W_s·F_fuse) i +b_s) yields the score sequence. i Where W_s and b_s are learnable parameters;

[0041] According to score i The top K candidate targets are selected in descending order, and this is denoted as the recommendation strategy.

[0042] Preferably, it further includes:

[0043] The optimal push recommendation strategy is obtained by refining the recommendation strategy based on the acquired user interaction behavior. A higher education career planning information push system based on a particle swarm optimization large language model includes:

[0044] The dataset construction unit is used to acquire student registration data from the Ministry of Education, job posting data from recruitment platforms, industry association development reports, and enterprise employment standards in real time to obtain multi-source heterogeneous datasets.

[0045] A vector association unit is used to associate the multi-source heterogeneous datasets to establish node representation vectors;

[0046] The vector generation unit is used to generate student profile vectors, occupational feature vectors, and school resource vectors based on the node representation vectors using the improved BERT-DeepWalk algorithm.

[0047] The model optimization unit is used to take the student profile vector, the occupational feature vector and the college resource vector as the input embedding layer of the preset large language model, and simultaneously set random particle swarm and guided particle swarm based on historical optimal solutions in the parameter space of the preset large language model to perform quantum particle swarm optimization to obtain the optimized large language model.

[0048] The reasoning unit is used to jointly reason with the output of the optimized large language model and the node representation vector to obtain a recommendation strategy.

[0049] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] This invention provides a method and system for pushing career planning information in higher education based on a particle swarm optimization (PSO) large language model. The method includes: acquiring real-time student registration data from the Ministry of Education, job postings from recruitment platforms, industry association development reports, and enterprise recruitment standards databases to obtain a multi-source heterogeneous dataset; associating the multi-source heterogeneous datasets to establish node representation vectors; generating student profile vectors, career feature vectors, and institution resource vectors based on the node representation vectors using an improved BERT-DeepWalk algorithm; using the student profile vectors, career feature vectors, and institution resource vectors as the input embedding layer of a pre-defined large language model, and simultaneously setting up random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions within the parameter space of the pre-defined large language model to obtain an optimized large language model; and jointly reasoning with the output of the optimized large language model and the node representation vectors to obtain a recommendation strategy. This invention achieves multi-source data fusion, dynamic optimization, accurate matching, and adaptive push for career planning in higher education, significantly improving the scientific rigor, personalization, and timeliness of career planning information recommendations. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The purpose of this invention is to provide a method and system for pushing career planning information in higher education based on a particle swarm optimization large language model. This method achieves multi-source data fusion, dynamic optimization, accurate matching, and adaptive pushing for career planning in higher education, significantly improving the scientific nature, personalization, and timeliness of career planning information recommendations.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a method for pushing higher education career planning information based on a particle swarm optimization large language model, including:

[0058] Step 100: Obtain real-time student registration data from the Ministry of Education, job posting data from recruitment platforms, industry association development reports, and enterprise hiring standards databases to obtain a multi-source heterogeneous dataset;

[0059] Step 200: Associate the multi-source heterogeneous datasets to establish node representation vectors;

[0060] Step 300: Based on the node representation vector, generate student profile vector, occupational feature vector, and school resource vector using the improved BERT-DeepWalk algorithm;

[0061] Step 400: Use the student profile vector, the occupational feature vector, and the school resource vector as the input embedding layer of the preset large language model, and simultaneously set up random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions in the parameter space of the preset large language model to obtain the optimized large language model.

[0062] Step 500: Perform joint reasoning between the output of the optimized large language model and the node representation vector to obtain the recommendation strategy.

[0063] Specifically, in a preferred embodiment of the present invention, step 100 includes: the system establishing API or data interface connections with the Ministry of Education's student registration management platform, major recruitment service platforms, data interfaces of major industry associations, and enterprise recruitment standard databases, respectively, and pulling various types of raw data in real time through a scheduled task module. Student registration data uses the student's unique ID as the primary key to collect personal identity, academic history, and skills certification information; job data includes text and structured fields such as job codes, job descriptions, required skills, and job requirements; industry association development reports automatically synchronize industry classifications, development trends, and key event data quarterly; and the enterprise recruitment standard database compiles job competency matrices and recruitment requirements released by major enterprises quarterly. After collection, all raw data is first stored in a heterogeneous data buffer through a data access layer, with the data source and collection time marked, achieving real-time aggregation of multi-source heterogeneous data.

[0064] To address the aforementioned multi-source heterogeneous data and ensure its standardization and fusion capabilities, this embodiment employs a unified format conversion module to perform structured processing on various data types after data access. For example, student registration data and job posting data are standardized into JSON objects containing fields such as primary key ID, category label, and timestamp, while text content undergoes word segmentation and noise reduction. The converted data is then sent to the data fusion layer, where synchronized timestamp information and source identifiers are integrated for subsequent construction of the 3D knowledge graph and node representation vectors. This process ensures that the present invention can efficiently and stably acquire and manage multi-source heterogeneous datasets, laying a data foundation for subsequent career planning information push algorithms.

