Imbalanced Classes Treatment in Deep Learning Multi-label Aspect Classification using Oversampling and Under-sampling

Categorising aspects is one of the research in aspect-based sentiment analysis (ABSA). The aspect classification will categorise the issues addressed by the review content. Multi-label categorisation incorporates aspects because each instance will have a unique set of labels or aspects. Aspect classification is one of the ABSA jobs that focuses on classifying the features related to the review text. The most used categorisation strategy relies on balanced data. Unbalanced data can be dealt with in various ways, including algorithm-level and data-balanced methods. This study addresses unbalanced data using three techniques: classification components of SMOTE, ADASYN, and undersampling. This study demonstrates that SMOTE and ADASYN can enhance model performance. However, sampling needs to be more appropriate for LSTM and Bi-LSTM techniques. The results indicate that SMOTE is more suitable than ADASYN for multi-label aspect classifications that employ LSTM or BiLSTM. Comparing the under-sampling approach to using SMOTE or ADASYN, all evaluation scores on LSTM and BiLTSM have been reduced. According to this research, the best model performance for multi-label aspect classification is LSTM with SMOTE implementation for data balancing.