Vladimir Ivančević

Using association rule mining to identify risk factors for early childhood caries

BACKGROUND AND OBJECTIVE Early childhood caries (ECC) is a potentially severe disease affecting children all over the world. The available findings are mostly based on a logistic regression model, but data mining, in particular association rule mining, could be used to extract more information from the same data set. METHODS ECC data was collected in a cross-sectional analytical study of the 10% sample of preschool children in the South Bačka area (Vojvodina, Serbia). Association rules were extracted from the data by association rule mining. Risk factors were extracted from the highly ranked association rules. RESULTS Discovered dominant risk factors include male gender, frequent breastfeeding (with other risk factors), high birth order, language, and low body weight at birth. Low health awareness of parents was significantly associated to ECC only in male children. CONCLUSIONS The discovered risk factors are mostly confirmed by the literature, which corroborates the value of the methods.

Building an ensemble from a single Naive Bayes classifier in the analysis of key risk factors for Polish State Fire Service

In this paper, we describe our solution in a competition that required performing data mining to identify key risk factors for the State Fire Service of Poland. The goal was to create an ensemble of Naive Bayes classifiers that could predict incidents involving firefighters, rescuers, children, or civilians. To this end, we first created a single Naive Bayes classifier and then partitioned the set of attributes used in that classifier. The attribute subsets were used to create new Naive Bayes classifiers that would form an ensemble, which generally performs better than both the single classifier and ensemble obtained by searching over all attributes considered when creating the single classifier. The application of our approach yielded a solution that ranked third in the competition.

Adaptive Testing in Programming Courses Based on Educational Data Mining Techniques

Designers of student tests, often teachers, primarily rely on their experience and subjective perception of students when selecting test items, while devoting little time to analyse factual data about both students and test items. As a practical solution to this common issue, we propose an approach to automatic test generation that acknowledges required areas of competence and matches the overall competence level of target students. The proposed approach, which is tailored to the testing practice in an introductory university course on programming, is based on the use of educational data mining. Data about students and test items are first evaluated using the predictive techniques of regression and classification, respectively, and then used to guide the test creation process. Besides a genetic algorithm that selects a test most suitable to the aforementioned criteria, we present a concept map of programming competencies and a method of estimating the test item difficulty.

National university rankings based on open data: A case study from Serbia

Abstract We investigate the potential of using open data about higher education and research activities as a basis for constructing university rankings at the national level. In our case study, open data from the Ministry of Education, Science, and Technological Development of Serbia served as a foundation for deriving indicators of university performance and calculating ranks of universities from Serbia. In addition to reviewing notable international rankings of universities, we extracted the international standings of universities from Serbia and discussed the national university rankings that were generated during our investigation.

Applying Domain Knowledge for Data Quality Assessment in Dermatology

The Dermatology Clinic at the Clinical Center of Vojvodina, Novi Sad, Serbia, has actively collected data regarding patients’ treatment, health insurance and examinations. These data were stored in documents in the comma-separated values (CSV) format. Since many fields in these documents were presented as free form text or allow null values, there are many data records that are inconsistent with the real-world system. Currently, there is a large need for an analytic system that can analyze these data and find relevant patterns. Since such an analytic system would require clean and accurate data, there is a need to assess data quality. Therefore, a data quality system should be designed and built with a goal of identifying inaccurate records so that they can be aligned with the real-world state. In our approach to data quality assessment, the domain knowledge about data is used to define rules which are then used to evaluate the quality of the data. In this paper, we present the architecture of a data quality system that is used to define and apply these rules. The rules are first defined by a domain expert and then applied to data in order to determine the number of records that do not match the defined rules and identify the exact anomalies in the given records. Also, we present a case study in which we applied this data quality system to the data collected by the Dermatology Clinic.

Decision Trees as Readable Models for Early Childhood Caries

Assessing risk for early childhood caries (ECC) is a relevant task in public health care and an important activity in fulfilling this task is increasing the knowledge about ECC. Discovering important information from data and sharing it in an understandable format with both experts and the general population could be beneficial for advancing and spreading the knowledge about this disease. After having experimented with association rule mining, we investigate the possibility of using decision trees as readable models in risk assessment. We build various decision trees using different algorithms and splitting criteria, favouring compact decision trees with good predictive performance. These decision trees are compared to the previous ECC models for the same analyzed population, namely a logistic regression model and an associative classifier, as well as to decision trees for caries from other studies. The results indicate flexibility and usefulness of decision trees in this context.

Human Friendly Associative Classifiers for Early Childhood Caries

Early childhood caries (ECC) is a widespread disease that may lead to serious complications and impact the whole society. For these reasons, we look for a predictive model that could be easily applied whenever and wherever necessary, especially in poor environments. As a result, we create human friendly classifiers for ECC that could be utilized in prevention programs. These classifiers are rule-based, with a few rules, easy to use even without computers, and without a loss in predictive performance. For this purpose, we mined association rules and clustered them by their contents. Next, we employed a genetic algorithm to assemble a classifier using dissimilar association rules. The proposed approach was tested on a data set about ECC in the South Backa area (Vojvodina, Serbia). We compared the performance of the resulting classifiers to that of the logistic regression model built around the previously identified risk factors.