Adaptive Machine Learning Recommendation Model Improves College English Scores and Efficiency

In an eight-week quasi-experimental study conducted within a compulsory undergraduate English curriculum, an adaptive machine-learning recommendation model was associated with significant improvements in student academic achievement and task efficiency compared to conventional uniform instructional delivery. Conducted by researcher Ruochen Song and published in Discover Internet of Things, the investigation followed 124 first-year non-English major undergraduates enrolled in a compulsory College English II course at a public undergraduate university in China during the spring semester of the 2024–2025 academic year. The student cohort was divided by intact administrative classes into an experimental group of 63 learners who received algorithmically personalized learning materials and a control group of 61 learners who utilized standard uniform platform resources. Following the eight-week intervention, the experimental group achieved a mean total score increase of 5.6 points, rising from a baseline of 68.7 to 74.3 points, whereas the control group gained 1.7 points, rising from 69.4 to 71.1 points. This between-group score gain difference of 3.9 points was statistically significant, with a 95% confidence interval spanning [2.84, 4.96], p < 0.001, and an effect size of Cohen’s d = 1.31.

The pedagogical rationale for developing this adaptive framework arises from an ongoing challenge in modern higher education: despite substantial digital repositories across learning management platforms, universities continue to rely on uniform content distribution that overlooks learner diversity. To bridge this gap between technological capability and classroom pedagogy, the study combined mature machine learning methods with established educational and linguistic theories. Cognitive Load Theory informed the system’s ability-difficulty matching mechanism to avoid cognitive overload from excessively difficult materials or reduced efficiency from overly simple content. Vygotsky’s Zone of Proximal Development supported progressive difficulty adjustments to ensure students consistently encounter materials situated slightly beyond their current competence. Self-Regulated Learning Theory guided weekly dynamic learning state updates, while Constructivist Learning Theory underpinned personalized learner profiles. In second language acquisition, the framework drew upon Krashen’s Input Hypothesis regarding comprehensible input and Swain’s Output Hypothesis emphasizing structured language production opportunities. This multi-theoretical synthesis was designed to balance curriculum requirements with diverse language proficiencies.

To evaluate the system under authentic classroom conditions, the quasi-experiment maintained identical course schedules, syllabus requirements, textbook units, assessment standards, and instruction across both classes. A single instructor taught both groups, minimizing instructor-related bias across the research period. The instructional repository contained 186 resources aligned with curriculum objectives, comprising 64 reading texts, 48 listening and audiovisual files, 42 supporting exercises and assessment tests, and 32 supplementary micro-lessons focusing on grammar explanations and targeted vocabulary drills. During the eight-week period, the digital platform logged 4,512 valid learner-resource interactions from the 124 students without interfering with normal study routines. Across the full interaction matrix of 23,064 potential learner-resource pairings, these observations yielded an interaction density of 19.56 percent and a matrix sparsity of 80.44 percent. Telemetry tracked resource visits, individual browsing durations, task completion records, and post-resource test scores, establishing an empirical behavioral foundation for subsequent learner modeling, resource profiling, and recommendation training.

The recommendation model relies on systematic, multi-dimensional feature engineering across instructional materials and student behaviors. Resource difficulty was quantified using a weighted composite of lexical complexity, syntactic complexity, and information density. Lexical complexity was calculated via the Lexical Complexity Analyzer using word frequency from log-transformed CELEX data, type-token ratios, and academic word list proportions. Syntactic complexity was measured using the L2 Syntactic Complexity Analyzer through mean sentence length, clauses per sentence, and passive voice frequency. Information density was assessed with TextLens software by evaluating concept counts and named entity density per 100 words. These three dimensions were weighted at 0.40, 0.35, and 0.25 based on a pilot survey of 15 college English teachers. Furthermore, three experienced teachers categorized resources across four capability dimensions—listening, reading, language knowledge, and comprehensive application—achieving an inter-rater agreement of kappa = 0.84. Semantic vectors were generated using term frequency-inverse document frequency across a 5,000-word vocabulary mapped into a 200-dimensional dense representation.

Learner modeling integrated baseline ability, behavioral preferences, and dynamic learning trajectories into unified feature vectors. Baseline English ability was formulated from pre-test reading comprehension, listening comprehension, and language knowledge scores weighted at 0.40, 0.35, and 0.25, respectively, and min-max normalized. Behavioral preference vectors captured the relative frequency of visits to each resource category, reflecting individual learning habits and thematic preferences. Because student competence evolves over time, the model applied a weekly dynamic updating mechanism where recent performance received greater weight through exponential smoothing, preventing recommendation lag. The study also recorded behavioral indicators of motivation, affective engagement, and metacognition, such as active resource click ratios, optional task completion rates, session interruptions, review counts, and re-attempt frequencies. However, the author emphasized that these platform metrics serve as behavioral proxies rather than validated psychometric measures, as automated logs cannot distinguish between intrinsic and extrinsic motivation or capture emotional states like boredom and frustration.

