Deep learning to predict user rating in imbalance classification data incorporating ensemble methods
Hendry Hendry, Rung-Ching Chen, Chung-Yen Liao · 2018 IEEE International Conference on Applied System Invention (ICASI) · 2018
In recent years, machine learning gets more attention in research field with the successful breakthrough of deep learning. Deep learning has demonstrated some remarkable success through its capacity to create detailed (deep) models of complex multivariate in structured data. Deep learning can be characterized in several different ways, but the most important that deep learning can learn higher-order interactions among features using a cascade of many layers. Despite the successful breakthrough of deep learning, it still faces big challenges in computation complexity with hyper-parameter and time-consuming process to calculate many fully connected layers. We propose an alternate way to learn deep learning model in imbalance dataset. We incorporate ensemble models to learn every class target (user rating) with one base deep learning classifier models from user comments, in this way we will learn every rating with its fitting model. The main idea for ensemble model is to give an alternate solution for base classifiers to select the best results. In ensemble model, we could use different feature selection for input layer. Using all features in the same model could cause computation complexity and make the running process consume much time. Also, some features could be assessed whether significant or not. Less significant feature later could be prune in classifier or could be substituted with other features which are extracted from sampling data. This classification is based on two dimensions: how predictions are combined (rule-based and Meta-learning) and how the learning process is done (parallel or sequential). In rule-based approach, predictions will be combined using rule to average the performance. Meta-learning techniques use predictions from component classifiers as features for a Meta-learning model. In output layer, we apply voting methods to select user rating as output from each base classifier to generate user rating predicting results with more extensive results.