[0065] Preferably, the multi-source heterogeneous datasets are associated to establish node representation vectors, including:

[0066] The multi-source heterogeneous dataset is constructed into a heterogeneous network containing three types of nodes: student nodes, job nodes, and industry event nodes. The student nodes are associated with acquired skills and the time they have mastered them. The job nodes are associated with required skills and their weights in job descriptions. The industry event nodes are associated with the affected industry sectors and their duration of influence.

[0067] The heterogeneous network is defined into three types of connections: the mastery relationship between students and skills, the demand relationship between job positions and skills, and the influence relationship between industry events and job positions.

[0068] Data time slices are divided into quarters, and time-series connections are established for the same nodes across data time slices. The connection weight decreases exponentially as the time interval increases.

[0069] Subnetworks are constructed under each industry category, and spatial attention weights between nodes are calculated to generate node representation vectors with spatiotemporal coordinates; the attention weights are determined by the similarity of node features; the temporal dimension value of the node representation vector is calculated based on the logarithmic decay relationship between the node creation time and the current time; the spatial dimension value is generated through a hash mapping of the industry category code.

[0070] In a preferred embodiment of the present invention, the real-time collected data from the Ministry of Education's student registration system, recruitment platform job data, industry association development reports, and enterprise hiring standards databases are first preprocessed and uniformly structured to construct a heterogeneous network containing three types of nodes (i.e., student nodes, job nodes, and industry event nodes). Specifically, student nodes are accompanied by attribute tags including the names of the skills they have acquired and the mastery time of each skill; job nodes are associated with the required skills and the weight coefficient of each skill in the job description; and industry event nodes are associated with the industry category number of their impact coverage and the period range of event effectiveness / ineffectiveness based on report details. All nodes in the above heterogeneous network store unique identifiers and data source information of the original data, achieving full connectivity between entity data.

[0071] Secondly, based on the key business relationships that actually exist between the three types of nodes, three connection methods are defined within the heterogeneous network: First, the "mastery relationship" between students and skill nodes, where "mastery relationship" represents a student's acquisition or certification of a certain skill; second, the "demand relationship" between positions and skill nodes, which extracts the required skills through parsing and weighting of job description texts; and third, the "influence relationship" between industry events and position nodes, which establishes directional edges that influence positions by matching the domain type of industry events with the industry category of the positions involved. Each connection relationship is labeled with strength or weight, reflecting the actual importance or degree of relevance of the association.

[0072] Building upon the aforementioned heterogeneous network, time slices are established for the entire network data quarterly. Instances of the same node in different quarterly time slices are then temporally connected, with connection weights decaying exponentially with time intervals, modeling the temporal characteristics of nodes. Furthermore, sub-networks are independently constructed under each industry category label. Spatial attention weights between nodes are calculated using methods such as cosine similarity or Euclidean distance between node attributes, yielding weighted coefficients reflecting the degree of node feature aggregation. Finally, the system uses the creation time of each node, logarithmically decayed with the current system time, as its temporal coordinate value. Spatial coordinates are written into the node representation using a hash mapping of the industry category number, realizing a unique spatiotemporal embedding vector for each node. This provides a richly structured spatiotemporally semantic foundation for subsequent intelligent reasoning and recommendation algorithms based on the knowledge graph.

[0073] Preferably, the formula for calculating the node representation vector is:

[0074]

[0075] Where K is the total number of related skills, w k Let φ(s) be the composite weight of the k-th skill. k ) represents the skill semantic encoding value, σ 2 Let be the mean variance of the skill vectors within the industry cluster, ∈ be the smoothing constant to prevent zero, η be the dynamic scaling factor, Δt be the node survival time, τ be the industry benchmark update cycle, λ be the decay rate of market popularity correlation, C be the number of related industry events, and I be the mean variance of the skill vectors within the industry cluster. c Let ψ(h) represent the influence intensity of the c-th event. c ) represents the industry fractal coding value, d represents the distance between industry cluster centers, and γ represents the spatial decay index.

[0076] The calculation formula for the node representation vector in this invention integrates multiple factors such as skill weights, industry characteristics, temporal dynamics, and complex relationships of industry events, demonstrating the following innovations: First, the formula not only considers the hierarchical skill structure of entities such as students and positions, but also introduces the semantic distance between skills, the spatial characteristics of industry clustering, and the dynamic impact of industry events. Among these factors, the skill composite weight (wk) and semantic encoding... Mean variance (σ) 2 The combined regulation mechanism of parameters such as industry fractals and spatial decay can take into account both static attributes and dynamic changes, providing an interpretable high-dimensional feature expression for personalized career path recommendations. Secondly, the fusion of logarithmic and exponential decay of survival time, industry benchmark update cycle, and market popularity ensures that node representations are timely, adapting to market and industry trends and continuously optimizing model expression capabilities in real-time recommendations. Therefore, this vector calculation method combines multi-source knowledge integration, spatiotemporal dynamic modeling, and refined regulation, breaking through the limitations of existing industry career recommendation technologies based on single features or static weights.

