1. Introduction

In the current information age, college graduates, as the fresh blood of the society, the comprehensive improvement of their abilities and qualities is a key factor in realizing the modernization of education and promoting economic and social development. With the rapid development and wide application of digital intelligence technologies such as artificial intelligence, big data and cloud computing, the enterprise recruitment market and the employment environment of college graduates are experiencing unprecedented changes. In this context, it has become an important issue in the field of education and human resource management to continuously optimize the talent resource management model of college graduates and improve the employment quality of graduates and the efficiency of enterprises in employing them ^[1,2,3]. The rise of digital intelligence technology has brought new opportunities and challenges for college graduate talent management. Talent information collection and analysis and processing capabilities under the support of big data have been significantly enhanced, and the demand for personalized and intelligent services has become increasingly prominent. However, the traditional university talent resource management model is often difficult to accurately respond to the rapid dynamics of the diversified employment market due to the lack of efficient data processing mechanisms and dynamic optimization strategies. Insufficient matching accuracy of enterprises in selecting talents and lack of scientific guidance for career development planning of graduates have, to some extent, limited the effective allocation of senior talent resources and the maximization of social value ^[4,5,6]. Among the many studies, the exploration of attempting to apply digital intelligence technology to the management of graduated talent resources in colleges and universities is gradually increasing. AI-assisted career planning, big data analysis of job market trend prediction and other research work has been carried out one after another, but little targeted pursuit of model optimization and technological integration, especially the lack of in-depth excavation of algorithmic performance enhancement and innovative applications. In addition, how to combine students' individualized needs and market changes to build an intelligent resource management system with stronger adaptability and higher accuracy is still a problem to be solved ^[7,8,9]. Aiming at the shortcomings of the existing research and the actual needs of the industry development, the study proposes a new type of university graduate talent resource management model under the view of digital intelligence. The model will integrate Adaptive variation improvement factor (AVI) and Genetic Algorithm (GA), aiming to optimize the performance of the genetic algorithm to improve the efficiency and accuracy of the talent resource management model. The innovation of the research lies in the combination of AVI and GA algorithms, which is expected to enhance the adaptability and dynamic adjustment ability of the model during the optimization process. Meanwhile, due to the need for higher levels of intelligence and adaptability in talent resource management in the digital age. The construction of this model not only meets the modern society's demand for precise employment management, but also utilizes adaptive and optimization algorithms to enhance the dynamic and real-time nature of the management model, enabling it to better respond to the ever-changing job market and graduate demand. Combining the current technological development trend and drawing on the latest theoretical and practical achievements in related fields, the study is expected to provide a new optimization path for the management of graduated talent resources in colleges and universities to achieve the optimal allocation of resources, which will in turn promote the improvement of education quality and the overall progress of graduates' career development.

The first part of the study summarizes and explains the relevant human resource management as well as genetic algorithms, the second part is the implementation of the proposed methodology, the third part is the validation of the proposed methodology, and the fourth part summarizes the results of the study as well as the outlook.

2. Related Work

In the context of digital intelligence, the optimization of talent resource management models for college graduates has become a hot research topic in recent years. I Alabri's team, based on resource-based theory, explored the relationship between human resource management practices and adaptive employee performance and examined the moderating role of transformational leadership in it. The findings revealed that performance appraisal, training, job enrichment, and job enlargement had a significant impact on enhancing adaptive employee performance. Transformational leadership further moderated the relationship between employee engagement and adaptive performance ^[10]. The Oseghale team, on the other hand, explored how institutional and cultural factors influence the reproduction of HRM practices and the selection of delivery mechanisms when transferring HRM practices between multinational corporations and their subsidiaries. The study suggests that organizational culture is the mechanism of reproduction and inhibition. The study provides HR managers with a conceptual framework for understanding how to reproduce transferred practices in developing countries ^[11]. Wu et al. on the other hand, proposed a clustering-based incremental association rule mining algorithm to improve data mining efficiency. Using database development tools, the system setup and programming of the algorithm for efficient large-scale database mining were realized, and it was successfully applied to the human resource management system of the university to realize the broadcasting of association rules and complete the visual display of information ^[12]. Liu et al. on the other hand, constructed a model of the relationship between human resource management activities and performance based on the LMBP algorithm to accurately predict the fluctuation of corporate performance risk of corporate performance fluctuation. The study shows that LMBP algorithm optimizes the accuracy and successfully predicts the risk of performance fluctuation under the HRM activities of enterprises, and builds the correlation model between HRM activities and performance, and the experiments show that LMBP algorithm can more accurately reflect the relationship between HRM and performance of enterprises ^[13].

