Multi-label classification is a general type of classification that has attracted many researchers in the last two decades due to its applicability to many modern domains such as scene classification, bioinformatics and text classification among others. Class labels ranking is a crucial problem in multi-label classification research because it directly impacts the performance of the final classifiers as labels, with high ranks getting a higher chance of being applied. This paper presents a new multi-label ranking algorithm called Multi-label Ranking based on Positive Correlations among labels (MLR-PC). MLR-PC captures positive correlations among labels to reduce the large search space and assigns the true rank per class label for multi-label classification problems. More importantly, MLR-PC utilizes novel problem transformation methods that facilitate exploiting accurate positive correlations among labels. This improves the predictive performance of the classification models derived. Empirical results using different multi-label datasets and three evaluation metrics reveal that the MLR-PC is superior to other commonly existing classification algorithms.
Key words: Prediction, machine learning, multi-label ranking, multi-label classification, problem transformation methods, class ranking methods.