[0077] The acquisition and calculation of each parameter are both data-operable and algorithmically innovative: the total number of associated skills (K), node survival time (Δt), and number of industry events (C) can be directly statistically analyzed using a three-dimensional knowledge graph; the composite weight (wk) of each skill can be weighted based on the skill acquisition channels recorded in the student database, the frequency of job requirements, and the semantic matching score; the skill semantic encoding value (φ(s)) k The industry cluster variance mean (σ) is obtained by embedding skill terms using deep language models such as BERT. 2The algorithm automatically calculates the variance average of skill vectors within the same industry. The smoothing constant (ε), dynamic scaling factor (η), and spatial decay index (γ) are preset and iteratively optimized through cross-validation and Bayesian tuning. Spatial parameters such as industry fractal coding and cluster center distance (d) can be directly calculated from industry label hash mapping and feature cluster center distance. The industry event impact strength (Ic) and market popularity correlation rate (λ) are dynamically allocated based on the latest industry reports and recruitment dynamics according to sovereign weights. The industry benchmark period (τ) is periodically set by industry associations and historical data. These parameters and structural design together ensure the authenticity, timeliness, and industry foresight of the recommended node vectors.

[0078] Preferably, based on the node representation vector, an improved BERT-DeepWalk algorithm is used to generate student profile vectors, occupational feature vectors, and school resource vectors, including:

[0079] The node representation vectors are divided into three types of entities: student entities, occupational entities, and school entities. The student entities carry student ID numbers and skill mastery time series. The occupational entities include job codes, competency requirement matrices, and industry-related paths. The school entities include associated course codes, teaching resource distribution, and employment matching index.

[0080] A type identifier code is created for each entity type, and six heterogeneous path patterns are defined. These six heterogeneous path patterns include: SSO path, CIC path, CEC path, SFC path, CTC path, and CAC path. The SSO path's transmission direction is: Student → Skill ← Career; the weight of the SSO path is the product of skill matching degree and time decay factor. The CIC path's transmission direction is: Career → Industry → Institution; the weight of the CIC path is the product of industry relevance and resource abundance index. The CEC path's transmission direction is: Institution → Course → Career; the weight of the CEC path is the product of course articulation coefficient and employment rate. The SFC path's transmission direction is: Student ← Feedback → Career; the weight of the SFC path is the product of satisfaction score and feedback timeliness coefficient. The CTC path's transmission direction is: Career ← Trend → Career; the weight of the CTC path is job evolution similarity. The CAC path's transmission direction is: Institution → Alumni → Career; the weight of the CAC path is alumni achievement index.

[0081] The semantic similarity between nodes is calculated based on the node representation vectors, and a three-layer weight matrix is ​​generated. The three-layer weight matrix includes a structural weight matrix, a semantic weight matrix, and a time-dependent weight matrix. The structural weight matrix is ​​constructed based on the occurrence frequency of meta-paths. The semantic weight matrix is ​​constructed based on the node similarity calculation results. The time-dependent weight matrix is ​​constructed based on the exponential decay value of the time axis coordinate difference of the node representation vectors. Based on the defined six types of heterogeneous meta-path patterns, a heterogeneous random walk is performed to obtain a walk sequence.

[0082] The walking sequence is input into the improved BERT model to obtain the student profile vector, the occupational feature vector, and the school resource vector.

[0083] Optionally, this embodiment first divides the constructed node representation vectors into three categories according to entity type: student entities, occupational entities, and school entities. Student entities include student registration numbers and time-series information on their skill mastery; occupational entities correspond to job codes, competency requirement matrices, and path information of associated industries; school entities include attributes such as course codes, distribution of on-campus teaching resources, and employment matching index. Each type of entity is assigned a unique type identifier code in the system to achieve efficient retrieval and management of entity data.

[0084] Subsequently, to achieve in-depth mining of network structure and relational semantics, six heterogeneous meta-path patterns were defined for the different entity categories mentioned above: SSO, CIC, CEC, SFC, CTC, and CAC paths. Each meta-path specifies a particular entity jump order and its business-driven weight function. For example, the SSO path moves along the "student → skill ← career" direction, with the weight being the product of skill matching degree and time decay factor. The CEC path is used for the connection analysis of "institution → course → career," and its weight depends on the product of course connection coefficient and employment rate. The role of all meta-paths is to reveal the multi-faceted connections between students, positions, and institutions in the complex education-employment network and to provide structural guidance for the probability distribution of wandering paths.

