Quantitative Analysis of the Impact of Cloud Computing Service Models on the Employment Structure of College Graduates

Applying the original random forest algorithm to the predictive analysis of the employment structure of college graduates can identify the factors and their interrelationships that have a significant impact on graduate employment, thereby helping to predict future employment trends and structural changes. However, when dealing with college graduate employment data, the challenge lies in the high dimensionality and diverse types of data, including dimensions such as educational background, skill proficiency, and industry demand. The complexity of such data makes it difficult for the original random forest algorithm to effectively process these high-dimensional data while maintaining a high accuracy rate. Additionally, using traditional methods to equally vote across all attributes might overlook the influence of certain key factors on employment outcomes, thus reducing the accuracy of predictions.

To effectively address these challenges, this paper introduces the Spark-IRF algorithm for the predictive analysis of the employment structure of college graduates. This algorithm, based on the powerful Spark cloud computing platform, optimizes the original random forest algorithm. First, when handling high-dimensional employment data, the algorithm enhances data processing efficiency and execution speed through dimensionality reduction. Secondly, in the ensemble process of the trees, not every tree is given the same weight for voting; instead, different weights are assigned based on the performance of the trees, which helps to reduce model bias and increase the accuracy of prediction results.

The employment structure data for college graduates covers multiple dimensions, such as employment status, industry classification, and work location, which are often inconsistent in format and standards. For this purpose, it is first necessary to standardize the employment status (e.g., employed, unemployed) to "Y" (yes) and "N" (no); industry classification may involve multiple levels, such as primary and secondary industry classifications, which need a unified format (e.g., IND1001, IND1002); city categories, including first-tier cities, second-tier cities, etc., also require uniform coding conversion (e.g., WP0001, WP0002). Additionally, specific work addresses, due to privacy and accessibility issues, can be converted to corresponding geographic codes or postal codes.

Considering the characteristics of college graduate employment structure data, the process of the Spark-IRF algorithm is divided into three stages: data preprocessing, dimensionality reduction of communicable disease data, and weighted voting method (WVM).

Data preprocessing aims to transform the original employment data into a consistent format for more accurate analysis. For example, employment status can be standardized to "employed" or "unemployed"; industry categories are divided into primary and secondary industry codes; work locations include specific cities and districts. For geographic location data, converting specific work place addresses to postal codes can be achieved using the cloud computing service's map API, which not only protects individual privacy but also facilitates uniformity and standardization in the dataset.

In terms of data dimensionality reduction, this study adopts a dimension reduction method based on the importance of feature attributes of employment data, as shown in Figure 1. During the training process of each decision tree, the information gain ratio of each feature attribute variable in the training data subset is calculated and these variables are arranged in descending order of their information gain ratios. Subsequently, the top k variables with the highest information gain ratios are selected as important variables, and a random selection of (l-j) variables from the remaining (L-j) is made to participate in subsequent model training. This reduces the dataset's dimensionality from the original L to l. This method not only significantly reduces data complexity but also retains the most critical information for the predictive model, thereby effectively enhancing the accuracy and efficiency of the employment structure predictive analysis.

During the dimensionality reduction process, the first focus is on assessing the importance of feature attributes for node splitting during the decision tree training process. Specifically, this process begins by calculating the entropy of each feature attribute variable before each node split. Entropy, a fundamental concept in information theory, is used to measure the uncertainty of a variable. Here, by calculating the entropy of each feature attribute under different employment statuses, the ability of each feature to distinguish between different employment outcomes can be quantified.

Suppose the number of different target variable attribute values in the dataset T_u is represented by f1, and the probability value of type x in all types within the target variable subset is represented by o_x. The following equation defines the entropy of the target variable in the training data subset T_u(u=1,2...j):

$E N\left(T_u\right)=\sum_{x=1}^{f 1}-o_x \log o_x$             (1)

11.png

Figure 1. Dimensionality reduction process for employment structure data of college graduates

2.png

Figure 2. Configuration of the spark cloud computing experimental platform

Next, for each training data subset T_u in the decision tree, the entropy value b_uk of each input variable relative to the target variable is calculated. This step further analyzes the correlation strength between each input variable (such as major, skills, education, etc.) and the final employment outcome. By assessing the entropy values of these input variables, the algorithm can identify the variables that most influence the prediction of employment status. The following equation provides the entropy calculation formula for each input variable b_uk:

$E N\left(b_{u k}\right)=\sum_{n \in N\left(b_{u k}\right)} \frac{\left|T_{(n, u)}\right|}{\left|T_u\right|} E N\left(n\left(b_{u k}\right)\right)$              (2)

Finally, the final step of the dimensionality reduction process is completed by calculating the self-split information of each input variable. Self-split information quantifies how much the entropy of the target variable is reduced under a given input variable, thus providing a measure of the quality of feature splitting. Variables that can maximize information gain, i.e., those that can most reduce the uncertainty of employment status, are selected as key features for building the decision tree. Suppose the number of different values of b_uk is represented by f2, and the probability value of type x in all types within variable b_k is represented by o_(x,k). The following equation provides the calculation formula for the self-split information U(b_uk) of each input variable:

$U\left(b_{u k}\right)=\sum_{x=1}^{f 2}-o_{(x, k)} \log _2\left(o_{(x, k)}\right)$            (3)

Let n(b_k)ÎN(b_k), the following equation gives the information gain of each feature attribute variable:

$\begin{aligned} & H\left(b_{u k}\right)=E N\left(T_u\right)-E N\left(b_{u k}\right)  =E N\left(T_u\right)-\sum_{n \in N\left(b_{u k}\right)} \frac{\left|T_{(n, u)}\right|}{\left|T_u\right|} E N\left(n\left(b_{u k}\right)\right)\end{aligned}$       (4)

The dimensionality reduction process for college graduate employment data crucially relies on the precise evaluation of the importance of feature variables. Although directly using information gain as a measure of feature importance can effectively identify features that significantly contribute to classification, this method is prone to overfitting, especially when dealing with complex and high-dimensional employment data. Overfitting results in excellent model performance on the training set but significantly reduced generalization ability on unknown data. To alleviate this issue and further improve the model's accuracy and generalization ability when processing college graduate employment data, using the information gain ratio instead of pure information gain to assess the importance of feature variables has become a necessary strategy.

The information gain ratio selects features that not only provide the maximum information gain but also avoid overfitting due to the high dispersion of the features themselves by considering the amount of information inherent in the features (i.e., the rate of change in information before and after splitting). The definition of the information gain ratio for b_uk is given by the following equation:

$H E\left(b_{u k}\right)=\frac{H\left(b_{u k}\right)}{U\left(b_{u k}\right)}$           (5)

Since college graduate employment data usually includes various factors such as personal skills, educational background, and industry demands, the importance of these factors is not consistent. By calculating the information gain ratio of each feature variable attribute to assess its importance, the features that most influence the prediction of changes in employment structure can be identified. To optimize model performance and improve prediction accuracy, this paper further defines the importance of feature variable attributes and performs data dimensionality reduction accordingly. The importance NU(b_uk) of b_uk is defined as follows:

$N U\left(b_{u k}\right)=\frac{H E\left(b_{u k}\right)}{\sum_{(x=1)}^L H E\left(b_{(u, x)}\right)}$             (6)

In the flow of the Spark-IRF algorithm for predictive analysis of the employment structure of college graduates, the WVM is a key step used to enhance the model's prediction accuracy. Employment structure prediction focuses on a comprehensive consideration of different industries, skills, and educational backgrounds, where the predictive performance of each decision tree may vary based on different features. Therefore, the Spark-IRF algorithm introduces WVM to assign different weights to the voting results of each tree, based on each tree's classification or regression accuracy during training. WVM ensures that predictions on the test dataset are not merely a simple majority vote but consider the weighted accuracy of each tree's prediction, effectively reducing the impact of noisy data and enhancing the accuracy and reliability of the model's predictions for changes in the employment structure of college graduates.