Genetic algorithm has been widely used in many fields since it was proposed in 1992. Its application in the field of human resource management focuses on solving the problems of organizational structure optimization, talent selection and job matching. Qin addresses the problems in the process of human resource scheduling and optimization in construction projects, establishes the basic mathematical model of the human resource scheduling problem for resource-constrained construction projects and the multi-project equilibrium problem, and puts forward the resource-constrained multi-project multi-skilled human resource scheduling problem and the generalized priority relationship under the integer planning mathematical model, and the accuracy of the proposed algorithm and model is verified by simulation results ^[14]. Gentile team, on the other hand, investigates the scheduling problem of satellite tracking by a heterogeneous ground station network under the consideration of the uncertainty of the allocated resources to minimize the final estimation uncertainty, and adopts the optimization method to efficiently select the best tracking plan. The results show that variable-length genetic algorithms consistently outperform the fixed-length algorithms used as comparisons, and the structured-chromosome genetic algorithm finds significantly better plans under strict budget constraints ^[15].

Researchers such as Wu analyzed the interrelationship between sustainable development goals and information and communication technology, and discussed the role of communication technology in achieving sustainable development goals. After literature review, it was found that technology has made significant contributions to the sustainable development goals, but there are shortcomings in the perspective of social welfare. Therefore, it is necessary to innovate and develop communication technology ^[16]. Regarding the connection between big data and green challenges, scholars such as Wu have revealed the issue of greening the lifecycle of big data systems through comprehensive literature review and discussion. The method includes analyzing the application and challenges of big data technology in achieving green goals. The results show that big data technology not only promotes the trend of green revolution, but also provides new possibilities for improving resource utilization efficiency and reducing environmental impact ^[17].

To summarize, these algorithms are often used independently, and the comprehensive advantages of multi-algorithm fusion are seldom considered. In the face of the growing number of college graduates and enterprises' individualized demands, there is an urgent need for more efficient and dynamically adaptable management models. Moreover, how to fuse AVI factors with GA to enhance the adaptability and efficiency of the model, as well as how to respond quickly to market changes, are still urgent research issues to be solved. Based on the existing research foundation, the study will explore an optimization method of college graduation talent resource management model by integrating AVI factors and GA with practical application scenarios.

3. AVI Factor and GA Algorithm of College Graduation Talent Resource Management Model Construction

The talent management system for college graduation can provide personalized job recommendations for college students, and can select suitable talents for enterprises to hire. The study incorporates AVI factors into the genetic algorithm and optimizes the genetic algorithm to be applied to the human resource management model for talent recommendation and management.

3.1 Constructing the Framework of Human Resources Management for Graduates of Universities and Colleges

In the context of rapid development of digitalization and information technology, strengthening the management of graduated talent resources in colleges and universities has become an important way to realize the modernization of education and improve the quality of human resources. This study optimizes the design of the university graduate talent resource management system and builds a comprehensive management framework that takes into account personalized service and efficient matching to meet the diversified needs in the context of digitization and intelligence ^[18,19]. The basic framework is shown in Fig. 1.

In Fig. 1, the resource management framework has five parts, and in the data collection layer, the key task is to obtain comprehensive, accurate and real-time data resources. Diversified collection means are used for different data sources, including online questionnaires, interface docking of the teaching system, human resource market research, feedback from enterprise cooperation and other methods. Student data is not only limited to basic education information, but also includes multi-dimensional information such as career assessment results, internship experiences, participation in innovation and entrepreneurship programs, and career planning intentions.

Fig. 1. Resource management framework.

The data processing layer is designed to work on transforming the voluminous data collected into information of analytical value. First, the data is cleaned by automated tools to eliminate erroneous, duplicate or irrelevant data points. Next, data standardization and normalization are performed so that data from different sources and formats can be compared and connected to each other. In this case, the data normalization formula can be expressed as Eq. (1).