[0085] Furthermore, based on the node representation vectors, the system first calculates the semantic similarity between nodes, and then generates three types of weight matrices: a structural weight matrix (recording the frequency of occurrence of each type of metapath), a semantic weight matrix (automatically adjusting the walk probability based on the similarity between nodes), and a temporal weight matrix (applying exponential decay with the time coordinate difference in the node representation vector as the independent variable), achieving simultaneous modeling of spatial and temporal dimensions. These weight matrices provide multi-layered regulation for path selection and sequence sampling during subsequent walks, realizing a unified network representation foundation integrating structure, semantics, and temporality.

[0086] Finally, based on the defined six categories of heterogeneous metapaths, the system employs a heterogeneous random walk algorithm to collect multi-category walk sequences and transforms these sequences into training samples for input into the improved BERT model. The BERT model not only introduces spatiotemporal coordinate attention gating in the Transformer layer but also jointly performs node type mask prediction, metapath classification, and spatiotemporal regression tasks during the pre-training phase. The trained BERT model can generate high-dimensional vector representations for three types of entities: students, professions, and institutions—namely, student profile vectors, profession feature vectors, and institution resource vectors—providing strong support for downstream recommendation and personalized push notifications.

[0087] Preferably, the heterogeneous random walk specifically involves: when encountering a student node, selecting the SSO and SFC paths; when encountering a professional node, emphasizing the CTC and CEC paths; and when encountering a college node, strengthening the CIC and CAC paths.

[0088] Preferably, the improved BERT model adds spatiotemporal coordinate attention gating to the Transformer layer; the improved BERT model introduces multi-task learning in the pre-training stage; the multi-task learning includes mask node type prediction, meta-path pattern classification, and spatiotemporal coordinate regression.

[0089] Furthermore, the core structure of the improved BERT model adds a spatio-temporal attention gating mechanism to the standard Transformer layer. Gate), the gate unit can adjust the weights of the time coordinates (such as the node’s historical creation time, event occurrence time, etc.) and spatial coordinates (such as industry labels, school affiliation, etc.) of each pair of node inputs during self-attention calculation, and dynamically adjust the contribution of different node information fusion, thereby enhancing the model’s ability to model the spatiotemporal relationship between nodes; In the pre-training stage, the BERT model jointly introduces multi-task learning objectives, including (1) a mask node type prediction task, that is, randomly masking the node type label part in the input sequence, training the model to automatically predict the node type under context inference, and improving the model’s generalization ability and semantic understanding ability; (2) a meta-path pattern classification task, which enables the model to automatically identify various heterogeneous path patterns by classifying and training each sampled meta-path structure, thereby enhancing the semantic modeling effect of path-level structure; (3) a spatiotemporal coordinate regression task, which requires the model output to simultaneously predict the time coordinates and spatial codes of each node, and improve the model’s ability to perceive and represent temporal and spatial distribution features. Through the above structure and multi-task joint training, the final entity vectors can fully integrate the content, structure, and spatiotemporal semantic features of heterogeneous networks, significantly improving the accuracy and intelligence of subsequent career recommendations and information push.

[0090] Preferably, the student profile vector, the occupational feature vector, and the institution resource vector are used as the input embedding layer of a preset large language model. Simultaneously, random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions are set within the parameter space of the preset large language model to perform quantum particle swarm optimization, resulting in an optimized large language model, including:

[0091] The student profile vector, the occupational feature vector, and the school resource vector are concatenated according to their dimensions to form a joint feature vector;

[0092] An adjustable parameter interface is inserted between the input embedding layer and the attention layer of the preset large language model. The interface dimension is aligned with the particle swarm position encoding dimension. N particle positions are sampled in a uniform distribution within the parameter space of the fully connected layer of the preset large language model.

[0093] Extract a set of particles that meet preset conditions from historical optimization records; the preset conditions include: elite particles with the top 10% fitness ranking, particle distribution covering the parameter space of the most recent 3 training cycles, and particles carrying a time decay factor.

[0094] The cosine similarity between the student profile vector and the occupation feature vector is calculated to obtain the ability matching degree. Based on the temporal attention weight, the area under the demand growth curve of the target position in the next 36 months is predicted by LSTM to obtain the development potential value. The market heat index is obtained by statistically analyzing the job increment and the rate of change of industry event influence intensity in the target industry in the last 24 hours.

[0095] A time-varying fitness function is constructed based on the capability matching degree, the development potential value, and the market popularity index: The calculation formula of the time-varying fitness function F(t) is: F(t)=α(t)·Capability matching degree +β(t)·Development potential value +γ(t)·Market popularity index; where α(t), β(t), and γ(t) are preset dynamic weight coefficients;

[0096] The position and velocity of each particle are encoded into Bloch spherical coordinates, and quantum rotation gate operation and quantum entanglement optimization are performed to establish entanglement pairs between the top 5% of the fitness particles in the guided particle swarm and the bottom 20% of the random particle swarm in each iteration, sharing the position update gradient.

