Towards Student Success PredictionHana Bydžovská, Michal Brandejs
článek na konferenci Knowledge Discovery and Information Retrieval 2014 (21. - 24. října 2014)
Keywords: Recommender System, Social Network Analysis, Data Mining, Prediction, University Information System.
Abstract: University information systems offer a vast amount of data which potentially contains additional hidden information and relations. Such knowledge can be used to improve the teaching and facilitate the educational process. In this paper, we introduce methods based on a data mining approach and a social network analysis to predict student grade performance. We focus on cases in which we can predict student success or failure with high accuracy. Machine learning algorithms can be employed with the average accuracy of 81.4%. We have defined rules based on grade averages of students and their friends that achieved the precision of 97% and the recall of 53%. We have also used rules based on study-related data where the best two achieved the precision of 96% and the recall was nearly 35%. The derived knowledge can be successfully utilized as a basis for a course enrollment recommender system.
Educational systems are focused on providing a high educational standard. Learning analytics is used to enhance the teaching and the learning. One of the most important issues often solved in educational environment is finding what influences student performance.
We are interested in designing a course enrollment recommender system (Jannach et al., 2011) that will help students with selecting courses to enroll in. For such recommender system, it is crucial not to recommend difficult courses for particular students. The student can fail to meet the minimum requirements and subsequently to discontinue the study.
It is equally important to advise students on mandatory courses that usually cause problems. The task is to identify such courses precisely. Initially, we intend to properly define the similarity between students based on their achievements as the course difficulty can be estimated from the results of similar students enrolled in earlier.
Our novel approach utilizes not only student-related data but also structured data inferred from social behavior of students. In (Bydžovská et al. II, 2014), we have confirmed that the impact of social ties among students on their study results is really not negligible.
The aim of the proposed method is to predict the success or failure of students in the selected courses. The preliminary work was published in (Bydžovská et al., 2014). In comparison to it, this follow-up work contains the description of new social feature extraction and their usage to improve results. Association rules and decision trees algorithms are used for a subgroup discovery. The experiment is also performed on a representative data sample that provides interesting observations.
In the following section, we give an overview of related work. In Section 3, we describe the course enrollment recommender system. The data used in our experiments is introduced in Section 4. In Section 5, we describe the experiments and the next section brings results. The discussion of results can be found in Section 7. The summary and directions for future work can be found in the last section.
2 RELATED WORK
Exploring student potential is an interesting task in the educational environment. Researches explore what influences students and what can be done to help them to improve their achievements.
A typical way for discovering regularity in data is using data mining techniques (Peña-Ayala, 2014) suitable for exploring a vast amount of data. It allows us to build predictive models by defining valid and exact rules.
Researchers mostly examine study-related data stored in university information systems which contain for example grades, gender, field of study, or age. However, exploring the influence of social ties to students’ performance is an attractive topic. Authors in (Poldin at al., 2014) obtain data about students’ friendships from questionnaires. They prove that such data can improve predictions of student failure. Unlike Poldin et al. (2014) whose research depends on answers from questionnaires, we use data obtained from an information system that can be computed from user requests, e.g. the publication co-authoring, statistics about e-mail conversation, or files uploaded into someone else's depository.
Sometimes researches are faced with unbalanced classes. This problem often encountered when analyzing educational data, also mentioned in (Thammasiri et al., 2014). Using some oversampling techniques could overcome this problem. Authors improve classification accuracy using SMOTE, an algorithm that generates new examples by interpolating among existing minority examples.
Authors in (Viliardi et al., 2009) aim at selecting courses for students in order to obtain good exam results. Difficulties of courses are compared with student potentials. Both variables are computed from grades. The work extension can be found in (Viliardi et al., 2010) where the analysis is based on the profile similarity.
We further extend the method from (Viliardi et al., 2009) by addition of social data. In this way we are able to compare students' data together with the information about their friends. Therefore, we can increase the prediction accuracy.
3 RECOMMENDER SYSTEM DESCRIPTION
Students are interested in information resources and learning tasks that could improve their skills and knowledge. Therefore the recommender system should monitor their duties and show them either an easy or an interesting way to graduate.
The current version of the system contains two modules: an extraction module and an analytical module. Figure 1 presents the data flow between the university information system, the recommender system and external tools such as Weka and Pajek.
The results of exploration can be used for a warning against too difficult courses in which students are enrolled in. When courses are mandatory, the canceling the enrollment is not possible but the knowledge is useful for students and they should study hard. When courses are selective or optional, students can revise their choices. They can select other courses recommended by the system. The system will recommend courses with respect to students' potentials and interests. A short explanation of the selection will be available for students to be able to verify the recommendations.
