Abstract
The increasingly large employment group in colleges and universities has brought huge employment pressure to the society. Therefore, improving the level of professional ability of college students, providing students with personalized career development direction planning, and helping students to establish a correct concept of career selection are feasible ways to solve the structural contradiction of college students’ employment. The arrival of big data era provides new development opportunities for employment work in colleges and universities. Big data drives the high-quality development of precise employment in colleges and universities, which is inevitably required to reach the precise requirements of work development, information mastering, platform docking, information pushing, and help work. The traditional employment guidance work is backward in means and poor in timeliness, which leads to poor intervention effect on students’ employment expectation. With the deepening of digital campus construction and the development of big data technology, massive educational data has been accumulated but not reasonably utilized. Therefore, in this paper, we take the massive student data accumulated in digital campus as the research object, based on the research method of convolutional neural network, to explore the hidden personalized information of students behind it, predict their future career development direction, and provide scientific and technological support for the work of university education and teaching. Thus, the architecture classical DenseNet was improvised to avoid gradient disappearance and guarantee the classification accuracy, thereby targeting to reduce the number of connections. The study also proposed a 2-DenseNet model, wherein the attention module was embedded into 2-DenseNet followed by the conduction of training and validation on the university employment dataset. The experimental results show that the method proposed in this paper could predict the appropriate career development direction timely and effectively based on multiple perspectives of students’ comprehensive quality. This would enable the students to plan and adjust their employment expectations leading to disseminating of efficient employment guidance in colleges and universities.
1. Introduction
Employment is the basis of people’s livelihood, and youth employment is the top priority of employment work. In recent years, the scale of college graduates in China has been rising year by year. The data from the Ministry of Education shows that the scale of college graduates is expected to reach 10.76 million in 2022 [1, 2]. Consequently, the impact of the new crown pneumonia epidemic and the severe macroeconomic situation at home and abroad have also caused a big impact on the employment of college graduates in China. As a result, the research on the employment of college students has become a hot issue in the research. The market environment facing the employment of college students today is changing significantly, and the slowdown of economic development has objectively led to the reduction of the employment demand of college students. Due to the differences in economic development and market environment in different parts of China, the cities in different regions differ greatly in their ability to absorb the employment quantity of graduates. To a certain extent, this also leads to the uneven distribution of graduates’ employment and makes the problem of difficult employment of college students show certain regional differences [3].
Since the outbreak of the new crown pneumonia, nearly half of college students have experienced increased employment anxiety due to the outbreak, and about one-third of college students have severe self-perceived employment anxiety [4]. Some studies have shown that there is a significant difference in the degree of employment anxiety among college students in terms of academic performance, and the degree of employment anxiety is significantly lower among college students who did not fail any classes during their school years than among those who did. College students from one-child families had significantly lower levels of employment anxiety than those from non-one-child families. Unlike expectations, there were no significant differences in employment anxiety among postepidemic college students by gender, family origin, and major [5]. The quality of employment of college students after graduation is always something that bothers college students and is always a key concern for families and is the core of school employment guidance services and a major issue for social development. In the new era of building an innovative development pattern, influenced by the high quality requirements of national development and the uncertainty of the external environment, the development and operation of the country and society have been affected to different degrees, and college students have been the first to experience the changes formed by the social industrial structure adjustment, and the employment concept of college students in the new era has undergone significant changes. Faced with such a realistic employment situation, especially when they realize that they exist as one of the millions of employed individuals, college students inevitably feel their own insignificance and the slim hope of employment, and they will be under great psychological pressure. When facing such pressure, some college students do not choose to struggle but choose to “give up,” only pressure but no motivation. From the personal subjective level, some college students are often unwilling to face the pressure of real work when they face graduation, unwilling to go out of the comfort circle of school life, and more willing to defer employment in the name of further study and research, rather than sincerely out of love for research. From the objective level of society, due to the downward pressure of domestic economic development, some enterprises’ layoffs and the reduction of recruitment are not able to provide sufficient jobs for the majority of graduates, and the professional knowledge learned by college students during the college period has a low degree of matching with the needs of society and other factors; the actual ability of individuals cannot be well played in the job market, so they have to defer employment by way of graduate school. This “roundabout” approach is a good way to delay employment. This kind of “roundabout” employment concept and employment choice has become more and more prominent after the emergence of the new crown pneumonia epidemic due to the increase of external employment pressure.