[0097] When the rate of change of the global optimal fitness is less than 1% and the proportion of guiding particles is greater than 60% in three consecutive iterations, it is determined to be converged, and the optimized large language model is obtained.

[0098] In a preferred embodiment of the present invention, student profile vectors, occupational feature vectors, and school resource vectors are first concatenated dimensionally to form a unified joint feature vector, which is then input into the embedding layer of a pre-defined large language model (such as a GPT-like model) to achieve deep fusion of multi-source information. To improve the model parameter optimization capability, an adjustable parameter interface is introduced between the input embedding layer and the first attention layer of the large language model structure. The dimension of this interface matches the particle swarm optimization position encoding, and N particles are sampled in a uniform distribution within the parameter space of the fully connected layer. Each particle represents a set of adjustable parameter vectors, and the parameter values ​​can be dynamically adjusted in each training round according to the particle swarm optimization algorithm mechanism.

[0099] Furthermore, this embodiment automatically selects a set of elite particles that meet preset conditions from the historical model optimization records, including: (1) the top 10% of historical best particles in terms of fitness; (2) particle distribution that can cover the parameter space of the most recent three complete training cycles; and (3) particle objects that need to carry parameters that reflect time decay characteristics. After the model calculates the current sample's ability matching degree (obtained by the cosine similarity between student profile and occupational feature vector), development potential value (using time-series attention weights plus LSTM sequence to predict the area of ​​the demand growth curve of occupational positions in the next 36 months), and market heat index (statistically calculating the 24-hour job increment and event impact intensity change rate of the target industry), a time-varying fitness function F(t) = α(t)·ability matching degree + β(t)·development potential value + γ(t)·market heat index is established based on these three core indicators. The three weights α(t), β(t), and γ(t) are coefficients that can be dynamically adjusted over time, so as to achieve adaptive adjustment of the optimization target for different time points and market environments.

[0100] In the quantum particle swarm optimization process, the position and velocity of each particle are encoded using Bloch spherical coordinates. Quantum rotation gate operations are used to adjust global and local search capabilities. At the same time, a quantum entanglement mechanism between particles is introduced. In each optimization iteration, the top 5% of particles in the guided particle swarm (historical elites, those with high fitness) are paired with the bottom 20% of particles in the random particle swarm to establish entanglement pairs. This enables cross-swarm sharing of parameter space gradients, effectively breaking through the bottleneck of traditional particle swarm optimization being prone to getting trapped in local optima, and improving the parameter optimization efficiency and multi-objective search capability of large language models.

[0101] Finally, this embodiment monitors the global optimal fitness in each optimization iteration. When the rate of change of the global optimal fitness is less than 1% in three consecutive iterations, and the proportion of guiding particles among all high-fitness particles exceeds 60%, it is considered that the current large language model parameters have sufficiently approached the global optimum, the quantum particle swarm optimization process converges, and the final optimized large language model is obtained. This model structure integrates diverse heterogeneous data information and dynamic, intelligent optimization mechanisms, significantly improving the matching accuracy and market adaptability of career planning information push.

[0102] Preferably, the output of the optimized large language model is jointly inferred with the node representation vector to obtain an initial recommendation strategy, including:

[0103] The semantic vector set and node representation vector set output by the optimized large language model are transformed to a common vector space of the same dimension through linear mapping.

[0104] Within the common vector space, using the set of semantic vectors as the query and the set of node representation vectors as the key / value pair, the cross-attention weight matrix is ​​calculated to obtain the fused semantic representation.

[0105] Multi-head attention is applied to the fused semantic representation, context dependencies are preserved, and the inter-node structural weight matrix is ​​invoked to perform graph convolution update on the set of node representation vectors;

[0106] The set of node representation vectors and the updated set of node representation vectors are concatenated as column vectors, and then passed through a fully connected layer to output the fused vector F_fuse. i ;

[0107] For the candidate target set, a scoring function is used. i =Softmax(W_s·F_fuse) i +b_s) yields the score sequence. i Where W_s and b_s are learnable parameters;

[0108] According to score i The top K candidate targets are selected in descending order, and this is denoted as the recommendation strategy.

[0109] In a preferred embodiment of the present invention, the set of semantic vectors output by the optimized large language model and the set of node representation vectors generated by the system are first mapped to a common vector space of the same dimension through independent linear mapping layers (such as fully connected layers). The mapping parameters are optimized synchronously during end-to-end training to ensure that data representations from different sources can be effectively aligned in the feature space and achieve homogeneous comparison and computation.