Another possibility is using the prediction of student success or failure to identify the best students or weak students for the faculty management. The best students represent candidates for tutoring, scholarships or grants. Weak students usually need help to be able to graduate successfully and heads of faculty can encourage them for example by providing the best students for tutoring them. It can be also useful to assist teachers to be able to organize students into seminar groups. On the other hand it can be tricky when teachers make conclusions based only on this information.
3.2 Extraction Module
The Information System of Masaryk University (IS MU) is a complex system used to support not only education at Masaryk University. Today, it serves many purposes and its functionality involves managing study-related records. The database of IS MU contains many types of data, e.g. students’ grades, exams, electronic questionnaires, course enrollments, discussion groups, shopping center, games, calendar, e-learning materials, publications, course management, etc. Furthermore, the system keeps the whole history of user requests.
The recommender system has its own database in order not to influence the functioning of the production system. The data we expect to define student characteristics is periodically imported to its database. The extraction module performs data pre-processing and provides time-consuming computations for study-related data.
The extraction module also manages social behavior data, converts it into the format suitable for Pajek (Nooy et al., 2011) which is a well-known tool for social network analysis. Pajek reads the data and creates sociograms where the nodes represent people we are interested in and ties represent relations between them. From such sociograms we can compute social attributes defining the importance of a person in the network. The extraction module also prepares batch files in the format for Pajek in order to launch the computations automatically. The resulting social attributes are also stored into the recommender system database to be ready for the next processing.
3.3 Analytical Module
The data prepared by the extraction module is further processed in the analytical module. This module makes use of feature selection algorithms to extract relevant features. It obtains also basic statistics about the features, and it is able to run machine learning algorithms from Weka (Witten et al., 2011).
The analytical module will be finally composed of two methods. The first one will provide finding interesting courses for each student and the second one will provide an estimation of the probability of passing the course.
The current version of the module contains three techniques described in this paper in Section 5 that lead to the realization of the second goal. Firstly, we use machine learning algorithms for mining historical data. The second approach is based on comparing student potential (student’s average grade) with course difficulty (average grade computed from all grades of students enrolled in a course). The last method discovers student subgroups for which the prediction can be more accurate. The obtained accuracy is always compared with the corresponding baseline, i.e. the case when all instances of the data are classified into the majority class.
The detailed description of these techniques can be found in Section 5, the results are in Section 6 and the discussion of interesting effects is situated in Section 7.
As it has been already mentioned, the data is obtained by the extraction module from IS MU.
4.1 Student Characteristics
Each student can be described by a set of attributes that precisely characterize student's qualities, potentials and interests. We use three types of data: the study-related data, the social behavior data and the data about previously passed courses. Some of these attributes were published in the previous work (Bydžovská et al., 2014) in Section 5. The most useful of them were: average of grades, weighted average of grades, number of credits to gain, gained credits and ratio between the last two attributes, programme and field of study, closeness centrality, degree, and weighted degree and betweenness centrality (computed from the network formed by students enrolled in the investigated course and all their friends).
4.1.1 New study-related attributes
In this section we present additional attributes generated for the follow-up experiments introduced in this paper. We add the following attributes:
- average number of enrolled/gained credits in a term,
- difference of previous attributes and 30 credits (recommended study workload per term),
- number of successful finished studies at the faculty/university,
- number of failed courses per term,
- number of courses previously failed that have to be passed in the future,
- number of successfully passed repeated courses,
- information about successfully/unsuccessfully utilized retakes (the second attempts to pass an exam),
- information if a course is mandatory, selective or optional for particular students with respect to their year of admission and field of study,
- number of times, when a teacher of an investigated course taught a particular student.
4.1.2 New social behavior attributes
Values defining relations between people already exist in IS MU and they are successfully used for personal search (Kasprzak et al., 2010). The value of the relation can be in the interval of [100, 200] and represents the measure of communication inside the system between two people. The higher the number, the stronger is the tie between the two people.
We used these values in the preliminary work (Bydžovská et al., 2014) to create a sociogram where students enrolled in the investigated course and all their friends were included. This approach had limitations: the sociogram was only a small part of the global network and all friends of students were included but not all of them really influenced corresponding students (with weak ties).
To overcome these limitations we process data about all people communicating in the information system in this work. To favor explicitly expressed friendships in IS MU, we add such relations to the sociogram with the value of 300 (the strongest tie). We also calculate different variants where rules influence which ties are included in the sociogram.
