Hybrid Artificial Intelligence Framework Evaluates Higher Education Teaching Effectiveness and Reduces Bias

A newly developed computational architecture that integrates deep neural learning, fuzzy logic inference, and metaheuristic optimization has demonstrated significant improvements in evaluating higher education faculty performance while mitigating demographic bias. The framework, designated as FAM-Teach by researcher Junjun Zheng, was formulated to address persistent structural limitations in conventional university assessment protocols, which predominantly depend on subjective student surveys, inconsistent departmental metrics, and static numerical aggregations. By processing heterogeneous educational records through a multi-stage analytical pipeline, the system extracts complex non-linear pedagogical patterns from student feedback and classroom indicators while systematically accommodating qualitative uncertainty. In rigorous experimental evaluations conducted on educational assessment data comprising hundreds of student records, the integrated triadic model outperformed conventional statistical indices, standalone fuzzy systems, and traditional machine learning classifiers across predictive accuracy, internal consistency, and algorithmic fairness metrics, establishing a robust computational paradigm for evidence-based instructional review.

Evaluating teaching effectiveness in modern higher education has become increasingly challenging due to the expansion of online and hybrid learning environments and the multifaceted nature of pedagogical delivery. Traditional evaluation schemes, typically structured around end-of-semester questionnaires, peer classroom observations, and administrative compliance checklists, frequently suffer from rater subjectivity, uncalibrated inter-departmental criteria, and delayed feedback cycles that offer little constructive guidance for professional development. Furthermore, standard statistical aggregation methods, such as weighted arithmetic means and analytical hierarchy process indices, enforce rigid linear assumptions that fail to capture the complex, higher-order interactions between teaching behaviors and student outcomes. Conventional algorithmic approaches also lack dynamic adaptability, relying on fixed numerical weights that cannot respond to contextual nuances across diverse disciplines. To bridge these methodology gaps, FAM-Teach combines three distinct computational paradigms into a unified decision architecture: a deep neural network for feature extraction, fuzzy inference systems for modeling linguistic ambiguity, and the Firefly Algorithm for global parameter calibration.

The deep learning component of the FAM-Teach framework is structured as a lightweight fully connected neural network specifically engineered to model non-linear relationships without overfitting limited institutional datasets. Operating on fewer than ten thousand trainable parameters, the network architecture incorporates an input layer configured for normalized academic indicators and natural language features, followed by two hidden layers containing sixty-four and thirty-two neurons, respectively. Both hidden layers utilize rectified linear unit activation functions to encourage sparse representation and prevent vanishing gradients, combined with batch normalization layers to stabilize internal feature distributions and mitigate covariate shifts during backpropagation. To ensure generalization stability across small sample sizes, the design applies a dropout rate of 0.3, early stopping after ten epochs of non-improving validation loss, and L2 weight regularization scaled at 0.001. Depending on whether the institutional objective requires continuous effectiveness ratings or discrete performance stratification, the output layer terminates in a linear activation unit or a sigmoid-activated classification neuron.

To capture qualitative insights from unstructured student commentary, FAM-Teach implements a targeted natural language processing pipeline that converts textual feedback into structured semantic features before neural integration. Rather than employing computationally intensive transformer models that could overparameterize smaller academic datasets, the feature-engineering pipeline extracts term frequency-inverse document frequency vectors across the top three hundred words, computes sentiment polarity scores, derives lexical variety indices, and calculates keyword frequencies for core pedagogical dimensions, such as instructional clarity, student engagement, responsiveness, and assessment fairness. To eliminate dimensional sparsity and avoid multicollinearity before combining textual representations with structured numerical indicators, the text-derived embedding space undergoes dimensionality reduction through principal component analysis. The resulting orthogonal semantic components are subsequently merged with normalized institutional records—including attendance rates, course completion metrics, and evaluation scores—into a unified hybrid input vector that feeds directly into the subsequent deep feature extraction and fuzzy reasoning stages.

Fuzzy logic provides the interpretability foundation of FAM-Teach, translating crisp numerical measurements and ambiguous linguistic feedback into formal mathematical representations of educational quality. Grounded in compensatory evaluation theory, the system evaluates faculty across four core competencies: student involvement, clear instruction, fair grading, and desired learning results. Each pedagogical criterion is mathematically modeled through continuous Gaussian membership functions defined by ideal reference values and variance parameters that dictate evaluation tolerance boundaries. The individual memberships are aggregated through an analytical expression termed the Holistic Fuzzy Teaching Performance Metric, which integrates a hyperbolic tangent function to reflect the diminishing marginal gains of teaching performance at advanced competency levels. Additionally, a Heaviside step threshold function acts as an institutional gating mechanism, ensuring that instructional performances falling below minimum acceptable standards contribute minimally to the overall score. The resulting continuous index, calculated on a hundred-point scale, maps directly into four practical administrative classifications: Needs Improvement, Average, Good, and Excellent.