[0110] Subsequently, within the common vector space, the set of semantic vectors output by the large language model is used as the query, and the set of node representation vectors are used as the key and value, respectively. An attention weight matrix is ​​calculated using the standard cross-attention mechanism, and the value is then weighted and summed accordingly to obtain the fused semantic representation. This fused semantic representation reflects both the large language model's high-order understanding of the input task and its ability to automatically learn dynamic relationships with the knowledge graph entity structure during interaction.

[0111] Building upon the fusion of semantic representations, the system further employs a multi-head attention mechanism to preserve contextual dependencies from different perspectives and enhance the expressive power of complex entity relationships. Simultaneously, utilizing a pre-constructed inter-node structural weight matrix, a graph convolutional network (GCN) update operation is performed on the node representation vector set, achieving adaptive aggregation and feature enhancement of node representations to graph structural information. This effectively captures the implicit path relationships between entities and the aggregate semantics of adjacent nodes.

[0112] Finally, the original set of node representation vectors and the updated set of node representation vectors after graph convolution are concatenated column-wise to form the final fused vector. For each candidate target entity, the fused vector is used as input to calculate the ranking score of the candidate target through a scoring function. The top K targets are selected in descending order of score to generate an initial recommendation strategy, providing users with optimized personalized recommendation results.

[0113] Preferably, it further includes:

[0114] By refining the recommendation strategy based on the user's interaction behavior, the optimal push recommendation strategy can be obtained.

[0115] In a preferred embodiment of the present invention, after the recommendation strategy is pushed, the system continuously collects various user interaction behavior data, including but not limited to clicks, favorites, applications, feedback, dwell time, and interest tag switching. Each interaction behavior is assigned a pre-set weight coefficient, and the system will automatically generate a behavior feature vector based on the user behavior sequence over a recent period, and dynamically trace back historical interactions to build a rich user preference model and real-time interest shift trends.

[0116] Furthermore, based on actual user interaction feedback, the system adjusts the scoring distribution of all pushed targets in the initial recommendation strategy. Specifically, for targets with click-through rates or application rates significantly higher than the average, their weight scores are dynamically increased; for targets that are frequently skipped, ignored, or have explicit negative feedback, their subsequent priority in the recommendation list is appropriately reduced. Simultaneously, the system can also employ reinforcement learning methods (such as policy gradient or Bandit algorithms) to fine-tune the parameters of the scoring function online, achieving continuous adaptive optimization of the recommendation strategy parameters.

[0117] Finally, after user interaction data correction and parameter adaptive adjustment, this embodiment will generate and update the optimal push recommendation strategy for specific users in real time. This strategy not only fully integrates historical large models, knowledge graphs, and cold start information, but also keenly responds to users' personalized preferences and current behavioral dynamics, achieving higher user acceptance and satisfaction in subsequent push iterations, and providing closed-loop optimization support for career planning recommendation applications.

[0118] like Figure 2 As shown, corresponding to the above method, this embodiment also provides a higher education career planning information push system based on particle swarm optimization large language model, including:

[0119] The dataset construction unit is used to acquire student registration data from the Ministry of Education, job posting data from recruitment platforms, industry association development reports, and enterprise employment standards in real time to obtain multi-source heterogeneous datasets.

[0120] A vector association unit is used to associate the multi-source heterogeneous datasets to establish node representation vectors;

[0121] The vector generation unit is used to generate student profile vectors, occupational feature vectors, and school resource vectors based on the node representation vectors using the improved BERT-DeepWalk algorithm.

[0122] The model optimization unit is used to take the student profile vector, the occupational feature vector and the college resource vector as the input embedding layer of the preset large language model, and simultaneously set random particle swarm and guided particle swarm based on historical optimal solutions in the parameter space of the preset large language model to perform quantum particle swarm optimization to obtain the optimized large language model.

[0123] The reasoning unit is used to jointly reason with the output of the optimized large language model and the node representation vector to obtain a recommendation strategy.

[0124] The beneficial effects of this invention are as follows:

[0125] (1) This invention can acquire and integrate multi-source heterogeneous information such as the Ministry of Education's student registration data, recruitment data, industry reports and enterprise employment standards in real time. Through three-dimensional knowledge graph modeling, it can realize multi-dimensional and spatiotemporal joint representation of multiple entities such as students, positions and industry events, which greatly improves the objectivity and comprehensiveness of information push.

[0126] (2) The introduction of quantum particle swarm optimization large language model can dynamically adjust the model parameters during the career planning information push process, so that the recommendation results can evolve in real time according to market changes and user behavior, taking into account students' interests, abilities and job requirements, and greatly improving personalization and adaptability.

[0127] (3) This invention comprehensively evaluates and recommends strategies through multi-objective fitness functions such as ability matching degree, development potential value and market popularity index, effectively balancing student development, job requirements and industry trends, which is superior to the single or static decision-making method of traditional models.