(1)

$ Y=(X-\mu )/\sigma . $

Fig. 2. Decision support layer framework.

In Eq. (1),$Y$ is the normalized value,$X$ is the original value,$\mu $ is the data mean, and$\sigma $ is the data standard deviation. The analysis and inference layer is the core of realizing personalized service, and the framework is shown in Fig. 2.

As shown in Fig. 2, the match between individual abilities and market demand is analyzed in depth by using statistical analysis and other techniques. Talent ability models and job demand models are established, based on which intelligent matching and predictive analysis using various types of algorithms are conducted to identify different groups of career interests and abilities; association rules are used to mine the intrinsic connection between students' ability characteristics and successful employment cases, so as to provide references for students' employment guidance. Among them, the career interest clustering analysis can be represented by Eq. (2),

(2)

$ d\left(i,k\right) = \sqrt{\sum \left(x_{ij } - x_{kj} \right)^{2} }. $

In Eq. (2), $x_{ij} $ is the students' ability value or interest in the corresponding dimension, and$x_{kj} $ is the coordinates of the clustering center in the corresponding dimension. The decision support layer is dedicated to transforming the results of data analysis into concrete decision recommendations, as shown in Fig. 3.

Fig. 3. Decision support layer framework.

As shown in Fig. 3, a six-dimensional evaluation method is used to ensure the comprehensiveness and adaptability of the system when constructing the framework of the decision support layer for college graduate talent resource management. The evaluation dimensions range from the infrastructure elements of data analysis and processing to the performance indicators of personalized recommendation systems, and then include the accuracy of market demand forecasts and the completeness of talent cultivation strategies. At the same time, enterprise collaboration and feedback mechanisms and policy and regulatory compliance are optimized as key components of decision support to ensure compliance and continuous improvement of management systems. Using predictive modeling, talent supply forecasting reports can be designed for companies to help them plan their recruitment strategies earlier. The decision support system will also provide a dynamic adjustment mechanism to fine-tune the recommendation strategy based on real-time data streams to ensure that it continues to adapt to market and individual changes. Among them, the ability matching degree can be expressed by Eq. (3).

(3)

$ S_{ij} =\sum _{n}^{a}w_{k} \times (a_{ik} -r_{jk} )^{2} . $

In Eq. (3), $S_{ij} $ is the student and job matching score, $w_{k} $ is the ability dimension weight, $a_{ik} $ is the student's ability or achievement in the corresponding dimension, and $r_{jk} $ is the requirement of the corresponding job in the dimension. And the job recommendation score can be calculated by formula (4).

(4)

$ R_{ij} =\frac{1}{(1+e^{-z} )} . $

In Eq. (4), $R_{ij} $ is the student's recommendation score for the position, and $z$ is a linear combination of the match score and other factors. The talent supply prediction model can be calculated by Eq. (5).

(5)

$ P_{t} =\alpha \times P_{(t-1)} +\beta \times I_{t} +\gamma \times E_{t} +\varepsilon _{t} . $

In Eq. (5), $P_{t} $ is the predicted talent supply at the corresponding moment, $P_{(t-1)} $ is the talent supply at the time of $t-1$, $I_{t} $ is the market demand index at the corresponding moment, $E_{t} $ is the educational effectiveness index at a certain moment, $\alpha $, $\beta $, $\gamma $ are the model coefficients, and $\varepsilon _{t} $ is the error term. The performance feedback adjustment formula can be expressed by formula (6).

(6)

$ \Delta \Theta =\eta \times \frac{\partial E}{\partial \Theta } . $

In Eq. (6), $\Delta \Theta $ is the adjustment of the model parameter $\Theta $, $\eta $ is the learning rate, and $E$ is the evaluation indicator. The formula reflects how the model parameters affect the evaluation metrics. In the service interaction layer, user interfaces will be created to provide customized services for students, teachers, career planners, and corporate HR managers, respectively. Users will be able to access the system through web or mobile applications to get real-time personalized career planning advice, recommended positions, detailed information on corporate culture and values, etc. Among them, the talent resource management cloud diagram is shown in Fig. 4.

Fig. 4. Human resource management cloud map.

