A Weighted Two-Level Ensemble Model for Sentiment Analysis of Finance Minister of Indonesia

The performance of ensemble learning depends on the performance of each weak learner. Assuming all weak learners are equal may lead to poor results. Therefore, determining the weight for each ensemble component is proposed to maximize a two-level ensemble model performance. The proposed model is used to classify the sentiment of Finance Minister Sri Mulyani Indrawati, using Twitter data. The dataset is tested in individual models first called validation. The weight of the two-level ensemble model is determined based on training performance and validation performance by considering individual models’ precision, recall, and F1-Score values. For each metric of an individual model greater than average performance, the weight for that model will increase. Validation performance has eight times the contribution compared to training in weight determination. The comparison with default two-level ensemble learning is needed to show the gain of our proposed model. Experiments on the dataset show that our proposed model improves recall and F1-Score of testing set by 4% without sacrificing the prediction performance of training set. The rationale is that each model usually receives the same weight in ensemble learning, which differs from a weighted ensemble. A weighted ensemble learning highlights the contribution of stronger models and lessens the influence of poorer models by allocating varying weights to individual models based on their training and validation performances.