The algorithmic architecture employs a three-stage hybrid recommendation model that combines collaborative matrix factorization with explicit feature fusion. In the first stage, matrix factorization decomposed the sparse interaction matrix of 124 learners and 186 resources into 20-dimensional latent vectors by minimizing squared prediction error. This 20-dimensional latent representation was selected after comparing candidate dimensions of 8, 16, 20, and 32, balancing representation power against overfitting risks on a moderate dataset. In the second stage, explicit learner features—encompassing baseline ability, preference vectors, and dynamic states—and explicit resource attributes—including difficulty values, capability labels, and semantic vectors—were concatenated and passed through a fully connected neural layer with 64 hidden units and ReLU activation. In the third stage, latent collaborative predictions and feature-based scores were combined into a final matching value. The system was implemented in Python using PyTorch, optimized with Adam at a learning rate of 0.001, batch size 64, L2 regularization coefficient 0.001, and early stopping after 10 epochs.

In addition to algorithmic ranking, the framework implemented three pedagogical adaptation mechanisms to maintain cognitive load and task suitability. First, an ability-matching function evaluated the absolute difference between normalized student ability and resource difficulty, screening out resources below an ability-matching threshold of 0.70. Second, an advanced difficulty control mechanism adjusted resource difficulty incrementally based on recent test feedback, using a linear adjustment coefficient of 0.10 to ensure progressive advancement without excessive leaps. Third, a weekly learning-load index was calculated by dividing total study time by the number of completed tasks. When this index exceeded a set limit of five hours or ten tasks per week, the recommendation engine automatically lowered recommendation frequency and reduced high-intensity materials, multiplying the matching score by a load attenuation penalty factor. These pedagogical constraints ensured that recommended top-five resources supported appropriate instructional challenges while mitigating cognitive overload and learning fatigue.

Offline computational testing on partitioned interaction logs demonstrated that the hybrid model outperformed standard recommendation baselines. The 4,512 interaction records were split at the interaction level into training, validation, and test subsets at a 70:15:15 ratio, containing 3,158, 677, and 677 records, respectively. On the test set, the proposed model achieved Precision@5 = 0.73, Recall@5 = 0.66, NDCG@5 = 0.78, F1@5 = 0.69, MAP@5 = 0.71, AUC = 0.82, RMSE = 0.64, and MAE = 0.51. Paired bootstrap tests with 5,000 resampling iterations confirmed that its Precision@5 was significantly higher than all seven baseline algorithms (p < 0.001), including rule-based delivery, content-based filtering, collaborative filtering, matrix factorization, Random Forest, Support Vector Machines, and XGBoost. An ablation study showed that removing semantic vectors, dynamic learning states, or ability-difficulty matching rules caused the most substantial declines in recommendation precision, demonstrating the joint value of multi-source feature fusion and adaptive pedagogical constraints.

Classroom evaluation confirmed that personalized recommendations were associated with higher academic performance, enhanced task efficiency, and increased behavioral engagement. Analysis of covariance with pre-test scores as covariates demonstrated a significant adjusted mean difference in post-test total scores of 3.72 points favoring the experimental group, 95% CI [2.61, 4.83], p < 0.001, partial eta squared = 0.15. In reading comprehension, the experimental group achieved an adjusted post-test mean of 75.9 compared to 73.8 for the control group, representing an adjusted difference of 2.1 points, 95% CI [1.08, 3.12], p = 0.002, partial eta squared = 0.08, with adjusted reading and listening score gains reaching 3.46 and 2.67 points, respectively. Furthermore, although both groups spent similar weekly study hours, experimental students exhibited greater efficiency, averaging 18.6 minutes per task versus 24.8 minutes in the control group, and achieving a higher task efficiency ratio of 2.1 versus 1.4. Experimental learners also demonstrated more continuous study days (9.4 versus 6.8), higher active click ratios (0.76 versus 0.59), higher optional task completion rates (0.69 versus 0.48), and differentiated, personalized learning paths compared to uniform platform trajectories in the control cohort.

Despite these encouraging findings, the author highlighted several methodological limitations that prevent interpreting these results as definitive causal proof. The study relied on a quasi-experimental design using intact administrative classes rather than individual randomization, introducing potential selection bias despite baseline comparability. The Hawthorne effect presents another limitation, as experimental students may have modified their behavior simply because they were aware of receiving personalized resources. While instructor effects were minimized by having the same teacher teach both classes, teacher awareness may have created unintentional expectancy effects. Moreover, platform logs cannot capture internal psychological states or offline study, and model benefits differed across proficiency levels: students with lower baseline skills achieved an average score gain of only 3.10 points compared to 6.40 points among high-level peers, suggesting that personalization efficacy may be mediated by prior knowledge. Finally, the eight-week intervention took place in a single institution and a single College English II course, limiting generalizability to other curricular environments and highlighting the necessity for multi-center randomized trials.

Subject of Research: Machine learning recommendation and adaptation model for college English learning resources

Article Title: Accurate recommendation and adaptation model of college English teaching and learning resources assisted by machine learning

Article References: Song, R. (2026). Accurate recommendation and adaptation model of college English teaching and learning resources assisted by machine learning. Discover Internet of Things, 6(1), Article 268. https://doi.org/10.1007/s43926-026-00535-z

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00535-z

Keywords: college English instruction, educational recommendation systems, machine learning in education, cognitive load theory, zone of proximal development, matrix factorization, feature fusion, lexical complexity analyzer, syntactic complexity, learning analytics, precision at five, quasi-experimental design, task completion efficiency, learning engagement metrics, dynamic learning state, adaptive learning systems, second language acquisition