Evolutionary Optimization of Ensemble Learning to Determine Sentiment Polarity in an Unbalanced Multiclass Corpus

Abstract

Sentiment polarity classification in social media is a very important task, as it enables gathering trends on particular subjects given a set of opinions. Currently, a great advance has been made by using deep learning techniques, such as word embeddings, recurrent neural networks, and encoders, such as BERT. Unfortunately, these techniques require large amounts of data, which, in some cases, is not available. In order to model this situation, challenges, such as the Spanish TASS organized by the Spanish Society for Natural Language Processing (SEPLN), have been proposed, which pose particular difficulties: First, an unwieldy balance in the training and the test set, being this latter more than eight times the size of the training set. Another difficulty is the marked unbalance in the distribution of classes, which is also different between both sets. Finally, there are four different labels, which create the need to adapt current classifications methods for multiclass handling. Traditional machine learning methods, such as Naïve Bayes, Logistic Regression, and Support Vector Machines, achieve modest performance in these conditions, but used as an ensemble it is possible to attain competitive execution. Several strategies to build classifier ensembles have been proposed; this paper proposes estimating an optimal weighting scheme using a Differential Evolution algorithm focused on dealing with particular issues that multiclass classification and unbalanced corpora pose. The ensemble with the proposed optimized weighting scheme is able to improve the classification results on the full test set of the TASS challenge (General corpus), achieving state of the art performance when compared with other works on this task, which make no use of NLP techniques.

Experimental Study and Comparison of Imbalance Ensemble Classifiers with Dynamic Selection Strategy

“…It occurs when the number of instances for different classes are significantly out of proportion. The minority classes with fewer instances usually contain the essential information, which has been observed in broad application areas, such as medical diagnosis [ 1 , 2 , 3 , 4 , 5 , 6 ], sentiment or image classification [ 7 , 8 ], fault identification [ 9 , 10 ], etc. Many typical classifiers may generate unsatisfactory results due to a concentration on global accuracy while ignoring the identification performance for minority samples.…”

Statistics-Based Outlier Detection and Correction Method for Amazon Customer Reviews

“…Several studies were conducted on sentiment analysis [ 28 ] and its application on e-commerce. With the increase in online consumption, e-commerce enhancement has become a hot topic for research.…”

Evolutionary bagging for ensemble learning

“…evolving bagged training samples in an class imbalanced scenario [18,28,29], 2.) assigning weights for individual learners [30,31,32], 3.) maintaining diversity between learners [33,34], and 4.)…”

“…We note that the methods that did not alter the training samples in each bag had the advantage of optimizing over a smaller search space [30,31,33,34,35]; e.g. the set of weights for individual learners or the subsets of features [30,31,32]. However, these methods lacked the ability to optimize the set of training samples in each bag as these remained the same throughout the evolution process.…”