[0128] (4) This invention continuously corrects the recommendation strategy through real-time user interaction feedback, realizes the adaptive iteration of the push system, effectively avoids information redundancy and recommendation rigidity, and improves the continuous intelligent evolution capability of the recommendation system in practical applications.

[0129] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0130] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for pushing career planning information in higher education based on a particle swarm optimization large language model, characterized in that, include: Real-time acquisition of student registration data from the Ministry of Education, job posting data from recruitment platforms, industry association development reports, and enterprise hiring standards databases yields multi-source heterogeneous datasets. The multi-source heterogeneous datasets are correlated to establish node representation vectors; Based on the node representation vector, the improved BERT-DeepWalk algorithm is used to generate student profile vectors, occupational feature vectors, and school resource vectors. The student profile vector, the occupational feature vector, and the college resource vector are used as the input embedding layer of the preset large language model. Random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions are set simultaneously in the parameter space of the preset large language model to obtain the optimized large language model. The output of the optimized large language model is combined with the node representation vector for joint reasoning to obtain the recommendation strategy.

2. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 1, characterized in that, Associating the aforementioned multi-source heterogeneous datasets to establish node representation vectors includes: The multi-source heterogeneous dataset is constructed into a heterogeneous network containing three types of nodes: student nodes, job nodes, and industry event nodes. The student nodes are associated with acquired skills and the time they have mastered them. The job nodes are associated with required skills and their weights in job descriptions. The industry event nodes are associated with the affected industry sectors and their duration of influence. The heterogeneous network is defined into three types of connections: the mastery relationship between students and skills, the demand relationship between job positions and skills, and the influence relationship between industry events and job positions. Data time slices are divided into quarters, and time-series connections are established for the same nodes across data time slices. The connection weight decreases exponentially as the time interval increases. Subnetworks are constructed under each industry category, and spatial attention weights between nodes are calculated to generate node representation vectors with spatiotemporal coordinates; the attention weights are determined by the similarity of node features; the temporal dimension value of the node representation vector is calculated based on the logarithmic decay relationship between the node creation time and the current time; the spatial dimension value is generated through a hash mapping of the industry category code.

3. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 2, characterized in that, The formula for calculating the node representation vector is: Where K is the total number of related skills, w k Let φ(s) be the composite weight of the k-th skill. k ) represents the skill semantic encoding value, σ 2 Let be the mean variance of the skill vectors within the industry cluster, ∈ be the smoothing constant to prevent zero, η be the dynamic scaling factor, Δt be the node survival time, τ be the industry benchmark update cycle, λ be the decay rate of market popularity correlation, C be the number of related industry events, and I be the mean variance of the skill vectors within the industry cluster. c Let ψ(h) represent the influence intensity of the c-th event. c ) represents the industry fractal coding value, d represents the distance between industry cluster centers, and γ represents the spatial decay index.

4. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 1, characterized in that, Based on the node representation vectors, the improved BERT-DeepWalk algorithm is used to generate student profile vectors, occupational feature vectors, and school resource vectors, including: The node representation vectors are divided into three types of entities: student entities, occupational entities, and school entities. The student entity carries a student ID number and a skills mastery time series. The occupational entity includes a job code, a competency requirement matrix, and an industry-related path. The school entity includes associated course codes, teaching resource distribution, and an employment matching index. A type identifier code is created for each entity type, and six heterogeneous path patterns are defined. These six heterogeneous path patterns include: SSO path, CIC path, CEC path, SFC path, CTC path, and CAC path. The SSO path's transmission direction is: Student → Skill ← Career; the weight of the SSO path is the product of skill matching degree and time decay factor. The CIC path's transmission direction is: Career → Industry → Institution; the weight of the CIC path is the product of industry relevance and resource abundance index. The CEC path's transmission direction is: Institution → Course → Career; the weight of the CEC path is the product of course articulation coefficient and employment rate. The SFC path's transmission direction is: Student ← Feedback → Career; the weight of the SFC path is the product of satisfaction score and feedback timeliness coefficient. The CTC path's transmission direction is: Career ← Trend → Career; the weight of the CTC path is job evolution similarity. The CAC path's transmission direction is: Institution → Alumni → Career; the weight of the CAC path is alumni achievement index. The semantic similarity between nodes is calculated based on the node representation vectors, and a three-layer weight matrix is ​​generated. The three-layer weight matrix includes a structural weight matrix, a semantic weight matrix, and a time-dependent weight matrix. The structural weight matrix is ​​constructed based on the frequency of occurrence of meta-paths. The semantic weight matrix is ​​constructed based on the node similarity calculation results. The time-dependent weight matrix is ​​constructed based on the exponential decay value of the time axis coordinate difference between the node representation vectors. Based on the six types of heterogeneous path patterns defined above, a heterogeneous random walk is performed to obtain the walk sequence; The walking sequence is input into the improved BERT model to obtain the student profile vector, the occupational feature vector, and the school resource vector.

5. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 4, characterized in that, The heterogeneous random walk specifically involves: when encountering a student node, selecting the SSO and SFC paths; when encountering a professional node, prioritizing the CTC and CEC paths; and when encountering a college node, strengthening the CIC and CAC paths.

6. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 4, characterized in that, The improved BERT model adds spatiotemporal coordinate attention gating to the Transformer layer; the improved BERT model introduces multi-task learning during the pre-training stage; The multi-task learning includes mask node type prediction, meta-path pattern classification, and spatiotemporal coordinate regression.

7. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 1, characterized in that, The student profile vector, the occupational feature vector, and the institution resource vector are used as the input embedding layer of a pre-defined large language model. Within the parameter space of the pre-defined large language model, both random particle swarm optimization and guided particle swarm optimization based on historical optimal solutions are simultaneously implemented to perform quantum particle swarm optimization, resulting in an optimized large language model, including: The student profile vector, the occupational feature vector, and the school resource vector are concatenated according to their dimensions to form a joint feature vector; An adjustable parameter interface is inserted between the input embedding layer and the attention layer of the preset large language model. The interface dimension is aligned with the particle swarm position encoding dimension. N particle positions are sampled in a uniform distribution within the parameter space of the fully connected layer of the preset large language model. Extract a set of particles that meet preset conditions from historical optimization records; the preset conditions include: elite particles with the top 10% fitness ranking, particle distribution covering the parameter space of the most recent 3 training cycles, and particles carrying a time decay factor. The cosine similarity between the student profile vector and the occupation feature vector is calculated to obtain the ability matching degree. Based on the temporal attention weight, the area under the demand growth curve of the target position in the next 36 months is predicted by LSTM to obtain the development potential value. The market heat index is obtained by statistically analyzing the job increment and the rate of change of industry event influence intensity in the target industry in the last 24 hours. A time-varying fitness function is constructed based on the capability matching degree, the development potential value, and the market popularity index: The calculation formula of the time-varying fitness function F(t) is: F(t)=α(t)·Capability matching degree +β(t)·Development potential value +γ(t)·Market popularity index; where α(t), β(t), and γ(t) are preset dynamic weight coefficients; The position and velocity of each particle are encoded into Bloch spherical coordinates, and quantum rotation gate operation and quantum entanglement optimization are performed to establish entanglement pairs between the top 5% of the fitness particles in the guided particle swarm and the bottom 20% of the random particle swarm in each iteration, sharing the position update gradient. When the rate of change of the global optimal fitness is less than 1% and the proportion of guiding particles is greater than 60% in three consecutive iterations, it is determined to be converged, and the optimized large language model is obtained.

8. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 1, characterized in that, The output of the optimized large language model is jointly inferred with the node representation vector to obtain an initial recommendation strategy, including: The semantic vector set and node representation vector set output by the optimized large language model are transformed to a common vector space of the same dimension through linear mapping. Within the common vector space, using the set of semantic vectors as the query and the set of node representation vectors as the key / value pair, the cross-attention weight matrix is ​​calculated to obtain the fused semantic representation. Multi-head attention is applied to the fused semantic representation, context dependencies are preserved, and the inter-node structural weight matrix is ​​invoked to perform graph convolution update on the set of node representation vectors; The set of node representation vectors and the updated set of node representation vectors are concatenated as column vectors, and then passed through a fully connected layer to output the fused vector F_fuse. i ; For the candidate target set, a scoring function is used. i =Softmax(W_s·F_fuse) i +b_s) yields the score sequence. i Where W_s and b_s are learnable parameters; According to score i The top K candidate targets are selected in descending order, and this is denoted as the recommendation strategy.

9. The method for pushing higher education career planning information based on particle swarm optimization large language model according to claim 1, characterized in that, Also includes: By refining the recommendation strategy based on the user's interaction behavior, the optimal push recommendation strategy can be obtained.

10. A higher education career planning information push system based on a particle swarm optimization large language model, characterized in that, include: The dataset construction unit is used to acquire student registration data from the Ministry of Education, job posting data from recruitment platforms, industry association development reports, and enterprise employment standards in real time to obtain multi-source heterogeneous datasets. A vector association unit is used to associate the multi-source heterogeneous datasets to establish node representation vectors; The vector generation unit is used to generate student profile vectors, occupational feature vectors, and school resource vectors based on the node representation vectors using the improved BERT-DeepWalk algorithm. The model optimization unit is used to take the student profile vector, the occupational feature vector and the college resource vector as the input embedding layer of the preset large language model, and simultaneously set random particle swarm and guided particle swarm based on historical optimal solutions in the parameter space of the preset large language model to perform quantum particle swarm optimization to obtain the optimized large language model. The reasoning unit is used to jointly reason with the output of the optimized large language model and the node representation vector to obtain a recommendation strategy.

Images