The use of big data technology to build a new model of college student employment guidance is not only an effective way to break through the employment problems of college students but also an effective means to cultivate college students’ interconnected thinking, enhance their social suitability and broaden their career, and provide a boost for college students to become a workplace elite [6]. In this regard, it is necessary to explore more advanced and scientific employment guidance paths based on big data technology from the employment challenges faced by contemporary college students and strive to make an effective breakthrough in the employment problems of college students. For college students, their intellectual level, learning ability, and professional strengths are worthy of recognition. However, due to the limit of life experience and high level of thought, college students are inevitably short-sighted when considering employment issues, resulting in less accurate and reasonable employment expectations, which may adversely affect their future career development. To address this problem, university career guidance teachers can use big data technology to analyze students’ career development scientifically and accurately and help them to make a career development plan that matches their personal situation. After that, teachers and students discuss and exchange with each other about the content of the plan, and teachers listen carefully to students’ personal ideas. Finally, with the teachers’ career sense and experience, they provide students with professional career guidance, point out the future career development direction for them, help them clarify their career development goals, adjust their employment expectations, and build up firm career beliefs. Through this way, we can avoid blind job hunting, random job hopping, and lightly giving up because of eagerness and anxiety. In the industry that suits you, the job is stable and steady, down-to-earth work. With the support of artificial intelligence and big data technology, colleges and universities can create an intelligent platform for college students’ employment guidance and promote the overall improvement of the effectiveness of employment guidance. The platform should at least have the following functions: first, self-assessment. On this platform, college students can carry out self-assessment of their own vocational ability, understand what level their professional level and employability are at, and what gaps exist between them and the actual needs of employers, so as to carry out targeted self-improvement and ensure that their own professionalism meets the needs of society. Second, subdivision of employment tendency. Each major has its own employment tendency, and the intelligent platform of college students’ employment guidance can help students further clarify their employment tendency by combining their own majors and existing conditions and help them make career plans, give correct guidance to their current view of career selection and career development, and promote the overall improvement of their ideological consciousness. For example, for technical students, the platform can recommend engineers, technicians, designers, and other positions to them.
Deep learning is one of the mainstream algorithms in the era of big data, and there are already many recommendation algorithms based on deep learning applied in various industries [7]. In this paper, we propose a recommendation model based on deep learning to guide students to adjust their employment expectations by considering students’ comprehensive quality values from multiple dimensions.
Thus, the unique contribution of the work includes (i)A massive student data was considered which was collected in the digital campus and CNN was used to explore the hidden personalized information of the students(ii)The classical DenseNet model was improved to avoid gradient disappearance and achievement of enhanced accuracy(iii)The model helped in predicting appropriate career development direction of the students considering multiple aspects
2. Related Works
2.1. The Current Employment Situation of College Students in the Era of Big Data
In reality, college graduates generally have poor career awareness and incorrect job hunting methods, which lead to unsuccessful bidding, unbalanced supply and demand of talents, and finally three major problems of “difficult employment for college graduates, difficult recruitment for employers and difficult education for schools” [8]. With the rapid development of science and technology, “big data+employment” has become an emerging employment promotion mode, which greatly solves the dilemma of “nowhere to bid” for college graduates, reduces the difficulty of employment, and improves the recruitment effectiveness of employers. In order to ensure full employment, the Ministry of Education explicitly requires universities to use big data technology to establish precise employment support mechanisms, provide graduates with matching job information, and carry out targeted personalized and informative employment service support, so that college students can fully plan their careers based on their own strengths and interests [9].
