Machine Learning Model Cracks the Black Box of US News Mechanical Engineering Rankings

The U.S. News & World Report ranking of mechanical engineering graduate programs has long functioned as a black box: departments know their position each spring, but not the recipe of institutional characteristics that produced it. A new study published in Neural Computing and Applications reports that this box can be modeled. A research team built a machine learning classifier that predicts which ranking interval a program will occupy, using nine measurable academic and institutional indicators, and it does so with a test accuracy of 0.91 on the 2024 rankings.

The study, led by Faeze Khademi, Nikith Rachakonda, Matias Salazar Ismodes and Houshang Darabi, is the first, according to its authors, to forecast the U.S. News mechanical engineering graduate program ranking. The team gathered rankings for 2023, 2024 and 2025 from U.S. News & World Report, then collected 59 candidate attributes from department websites, American Society for Engineering Education datasets and other academic sources. Because rankings are driven largely by a reputational survey whose internal weighting remains opaque, the researchers framed the problem as one of reverse engineering: can a model trained only on controllable program characteristics reproduce the outcomes of the mysterious evaluation process?

The answer required some compromises. An early ambition to predict exact rank positions proved infeasible given the dataset’s size and structure, so the team adopted a classification strategy, sorting programs into four intervals: 1–20, 21–50, 51–100 and 101–150. The final modeling dataset comprised 148 universities, distributed as 22 programs in the top interval, 32 in the second, 45 in the third and 49 in the fourth. The counts exceed the numerical width of some ranges because of ties; six universities were tied at rank 17 in the 2024 ranking, which pushed more than 20 institutions into the top-20 category. Programs ranked beyond 150 were excluded.

Data preparation was far from trivial. String-formatted entries were converted to numeric values with a custom extraction function, and missing values were filled using mean imputation, a choice that allowed the researchers to keep a broader set of universities in the model rather than discarding incomplete records. Because the ranking classes are unevenly populated, the team used random oversampling, though notably only during estimation of the AUC confidence interval, not to train the classifier itself. Feature values spanning very different scales were standardized with Z-score normalization, with the scaler fitted exclusively on training data to avoid leakage into the test set.

Of the 59 collected variables, nine predictors survived an iterative subset-evaluation process performed on training data alone. The final set includes the numbers of full, tenured, total and emeritus faculty in the program during the prior year, prior-year full-time enrollment, total university-level funds from two and three years before, prior-year average quantitative GRE scores, and the previous year’s U.S. News ranking itself. The two-year lag on funding reflects a careful timing argument: the rankings are released shortly after the start of each year, before the prior year’s funding outcomes could plausibly have shaped national visibility or reputation. The historical ranking was retained deliberately, the authors explain, because the U.S. News evaluation is reputation-based and temporally dependent; dropping it weakened the models substantially. They interpret it as a predictor of standing persistence, not as evidence of a causal arrow from one year’s rank to the next.

Three algorithms were compared: Decision Tree, Gradient Boosting and Random Forest, all built in scikit-learn with a 70/30 stratified train-test split and a fixed random seed for reproducibility. Random Forest, configured with 200 trees, a maximum depth of 10, and half of the features considered at each split, and validated with 5-fold cross-validation, consistently outperformed the alternatives. On the 2024 test set it achieved accuracy of 0.91, precision of 0.91, recall of 0.90 and an F1-score of 0.91, all macro-averaged across the four classes so that each ranking interval contributed equally regardless of size. The multiclass AUC, computed with a one-vs-one approach and a bootstrap procedure of 1,000 resamples for the 95 percent confidence interval, came in high with a narrow interval, indicating strong discrimination between ranking tiers.

Confusion matrices revealed remarkably few errors. The top-20 category on the test set contained a single misclassification, and the 101–150 group was classified perfectly. The 21–50 interval proved slightly harder, with two programs misplaced, one into the top-20 class and one into the 51–100 class. Feature importance analysis showed the prior-year ranking dominating at roughly 52 percent of total importance, but nearly half of the model’s predictive weight rested on modifiable characteristics: enrollment, faculty counts, research funding and GRE scores. The model was then applied unchanged to the 2025 dataset, where it maintained strong performance across all metrics, which the authors take as evidence of consistency and adaptability across consecutive academic years.

The most provocative output is a sensitivity analysis delivered through a companion website where users can enter hypothetical input values and see the corresponding predicted rank interval. In a case study, the model predicted that Duke University’s Pratt program could move from the 21–50 category into the top 20 under a scenario involving a 1.78-fold increase in faculty from 50 to 89, roughly 1.8-fold increases in total university funds across two consecutive years, a 1.45-fold rise in full-time enrollment from 2,018 to 2,926, and marginal gains in quantitative GRE score and professor counts. The authors are careful, at length, to emphasize that the model is predictive rather than causal: the analysis identifies combinations of program characteristics associated with different predicted intervals, not guaranteed interventions. Changing a single variable such as faculty size, they note, would not necessarily shift the actual U.S. News ranking, and establishing causal effects would require a fundamentally different study design given the reputation-driven nature of the process.

The sensitivity analysis also carries an internal caveat discovered through comprehensive testing of all 150 rank positions. The model behaved consistently only for programs in the top nine ranks within each interval. A program ranked 60 to 100, for example, may not receive a trustworthy strategy for leaping into the 21–50 category. The recommended workaround is stepwise: improve the underlying inputs until the actual rank enters the top nine of its interval, then use the model to plan the transition to a better category. Even so, the team’s experiments found that positively modifying input variables consistently yielded equal or better predicted rankings.

The study’s limitations are acknowledged in its own framing: it covers only 148 U.S. programs ranked between 1 and 150 over three consecutive years, it predicts intervals rather than exact positions, and its predictors were hand-collected through laborious manual verification. The authors plan to refresh the analysis annually with new U.S. News data beginning in 2026, to enrich the website’s underlying dataset, and to extend the same approach to industrial engineering graduate programs. For the moment, the work offers mechanical engineering departments a data-driven simulator of a process that U.S. News has not opened to inspection, along with a sober statistical warning that reputation, embodied in last year’s rank, still accounts for roughly half of where a model places a program.

Subject of Research: Machine learning prediction of U.S. News mechanical engineering graduate program rankings

Article Title: A machine learning model for the prediction of the US news & world report ranking of the mechanical engineering graduate programs in the United States

Article References: Khademi, F., Rachakonda, N., Ismodes, M. S., & Darabi, H. (2026). A machine learning model for the prediction of the US news & world report ranking of the mechanical engineering graduate programs in the United States. Neural Computing and Applications, 38(20), Article 832. https://doi.org/10.1007/s00521-026-12636-5

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12636-5

Keywords: machine learning, Random Forest, university rankings, mechanical engineering, U.S. News & World Report, graduate programs, classification, sensitivity analysis, higher education, predictive modeling, machine, learning