Improving the Economy of Universities of Applied Sciences using Artificial Intelligence

Finnish economy has been suffering for almost a decade now, and the field of education has also had to participate in the required savings of public resources. Focusing the reduced resources efficiently on targets and activities that best improve the measured results has become essential. Applications of artificial intelligence could help in identifying these targets and activities.

Author: Antti Salopuro

Battlefield of Universities of Applied Sciences

For the Finnish universities of applied sciences (UAS), the prolonged depression has caused a reduction of public funding by 22% (Arene 2018). In the current UAS funding model, 85% of the total funding depends on the UAS performance in educational measures such as number of degrees granted and progress of the studies (Opetushallitus 2014). The UASs have thus had to focus on making the students proceed faster with their studies and eventually graduate in higher numbers. As the motivation and capabilities of a student are important in enabling a successful study path on a personal level, a successful assessment of these measures and using them as criteria in student selection could be an asset for any UAS. Equally, identifying the students that most probably will have difficulties with their studies at an early stage could help in targeting supportive activities to correct individuals.

Improving selection of students

A successful match of students with different study programs has always had an impact, not only on UASs, but also on the success of the whole nation, and, therefore, it has been a topic of research. Rantanen (2001) identified this being a difficult problem in general, and specifically for Finnish UASs. The power of the entrance test in forecasting study success was found to be much poorer than previously assumed. In his follow-up research, Rantanen (2004) created stepwise regression models to explain success in studies with all information available at the time of applying to a UAS. In general, the average grade of the high school final certificate was found to be the most significant single factor (r2 = 0.18) in explaining the success. Also, matriculation exam grades were found to correlate to success in studies more significantly than the entrance exam scores. One impact of this can be seen in the methodology of the future application process of UASs. Starting in 2020, the principle of student selection will mainly be based on the grades of the certificate of the applicants from the previous education (Opintopolku 2018). The admission score of an individual applicant will be calculated as a weighted sum of a set of grades, which is both a practical and a transparent method. The weights applied are the same in almost all programs with only a few exceptions.

This new approach of student selection is thus even more straightforward and simpler than the current one and no longer allows individual UASs or their programs to decide on the criteria independently. But could a better solution for building such scoring models be found by applying artificial intelligence (AI) and by, again, acknowledging the requirements of the study program? Kabakchieva (2013) experimented with six different classifiers to model the university performance (output) of a set of Bulgarian students using some pre-university data (input). In addition to the secondary school final grades and entrance test scores , some background information such as the place and profile of the pre-university education was also used as input data. In general, the applied classifiers were able to predict with an accuracy of 60% or more only on three levels (1, 3 and 4) of the five-level scale. The classification of the students falling in either ‘average’ (2) or ‘excellent’ (5) classes were identified poorly, with accuracies of less than 10%. On the other hand, the ability to identify students that would finally perform the weakest were identified with an accuracy of 80%. It is exactly this kind of ability that would be helpful for a UAS to filter out the worst performing students from the group of applicants.

Identifying students in need of help

While the renewal of the entrance criteria, again, will most likely not be possible in the near future, another question could be made if there were any ways to make it easier for a UAS to identify the students that managed their ways into a program but have a high risk of being left behind or even drop out of the studies? This would allow, already in an early stage, focusing the limited student counselor resources on students for those it would be the most helpful. Huang and Fang (2013) studied the ability of four different methods, each considered to be an elementary part of the standard AI toolkit, to predict the student performance in the final exam of an engineering dynamics course. Six combinations of a set of input variables, consisting of the total sum of grade points collected so far and grades of seven previously studied courses, were applied. The results were promising: for the 24 different models (combinations of the methods and the variables) assessed, the authors report a minimum average prediction accuracy of 86.5%. This implies that even with a simple method (e.g. linear regression) and a single predictor (e.g. grade of single previous course) it is possible to foresee student success in a future exam accurately.

