Imbalanced Data Handling in Multi-label Aspect Categorization using Oversampling and Ensemble Learning

In sentiment analysis, aspect based sentiment analysis (ABSA) provides detailed information of user sentiment for a product rather than document level and sentence level. Aspect categorization is one of ABSA tasks, which focuses on categorizing which aspects are related to a review text. This task working on multilabel data that usually have uneven distribution of aspect occurrences or imbalanced data condition. This paper uses 9284 data from user review text in the hotel domain. We employ 3 techniques to address imbalanced multilabel data, namely cross-coupling aggregation (COCOA), multilabel synthetic minority oversampling technique (MLSMOTE), and multilabel synthetic oversampling approach based on the local distribution of labels (MLSOL). Convolutional Neural Network (CNN)-Classifier Chain (CC)-Extreme Gradient Boosting (XGBoost) is employed as a baseline and base architecture to be applied into those 3 techniques of handling imbalanced multilabel dataset. COCOA and MLSMOTE are the best performers. COCOA achieved F1-Macro of 0.9272, F1 macro MLSMOTE is 0.9276 and F1-Macro baseline is 0.9261. The best performer of COCOA is configured using 4 parameters: binary relevance mode is smote-oversampling, multiclass mode is smote-oversampling, random state =10, and binary relevance ratio =0.5. The best performer of MLSMOTE is configured using 2 parameters: number of neighbors =5, and random state =42.