Machine Learning Model Cracks the Black Box of US News Mechanical

For years, the U.S. News & World Report rankings of mechanical engineering graduate programs have been a mystery. Every spring, universities find out where their programs rank, but they do not know exactly how those rankings are decided. A new study published in Neural Computing and Applications shows that a computer model can help explain and predict these rankings. Researchers created a machine learning model that uses nine measurable facts about universities and their programs to predict which ranking group a program will fall into. When tested on the 2024 rankings, the model was correct 91% of the time.

The study was led by Faeze Khademi, Nikith Rachakonda, Matias Salazar Ismodes and Houshang Darabi. According to the researchers, theirs is the first study to predict the U.S. News rankings of mechanical engineering graduate programs. They collected rankings from U.S. News & World Report for 2023, 2024 and 2025. They also gathered information about 59 possible factors from university department websites, data from the American Society for Engineering Education and other academic sources. U.S. News rankings depend heavily on surveys asking people about a program's reputation, but the exact way those opinions are used to produce the rankings is not clear. The researchers wanted to work backward to understand the results. They asked whether a computer model could use things that university programs can control to predict the results of a ranking system whose inner workings are mostly hidden.

The researchers had to change their original plan. At first, they wanted to predict each program's exact ranking. But the amount and type of data available made this too difficult. Instead, they divided programs into four ranking groups: 1–20, 21–50, 51–100 and 101–150. Their final dataset included 148 universities. Of these, 22 programs were in the first group, 32 in the second, 45 in the third and 49 in the fourth. Some groups contained more programs than their ranking ranges might suggest because several universities shared the same rank. For example, six universities were tied at number 17 in 2024, meaning more than 20 universities were included in the top-20 group. Programs ranked below 150 were not included.

Getting the data ready was also a difficult job. Some information was stored as text, so the researchers created a special function to turn it into numbers. When information was missing, they filled in the gaps using the average of the available values. This method, called mean imputation, allowed them to include more universities instead of removing those with incomplete information. Because the four ranking groups contained different numbers of programs, the researchers also used a method called random oversampling to help calculate how certain they could be about one of their performance measurements, the AUC. Importantly, they did not use this method to train the prediction model itself. They also adjusted the numbers so that measurements with very different sizes could be compared fairly. This process is called Z-score normalization. They calculated the adjustment using only the training data, so information from the test data could not accidentally influence the model.

Out of the original 59 factors, the researchers selected nine that worked best for predicting rankings. They tested different combinations using only the training data. The final nine factors were the number of full professors, professors with tenure, all faculty members and retired professors who still hold an academic title, all measured from the previous year; the number of full-time students enrolled the previous year; the university's total funding from two years earlier and three years earlier; the previous year's average score on the math section of the GRE graduate school entrance exam; and the program's U.S. News ranking from the previous year. The researchers used funding information from two and three years earlier for a reason. Rankings come out shortly after each year begins, so funding from the previous year might not yet have had enough time to affect how well known or respected a university is across the country. They also deliberately included the previous year's ranking because U.S. News relies heavily on reputation, which usually changes slowly. When they removed this factor, the models became much less accurate. The researchers explained that an earlier ranking helps predict whether a program will stay in a similar position. It does not prove that last year's ranking directly causes the next year's ranking.

The team compared three machine learning methods: Decision Tree, Gradient Boosting and Random Forest. They built all three using the Python software library scikit-learn. They used 70% of the data to train the models and 30% to test them, while keeping the proportions of the four ranking groups similar in both sets. They also used a fixed random seed so the experiment could be repeated with the same results. Random Forest performed better than the other methods. It used 200 decision trees, allowed each tree to reach a maximum depth of 10 and considered half of the available factors at each decision point. The researchers also checked its performance using 5-fold cross-validation, which tests the model across five different divisions of the training data. On the 2024 test data, Random Forest achieved an accuracy of 0.91, precision of 0.91, recall of 0.90 and an F1-score of 0.91. These are different ways of measuring how well a model makes predictions. The researchers calculated the average for each measurement by giving all four ranking groups equal importance, no matter how many programs each group contained. They also measured how well the model could tell the ranking groups apart using a multiclass AUC score. This calculation compared groups two at a time and used 1,000 repeated samples to estimate a 95% confidence interval. The AUC was high, and its confidence interval was narrow, suggesting that the model was good at telling the different ranking groups apart.

A closer look at the predictions showed that the model made very few mistakes. In the test data, it placed only one program from the top-20 group in the wrong category. It correctly classified every program in the 101–150 group. The 21–50 group was a little more difficult. Two programs were placed in the wrong groups: one was predicted to be in the top 20, and another was predicted to be in the 51–100 group. The researchers also studied which factors mattered most to the model. The previous year's ranking was by far the most important, accounting for about 52% of the model's total measure of factor importance. However, almost half of its predictive importance came from things universities may be able to change, such as student enrollment, the number of faculty members, research funding and GRE scores. The researchers then used the same model, without changing it, to predict rankings in the 2025 dataset. It continued to perform well across all the measurements. The authors believe this shows that the model can give consistent results and work across different academic years.

One of the study's most interesting results comes from a test of how changing different factors affects the predicted rankings. The researchers created a companion website where people can enter possible values for a university's characteristics and see which ranking group the model predicts. In one example, they looked at Duke University's Pratt program. The model suggested that the program could move from the 21–50 group into the top 20 under a scenario with several changes. These included increasing faculty numbers by 1.78 times, from 50 to 89; increasing total university funding by about 1.8 times in each of two consecutive years; increasing full-time enrollment by 1.45 times, from 2,018 to 2,926; and making small improvements in average quantitative GRE scores and professor numbers. However, the researchers strongly warned that these results are predictions, not proof of cause and effect. The model shows which combinations of university characteristics are linked to better predicted ranking groups. It does not guarantee that making those changes will improve a university's real ranking. For example, simply hiring more faculty members would not necessarily move a program higher in the U.S. News rankings. To prove that a particular change directly improves rankings, researchers would need a very different kind of study, especially because reputation plays such a large role in the ranking process.

The researchers also found an important weakness in this method of testing possible changes. After checking all 150 ranking positions, they discovered that the model gave consistent results only for programs ranked within the top nine positions of each ranking group. For example, a program ranked anywhere from 60 to 100 might not get reliable advice from the model about how to move into the 21–50 group. The researchers suggested improving a program's characteristics in stages. First, a university should work on improving the factors that may help its real ranking move into the top nine positions of its current group. Once it reaches that point, it could use the model to explore possible changes that might help it move into the next, better ranking group. Even with this limitation, the researchers found that improving the input values in their experiments always produced predictions that were either the same or better than before.

The study also has several broader limitations. It includes only 148 U.S. mechanical engineering graduate programs ranked between 1 and 150, and it covers just three consecutive years. The model predicts ranking groups rather than exact positions. The researchers also had to collect the information about the programs by hand and carefully check it, which took considerable effort. They plan to update their analysis every year using new U.S. News rankings, starting in 2026. They also want to add more information to the companion website and use a similar method to predict rankings for industrial engineering graduate programs. For now, the study gives mechanical engineering departments a way to explore how measurable program characteristics relate to a ranking process that U.S. News has not fully explained. At the same time, it offers an important reminder: a program's reputation, represented by its ranking from the previous year, still accounts for about half of what determines where this model predicts the program will rank.

Subject of Research: Using machine learning to predict U.S. News rankings of mechanical engineering graduate programs

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: Created using AI

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