We generate two types of a sociogram. The first one contains data only about students of the Faculty of Informatics (people who have an active study or have finished a study in the last two years at the faculty). The second one contains students, teachers (people who taught in the previous years at the faculty), academics and faculty staff.
For both sociogram types we build variants with the following reductions of ties between people:
- all ties,
- ties stronger than the third of all communication values,
- ties stronger than the median of all communication values,
- ties stronger than the value 150,
- ties that were explicitly expressed (value 300).
For these 10 sociogram variants we extract the following social features: degree, weighted degree, centrality, betweenness. We also compute grade averages of all neighbors.
We also extract the following information from IS MU:
- the strongest tie with a teacher of the investigated course,
- course marked as favorite,
- course attendance disclosure,
- course seminar group disclosure,
- course examination date disclosure,
- application for the study disclosure.
4.1.3 Previously passed courses
The subset of courses is selected for each investigated course with respect to the rule that almost 20 students are enrolled in the course. This helps to reduce the number of courses and to select only reliable ones. We added these attributes to datasets to find out if there is a correlation between these courses and the investigated course.
4.2 Used data
We took 62 courses offered at the Masaryk University for bachelor and master programmes. We selected courses where many students were enrolled in. The experiment comprised of 7457 enrolled students in the years 2010-2012 and their 148750 grades. Figure 2 shows the course enrollment statistics. The most students are enrolled in the mandatory courses for bachelor programmes. The largest course in the data set is Mathematics I with 1767 students.
5 STUDENT SUCCESS PREDICTION
The aim is to precisely predict student success or failure in the investigated course based on the analysis of the historical data. Approaches using machine learning algorithms and subgroup discovery based on comparing grade averages, association rules and decision trees are presented in this section.
5.1 Prediction Using Machine Learning Algorithms
5.1.1 Data Sets
We extracted six datasets for each course for machine learning processing as follows:
- study-related data (SR),
- social behavior data (SB),
- study-related and social behavior data (SS),
- study-related data enriched with data about passed courses (SRC),
- social behavior data enriched with data about passed courses (SBC),
- study-related and social behavior data and data about passed courses (ALL).
5.1.2 Used ML algorithms
We utilized different machine learning algorithms implemented in Weka (Witten et al., 2011), namely naive Bayes (NB), support vector machines (SMO), instance based learning (IB1), classification rules (PART), one rule (OneR) and decision trees (J48), Random Forests, and also ensemble learning methods, namely AdaBoost and Bagging. AdaBoost algorithm achieved the best results in the combination with Decision Stump and Bagging with SMO or REPTree.
5.2 Subgroup discovery
We aim at finding subgroups of students for which the prediction could be more accurate than using ML algorithms.
5.2.1 Grade averages
The technique was inspired by Vialardi et al. (Vialardi et al., 2009 and 2010). The method was based on the comparison of the course difficulty and values defining student potentials. The course difficulty was defined as the average grade of all students enrolled in the investigated course.
We have already made a similar experiment in (Bydžovská et al., 2014) when we created an ensemble learner from 3 classifiers (weighted average of student grades, their friends’ weighted average grades and weighted average grades of their friends that attended the investigated course with the corresponding student). The results were satisfactory only for 4 of 5 investigated courses. That was the reason why we extended our research with new social attributes. As it was mentioned in Section 4.1.2, we computed average grades for each student from all or selected student's neighbors from all variants of sociograms. These values were able to define the student potential.
When the student's potential was lower than the course difficulty, it meant that the student or his or her friends had better study results than students attended the investigated course in the past. In this situation we predicted success otherwise we did not give any prediction.
5.2.2 Association rules
For subgroup discovery (Lavrač et al., 2002 and 2006) we combined the technique of finding interesting subsets of attribute values (by means of discretization for continuous attributes and by building subsets of values for categorical attributes) with two learning algorithms—decision trees (J48) and class association rules (Liu, 1998).
We created one data set for all investigated courses and all students from the training set (students enrolled in courses in 2010 and 2011) and we found interesting rules that could be applied to the test set (students enrolled in courses in 2012) with high accuracy. This approach allowed us to find general rules for the prediction of student success or failure regardless of the specific course. It depended only on student's study achievements. We also explored decision trees and association rules based on the Apriori algorithm for all courses in order to improve the results for the corresponding subgroup of courses.
Our goal was to precisely predict student success or failure in the investigated courses. The most important task was revealing such courses and such students for which the prediction could be highly confident.