The school employment function department widens the employment channels in all aspects and multiple channels and actively develops the effective employment market which is of great significance to the employment work. Big data has changed the traditional way of contacting employment recruitment work such as telephone, mail, and letter. Through the information technology platform, employers can directly register online, upload their qualifications and recruitment information, and flexibly choose and arrange the schedule of lecture and recruitment according to the venue provided by the university, which provides great convenience for employers to lecture and recruit [10]. According to the professional needs of school graduates, the employment function department will review the qualifications of employers and decide to agree or politely reject the requests of employers to participate in the job market, which minimizes the workload of employment staff in reviewing the qualifications of employers and recruitment information and provides efficiency and effectiveness of feedback. The use of big data technology can accurately reflect the demand situation of enterprises, scientifically grasp the trend of market supply and demand, help graduates understand and grasp the employment situation and employment trend of this major, and guide graduates to actively seek employment. By analyzing the data of 2021 spring campus job fair submission through big data, we can accurately grasp the graduates’ tendency degree towards industry, enterprise, position, salary, and so on [11]. The employment management department of the university collects various types of information on graduates’ majors, specialties, positions, salaries, working regions, etc. through the employment information technology platform and makes basic data sheets. The employment questionnaire survey is issued and collected through the network to understand and grasp the employment intention and employment status of graduates in a timely manner, which is used as the basic basis for employment guidance and assistance. According to graduates’ job-seeking intention and talent demand of employers, we provide targeted employment guidance to graduates. Through the big data technology, the employer’s job demand, job requirements, salary, geographic information and graduates’ job-seeking intention, professional strengths, vocational ability certificates, and other information are precisely matched to achieve accurate pushing of recruitment information, on the one hand, greatly reducing the blindness of graduates’ job search, reducing the economic burden caused by job search, and improving the success rate of graduates’ job applications. At the same time, employers can also recruit the graduates they like in the shortest possible time, which greatly reduces the workload of reviewing graduates’ application materials, improves recruitment efficiency, and reduces recruitment expenses. The employment quality of college graduates is an important reflection of the quality of education, teaching, and talent cultivation in colleges and universities. Using big data to provide all-round and multiperspective feedback on employment quality can find the direction of efforts to further improve employment quality, provide an important reference basis for schools to adjust the layout of majors, make development plans, reform education and teaching, optimize enrollment policies and measures, and effectively promote universities to better fulfill their historical mission of serving the society.
The employment guidance work of college students is mainly to provide high-quality employment guidance services to college students, and its main contents include common services and rough services [12]. In the process of employment guidance for college students, we analyze the employment policy and employment situation of college students and provide college students with enterprise recruitment information and cultivate college students’ job hunting and application skills. By providing these public information and rough information to college students, we can help them accomplish their employment plans and employment goals. However, in practice, most colleges and universities do not analyze the employment data information of college students deeply enough, and the staffs mostly only pass the shallow employment information they know to college students, without really understanding the employment situation of college students and the employment situation of enterprises. It is mainly because the big data technology mastered by college students’ employment guidance workers is not particularly adequate, the application and operation level of big data technology is low, and even a few college students’ employment guidance workers do not have the ability to apply big data technology at all, so they cannot accurately analyze the employment trend, and at the same time, they do not analyze the employment path and employment direction for the college students’ majors, and the college students’ employment guidance work cannot meet the personalized development needs of college students. Professional big data analysis requires certain professional knowledge of computer technology, Internet technology, statistics, digital technology, etc. It is not only simple data collection, processing, and analysis but also deep excavation and analysis of college students’ related information, enterprise recruitment information, and employment policy information by using big data technology [13]. Applying big data information to college students’ employment guidance work and providing students with employment guidance services are an inevitable demand for college students’ employment guidance work in the era of big data and also a major challenge for college students’ employment guidance work.
2.2. Status of Research on Recommendation Algorithms in Big Data Technology
With the increasing maturity of cloud computing, Internet of Things, and other technologies, the large amount of information contained behind big data is being paid more and more attention to, and the education field is also paying more and more attention to big data in education. In the context of digital campus construction, how to better serve students’ education work has become a key issue for educators to think about nowadays [14]. Applying data mining technology to education management systems and collecting and analyzing the representational contents of various types of student data and the interrelated information behind them can help teaching managers identify problems in teaching, accelerate the improvement of teaching quality, and help optimize education reform. Based on this, we aim to use data mining algorithms to explore the hidden personalized information of students based on the massive student data accumulated by the digital campus construction, so as to provide college educators with a basis for career guidance, scientific prediction of students’ future career development direction, and further personalized career guidance and planning for students. It helps students to make clear development direction, precise career positioning, and improve their comprehensive ability of career literacy when they are choosing their life near graduation.
In recent years, although many scholars have studied topics such as prediction and planning of college students’ career development direction based on data mining and artificial intelligence [15, 16] and proposed some research methods there are still shortcomings. For example, in the current study, a large proportion of students’ potential graduation destinations are derived from questionnaires, and career counseling teachers, then, plan the corresponding career development directions for students based on their graduation destination choices. Alternatively, students’ career paths are predicted in a single dimension based on their school performance data, and then further career guidance is provided. There is relatively little research on predicting students’ career development direction based on their comprehensive school data and educational big data mining to help students’ career guidance work more scientifically. Therefore, the current research method can continue to be improved, the impact dimension of the prediction model needs to be expanded, and the accuracy rate still has some room for improvement.
