Researchers at Chinese universities have developed a deep learning framework that combines fuzzy logic to predict the quality of employment for college graduates. The study, published in Discover Internet of Things, analyzes data from 5,623 graduates across six institutions to identify key factors influencing high-quality job outcomes. The model, designated as DFCN, integrates multi-channel data encoding with a self-organizing fuzzy inference system to handle the complex, nonlinear, and high-dimensional nature of employment data. This approach aims to address the limitations of traditional statistical models, which often struggle with the uncertainty and vagueness inherent in student career trajectories and labor market dynamics.
The research team, led by Jiayi Shi and Nan Wang, constructed a dual-channel feature extractor to process both structured and unstructured data. Structured data includes static attributes such as gender, family background, academic performance, and standardized test scores. Unstructured data encompasses behavioral sequences, such as course selection histories, as well as free-text documents like resumes, internship descriptions, and self-evaluations. By using a residual fully connected network for structured features and a bidirectional long short-term memory network for course sequences, the model captures the temporal dependencies in student academic progress. A pre-trained BERT model is employed to encode textual data, allowing the system to extract semantic information from resumes and career aspiration statements.
A critical component of the framework is the cross-modal attention fusion mechanism, which dynamically adjusts the weighting between structured and unstructured features. This mechanism uses a gating network to determine the fusion ratio, ensuring that the model prioritizes relevant information based on the specific student profile. For instance, for students with strong academic records but limited practical experience, the model assigns higher weight to structural features. The fused feature vector is then input into the self-organizing fuzzy inference system, which maps the high-dimensional features into a fuzzy semantic space. This system uses Gaussian membership functions and a data-driven rule generation algorithm to create interpretable fuzzy rules that guide the final decision-making process.
The experimental evaluation was conducted on a dataset comprising employment records from graduates between 2018 and 2023. The target variable, employment quality score, is a continuous value ranging from 0 to 10, derived from a composite of five dimensions: starting salary, job-education match, career development potential, job stability, and job satisfaction. For binary classification tasks, scores of 7.5 or higher were labeled as high-quality employment. The dataset was split into training, validation, and test sets using stratified sampling to maintain class proportions. The model was trained using the Adam optimizer with a learning rate of 3e-4 and early stopping based on validation loss. Hyperparameters were optimized using a combination of grid search and Bayesian optimization to ensure the best possible configuration for the model.
The results demonstrate that the DFCN model significantly outperforms traditional statistical and machine learning baselines. In regression tasks, the model achieved a mean squared error of 0.831 and a coefficient of determination of 0.892, indicating a strong explanatory power for job quality. For classification tasks, the model reached an accuracy of 0.865 and an F1 score that surpassed other models, including random forests and multilayer perceptrons. The kappa coefficient of 0.791 reflects high consistency in the model’s classification results. Ablation experiments confirmed that the self-organizing fuzzy inference system is the most critical component, with its removal leading to an 18.3% increase in mean squared error. This highlights the importance of fuzzy reasoning in handling the uncertainty associated with employment outcomes.
Feature importance analysis using SHAP values revealed that internship duration and GPA are the core determinants of high-quality employment. Course difficulty also ranked highly, suggesting that challenging academic work has long-term value for career prospects. In contrast, family background ranked 12th, indicating a limited impact on the model’s predictions. This finding suggests that the model focuses more on individual efforts and academic achievements than on socioeconomic factors. The similarity between the model’s high-quality employment rating and expert ratings was as high as 0.93, demonstrating a high degree of consistency with industry standards. The model’s ability to capture nonlinear and fuzzy features allows it to provide a more nuanced assessment of student employment potential compared to traditional linear models.
The study also examined the model’s performance across different professional categories. The model showed particularly strong performance in science, engineering, and medicine, achieving accuracy rates of 89.7% and 88% respectively. This is attributed to the highly standardized career paths and skill indicators in these fields. In contrast, students in liberal arts, law, and the arts, who have more diverse and less standardized career paths, showed a slightly lower accuracy rate of 82.3%. Despite this, the model still outperformed traditional random forest algorithms by over 5.3% across all majors. The robustness of the model was further demonstrated by its performance in small sample sizes, such as in the medicine category, where it maintained its predictive advantage.
Uncertainty modeling was a key focus of the evaluation, with the model demonstrating superior capability in generating reliable prediction intervals. The prediction interval coverage, which measures the probability of true values falling within the 95% confidence interval, reached 92.1% for the DFCN model, compared to 77% for traditional deep learning models. The normalized interval width was also the narrowest at 1.48, indicating that the model achieves high coverage while maintaining high accuracy. This ability to accurately reflect the confidence level of predictions is crucial for career counseling, as it provides students and advisors with a clearer understanding of the potential outcomes and risks associated with different career paths.
Sensitivity analysis and fairness assessment revealed that the model exhibits relatively balanced performance across gender groups, with a minimal performance gap of 1.3% in F1-score. However, a moderate performance disparity was observed across income levels, with the model performing best for high-income students and worst for low-income students. The gap was reduced from 5.3% to 2.8% after applying fairness-aware mitigation strategies, such as sample re-weighting and threshold calibration. The school-tier gap also narrowed from 4.8% to 2.1% after these adjustments. These results suggest that while the model may initially reflect existing social inequalities, it can be effectively debiased with minimal loss of predictive performance, making it suitable for deployment in contexts where fairness is a priority.
The study acknowledges several limitations, including the potential for algorithmic discrimination due to the reliance on socioeconomic features. The model is intended to serve as a decision-support tool for career counselors rather than an autonomous decision-making system. Data privacy concerns were addressed by anonymizing all personal identifiers and complying with China’s Personal Information Protection Law. The generalizability of the model is limited to the Chinese educational and labor market context, and replication studies in other regions are necessary to establish cross-cultural validity. The normative construct of employment quality, defined by the composite score used in this study, may vary across different cultural and social norms. The model’s interpretability, while superior to conventional deep learning approaches, remains a challenge for end-users without technical expertise. Future work will focus on developing visual analytics dashboards to present the model’s reasoning process in an intuitive format, highlighting the most influential factors for individual students and suggesting targeted interventions.
Subject of Research: Deep learning and fuzzy logic model for predicting college graduate employment quality
Article Title: Analysis of factors affecting high-quality employment of college students based on deep learning and fuzzy algorithm
Article References: Shi, J., & Wang, N. (2026). Analysis of factors affecting high-quality employment of college students based on deep learning and fuzzy algorithm. Discover Internet of Things, 6(1), Article 245. https://doi.org/10.1007/s43926-026-00502-8
Image Credits: AI Generated
DOI: 10.1007/s43926-026-00502-8
Keywords: deep learning, fuzzy logic, employment prediction, college graduates, machine learning, career counseling, data analysis, artificial intelligence, Analysis, factors, affecting, high-quality
News Source: Blake Davidson. (November 10, 2026). Deep Learning Model Predicts College Employment Quality. Scienmag.