Through interaction design, the service layer will ensure superior representation and communication of information, including interactive data views, graphical presentation of recommendation reports and intuitive operational flows. System feedback and help navigation features will also be provided to allow users to make inquiries and suggest changes, further enhancing the transparency and interactivity of the system.

3.2 Talent Resource Optimization Model Construction by AVI Factor Fusion GA Algorithm

Facing the growing number of college graduates and the rapid change of enterprise demand, the traditional college talent resource management model appears to be overwhelmed. The study is to improve the efficiency and accuracy of talent resource allocation through the introduction of digital intelligence. To this end, combines AVI with GA in pursuit of a more reliable and dynamically adaptable solution to the problem of optimizing the allocation of talent resources.

GA draws on the mechanism of biological evolution to approximate the global optimal solution in a multi-generation iterative process by initializing, fitness evaluation, selection, crossover and mutation steps for the candidate solution set. Among them, the inclusion of AVI factor aims to improve the performance of GA in complex search space and make adaptive adjustments to the mutation step of traditional GA. The model optimization construction is divided into the following key steps, first, defining the candidate solution set, i.e., the individual representation. In the context of university graduate talent resource management, individuals represent different graduate-enterprise matching solutions. The population definition formula is shown in Eq. (7).

(7)

$ P=\{ p_{1} ,~p_{2} ,~p_{3} ,~\cdots,~ p_{n} \}. $

In Eq. (7), $P$ is the population and $p_{i} $ is the corresponding candidate solution, i.e., the graduate and enterprise matching program. And the fitness function can be expressed by Eq. (8).

(8)

$ F(p_{i} )=\alpha *G_{c} (p_{i} )+\beta *G_{e} (p_{i} ) . $

In Eq. (8), $F(p_{i} )$ is the fitness of the candidate solution, $G_{c} (P_{i} )$ is the synthetic degree of matching between the candidate and the job, $G_{e} (p_{i} )$ is the satisfaction degree of the enterprise to the candidate, and $\alpha $ and $\beta $ are the corresponding weight coefficients.

The core steps of the genetic algorithm, i.e., selection, crossover and mutation operations, in which the crossover method of the personnel assignment matrix is realized by using consecutive real number coding in the form of Eqs. (9) and (10).

(9)

$ child1(i,j)=\alpha (i,j)*parent1(i,j) \\ \quad +(1-\alpha (i,j))*parent2(i,j). $

In Eq. (9), $child1(i,j)$ is the element at the corresponding row and column position of the cross-generated child individual, and $parent1(i,j)$ is the element at the corresponding row and column position of the parent individual.

(10)

$ child2(i,j)=\alpha (i,j)*parent2(i,j) \\ \quad +(1-\alpha (i,j))*parent1(i,j). $

In Eq. (10), $child2(i,j)$ is the element in the corresponding row and column position of the child individual produced by the crossover, and $parent2(i,j)$ is the element in the corresponding row and column position of the parent individual. The schematic diagram of the crossover of the personnel assignment matrix is shown in Fig. 5.

Fig. 5. Cross diagram of personnel assignment matrix.

In Fig. 5, it is the crossover process of the personnel allocation matrix, and the child individuals can be obtained from the $\alpha $ matrix after the operation with the parent individuals, and the calculation process is shown in Eqs. (9) and (10). The variation formula of the personnel allocation matrix is shown in Eq. (11).

(11)

$ M_{new} =M+\alpha *randM . $

In Eq. (11), $\alpha $ is the mutation step size and$randM$ is the random number under normal distribution. However, the demand for mutation to produce new individuals at different stages of the iterative process of population evolution is not considered in the base GA algorithm. Therefore, the adaptive mutation improvement factor is introduced to optimize the GA algorithm by setting the observer variable as the dynamic mutation odds as shown in Eq. (12).

(12)

$ p_{m} =0.1*e^{\frac{count*\ln (10)}{0.1*t_{\max } } } . $

In Eq. (12), $p_{m} $ is the mutation chance, $e$ is the natural constant, $t_{\max } $ is the total number of iterations, and $count$ is the variable value. In the GA algorithm, the second layer of chromosomes sets the adaptive variation length when mutating, and the calculation formula is shown in Eq. (13).