Employment recommendation algorithms are an applied branch of deep learning-based recommendation algorithms in big data technology. Recommendation systems [17] can effectively filter information to help users retrieve information resources that meet their needs in a personalized manner and alleviate the problem of information overload. Recommendation techniques have been continuously developed and updated and have been widely used in education, healthcare, e-commerce, social networks, and other fields. After collaborative filtering algorithm was proposed, recommendation system gradually became a new research hotspot and also faced the data sparsity problem (the number of users’ ratings for recommendation items is too small) and cold start problem (no rating data for new recommendation items and new users). Deep learning is a machine learning algorithm with recognition, analysis, and computation, which brings new opportunities to alleviate the data sparsity and cold start problems. Since 2015, the algorithm has been widely used in target retrieval [18], face recognition [19], and text generation [20], and the gradual maturity of deep learning models brings new opportunities for the development of recommendation systems.
The deep neural network model was incorporated into the field of video recommendation by Deldjoo et al. [21], which chose the YouTube video site for simulation experiments. YouTube video site is characterized by a large number of registered users, faster video updates, and varying length and number of videos, which makes it difficult for the traditional recommendation algorithm to recommend video content that matches the user’s preferences. The recommendation process is divided into 2 stages: candidate set generation and video sorting. The candidate set generation stage can be regarded as a video screening process, i.e., selecting a collection of videos similar to the user’s viewing history from hundreds of existing videos based on the user’s viewing history as the candidate videos for the next recommendation, and in this stage they treat the video recommendation problem as a multiclassification problem, using deep neural networks to users and videos modeled by a prediction function to calculate the probability of a user watching a video type at a certain moment. Huang et al. [22] proposed a wide and deep (Wide & Deep) model to solve a large-scale online recommendation problem, which is a model combining a single-layer Wide part and a multilayer Deep part. The model mainly uses the Wide part to learn the features of the target user and the Deep part to generalize similar recommendation items, which can be trained on 500 billion samples and can effectively alleviate very sparse data. The model can also be used for classification, regression, and finding problems, and the shortcoming of the model is that it requires human feature engineering.
Recurrent Neural Network (RNN), which includes bidirectional recurrent neural network and Long Short Term Memory (LSTM) network, is used in deep neural networks, where a model is trained and given an in the input layer; a specific is obtained in the output layer, but it is only suitable for sequences with no relationship between the front and back inputs at all. In recommendation, LSTM [23] and Gated Recurrent Unit (GRU) [24] are usually used to handle long sequential information in recommendation problems. Convolutional neural networks (CNNs) are also commonly used in recommendation algorithms due to their robust feature representation capability [25].
The current recommendation system is more about the user’s rating or feedback on the recommended item but ignores the characteristics of the user and the recommended item itself, and the current research lacks appropriate modeling methods to extract the features, linear and nonlinear relationships between the user, and the recommended item in multiple dimensions. Therefore, the next research needs to introduce more diverse ways to extract the features of users and recommendation objects.
3. Algorithm Design
In this paper, DenseNet [1, 2, 26] is used as an employment recommendation model to predict the conforming employment direction and timely intervention in employment expectation by analyzing the comprehensive quality of students from multiple perspectives and multiple dimensions. In addition, to address the problem of poor user feature representation, this paper introduces an attention mechanism to amplify important features.
3.1. DenseNet-Based Multiperspective Employment Recommendation Model
Studies have shown that the accuracy of convolutional neural networks for image recognition can be improved by increasing the depth or width of the network, but when the depth or width of the network increases to a certain level, a problem arises: the input information or gradient information may disappear after passing through many layers. In order to solve the gradient disappearance, maintain the feed-forward property and ensure the maximum information flow between layers; each layer of DenseNet is connected to all previous layers, and the feature mapping of the current layer is used as the input of all subsequent layers. The DenseNet architecture has various advantages over other traditional models. DenseNet architecture provides relatively the best representation of images, and the enhanced efficiency of the parameters makes it easier to train the network model. Most importantly DenseNet improves the declining accuracy caused by vanishing gradients in case of higher level neural networks. Various studies have been conducted implementing the DenseNet-based deep learning model. As an example, the study in [27, 28] implemented a visualization-based technique, wherein malware binaries were presented as two-dimensional images which were classified using a deep learning model. The framework used a reweighted class-balanced loss function in the last layer of the DenseNet model to achieve enhanced performance in the classification of malware by handling imbalanced data issues. The study in [29] proposed a metastatic cancer image classification model using DenseNet block which could efficiently detect metastatic cancer in small image patches captured from larger digital pathology scans. The model yielded better accuracy in comparison to ResNet34 and VGG19.