Specifically, after the model training phase, the performance of each tree is tested using Out-of-Bag (OOB) data. Each decision tree g_u tests its performance using the OOB data OOB_u not involved in training, thereby calculating each tree's classification accuracy ZX_u. Classification accuracy is defined as the proportion of correctly classified samples out of all the OOB samples tested by the tree, including correctly classified and misclassified categories. This ratio reflects each tree's effectiveness in dealing with unseen data, providing a basis for subsequent weighted voting. Suppose the indicator function is represented by U(·), the correct classification value by b, and the incorrect classification value (c≠b) by c, then the calculation formula is:

$Z X_u=\frac{U\left(g_u(a)=b\right)}{U\left(g_u(a)=b\right)+\sum U\left(g_u(a)=b\right)}$             (7)

In the prediction phase, the entire random forest model predicts each record in the test dataset A. When the target variable is continuous, the random forest is configured as a regression model. In this case, the prediction result consists of the weighted average of all decision trees' predictions, where the weighting is based on each tree's OOB classification accuracy, directly relating each tree's voting weight to its prediction accuracy. Suppose the voting weight of decision tree g_u is represented by q_u, then the definition formula for the weighted regression result G_e(A) is:

$G_e(A)=\frac{1}{j} \sum_{u=1}^j\left[q_u \times g_u(a)\right]=\frac{1}{j} \sum_{u=1}^j\left[Z X_u \times g_u(a)\right]$          (8)

For discrete target variables, the prediction result is the classification result obtained through the weighted majority vote of all decision trees. Each tree gains its voting weight based on its classification accuracy on OOB data. Thus, for each sample in the test dataset A, the algorithm aggregates the weighted voting results of all decision trees to determine the final category. The definition formula for the weighted classification result G_z(A) is provided below:

$G_z(A)=M A_{u=1}^j\left[q_u \times g_u(a)\right]=M A_{u=1}^j\left[Z X_u \times g_u(a)\right]$          (9)

Figure 2 presents a diagram showing the configuration of the Spark cloud computing experimental platform used in this study. In the process of applying the Spark-IRF algorithm for predictive analysis of the employment structure of college graduates, a key step is using Spark's distributed computing capability to process and analyze large-scale employment datasets. Taking "field of study" as an example feature attribute, different fields of study impact graduates' employment outcomes differently. Initially, a Resilient Distributed Dataset (RDD) generated by the field of study attribute is used to process data in parallel and calculate the data volume under various fields of study. In the Spark environment, a new RDD is generated through filtering operations, and then the count() function is used to calculate the total number of data items in this new RDD, thereby determining the sample size for each field of study. The Spark cluster assigns a task to each Partition of the RDD to calculate the data items in each partition in parallel. After completing these partition calculations, the cluster controller aggregates these results, ultimately obtaining the data count results for specific fields of study. This processing method fully utilizes Spark's parallel computing features, significantly enhancing the efficiency of processing large-scale employment datasets, and providing effective data support for predictive analysis of the employment structure of college graduates.

From the results processed in Table 1, it is evident that this study successfully organized the employment data of college graduates into structured information, encompassing multiple dimensions including, but not limited to, educational level, field of study, industry classification, city of residence, academic performance, and internship experience. This structured processing of data provides a solid foundation for subsequent analyses. Particularly, by introducing the "employment intention" indicator, the study is able to more intricately explore the relationship between graduates' employment preferences and market demands. Utilizing the Spark-IRF algorithm to handle these large-scale datasets not only demonstrates the capability to manage large data sets but also ensures the efficiency and accuracy of the analysis, thereby making predictive analysis of the employment structure of college graduates possible. The experimental results show that through the application of the cloud computing service model and Spark-IRF algorithm, the study effectively quantifies the impact of various factors on the employment structure of college graduates and reveals the complex connections between educational backgrounds, professional skills, and job market demands.

Table 2 and Figure 4 present the Mean Decrease Accuracy (MDA) and Mean Decrease Gini (MDG) metrics for variables affecting the employment structure of college graduates, as analyzed using the proposed method. The data indicate that academic performance, field of study, internship experience, and graduation degree are the main factors affecting employment structure. Academic performance shows the highest importance in both MDA and MDG metrics, indicating that graduates with excellent academic records are more likely to secure better employment opportunities. The importance of the field of study ranks second, reflecting significant differences in job market demands across different fields. Internship experience is also an important factor, suggesting that practical experience significantly enhances graduates' employment competitiveness. The impact of graduation degree on employment structure is also significant but relatively lower in importance. Employment intention, skills and abilities, city, and job market conditions have relatively minor impacts, but are still not to be ignored. These experimental results emphasize the effectiveness of the cloud computing service model used in this paper for quantifying the analysis of factors influencing the employment structure of college graduates. By processing large-scale job market data using the proposed methods, not only are the key drivers of changes in employment structure revealed, but also deeper insights into subtle changes in the job market are provided.