(13)

$ P_{length} =(count/t_{\max } )*N_{j}. $

In Eq. (13), $N_{j} $ is the number of variants. And the incremental function of AVI factor can be expressed in Eq. (14).

(14)

$ adap(\Delta F)=\gamma *\exp (-\delta *\Delta F) . $

In Eq. (14), $\gamma $ and $\delta $ are the parameters that determine the response sensitivity and curve of the adaptive variability rate. The population diversity maintenance strategy, on the other hand, can be expressed in Eq. (15).

(15)

$ D(P)=(1/\left|P\right|)*\sum (\left\| p_{i} -P_{avg} \right\| ) . $

In Eq. (15), $\left\| p_{i} -P_{avg} \right\| $ is the Euclidean distance between the corresponding individual in the population and the average individual in the population, and $P_{avg} $ is the average position of all individuals. And the calculation of the population mean position can be expressed by Eq. (16).

(16)

$ P_{avg} =(1/\left|P\right|)*\sum p_{i} . $

In Eq. (16),$P$ is the population diversity.

4. Analysis of the Results of the AVI Fusion GA-based University Graduate Talent Resource Management Model

To verify the effectiveness and practicality of the proposed model, this study constructs a genetic algorithm optimization model of college graduate talent resource management containing AVI factors. The performance of the model in the simulated college graduate job market is evaluated through experimental simulation. The experimental design includes different sizes of graduated talents and enterprise hiring demand datasets, aiming to explore the model's adaptability and optimization ability under different complexity conditions. Among them, the hardware and software configuration table are shown in Table 1.

Table 1 shows the configuration parameters of the system's hardware and software environments. The combination of high-performance Intel Xeon Gold CPUs and expanded memory ensures efficient execution of complex computing tasks. Meanwhile, the integration of NVIDIA GTX 3080Ti provides a great improvement in graphics and parallel processing power, while the updated software environment further guarantees system security and application compatibility.

From the convergence speed comparison in Fig. 6, it can be seen that AVI-GA shows a significantly better convergence speed than the other algorithms at the beginning of the iterations. During 150 iterations, AVI-GA has an adaptation score of 0.1 in the initial 10 iterations, and then steadily increases to reach a score of 0.87 in 150 iterations. In contrast, the classical genetic algorithm has an initial score of 0.05 and reaches a fitness score of 0.47 after 150 iterations, indicating that its convergence speed and optimization effectiveness are not as good as that of AVI-GA. Within the same iteration stage, the particle swarm optimization algorithm and the differential evolution algorithm grow gradually from initial scores of 0.07 and 0.08 to 0.61 and 0.69, respectively, reflecting a moderate level of optimization effect. In contrast, the GA has the slowest growth in fitness score, demonstrating relatively poor optimization efficiency. The particle swarm optimization algorithm and the differential evolution algorithm exhibit moderately fast convergence behavior.

In addition, the performance improvement of the simulated annealing algorithm fluctuates widely. Fig. 7 shows the fitness scores of different optimization algorithms in five independent runs to measure their performance in specific tasks. In the iterative test, the AVI-GA algorithm shows high fitness scores of 0.82, 0.84, 0.86, 0.85 and 0.87, respectively, indicating that it is well adapted to the optimization requirements. The stability of GA algorithm is 0.65, 0.66, 0.67, 0.65 and 0.68 respectively, while the simulated annealing scores are 0.60, 0.59, 0.62, 0.61 and 0.64. The scores of genetic simulated annealing algorithm are relatively stable. In five runs, the PSO algorithm scored between 0.78 and 0.81, and the highest score failed to beat the research optimization algorithm. The DE algorithm scored even lower, with an average score of 0.74, which was 10.8 lower than the average score of the research algorithm. The highest score is only 0.76, and the fitness score of SA algorithm is between 0.6 and 0.6, which is a low overall level. On the whole, AVI-GA shows high stability in all algorithms, showing superior and stable performance. Fig. 8 shows the CPU usage of different algorithms.