6.1 Prediction Using Machine Learning Algorithms
Many experiments mentioned in Section 2 mined study-related attributes for prediction. We enriched data sets with social behavior data and data about previously passed courses. We used 10-fold cross validation. The prediction accuracy was 81.4% in average.
Figure 3 presents the contribution of additional data on the investigated courses. In comparison with baseline, the accuracy was improved by 8% in average. In comparison with study-related data, the accuracy was improved by 2% in average. The maximal improvement in accuracy was 27% in comparison with baseline and more than 5% with using only study-related data.
The distribution of best results of mining different datasets (defined in Section 5.1.1) is presented in Figure 4. The best results were often obtained on study-related and social behavior data (25 courses) and only on study-related data (24 courses). We did not improve accuracy for 8 courses.
The distribution of best ML algorithms (defined in Section 5.1.2) is presented in Figure 5. The most accurate algorithms were SMO (22 courses), Bagging (11 courses), and Random Forests (8 courses).
We also tested the data split to the training data (students enrolled in courses in 2010 and 2011) and the test data (students enrolled in courses in 2012). Based on the training data, we defined three rules to obtain course subgroup for which predictions using machine learning algorithms were reliable:
- We were not capable of making prediction in courses where conditions were changing (different teacher, different students’ evaluation). The difference between success rate of the test set and the training set was more than 11.5%.
- For courses with more than 81% successful students in the training set, the prediction is not suitable. In comparison with baseline the improvements were only about 1% in average while the average accuracy for these courses was almost 87% in the test set.
- If the course in the training set had success rate lower than 61.5%, the algorithms improved prediction by more than 16% in average but the final accuracy was still lower than 80% in the test set which was not good enough.
6.2 Subgroup discovery
6.2.1 Grade averages
We constructed 10 variants of the sociogram described in 4.1.2 to be able to compare the influence of different people to a student.
For all student averages (student's own, friends or schoolmates with unlimited/limited ties) we computed the precision and the recall. We were interested in the highest precision. We observed that the most frequently selected attributes for the comparison with course difficulties were the following:
- student weighted average grade,
- student average grade,
- weighted average of explicitly expressed schoolmates,
- weighted average of schoolmates with ties higher than 150,
- weighted average of explicitly expressed friendship,
- weighted average of friends with ties higher than 150.
Figure 6 represents the distribution of the best results of the ties. We improved the results for more than one fourth of all courses.
This indicates that the student potential calculated from closer friend ties was more accurate than when all ties were considered.
The second observation was that this technique did not meet the expectations for courses with the baseline lower than 60%. These courses were very difficult and the student's or the student friends' knowledge might not be sufficient.
For all other courses the average precision reached 92.86%. On the other hand, these classifiers' recall reached only 53%.
6.2.2 Association rules
Based on the previous findings a rule with a high accuracy for subgroups of students was found. The weighted average grade and average grade were the most frequently selected as the beneficial attributes defining the student potential. Because we calculated the weighted average grade from all grades regardless of studies, it was lower than if only the particular study was considered because students always had to pass difficult courses in the first years of bachelor studies. The first defined rule:
- weighted average grade ≤ 2.4 → success.
The distribution of students' weighted average grades can be seen in Figure 7. Grade 1 was the best grade, grade 4 was the worst. The precision for this rule was 96% and recall 24.86%.
We also wanted to fill the gap for improvements of very successful courses. The second defined rule:
- weighted average grade ≤ 3 AND ratio of gained credits to credits to gain > 0.8 AND courses with success rate > 80% → success.
When we considered the whole data set, the precision for this rule was 96.85% and the recall 23.78%. When we omitted students' success predicted by the first rule, the precision for this rule was 94.8% and recall 11%.
The best fitting data set for machine learning algorithms was study-related and social behavior data together. The next observation was that there were no noticeable improvements when we added data about previously passed courses into the former data sets except for five courses which will be objects of our future interests. Based on the results published in Bydžovká et al., 2014, we suppose that these attributes have a stronger influence for improving good/bad/failure prediction or the exact grade prediction than for success/failure prediction based on the previous experiments.
We explored many different rules with high precision but they had small recall. For example:
- weighted average grade ≤ 3 AND successfully finished studies at the university ≥ 2 AND mandatory course for the specific field of study = 'AP' → success (Precision: 97.2%, Recall: 6.92%),
- successfully finished studies at the faculty ≥ 2 AND repeating the course → success (Precision: 97.8, Recall: 0.8%).
However, the next step will be to apply such precise rules in sequence. We can get more accurate prediction and improve the global recall.
If we concerned a particular course, there were plenty of rules explored for the specific course, for example
- IA008: schoolmates attended the same course in the past AND an average grade < 2.5 AND a tie > 150 to a student → success (Precision: 93.75%, Recall: 7.8%).