Table 1. Results after data preprocessing

Table 2. MDA and MDG of variables affecting the employment structure of college graduates using the methodology in this study

(Note: MDA - Mean Decrease Accuracy, MDG - Mean Decrease Gini)

4.png

Figure 4. Importance ranking chart of independent variables in employment structure data of college graduates using the proposed method

5.png

Figure 5. Average accuracy of different predictive methods for college graduate employment structure under varying numbers of decision trees

Figure 5 demonstrates the average accuracy of different predictive methods, including the proposed method, traditional Random Forest (RF), Improved Random Forest (DRF), and Spark Machine Learning Random Forest (Spark-MLRF), in predicting the employment structure of college graduates as the number of decision trees changes. The data show that as the number of decision trees increases, the predictive accuracy of all methods improves, but the proposed method consistently outperforms the others. Particularly when the number of decision trees reaches 9, the accuracy of the proposed method reaches 0.912, while the accuracies of RF, DRF, and Spark-MLRF are 0.82, 0.836, and 0.882, respectively. This indicates that increasing the number of decision trees can enhance predictive performance, but the proposed method demonstrates higher accuracy at each decision tree count level. These experimental results fully confirm the effectiveness and superiority of the Spark-IRF algorithm proposed in this paper for the predictive analysis of the employment structure of college graduates. By comparing with RF, DRF, and Spark-MLRF, the proposed method not only shows efficient processing capability when dealing with large data sets but also has significant advantages in predictive accuracy. This high-performance predictive analysis capability highlights the strong potential of cloud computing technology in data processing and analysis, particularly in understanding and predicting complex changes in employment structures.

6.png

Figure 6. Recall rates of different predictive methods for college graduate employment structure under varying data volumes

Figure 6 shows the comparison of recall rates of different predictive methods for college graduate employment structure as the data volume increases. The Spark-IRF algorithm proposed in this paper exhibits significant stability and superior performance, especially when the data volume exceeds 100 records, maintaining a recall rate above 0.89. Compared to other methods, particularly at higher data volumes (such as 2000 and 3000 records), this method's recall rate remains around 0.91, showing high stability and accuracy. In contrast, the recall rates of RF and DRF fluctuate more across different data volumes and overall perform less effectively than the proposed method. The Spark-MLRF performs similarly to the proposed method at smaller data volumes (1 to 50 records) but begins to decrease in recall rate as data volume increases, especially underperforming at larger data volumes. By demonstrating high recall rates at different data volumes, the proposed method not only effectively processes and analyzes large volumes of employment market data but also ensures the accuracy and reliability of the predictive results.

7.png

Figure 7. Accuracy of different college graduate employment structure prediction methods under various lag times

Figure 7 displays the comparison of accuracy between the proposed method and the ARIMA model in predicting the employment structure of college graduates at different lag times. The proposed method exhibits higher accuracy at lower lag times (2 to 10), particularly peaking at a lag time of 4 with an accuracy of 0.885. However, as the number of lags increases, the accuracy of the proposed method shows a decreasing trend, particularly when the lag exceeds 12, with the accuracy significantly dropping to 0.62 at 18 lags. In contrast, the ARIMA model shows generally lower accuracy across all lag times and experiences greater fluctuations, though it peaks at a lag of 6 with an accuracy of 0.77, subsequently decreasing as the number of lags increases. These results indicate that the proposed algorithm, especially at lower lag times, provides higher prediction accuracy compared to the traditional ARIMA model, demonstrating its effectiveness and superiority in predicting the employment structure of college graduates. Although the accuracy of the proposed method declines with increasing lag times, it still maintains a relatively high level, possibly due to the Spark-IRF algorithm's better capability to handle and analyze complex relationships in time series data, thereby mitigating the impact of lag effects to some extent.