Investigating structural metrics for understandability prediction of data warehouse multidimensional schemas using machine learning techniques

Data warehouse (DW) quality metrics help in evaluating quality attributes and building classification models for predicting multidimensional (MD) schemas as understandable/non-understandable, thereby assisting in DW maintenance. To evaluate DW MD schema quality, we have earlier proposed a set of metrics based on some important aspects of dimension hierarchies and its sharing (like sharing of few hierarchy levels within a dimension; sharing of few hierarchy levels between dimensions, within and across facts) which may lead to structural complexity of MD schemas, thereby affecting its quality. The preliminary empirical validation of these metrics using classical statistical techniques (correlation and linear regression) indicated some of them as possible understandability indicators. However, machine learning (ML) techniques can model the complex associations between DW structural metrics and their quality attributes in a better way. Therefore, this work employs five ML classifiers [J48, partial decision trees (PART), Naïve Bayes, support vector machines (SVM) and logistic regression] to empirically investigate whether accurate prediction models can be built, based on our structural metrics, to be used as understandability predictors. The obtained results reveal that four of our metrics are good predictors of understandability of DW MD schemas. The experimentation further involved comparing the classifiers using mainly five performance measures: accuracy, precision, sensitivity, specificity and area under the receiver operating characteristic curve. The study confirmed the predictive capability of ML techniques for understandability prediction of DW MD schemas. The results also suggest that the SVM and Naïve Bayes classifiers perform better than other classifiers included in the study. Further, the typically used logistic regression technique gave results that were reasonably competitive with the more sophisticated techniques. However, the tree-based (J48) and rule-based (PART) techniques performed significantly worse than the best performing techniques.