Adaptive parameter optimization within the FAM-Teach pipeline is executed by the Firefly Algorithm, a nature-inspired metaheuristic search mechanism that fine-tunes indicator importance weights and fuzzy membership bounds. Unlike genetic algorithms that utilize stochastic crossover operators or particle swarm optimization routines prone to premature local convergence, the Firefly Algorithm relies on distance-based light attractiveness decay to balance global exploration and local parameter exploitation across multimodal fitness landscapes. Configured with a population of twenty-five fireflies evaluated over one hundred iterations, the optimization engine initializes attractiveness at 1.0, sets the light absorption coefficient to 0.50, and employs an adaptive randomization factor that cools from 0.20 to refine weight vectors toward global minima. The multi-objective loss function minimized by the swarm combines empirical mean squared error with parameter regularization and an entropy-regularized indicator correlation constraint. By penalizing anomalous covariance matrix determinants, this constraint prevents learned parameter configurations from collapsing into degenerate weight distributions that misrepresent the multidimensional structure of teaching.

The empirical validation of FAM-Teach was conducted using the publicly accessible Kaggle Teaching Quality Evaluation Dataset, derived from an initial cohort of 1,200 students across five academic disciplines. Prior to computational modeling, the raw records underwent rigorous preprocessing to ensure data consistency, including structural integrity verification, duplicate removal, and interquartile range filtering at a 1.5 threshold to eliminate distortive outliers. For missing data, variables exhibiting less than five percent missingness were completed using departmental stratified median imputation, while attributes with five to fifteen percent missingness received k-nearest neighbor imputation; records missing more than twenty percent of values were discarded. This quality filtering produced a balanced final dataset of 650 valid instances characterized by low sparsity below 0.02, minimal Kullback-Leibler divergence of 0.03, and variance inflation factors below 3.8, confirming an absence of adverse multicollinearity across predictors.

Experimental testing under a stratified ten-fold cross-validation scheme demonstrated that FAM-Teach achieved superior predictive accuracy compared to traditional statistical and machine learning baselines. The integrated architecture attained an overall classification accuracy of 94.8 percent, an F1-score of 0.94, a mean absolute error of 0.041, a root mean squared error of 0.057, and a coefficient of determination of 0.96. In contrast, the conventional statistical weighted-index baseline achieved only 84.2 percent accuracy with an F1-score of 0.82, while a standalone Mamdani fuzzy inference system recorded 88.6 percent accuracy and an F1-score of 0.87. Standard supervised baselines likewise trailed the proposed model: logistic regression posted 86.4 percent accuracy, random forest reached 91.3 percent, and a standalone deep neural network attained 92.1 percent. Paired t-tests verified that FAM-Teach’s performance gains were statistically significant, outperforming the pure deep learning baseline with a p-value of 0.003 and a ninety-five percent confidence interval spanning 1.6 to 4.8 percent.

Algorithmic fairness was explicitly evaluated and optimized by introducing demographic parity regularization into the metaheuristic objective function, addressing systemic demographic disparities in historical evaluation data. Incorporating moderate regularization at 0.15 balanced these proportions to 75 and 74 percent, slashing demographic parity difference to 0.012 and equal opportunity difference to 0.018 while preserving 94.8 percent predictive accuracy, demonstrating substantial bias mitigation with minimal accuracy trade-off.

To provide model interpretability for academic administrators, the framework incorporated post-hoc explainability techniques alongside its transparent fuzzy rule base, utilizing Shapley additive explanations and local interpretable model-agnostic explanations. Across sample evaluations, Shapley analysis revealed that student feedback sentiment exerted the greatest absolute impact on effectiveness predictions with a magnitude of 0.104, followed by student attendance rates at 0.080 and instructional clarity ratings at 0.060. However, the Firefly optimization routine accounted for 6.84 seconds, or 89.4 percent of runtime, identifying parameter optimization as the primary computational bottleneck. The author noted that while the framework scaled near-linearly up to 10,000 instances, future implementations should explore adaptive swarm termination and parallel fitness evaluations to deploy the system effectively on resource-constrained educational platforms.

Subject of Research: Higher education teaching effectiveness evaluation using artificial intelligence and fuzzy metaheuristics

Article Title: Evaluation of higher education teaching effectiveness based on fuzzy algorithms and artificial intelligence

Article References: Zheng, J. (2026). Evaluation of higher education teaching effectiveness based on fuzzy algorithms and artificial intelligence. Discover Artificial Intelligence, 6(1), Article 1810. https://doi.org/10.1007/s44163-026-02420-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02420-8

Keywords: artificial intelligence, fuzzy logic, metaheuristic optimization, teaching effectiveness, higher education, machine learning, firefly algorithm, algorithmic fairness, Evaluation, higher, education, teaching