A natural follow up to the works referenced above would be to combine the prediction power of the pre-university grades with that of the university course grades in seeking for a better prediction of the university graduation performance. Asif, Merceron and Pathan (2015) have done exactly this: in addition to the pre-university grades (as in Rantanen 2001 and 2004), some marks from the first two years in the university were also included in the set of predictor variables as they predicted the overall success of students in the university level studies. The study applied two separate datasets, both consisting of data from two subsequent academic years (four full years covered), thus simulating a real life situation where the data of the previous year is used to build a model to forecast the performance of the following year. Splitting the data (N = 347) into four subsets has, however, resulted in relatively small sizes of both the predictor and the test data sets. This has, very likely, reduced the predictive power of the models created. Despite this, the paper reports some sufficiently high precision values, especially with respect to the use case of identifying the students that would best benefit from specific supportive activities. The students whose final grade would most likely be one of the two lowest (D or E) are, at best, identified with an accuracy of 89.5%. The best performances are achieved with models applying a decision tree with different hyperparameters. The study concludes that including both pre- and past-admission marks improves the prediction power of the graduation level grades significantly compared to cases where only one of these is used as a predictor.


From these few referenced examples it is possible to conclude that applying statistical models and artificial intelligence has potential in helping to recognize the applicants that most probably do not have the capabilities required for graduation and especially the students with the highest need for an intervention by a study counselor. Not selecting applicants that have a high risk of not ever finishing the degree would -even if it meant groups of program students being smaller- benefit all stakeholders. It would leave more resources to be targeted on students with good enough acquirements which would serve both students and teaching personnel. Accurate identification of students in most need of help would allow focusing the supportive resources such that it optimally improves the graduation rate and thus the most important measured result of UASs. Therefore, both these applications of AI would further help the Finnish UASs to manage themselves in the world with less public funding.


Arene.  2018. Osaaminen ja työllisyys kasvu-uralle – Amk-rahoitus kuntoon. [Cited 20 Oct 2018]. Available at:

Asif, R., Merceron, A. & Pathan, M. K. 2015. Predicting Student Academic Performance at Degree Level: A Case Study. I.J. Intelligent Systems and Applications.  Vol. 7 (1), 49 – 61. [Cited 20 Oct 2018]. Available at:

Huang, S. & Fang, N. 2013. Predicting student academic performance in an engineering dynamics course: A comparison of four types of predictive mathematical models. Computer and Education. Vol. 61, 133-145. [Cited 20 Oct 2018]. Available at:

Kabakchieva, D. 2013. Predicting Student Performance by Using Data Mining. Cybernetics and Information Technologies. Vol. 13 (1), 61 – 72. [Cited 20 Oct 2018]. Available at:

Opetushallitus. 2014. Yksikköhinnat vuodelle 2014. [Cited 20 Oct 2018]. Available at:

Opintopolku. 2018. Mikä korkeakoulujen opiskelijavalinnoissa muuttuu vuoteen 2020 mennessä? [Cited 20 Oct 2018]. Available at:

Rantanen, P. 2001. Valintakoe vai ei? Ammatillisen Koulutuksen ja ammattikorkeakoulujen opiskelijavalinnan tarkastelua. Helsinki: Opetusministeriö. Koulutus ja tiedepolitiikan osaston julkaisusarja, 83.

Rantanen, P. 2004. Valinnasta työelämään. Ammatillisen koulutuksen ja ammattikorkeakoulujen opiskelijavalinnan tarkastelua. Helsinki: Opetusministeriö. Opetusministeriön julkaisuja 2004:19.[Cited 20 Oct 2018]. Available at:


Antti Salopuro is Senior Lecturer of Business Information Technology at Lahti University of Applied Sciences, private practitioner at Tietokutomo and student of Data Sciences at University of Helsinki. He believes that using natural wisdom could sometimes protect against problems that can be solved using artificial intelligence.

Illustration: (CC0)

Published 22.10.2018

Reference to this publication

Salopuro, A. 2018. Improving the Economy of Universities of Applied Sciences using Artificial Intelligence. LAMK Pro. [Cited and date of citation]. Available at:


Sähköpostiosoitettasi ei julkaista. Pakolliset kentät on merkitty *