This rule could not be applied for all courses because of the low precision.
The next step will be finding subgroups of courses for which such rules are suitable.
CONCLUSION AND FUTURE WORK
The main contribution of this work is the presentation of different approaches for student success or failure prediction. The paper brings results and discusses advantages and disadvantages of these methods.
Machine learning algorithms can be successfully employed with the presented data set with the average accuracy 81.4%.
We also split the data to the training and the test set to identify courses for which ML cannot be successfully applied to courses with more than 81% or less than 61.5% successful students in the training set. The results in the test set were also not so convincing when there was a significant difference between the training set and the test set.
On the other hand, we can apply the following discovered rule for easier courses. All students with the weighted average grade ≤ 3 and the ratio of gained credits to credits to gain > 0.8 are successful. This fills the gap.
We also defined rules based on the grade averages of students and their friends. One conclusion was that the prediction was more accurate when only close friends were considered. This approach offered the precision about 97% but decreased the recall to 53%.
In the future work, we intend to find the appropriate balance of using these methods and to combine precise association rules to get the most accurate predictions with a reliable recall. The courses evinced the relations with other courses will be explored. We also intend to enrich the data with temporal features that can improve the current results.
These predictions will constitute a part of the course enrollment recommender system which will help students to select courses and warn them against difficult courses they have to pass.
We thank Lubomír Popelínský, colleagues of Knowledge Discovery Lab, and also all colleagues of IS MU development team for their assistance. This work has been partially supported by Faculty of Informatics, Masaryk University.
Bydžovská, H. and Popelínský L. 2014. Course Recommendation from Social Data. Proceedings of the 6th International Conference on Computer Supported Education, Volume 1, pp. 268-275.
Bydžovská, H. and Popelínský L. 2014 II. The Influence of Social Data on Student Success Prediction. Proceedings of the 18th International Database Engineering & Applications Symposium, pp. 374-375.
Jannach, D. and Zanker, M. and Felfernig, A. and Friedrich, G. 2011. Recommender Systems An Introduction, Cambridge University Press.
Kasprzak, J. and Brandejs, M. and Čuhel, M. and Obšívač, T. 2010. Access rights in enterprise full-text search. In Proceedings of the 12th International Conference on Enterprise Information Systems, Volume 1, pp. 32–39.
Lavrač, N. and Železný, F. and Flach, P. A. 2002. RSD: Relational Subgroup Discovery through First-Order Feature Construction. In 12th International Conference on Inductive Logic Programming, pp. 149-165.
Lavrač, N and Kavšek, B. 2006. APRIORI-SD: Adapting Association Rule Learning to Subgroup Discovery. Applied Artificial Intelligence, pp. 543-583.
Liu, B. and Hsu, W. and Ma, H. 1998 Integrating classification and association rule mining. Proceedings of the 4th international conference on Knowledge Discovery and Data mining KDD'98, page 80-86. AAAI Press.
Nooy, W. and Mrvar, A. and Batagelj, V. 2011. Exploratory Social Network Analysis with Pajek. Structural Analysis in the Social Sciences. Cambridge University Press.
Poldin, O. and Valeeva D. and Yudkevich M. 2014. Friendship and study assistance ties of university students. In WP BRP Series: Sociology / SOC, pages 1-17.
Peña-Ayala, A. 2014. Educational data mining: A survey and a data mining-based analysis of recent works. In Expert Systems with Applications , pp. 1432–1462.
Thammasiri, D. and Delen, D. and Meesad, P. and Kasap N. 2014. A critical assessment of imbalanced class distribution problem: The case of predicting freshmen student attrition. In Expert Systems with Applications 41, pp. 321–330.
Vialardi, C. and Agapito, J. and Shafti, L. and Ortigosa, A. 2009. Recommendation in higher education using data mining techniques. In T. Barnes, M. Desmarais, C. Romero, and S. Ventura (eds.), Proceedings of the 2nd International Conference on Educational Data Mining (EDM’09), pages 191–199.
Vialardi, C. and Chue, J. and Barrientos, A. and Victoria, D. and Estrella, J. and Ortigosa, A. and Peche, J. 2010. A case-study: Data mining applied to student enrollment. In Proceedings of Third Educational Data Mining Conference, Pennsylvania, USA, pages 333–335.
Witten, I. and Frank, E and Hall, M. 2011. Data Mining: Practical Machine Learning Tools and Techniques, 3rd edition, Morgan Kaufmann Publishers.