Fig. 8 shows the CPU utilization data of the six algorithms at different running stages to evaluate the resource consumption of the algorithms. The CPU usage of the optimization algorithm increases gradually from 68% in the first run to 75% in the fifth run, showing a gradual increase and then a slight decrease. The classical genetic algorithm shows a steady increase from 75% to 80%, while the particle swarm optimization algorithm remains relatively stable between 68% and 71% during the run. The differential evolution algorithm and the genetic simulated annealing algorithm fluctuated between 73%-78% and 71%-74%, respectively, while the simulated annealing algorithm gradually increased from 80% to 85%, indicating a significant growth in its resource consumption. It can be seen that the optimization algorithms are able to maintain a low CPU load factor while the system is running. The ROC curves of different algorithms are shown in Fig. 9. As can be seen in Fig. 9, in Fig. 9(a), the ROC area of the optimization algorithm is more than 0.9, which has good recommendation and management effect. In the SA-GA algorithm in Fig. 9(b), although it has good prediction effect, the ROC curve area is lower than 0.9. In the DE model in Fig. 9(c), the ROC curve area is even smaller, and the actual judgment effect is average. In the PSO algorithm of Fig. 9(d), the fitting degree is medium, the deviation of the upper left corner is insufficient, and the prediction effect is much lower than that of the optimization algorithm. And the comparison of resource consumption of different algorithms is shown in Table 2.

From Table 2, it can be seen that the Adaptive Variation Improvement Factor Fusion Genetic Algorithm performs well in terms of memory consumption with only 700MB, while the number of disk reads and writes is 1200, the network traffic is 300KB, and the power consumption is only 0.8 Wh. The classical genetic algorithms, on the other hand, are higher in terms of resource consumption, which is especially notable in terms of 950MB of memory consumption and 1800 disk reads and writes. Particle Swarm Optimization Algorithm, Differential Evolutionary Algorithm and Genetic Simulated Annealing Algorithm show moderate resource requirements in all the listed metrics, with network traffic and power consumption maintained at moderate levels. The simulated annealing algorithm shows higher resource consumption than the other algorithms, especially in the memory consumption of 950MB and power consumption of 1.2Wh. The optimization system is actually used in a university and let the relevant personnel to score the model effect evaluation by percentage system, the evaluation results are shown in Table 3.

Table 3 shows the multidimensional social impact scores, which are used to measure the magnitude of the impact of each algorithm on socio-economic indicators. It can be seen that the research algorithm scored 90 points for employment growth, which is an excellent performance compared to the industry standard of 75 points and the target value of 85 points. On the income growth indicator, it scored 85 points, exceeding the industry standard by 20 points. Skills and Career Enhancement scored 92, showing a significant boost to talent capacity expansion. Other dimensions such as Social Communication and Network Expansion (88 points), Industry Technology Contribution and Innovation (90 points) and Environmental and Social Responsibility (89 points) all performed well. The algorithm's management model is also effective in talent supply prediction (91 points) and social welfare impact (93 points). The algorithm's overall score of 91.7 is well above the industry average, demonstrating its strong performance and social benefits in a number of key dimensions.

Fig. 6. Convergence rate comparison.

Fig. 7. Fitness scores of different optimization algorithms.

Fig. 8. CPU usage comparison.

Fig. 9. ROC curve comparison of different algorithms.

5. Conclusion

In order to enhance the efficiency and accuracy of college graduate talent resource management in the era of digital intelligence, research constructs a set of college graduate talent resource management optimization model by introducing adaptive variation improvement factors and combining with genetic algorithm. The results show that through 150 iterations of experiments, AVI-GA demonstrates a high adaptive score of 0.87, which is more significant in terms of optimization efficiency and performance compared to the 0.47 of the classical GA algorithm and the 0.61 to 0.69 scores of other algorithms. In five independently run stability tests, AVI-GA averages a score of 0.85, significantly higher than other algorithms, highlighting its stability. In terms of resource consumption, AVI-GA shows optimal performance in terms of memory usage and power consumption (only 700MB and 0.8Wh, respectively), and its optimization effectiveness and stability provide a feasible solution for number-wise talent resource management. The shortcoming of the study is that the adaptability under different industries and diversified demands has not been fully verified, so the model needs to be more widely applied and tested in different universities and employers in the future. To sum up, the AVI-GA optimization model proposed by and not only provides a new perspective for the research of intelligent algorithms in the application field of talent resource management in theory, but also confirms its high practical value and social significance in practice.