The DenseNet model consists of dense blocks and transition layers connecting the dense blocks. In a dense block of layers, the input of each layer is defined as ; then, the input of the -th layer can be represented by
Among them, the function is a composite function consisting of three consecutive operations, which perform Batch Normalization (BN), ReLU activation operation, and convolution operation, respectively; is the aggregation operator, which indicates that the feature maps of each layer are connected.
Although DenseNet alleviates the gradient disappearance problem, enhances feature reuse, and facilitates feature propagation, it suffers from overfitting and large memory consumption due to the direct connection between every two layers. -order Markov model is a statistical model whose current state depends only on the previous states and is independent of the states before the previous . represents the dependency relationship, which is defined as shown
The second-order Markov model is a submodel of the -th-order Markov model, whose current state is only related to the previous two states and has higher classification accuracy and generalization ability. Both DenseNet and second-order Markov models emphasize in principle that the current state depends on the previous states and exploits the previous states or features. Therefore, in this paper, we use the idea of second-order Markov to simplify DenseNet and propose 2-DenseNet, which reduces the number of connections by aggregating the inputs of the first 2 layers for use in subsequent layers, thus reducing the network parameters and improving the network convergence speed. The model has better classification accuracy and better generalization ability than the traditional neural network. In a dense block of 2-DenseNet, the input of the -th layer is only from the first 2 layers, which can be represented by
The similarity between DenseNet and 2-DenseNet is that the -th layer depends on the previous layers and achieves feature reuse; the difference is that the -th layer of the former needs to connect all the previous layers, while the connection of the latter is targeted and regular, and the -th layer only depends on the first two layers. To further explore the amount of operations in DenseNet and 2-DenseNet, the forward propagation operation for the -th layer in a dense block is defined as shown
There are connections in the DenseNet network and connections in the 2-DenseNet network. When , 2-DenseNet reduces the number of connections by 46.7% compared with DenseNet, and the number of parameters is also reduced. 2-DenseNet can achieve feature reuse and mitigate gradient disappearance, but also reduce the number of connections, which greatly reduces the amount of operations and makes the network converge faster and run shorter and solves the problems in DenseNet while ensuring the classification accuracy. The problem in DenseNet is solved with the guarantee of classification accuracy.
3.2. Attentional Mechanism
Recently, the attention mechanism has been widely used in the fields of speech recognition, image recognition, and natural language processing, and it can be used to guide the network to focus on the most discriminative information in the input data. In case of deep neural network architecture, the attention mechanism is used for handling challenges associated with natural language processing (NLP) such as summarization, comprehension, and story completion. The use of attention mechanism improved the encoder-decoder model for machine translation. The decoders were able to utilize the most important and relevant parts of the input sequence in a flexible manner using weighted combination of the encoded input vectors, wherein the most relevant vectors would be assigned the highest weights. To further enhance the network performance, this paper proposes a new attention mechanism module, which consists of a convolutional module and an attention module, as shown in Figure 1.
The attention mechanism takes the feature mapping of the main network 2-DenseNet as input and assumes that the input vector is . Figure 2 shows the convolution module on the left, denoted by , and the attention module on the right, denoted by . The convolution module is a convolution in a 2-DenseNet dense block, and the attention mechanism module consists of one global pooling layer and two convolution modules. By performing global pooling on the feature map of , a , one-dimensional feature vector is obtained, which represents the importance of each channel information. convolution module performs BN, ReLU, and convolution operations, which are designed to perform nonlinear transformation on the result after global pooling. In this paper, the output of the global pooling layer and the first convolutional module is used as the input of the second convolutional module, and the output of the second convolutional module is used to adjust the output of the original network, whose output feature map size is . Finally, the input feature map is fully multiplied with the feature map to obtain the weights of each channel, and the network will automatically focus on the channel with the larger weight and learn the useful features. The final result can be expressed as after the attention mechanism module.
3.3. Overall Framework
In this paper, after introducing the attention mechanism module in 2-DenseNet, the general framework is shown in Figure 3. The network contains three dense blocks, and the layer between the blocks is called the transition layer, which performs the convolution and merging operations, and the transition layer consists of a convolution layer and a average pooling layer. After each dense block, a global average pooling is performed, and then a Softmax classifier is attached, which is used to supervise the learning of the network model.
4. Experiments
4.1. Experimental Preparation
A total of 1,600 students in a major from the class of 2011 to the class of 2019 at a university were used as the study population, and comprehensive data from the students’ four years in school as well as personal technical information data were collected as the data sample for this experiment. Seven different career paths were studied in this paper: further education, study abroad, manufacturing, finance, information technology, operations, and administration. The data set is divided according to the training set, validation set, and test , 15%, 25%, as shown in Figure 2. The data sources of this paper are the database of basic information of students, the comprehensive database of academic performance in the four years of school, the database of student academic management, and the database of employment information of graduates. With the help of teachers and counselors from relevant departments of this college, the data related to a total of 1600 students from the class of 2011 to the class of 2018 were successfully obtained. It includes students’ personal information form, academic achievement form, school activity extra credit form, award and merit public announcement form, and employment information form. The corresponding student grades for each of the seven different career directions are shown in Table 1.
4.2. Analysis of Model Training Process
In this paper, the training model uses stochastic gradient descent optimizer with momentum set to 0.9 and weight penalty set to 0.00001. The initial learning rate is 0.1, and the output layer uses Softmax function for classification. The cosine learning rate is used and the termination learning rate is 0.001. 100 iterations of epoch are used for model training, and the loss curves on the training and validation sets are shown in Figure 4, and the accuracy variation is shown in Figure 5. It can be seen that the model in this paper can converge on the student employment dataset, and the model converges faster in the first 40 iterations and tends to be smooth thereafter. As the training proceeds, the model prediction accuracy gradually improves, and the recommendation performance is good.
4.3. Analysis of the Effect of Employment Referrals
We first validated the performance change of the model before and after improving DenseNet. The horizontal axis of the ROC curve is the False Positive Rate (FPR) and the vertical axis is the True Positive Rate (TPR). TPR is the probability of predicting positive samples as positive, and the larger the TPR value, the better the prediction. From the ROC curve in Figure 6, it can be seen that the ROC curve of the improved 2-DenseNet is closer to the upper left corner, thus proving that the improved prediction model has better prediction accuracy.
In addition, we investigated the effect of the attention mechanism on the model performance. Figure 7 shows the classification accuracy after inserting the attention module between the convolutional layers within different DenseBlock. A DenseBlock is a module in a CNN which connects all the layers having matching feature-map sizes with each other. In order to maintain the feed-forward nature of the network, each layer acquires additional input from all the preceding layers and passes its feature maps to the subsequent layers. The green color means that attention is inserted in multiple blocks (the current block and its previous convolutional layers), while the yellow color means that it is inserted in only one block. It can be noticed that the insertion within the second DenseBlock is significantly better. This is due to the important role of this layer in feature extraction. Table 2 shows the average prediction accuracy after adding attention between all the convolutional layers. Since the attention module amplifies the multidimensional features and important features, the prediction accuracy is also significantly better, providing a prediction model with realistic meaning for the employment rate and quality of college students.
We compared with the commonly used deep feedforward neural network (NN), support vector machine (SVM), and random forest algorithms. To avoid the effect of the size of the dataset, we evaluated the models using accuracy, recall, and F1-score, and the results are shown in Table 3. The prediction accuracy of the deep learning model proposed in this paper outperforms other methods.
5. Conclusions
Career guidance for college students is an important part of college education work, but its effectiveness has not been given full play. With the arrival of big data era, big data technology is more and more widely and deeply applied to many fields in all walks of life, and the employment guidance work of college graduates will also face new challenges and opportunities. The article analyzes the difficulties and problems in the current employment work of colleges and universities and discusses that using big data technology to innovate employment guidance work is of great importance to improve the employment service system, enhance the quality and effect of employment service, and promote the active employment of graduates. In this paper, we adopt the improved DenseNet algorithm integrating attention mechanism; input perspective including English proficiency, politics, science and technology innovation, art, social volunteering, and other six-dimensional influence dimensions; set seven categories of career development direction of further education, study abroad, manufacturing, finance, clerical, operation, and administration; and build a prediction model of college students’ career development direction. It provides universities with a more scientific and effective decision basis for career development direction, which helps college students plan their career development direction correctly and adjust their employment expectation according to their own situation in time. The experiments in this paper prove the validity of the prediction results. The study could be further evaluated in the future by comparing the same with other classical models like ResNet34, VGG19, and similar competitor techniques in association to the use of additional metrics like precision, specificity, and sensitivity.
Data Availability
The datasets used during the current study are available from the corresponding author on reasonable request.
Conflicts of Interest
The author declares that he has no conflict of interest.